main..87c2165 · crush · git.iko.soy

git.iko.soy/crush

on 87c2165 help sign in

.agents/skills/builtin-skills/SKILL.mddeleted−42

since splitsnapshots

unifiedsplit

@@ -1,42 +0,0 @@

1−---

2−name: builtin-skills

3−description:

4− Use when creating a new builtin skill for Crush, editing an existing builtin

5− skill (internal/skills/builtin/), or when the user needs to understand how the

6− embedded skill system works.

7−---

8−

9−# Builtin Skills

10−

11−Crush embeds skills directly into the binary via `internal/skills/builtin/`.

12−These are always available without user configuration.

13−

14−## How It Works

15−

16−- Each skill lives in `internal/skills/builtin/<skill-name>/SKILL.md`.

17−- The tree is embedded at compile time via `//go:embed builtin/*` in

18− `internal/skills/embed.go`.

19−- `DiscoverBuiltin()` walks the embedded FS, parses each `SKILL.md`, and sets

20− paths with the `crush://skills/` prefix (e.g., `crush://skills/jq/SKILL.md`).

21−- The View tool resolves `crush://` paths from the embedded FS, not disk.

22−- User skills with the same name override builtins (last occurrence wins in

23− `Deduplicate()`).

24−

25−## Adding a New Builtin Skill

26−

27−1. Create `internal/skills/builtin/<skill-name>/SKILL.md` with YAML frontmatter

28− (`name`, `description`) and markdown instructions. The directory name must

29− match the `name` field.

30−2. No extra wiring needed — `//go:embed builtin/*` picks up new directories

31− automatically.

32−3. Add a test assertion in `TestDiscoverBuiltin` in

33− `internal/skills/skills_test.go` to verify discovery.

34−4. Build and test: `go build . && go test ./internal/skills/...`

35−

36−## Existing Builtin Skills

37−

38−| Skill | Directory | Description |

39−| -------------- | ----------------------- | ------------------------------------------ |

40−| `crush-config` | `builtin/crush-config/` | Crush configuration help |

41−| `crush-hooks` | `builtin/crush-hooks/` | Authoring, configuring and debugging hooks |

42−| `jq` | `builtin/jq/` | jq JSON processor usage guide |

.agents/skills/shell-builtins/SKILL.mddeleted−72

@@ -1,72 +0,0 @@

1−---

2−name: shell-builtins

3−description: Use when creating a new shell builtin command for Crush (internal/shell/), editing an existing one, or when the user needs to understand how commands are intercepted in Crush's embedded shell.

4−---

5−

6−# Shell Builtins

7−

8−Crush's shell (`internal/shell/`) uses `mvdan.cc/sh/v3` for POSIX shell

9−emulation. Commands can be intercepted before they reach the OS by adding

10−**builtins** — functions handled in-process.

11−

12−## How Builtins Work

13−

14−Builtins live in `builtinHandler()` in `internal/shell/run.go`. This is an

15−`interp.ExecHandlerFunc` middleware registered in `standardHandlers()`

16−**before** the block handler, so builtins run even for commands that would

17−otherwise be blocked. The same handler chain is shared by the stateful

18−`Shell` type and the stateless `Run` entrypoint used by the hook runner,

19−so builtins are available identically in the `bash` tool and in hooks.

20−

21−The handler is a switch on `args[0]`. Each case either handles the command

22−inline or delegates to a helper function.

23−

24−## Adding a New Builtin

25−

26−1. **Add the case** to the switch in `builtinHandler()` in `run.go`.

27−2. **Get I/O from the handler context**, not from `os.Stdin`/`os.Stdout`.

28− This ensures the builtin works with pipes and redirections:

29− ```go

30− case "mycommand":

31− hc := interp.HandlerCtx(ctx)

32− return handleMyCommand(ctx, args, hc.Stdin, hc.Stdout, hc.Stderr)

33− ```

34−3. **Implement the handler** in its own file (e.g.,

35− `internal/shell/mycommand.go`). The function signature must accept a

36− `context.Context` as the first parameter, plus args, stdin, stdout, and

37− stderr:

38− ```go

39− func handleMyCommand(ctx context.Context, args []string, stdin io.Reader, stdout, stderr io.Writer) error {

40− // args[0] is the command name ("mycommand"), args[1:] are arguments.

41− // Write output to stdout, errors to stderr.

42− // Return nil on success, or interp.ExitStatus(n) for non-zero exit codes.

43− }

44− ```

45−4. **Poll `ctx` in every unbounded loop.** Builtins that iterate over

46− input, emit values in a generator-style loop, or do any other work

47− that can exceed a few milliseconds MUST check `ctx.Err()` on each

48− iteration and return it verbatim when non-nil. Hook timeouts rely on

49− this: an unbounded builtin that never polls ctx cannot be interrupted

50− by a hook's `timeout_sec`, and the hook runner will have to abandon

51− the goroutine (see `internal/hooks/runner.go`). Returning `ctx.Err()`

52− (not `interp.ExitStatus(n)`) lets callers distinguish "command exited

53− non-zero" from "we ran out of time".

54− ```go

55− for _, item := range items {

56− if err := ctx.Err(); err != nil {

57− return err

58− }

59− // ... process item

60− }

61− ```

62−5. **Return values**: return `nil` for success, `interp.ExitStatus(n)` for

63− non-zero exit codes, or `ctx.Err()` on cancellation. Write error

64− messages to `stderr` before returning.

65−6. **No extra wiring needed** — `builtinHandler()` is already registered

66− in `standardHandlers()`.

67−

68−## Existing Builtins

69−

70−| Command | File | Description |

71−|---------|------|-------------|

72−| `jq` | `jq.go` | JSON processor using `github.com/itchyny/gojq` |

.envrcdeleted−1

@@ -1 +0,0 @@

1−use flake

.github/CODEOWNERSnew file+1

@@ -0,0 +1 @@

1+*.go @charmbracelet/everyone

.github/cla-signatures.json−952

@@ -1143,958 +1143,6 @@

1143 "created_at": "2026-01-29T07:05:12Z",

1144 "repoId": 987670088,

1145 "pullRequestNo": 2043

1146− },

1147− {

1148− "name": "bittoby",

1149− "id": 218712309,

1150− "comment_id": 3824931235,

1151− "created_at": "2026-01-30T17:52:15Z",

1152− "repoId": 987670088,

1153− "pullRequestNo": 2065

1154− },

1155− {

1156− "name": "ijt",

1157− "id": 15530,

1158− "comment_id": 3832667774,

1159− "created_at": "2026-02-02T03:06:23Z",

1160− "repoId": 987670088,

1161− "pullRequestNo": 2080

1162− },

1163− {

1164− "name": "khalilgharbaoui",

1165− "id": 8024057,

1166− "comment_id": 3832796060,

1167− "created_at": "2026-02-02T04:04:04Z",

1168− "repoId": 987670088,

1169− "pullRequestNo": 2081

1170− },

1171− {

1172− "name": "acmacalister",

1173− "id": 1024755,

1174− "comment_id": 3837172797,

1175− "created_at": "2026-02-02T19:27:08Z",

1176− "repoId": 987670088,

1177− "pullRequestNo": 2095

1178− },

1179− {

1180− "name": "zhiquanchi",

1181− "id": 29973289,

1182− "comment_id": 3845838711,

1183− "created_at": "2026-02-04T07:39:06Z",

1184− "repoId": 987670088,

1185− "pullRequestNo": 2112

1186− },

1187− {

1188− "name": "inquam",

1189− "id": 1265038,

1190− "comment_id": 3849304908,

1191− "created_at": "2026-02-04T19:22:33Z",

1192− "repoId": 987670088,

1193− "pullRequestNo": 2124

1194− },

1195− {

1196− "name": "nickgrim",

1197− "id": 8376,

1198− "comment_id": 3852565144,

1199− "created_at": "2026-02-05T10:17:46Z",

1200− "repoId": 987670088,

1201− "pullRequestNo": 2131

1202− },

1203− {

1204− "name": "francescoalemanno",

1205− "id": 50984334,

1206− "comment_id": 3858464719,

1207− "created_at": "2026-02-06T07:16:50Z",

1208− "repoId": 987670088,

1209− "pullRequestNo": 2142

1210− },

1211− {

1212− "name": "biisal",

1213− "id": 153633053,

1214− "comment_id": 3866503536,

1215− "created_at": "2026-02-08T08:15:11Z",

1216− "repoId": 987670088,

1217− "pullRequestNo": 2164

1218− },

1219− {

1220− "name": "mishudark",

1221− "id": 211144,

1222− "comment_id": 3866939317,

1223− "created_at": "2026-02-08T10:27:09Z",

1224− "repoId": 987670088,

1225− "pullRequestNo": 2165

1226− },

1227− {

1228− "name": "portertech",

1229− "id": 149630,

1230− "comment_id": 3878650318,

1231− "created_at": "2026-02-10T15:39:14Z",

1232− "repoId": 987670088,

1233− "pullRequestNo": 2183

1234− },

1235− {

1236− "name": "wallacegibbon",

1237− "id": 22029486,

1238− "comment_id": 3890517245,

1239− "created_at": "2026-02-12T11:58:04Z",

1240− "repoId": 987670088,

1241− "pullRequestNo": 2203

1242− },

1243− {

1244− "name": "PHPCraftdream",

1245− "id": 14233546,

1246− "comment_id": 3893502046,

1247− "created_at": "2026-02-12T21:34:20Z",

1248− "repoId": 987670088,

1249− "pullRequestNo": 2212

1250− },

1251− {

1252− "name": "maxbrunet",

1253− "id": 32458727,

1254− "comment_id": 3905030983,

1255− "created_at": "2026-02-15T19:14:26Z",

1256− "repoId": 987670088,

1257− "pullRequestNo": 2229

1258− },

1259− {

1260− "name": "0xarcher",

1261− "id": 18182408,

1262− "comment_id": 3907297580,

1263− "created_at": "2026-02-16T09:16:57Z",

1264− "repoId": 987670088,

1265− "pullRequestNo": 2236

1266− },

1267− {

1268− "name": "julienrbrt",

1269− "id": 29894366,

1270− "comment_id": 3920039276,

1271− "created_at": "2026-02-18T10:37:53Z",

1272− "repoId": 987670088,

1273− "pullRequestNo": 2246

1274− },

1275− {

1276− "name": "erikstmartin",

1277− "id": 50041,

1278− "comment_id": 3937646631,

1279− "created_at": "2026-02-20T23:35:55Z",

1280− "repoId": 987670088,

1281− "pullRequestNo": 2274

1282− },

1283− {

1284− "name": "Jaylonnet",

1285− "id": 101283545,

1286− "comment_id": 3944717192,

1287− "created_at": "2026-02-23T13:12:50Z",

1288− "repoId": 987670088,

1289− "pullRequestNo": 2293

1290− },

1291− {

1292− "name": "mavaa",

1293− "id": 1224973,

1294− "comment_id": 3966463081,

1295− "created_at": "2026-02-26T12:56:51Z",

1296− "repoId": 987670088,

1297− "pullRequestNo": 2314

1298− },

1299− {

1300− "name": "aisk",

1301− "id": 699636,

1302− "comment_id": 3968065934,

1303− "created_at": "2026-02-26T17:23:44Z",

1304− "repoId": 987670088,

1305− "pullRequestNo": 2315

1306− },

1307− {

1308− "name": "detro",

1309− "id": 114508,

1310− "comment_id": 3976601253,

1311− "created_at": "2026-02-28T07:36:40Z",

1312− "repoId": 987670088,

1313− "pullRequestNo": 2326

1314− },

1315− {

1316− "name": "taoeffect",

1317− "id": 138706,

1318− "comment_id": 3979067985,

1319− "created_at": "2026-03-01T04:22:00Z",

1320− "repoId": 987670088,

1321− "pullRequestNo": 2333

1322− },

1323− {

1324− "name": "vmfu",

1325− "id": 80844805,

1326− "comment_id": 3980616480,

1327− "created_at": "2026-03-01T17:51:04Z",

1328− "repoId": 987670088,

1329− "pullRequestNo": 2337

1330− },

1331− {

1332− "name": "ZeitbyteRepo",

1333− "id": 187110998,

1334− "comment_id": 4025287397,

1335− "created_at": "2026-03-09T17:00:06Z",

1336− "repoId": 987670088,

1337− "pullRequestNo": 2390

1338− },

1339− {

1340− "name": "seroperson",

1341− "id": 3791221,

1342− "comment_id": 4047392328,

1343− "created_at": "2026-03-12T14:53:20Z",

1344− "repoId": 987670088,

1345− "pullRequestNo": 2401

1346− },

1347− {

1348− "name": "xulongwu4",

1349− "id": 16892446,

1350− "comment_id": 4062024320,

1351− "created_at": "2026-03-15T02:30:55Z",

1352− "repoId": 987670088,

1353− "pullRequestNo": 2410

1354− },

1355− {

1356− "name": "nghiant03",

1357− "id": 102637959,

1358− "comment_id": 4077213520,

1359− "created_at": "2026-03-17T18:41:50Z",

1360− "repoId": 987670088,

1361− "pullRequestNo": 2421

1362− },

1363− {

1364− "name": "whatnick",

1365− "id": 491396,

1366− "comment_id": 4102155868,

1367− "created_at": "2026-03-21T03:49:12Z",

1368− "repoId": 987670088,

1369− "pullRequestNo": 2449

1370− },

1371− {

1372− "name": "Alex-wuhu",

1373− "id": 57132813,

1374− "comment_id": 4106451941,

1375− "created_at": "2026-03-22T15:26:05Z",

1376− "repoId": 987670088,

1377− "pullRequestNo": 2454

1378− },

1379− {

1380− "name": "hongquan",

1381− "id": 314607,

1382− "comment_id": 4112122644,

1383− "created_at": "2026-03-23T16:51:09Z",

1384− "repoId": 987670088,

1385− "pullRequestNo": 2462

1386− },

1387− {

1388− "name": "faelis",

1389− "id": 91593249,

1390− "comment_id": 4113326932,

1391− "created_at": "2026-03-23T19:57:54Z",

1392− "repoId": 987670088,

1393− "pullRequestNo": 2465

1394− },

1395− {

1396− "name": "UnderLotus",

1397− "id": 84985759,

1398− "comment_id": 4118727062,

1399− "created_at": "2026-03-24T14:25:40Z",

1400− "repoId": 987670088,

1401− "pullRequestNo": 2475

1402− },

1403− {

1404− "name": "malikwirin",

1405− "id": 117918464,

1406− "comment_id": 4119816035,

1407− "created_at": "2026-03-24T16:51:44Z",

1408− "repoId": 987670088,

1409− "pullRequestNo": 2476

1410− },

1411− {

1412− "name": "fuleinist",

1413− "id": 1163738,

1414− "comment_id": 4133146531,

1415− "created_at": "2026-03-26T09:45:48Z",

1416− "repoId": 987670088,

1417− "pullRequestNo": 2492

1418− },

1419− {

1420− "name": "iceymoss",

1421− "id": 114280774,

1422− "comment_id": 4134659508,

1423− "created_at": "2026-03-26T13:13:25Z",

1424− "repoId": 987670088,

1425− "pullRequestNo": 2493

1426− },

1427− {

1428− "name": "srivilliamsai",

1429− "id": 191893106,

1430− "comment_id": 4144200398,

1431− "created_at": "2026-03-27T17:37:29Z",

1432− "repoId": 987670088,

1433− "pullRequestNo": 2511

1434− },

1435− {

1436− "name": "Dexterity104",

1437− "id": 173429049,

1438− "comment_id": 4153067126,

1439− "created_at": "2026-03-30T08:05:56Z",

1440− "repoId": 987670088,

1441− "pullRequestNo": 2525

1442− },

1443− {

1444− "name": "majiayu000",

1445− "id": 19658300,

1446− "comment_id": 4160074332,

1447− "created_at": "2026-03-31T05:44:38Z",

1448− "repoId": 987670088,

1449− "pullRequestNo": 2534

1450− },

1451− {

1452− "name": "afsuyadi",

1453− "id": 104284194,

1454− "comment_id": 4160431726,

1455− "created_at": "2026-03-31T07:10:20Z",

1456− "repoId": 987670088,

1457− "pullRequestNo": 2532

1458− },

1459− {

1460− "name": "owldev127",

1461− "id": 175169034,

1462− "comment_id": 4161299761,

1463− "created_at": "2026-03-31T09:44:52Z",

1464− "repoId": 987670088,

1465− "pullRequestNo": 2538

1466− },

1467− {

1468− "name": "MrRolie",

1469− "id": 80284543,

1470− "comment_id": 4168509138,

1471− "created_at": "2026-04-01T08:44:22Z",

1472− "repoId": 987670088,

1473− "pullRequestNo": 2544

1474− },

1475− {

1476− "name": "Kartik33",

1477− "id": 8198703,

1478− "comment_id": 4181386196,

1479− "created_at": "2026-04-03T02:02:31Z",

1480− "repoId": 987670088,

1481− "pullRequestNo": 2554

1482− },

1483− {

1484− "name": "talha7k",

1485− "id": 51679353,

1486− "comment_id": 4182808517,

1487− "created_at": "2026-04-03T09:56:29Z",

1488− "repoId": 987670088,

1489− "pullRequestNo": 2555

1490− },

1491− {

1492− "name": "avianion",

1493− "id": 37309215,

1494− "comment_id": 4186934644,

1495− "created_at": "2026-04-04T10:54:16Z",

1496− "repoId": 987670088,

1497− "pullRequestNo": 2557

1498− },

1499− {

1500− "name": "shahidshabbir-se",

1501− "id": 140221016,

1502− "comment_id": 4191066791,

1503− "created_at": "2026-04-06T08:11:08Z",

1504− "repoId": 987670088,

1505− "pullRequestNo": 2563

1506− },

1507− {

1508− "name": "smlx",

1509− "id": 861778,

1510− "comment_id": 4200126763,

1511− "created_at": "2026-04-07T15:22:15Z",

1512− "repoId": 987670088,

1513− "pullRequestNo": 2575

1514− },

1515− {

1516− "name": "gurnben",

1517− "id": 14117802,

1518− "comment_id": 4216941498,

1519− "created_at": "2026-04-09T19:22:39Z",

1520− "repoId": 987670088,

1521− "pullRequestNo": 2593

1522− },

1523− {

1524− "name": "dshumphr",

1525− "id": 9056521,

1526− "comment_id": 4229412078,

1527− "created_at": "2026-04-11T12:20:44Z",

1528− "repoId": 987670088,

1529− "pullRequestNo": 2601

1530− },

1531− {

1532− "name": "bensantora",

1533− "id": 274462811,

1534− "comment_id": 4229507403,

1535− "created_at": "2026-04-11T13:31:32Z",

1536− "repoId": 987670088,

1537− "pullRequestNo": 2602

1538− },

1539− {

1540− "name": "ahostbr",

1541− "id": 7461677,

1542− "comment_id": 4231674482,

1543− "created_at": "2026-04-12T14:14:41Z",

1544− "repoId": 987670088,

1545− "pullRequestNo": 2606

1546− },

1547− {

1548− "name": "KimBioInfoStudio",

1549− "id": 22641519,

1550− "comment_id": 4244355652,

1551− "created_at": "2026-04-14T13:43:55Z",

1552− "repoId": 987670088,

1553− "pullRequestNo": 2622

1554− },

1555− {

1556− "name": "taigong12",

1557− "id": 276221439,

1558− "comment_id": 4259818851,

1559− "created_at": "2026-04-16T11:50:25Z",

1560− "repoId": 987670088,

1561− "pullRequestNo": 2634

1562− },

1563− {

1564− "name": "pi128",

1565− "id": 77074839,

1566− "comment_id": 4271523344,

1567− "created_at": "2026-04-17T21:52:43Z",

1568− "repoId": 987670088,

1569− "pullRequestNo": 2643

1570− },

1571− {

1572− "name": "enrell",

1573− "id": 60587839,

1574− "comment_id": 4272340887,

1575− "created_at": "2026-04-18T01:47:04Z",

1576− "repoId": 987670088,

1577− "pullRequestNo": 2646

1578− },

1579− {

1580− "name": "gavmor",

1581− "id": 606529,

1582− "comment_id": 4285449550,

1583− "created_at": "2026-04-21T02:02:06Z",

1584− "repoId": 987670088,

1585− "pullRequestNo": 2612

1586− },

1587− {

1588− "name": "axeprpr",

1589− "id": 8937600,

1590− "comment_id": 4293455788,

1591− "created_at": "2026-04-22T04:15:18Z",

1592− "repoId": 987670088,

1593− "pullRequestNo": 2674

1594− },

1595− {

1596− "name": "iuga",

1597− "id": 1430179,

1598− "comment_id": 4298985284,

1599− "created_at": "2026-04-22T18:33:08Z",

1600− "repoId": 987670088,

1601− "pullRequestNo": 2681

1602− },

1603− {

1604− "name": "ne275",

1605− "id": 174412782,

1606− "comment_id": 4303280163,

1607− "created_at": "2026-04-23T09:29:30Z",

1608− "repoId": 987670088,

1609− "pullRequestNo": 2686

1610− },

1611− {

1612− "name": "flynn-eye",

1613− "id": 23251324,

1614− "comment_id": 4304378619,

1615− "created_at": "2026-04-23T12:33:57Z",

1616− "repoId": 987670088,

1617− "pullRequestNo": 2688

1618− },

1619− {

1620− "name": "georgeglarson",

1621− "id": 744063,

1622− "comment_id": 4341592963,

1623− "created_at": "2026-04-29T07:08:48Z",

1624− "repoId": 987670088,

1625− "pullRequestNo": 2749

1626− },

1627− {

1628− "name": "vorticalbox",

1629− "id": 10886065,

1630− "comment_id": 4344669740,

1631− "created_at": "2026-04-29T14:29:35Z",

1632− "repoId": 987670088,

1633− "pullRequestNo": 2751

1634− },

1635− {

1636− "name": "lloydzhou",

1637− "id": 1826685,

1638− "comment_id": 4346011097,

1639− "created_at": "2026-04-29T17:27:37Z",

1640− "repoId": 987670088,

1641− "pullRequestNo": 2752

1642− },

1643− {

1644− "name": "carlosgrillet",

1645− "id": 13719979,

1646− "comment_id": 4351020450,

1647− "created_at": "2026-04-30T08:48:52Z",

1648− "repoId": 987670088,

1649− "pullRequestNo": 2759

1650− },

1651− {

1652− "name": "SAY-5",

1653− "id": 240962040,

1654− "comment_id": 4357189565,

1655− "created_at": "2026-05-01T01:00:00Z",

1656− "repoId": 987670088,

1657− "pullRequestNo": 2691

1658− },

1659− {

1660− "name": "mkaaad",

1661− "id": 119158371,

1662− "comment_id": 4358493210,

1663− "created_at": "2026-05-01T08:25:06Z",

1664− "repoId": 987670088,

1665− "pullRequestNo": 2772

1666− },

1667− {

1668− "name": "pragneshbagary",

1669− "id": 41118746,

1670− "comment_id": 4359152162,

1671− "created_at": "2026-05-01T11:54:16Z",

1672− "repoId": 987670088,

1673− "pullRequestNo": 2773

1674− },

1675− {

1676− "name": "somjik-api",

1677− "id": 47082582,

1678− "comment_id": 4364741275,

1679− "created_at": "2026-05-02T21:17:55Z",

1680− "repoId": 987670088,

1681− "pullRequestNo": 2783

1682− },

1683− {

1684− "name": "ilgax",

1685− "id": 79969769,

1686− "comment_id": 4367211073,

1687− "created_at": "2026-05-03T21:25:46Z",

1688− "repoId": 987670088,

1689− "pullRequestNo": 2790

1690− },

1691− {

1692− "name": "ardevd",

1693− "id": 366855,

1694− "comment_id": 4377260996,

1695− "created_at": "2026-05-05T07:22:31Z",

1696− "repoId": 987670088,

1697− "pullRequestNo": 2801

1698− },

1699− {

1700− "name": "yeonuk-hwang",

1701− "id": 54208214,

1702− "comment_id": 4407666773,

1703− "created_at": "2026-05-08T15:33:03Z",

1704− "repoId": 987670088,

1705− "pullRequestNo": 2836

1706− },

1707− {

1708− "name": "sven2718",

1709− "id": 4741888,

1710− "comment_id": 4410098922,

1711− "created_at": "2026-05-08T21:47:18Z",

1712− "repoId": 987670088,

1713− "pullRequestNo": 2840

1714− },

1715− {

1716− "name": "acheong08",

1717− "id": 36258159,

1718− "comment_id": 4410600427,

1719− "created_at": "2026-05-08T23:37:33Z",

1720− "repoId": 987670088,

1721− "pullRequestNo": 2842

1722− },

1723− {

1724− "name": "smeinecke",

1725− "id": 7316652,

1726− "comment_id": 4411012852,

1727− "created_at": "2026-05-09T01:47:08Z",

1728− "repoId": 987670088,

1729− "pullRequestNo": 2844

1730− },

1731− {

1732− "name": "jan-xyz",

1733− "id": 5249233,

1734− "comment_id": 4432309551,

1735− "created_at": "2026-05-12T15:48:38Z",

1736− "repoId": 987670088,

1737− "pullRequestNo": 2886

1738− },

1739− {

1740− "name": "johnjansen",

1741− "id": 105938,

1742− "comment_id": 4434007350,

1743− "created_at": "2026-05-12T19:28:01Z",

1744− "repoId": 987670088,

1745− "pullRequestNo": 2887

1746− },

1747− {

1748− "name": "pablodz",

1749− "id": 28235457,

1750− "comment_id": 4435307961,

1751− "created_at": "2026-05-12T22:25:09Z",

1752− "repoId": 987670088,

1753− "pullRequestNo": 2892

1754− },

1755− {

1756− "name": "akhenakh",

1757− "id": 218430,

1758− "comment_id": 4446905473,

1759− "created_at": "2026-05-14T02:28:13Z",

1760− "repoId": 987670088,

1761− "pullRequestNo": 2912

1762− },

1763− {

1764− "name": "leonardoaraujosantos",

1765− "id": 898383,

1766− "comment_id": 3166308642,

1767− "created_at": "2025-08-08T01:40:17Z",

1768− "repoId": 987670088,

1769− "pullRequestNo": 644

1770− },

1771− {

1772− "name": "yookibooki",

1773− "id": 61198899,

1774− "comment_id": 4459665029,

1775− "created_at": "2026-05-15T12:16:26Z",

1776− "repoId": 987670088,

1777− "pullRequestNo": 2922

1778− },

1779− {

1780− "name": "CatMe0w",

1781− "id": 7774491,

1782− "comment_id": 4461676513,

1783− "created_at": "2026-05-15T16:58:04Z",

1784− "repoId": 987670088,

1785− "pullRequestNo": 2925

1786− },

1787− {

1788− "name": "Broderick-Westrope",

1789− "id": 55474351,

1790− "comment_id": 4469173328,

1791− "created_at": "2026-05-17T03:35:13Z",

1792− "repoId": 987670088,

1793− "pullRequestNo": 2937

1794− },

1795− {

1796− "name": "processtrader",

1797− "id": 232431073,

1798− "comment_id": 4480941961,

1799− "created_at": "2026-05-18T18:58:31Z",

1800− "repoId": 987670088,

1801− "pullRequestNo": 2950

1802− },

1803− {

1804− "name": "dcu",

1805− "id": 73729,

1806− "comment_id": 4484181213,

1807− "created_at": "2026-05-19T03:37:32Z",

1808− "repoId": 987670088,

1809− "pullRequestNo": 2954

1810− },

1811− {

1812− "name": "mei2jun1",

1813− "id": 285962643,

1814− "comment_id": 4495440988,

1815− "created_at": "2026-05-20T06:53:43Z",

1816− "repoId": 987670088,

1817− "pullRequestNo": 2964

1818− },

1819− {

1820− "name": "g2mt",

1821− "id": 166577174,

1822− "comment_id": 4513096006,

1823− "created_at": "2026-05-21T21:56:14Z",

1824− "repoId": 987670088,

1825− "pullRequestNo": 2979

1826− },

1827− {

1828− "name": "Ricardo-M-L",

1829− "id": 69202550,

1830− "comment_id": 4514494248,

1831− "created_at": "2026-05-22T02:37:08Z",

1832− "repoId": 987670088,

1833− "pullRequestNo": 2847

1834− },

1835− {

1836− "name": "Muttaqin86",

1837− "id": 69788027,

1838− "comment_id": 4520181780,

1839− "created_at": "2026-05-22T15:46:32Z",

1840− "repoId": 987670088,

1841− "pullRequestNo": 2984

1842− },

1843− {

1844− "name": "officialasishkumar",

1845− "id": 87874775,

1846− "comment_id": 4527874300,

1847− "created_at": "2026-05-24T08:44:14Z",

1848− "repoId": 987670088,

1849− "pullRequestNo": 2995

1850− },

1851− {

1852− "name": "yhyu13",

1853− "id": 19365678,

1854− "comment_id": 4528076561,

1855− "created_at": "2026-05-24T10:19:27Z",

1856− "repoId": 987670088,

1857− "pullRequestNo": 2996

1858− },

1859− {

1860− "name": "xulongzhe",

1861− "id": 10559793,

1862− "comment_id": 4569454380,

1863− "created_at": "2026-05-29T00:39:36Z",

1864− "repoId": 987670088,

1865− "pullRequestNo": 3036

1866− },

1867− {

1868− "name": "ardi1s",

1869− "id": 200318875,

1870− "comment_id": 4589172224,

1871− "created_at": "2026-06-01T02:48:53Z",

1872− "repoId": 987670088,

1873− "pullRequestNo": 3049

1874− },

1875− {

1876− "name": "cak3ninja",

1877− "id": 4338875,

1878− "comment_id": 4596570813,

1879− "created_at": "2026-06-01T21:14:16Z",

1880− "repoId": 987670088,

1881− "pullRequestNo": 3059

1882− },

1883− {

1884− "name": "GroovyCarrot",

1885− "id": 5589906,

1886− "comment_id": 4623915928,

1887− "created_at": "2026-06-04T16:01:30Z",

1888− "repoId": 987670088,

1889− "pullRequestNo": 3076

1890− },

1891− {

1892− "name": "CnsMaple",

1893− "id": 92523839,

1894− "comment_id": 4628808493,

1895− "created_at": "2026-06-05T06:24:00Z",

1896− "repoId": 987670088,

1897− "pullRequestNo": 3080

1898− },

1899− {

1900− "name": "yuseferi",

1901− "id": 1736832,

1902− "comment_id": 4548802420,

1903− "created_at": "2026-05-26T21:00:36Z",

1904− "repoId": 987670088,

1905− "pullRequestNo": 3019

1906− },

1907− {

1908− "name": "gmit3",

1909− "id": 26192089,

1910− "comment_id": 4651029942,

1911− "created_at": "2026-06-08T16:18:06Z",

1912− "repoId": 987670088,

1913− "pullRequestNo": 3096

1914− },

1915− {

1916− "name": "brianjlandau",

1917− "id": 3122,

1918− "comment_id": 4654867087,

1919− "created_at": "2026-06-09T00:37:02Z",

1920− "repoId": 987670088,

1921− "pullRequestNo": 3098

1922− },

1923− {

1924− "name": "kypkk",

1925− "id": 68454243,

1926− "comment_id": 4687904764,

1927− "created_at": "2026-06-12T05:55:06Z",

1928− "repoId": 987670088,

1929− "pullRequestNo": 3111

1930− },

1931− {

1932− "name": "warmjademe",

1933− "id": 257661268,

1934− "comment_id": 4720070206,

1935− "created_at": "2026-06-16T14:48:33Z",

1936− "repoId": 987670088,

1937− "pullRequestNo": 3131

1938− },

1939− {

1940− "name": "warmjademe",

1941− "id": 257661268,

1942− "comment_id": 4720177856,

1943− "created_at": "2026-06-16T14:58:33Z",

1944− "repoId": 987670088,

1945− "pullRequestNo": 3131

1946− },

1947− {

1948− "name": "jialudev",

1949− "id": 39230796,

1950− "comment_id": 4725593255,

1951− "created_at": "2026-06-17T03:26:00Z",

1952− "repoId": 987670088,

1953− "pullRequestNo": 3137

1954− },

1955− {

1956− "name": "knight110001",

1957− "id": 21118983,

1958− "comment_id": 4741392218,

1959− "created_at": "2026-06-18T11:16:02Z",

1960− "repoId": 987670088,

1961− "pullRequestNo": 3140

1962− },

1963− {

1964− "name": "ken-jo",

1965− "id": 42824840,

1966− "comment_id": 4755333253,

1967− "created_at": "2026-06-19T22:56:55Z",

1968− "repoId": 987670088,

1969− "pullRequestNo": 3158

1970− },

1971− {

1972− "name": "xorangekiller",

1973− "id": 5561926,

1974− "comment_id": 4758885426,

1975− "created_at": "2026-06-20T16:00:01Z",

1976− "repoId": 987670088,

1977− "pullRequestNo": 3163

1978− },

1979− {

1980− "name": "sithglan",

1981− "id": 1595655,

1982− "comment_id": 4761251782,

1983− "created_at": "2026-06-21T07:28:19Z",

1984− "repoId": 987670088,

1985− "pullRequestNo": 3168

1986− },

1987− {

1988− "name": "macro-ss",

1989− "id": 59944291,

1990− "comment_id": 4729268228,

1991− "created_at": "2026-06-17T11:09:03Z",

1992− "repoId": 987670088,

1993− "pullRequestNo": 3140

1994− },

1995− {

1996− "name": "andrinoff",

1997− "id": 175145001,

1998− "comment_id": 4772633817,

1999− "created_at": "2026-06-22T20:13:54Z",

2000− "repoId": 987670088,

2001− "pullRequestNo": 3184

2002− },

2003− {

2004− "name": "calyptobai",

2005− "id": 111788964,

2006− "comment_id": 4774213287,

2007− "created_at": "2026-06-22T23:56:09Z",

2008− "repoId": 987670088,

2009− "pullRequestNo": 3189

2010− },

2011− {

2012− "name": "truffle-dev",

2013− "id": 260125384,

2014− "comment_id": 4783906418,

2015− "created_at": "2026-06-23T22:07:28Z",

2016− "repoId": 987670088,

2017− "pullRequestNo": 3195

2018− },

2019− {

2020− "name": "chardoncs",

2021− "id": 44740171,

2022− "comment_id": 4786221902,

2023− "created_at": "2026-06-24T05:36:39Z",

2024− "repoId": 987670088,

2025− "pullRequestNo": 3197

2026− },

2027− {

2028− "name": "notno",

2029− "id": 3149184,

2030− "comment_id": 4793889316,

2031− "created_at": "2026-06-24T21:37:51Z",

2032− "repoId": 987670088,

2033− "pullRequestNo": 3202

2034− },

2035− {

2036− "name": "vismaytiwari",

2037− "id": 89745751,

2038− "comment_id": 4797847556,

2039− "created_at": "2026-06-25T09:38:55Z",

2040− "repoId": 987670088,

2041− "pullRequestNo": 3205

2042− },

2043− {

2044− "name": "feizhuzheng",

2045− "id": 130743669,

2046− "comment_id": 4809880810,

2047− "created_at": "2026-06-26T13:09:43Z",

2048− "repoId": 987670088,

2049− "pullRequestNo": 3211

2050− },

2051− {

2052− "name": "syf2211",

2053− "id": 132282304,

2054− "comment_id": 4817246953,

2055− "created_at": "2026-06-27T11:44:54Z",

2056− "repoId": 987670088,

2057− "pullRequestNo": 3221

2058− },

2059− {

2060− "name": "EduardF1",

2061− "id": 50618110,

2062− "comment_id": 4818319813,

2063− "created_at": "2026-06-27T14:15:50Z",

2064− "repoId": 987670088,

2065− "pullRequestNo": 3133

2066− },

2067− {

2068− "name": "sablea",

2069− "id": 52388727,

2070− "comment_id": 4851979164,

2071− "created_at": "2026-07-01T08:13:03Z",

2072− "repoId": 987670088,

2073− "pullRequestNo": 3235

2074− },

2075− {

2076− "name": "Osamaali313",

2077− "id": 86572800,

2078− "comment_id": 4870561813,

2079− "created_at": "2026-07-02T21:19:15Z",

2080− "repoId": 987670088,

2081− "pullRequestNo": 3244

2082− },

2083− {

2084− "name": "lockp111",

2085− "id": 28383475,

2086− "comment_id": 4878248994,

2087− "created_at": "2026-07-03T17:12:26Z",

2088− "repoId": 987670088,

2089− "pullRequestNo": 3247

2090− },

2091− {

2092− "name": "TheRAWagent",

2093− "id": 77039476,

2094− "comment_id": 4879276761,

2095− "created_at": "2026-07-03T20:52:00Z",

2096− "repoId": 987670088,

2097− "pullRequestNo": 3251

2098 }

1146 }

2099 ]

1147 ]

2100}

1148}

.github/dependabot.yml−1

@@ -24,7 +24,6 @@ updates:

24 - dependency-name: github.com/charmbracelet/lipgloss/v2

25 versions:

26 - v2.0.0-beta1

27− - dependency-name: mvdan.cc/sh/moreinterp

29 - package-ecosystem: "github-actions"

28 - package-ecosystem: "github-actions"

30 directory: "/"

29 directory: "/"

.github/labeler.yml−8

@@ -45,8 +45,6 @@

45 - "/(panic|crash|segfault)/i"

46"provider: anthropic claude":

47 - "/(anthropic|claude)/i"

48−"provider: avian":

49− - "/avian/i"

50"provider: aws bedrock":

48"provider: aws bedrock":

51 - "/(aws|bedrock)/i"

49 - "/(aws|bedrock)/i"

52"provider: azure":

50"provider: azure":

@@ -57,12 +55,8 @@

57 - "/gemini/i"

55 - "/gemini/i"

58"provider: google vertex":

56"provider: google vertex":

59 - "/vertex/i"

57 - "/vertex/i"

60−"provider: hyper":

61− - "/hyper/i"

62"provider: kimi":

58"provider: kimi":

63 - "/kimi/i"

59 - "/kimi/i"

64−"provider: minimax":

65− - "/minimax/i"

66"provider: ollama":

60"provider: ollama":

67 - "/llama/i"

61 - "/llama/i"

68"provider: openai chatgpt":

62"provider: openai chatgpt":

@@ -71,8 +65,6 @@

71 - "/openrouter/i"

65 - "/openrouter/i"

72"provider: qwen":

66"provider: qwen":

73 - "/qwen/i"

67 - "/qwen/i"

74−"provider: vercel":

75− - "/vercel/i"

76"provider: xai grok":

68"provider: xai grok":

77 - "/(xai|x\\.ai|grok)/i"

69 - "/(xai|x\\.ai|grok)/i"

78"security":

70"security":

.github/workflows/build.yml+3−6

@@ -1,8 +1,5 @@

1name: build

2−on:

2+on: [push, pull_request]

3− push:

4− branches: [main]

5− pull_request:

7permissions:

4permissions:

8 contents: read

5 contents: read

@@ -18,10 +15,10 @@ jobs:

18 os: [ubuntu-latest, macos-latest, windows-latest]

15 os: [ubuntu-latest, macos-latest, windows-latest]

19 runs-on: ${{ matrix.os }}

16 runs-on: ${{ matrix.os }}

20 steps:

17 steps:

21− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

18+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

22 with:

19 with:

23 persist-credentials: false

20 persist-credentials: false

24− - uses: actions/setup-go@924ae3a1cded613372ab5595356fb5720e22ba16 # v6.5.0

21+ - uses: actions/setup-go@7a3fe6cf4cb3a834922a1244abfce67bcef6a0c5 # v6.2.0

25 with:

22 with:

26 go-version-file: go.mod

23 go-version-file: go.mod

27 - run: go mod tidy

24 - run: go mod tidy

.github/workflows/cla.yml+2−2

@@ -22,14 +22,14 @@ jobs:

22 github.event.comment.body == 'recheck' ||

23 github.event.comment.body == 'I have read the Contributor License Agreement (CLA) and hereby sign the CLA.' ||

24 github.event_name == 'pull_request_target'

25− uses: contributor-assistant/github-action@58daaf842ac3e26d08c2bdf26770d6e9c5b965a9

25+ uses: contributor-assistant/github-action@ca4a40a7d1004f18d9960b404b97e5f30a505a08 # v2.6.1

26 env:

27 GITHUB_TOKEN: ${{ secrets.PERSONAL_ACCESS_TOKEN }}

28 with:

29 path-to-signatures: ".github/cla-signatures.json"

30 path-to-document: "https://github.com/charmbracelet/crush/blob/main/CLA.md"

31 branch: "main"

32− allowlist: charmcli,charmcrush,dependabot[bot],Copilot

32+ allowlist: charmcli,charmcrush,dependabot[bot]

33 custom-pr-sign-comment: "I have read the Contributor License Agreement (CLA) and hereby sign the CLA."

34 custom-notsigned-prcomment: "Thank you for your submission. We really appreciate it! Like many open-source projects, we ask that you sign our [Contributor License Agreement](https://github.com/charmbracelet/crush/blob/main/CLA.md) before we can accept your contribution. You can sign the CLA by just posting a Pull Request comment same as the below format."

35 lock-pullrequest-aftermerge: false

.github/workflows/labeler.yml+1−1

@@ -14,7 +14,6 @@ on:

15permissions:

16 issues: write

17− pull-requests: write

18 contents: read

17 contents: read

20jobs:

19jobs:

@@ -27,4 +26,5 @@ jobs:

27 enable-versioned-regex: 0

26 enable-versioned-regex: 0

28 include-title: 1

27 include-title: 1

29 include-body: 0

28 include-body: 0

29+ repo-token: ${{ secrets.PERSONAL_ACCESS_TOKEN }}

30 issue-number: ${{ github.event.inputs.issue-number || github.event.issue.number || github.event.pull_request.number }}

.github/workflows/lint.yml+1−2

@@ -1,7 +1,6 @@

1name: lint

2on:

3 push:

4− branches: [main]

5 pull_request:

4 pull_request:

7jobs:

6jobs:

@@ -9,5 +8,5 @@ jobs:

9 uses: charmbracelet/meta/.github/workflows/lint.yml@main

8 uses: charmbracelet/meta/.github/workflows/lint.yml@main

10 with:

9 with:

11 golangci_path: .golangci.yml

10 golangci_path: .golangci.yml

12− golangci_version: v2.11

11+ golangci_version: v2.4

13 timeout: 10m

12 timeout: 10m

.github/workflows/nightly.yml+1−1

@@ -11,7 +11,7 @@ jobs:

11 outputs:

12 should_run: ${{ steps.check.outputs.should_run }}

13 steps:

14− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

14+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

15 with:

16 fetch-depth: 1

17 - id: check

.github/workflows/release.yml+5−1

@@ -15,8 +15,10 @@ jobs:

15 goreleaser:

16 uses: charmbracelet/meta/.github/workflows/goreleaser.yml@main

17 with:

18− go_version: "1.26"

18+ go_version: "1.25"

19 macos_sign_entitlements: "./.github/entitlements.plist"

20+ # XXX: remove it after goreleaser 2.14.

21+ goreleaser_version: nightly

20 secrets:

22 secrets:

21 docker_username: ${{ secrets.DOCKERHUB_USERNAME }}

23 docker_username: ${{ secrets.DOCKERHUB_USERNAME }}

22 docker_token: ${{ secrets.DOCKERHUB_TOKEN }}

24 docker_token: ${{ secrets.DOCKERHUB_TOKEN }}

@@ -25,6 +27,8 @@ jobs:

25 fury_token: ${{ secrets.FURY_TOKEN }}

27 fury_token: ${{ secrets.FURY_TOKEN }}

26 nfpm_gpg_key: ${{ secrets.NFPM_GPG_KEY }}

28 nfpm_gpg_key: ${{ secrets.NFPM_GPG_KEY }}

27 nfpm_passphrase: ${{ secrets.NFPM_PASSPHRASE }}

29 nfpm_passphrase: ${{ secrets.NFPM_PASSPHRASE }}

30+ npm_token: ${{ secrets.NPM_TOKEN }}

31+ snapcraft_token: ${{ secrets.SNAPCRAFT_TOKEN }}

28 aur_key: ${{ secrets.AUR_KEY }}

32 aur_key: ${{ secrets.AUR_KEY }}

29 macos_sign_p12: ${{ secrets.MACOS_SIGN_P12 }}

33 macos_sign_p12: ${{ secrets.MACOS_SIGN_P12 }}

30 macos_sign_password: ${{ secrets.MACOS_SIGN_PASSWORD }}

34 macos_sign_password: ${{ secrets.MACOS_SIGN_PASSWORD }}

.github/workflows/schema-update.yml+3−4

@@ -6,21 +6,20 @@ on:

6 paths:

7 - "internal/config/**"

8 - "internal/agent/hyper/**"

9− workflow_dispatch:

11jobs:

10jobs:

12 update-schema:

11 update-schema:

13 runs-on: ubuntu-latest

12 runs-on: ubuntu-latest

14 steps:

13 steps:

15− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

14+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

16 with:

15 with:

17 token: ${{ secrets.PERSONAL_ACCESS_TOKEN }}

16 token: ${{ secrets.PERSONAL_ACCESS_TOKEN }}

18− - uses: actions/setup-go@924ae3a1cded613372ab5595356fb5720e22ba16 # v6.5.0

17+ - uses: actions/setup-go@7a3fe6cf4cb3a834922a1244abfce67bcef6a0c5 # v6.2.0

19 with:

18 with:

20 go-version-file: go.mod

19 go-version-file: go.mod

21 - run: go run . schema > ./schema.json

20 - run: go run . schema > ./schema.json

22 - run: go generate ./internal/agent/hyper/...

21 - run: go generate ./internal/agent/hyper/...

23− - uses: stefanzweifel/git-auto-commit-action@4a55954c782fc1ea30b9056cd3e7a2b40ca8887d # v5

22+ - uses: stefanzweifel/git-auto-commit-action@04702edda442b2e678b25b537cec683a1493fcb9 # v5

24 with:

23 with:

25 commit_message: "chore: auto-update files"

24 commit_message: "chore: auto-update files"

26 branch: main

25 branch: main

.github/workflows/security.yml+14−14

@@ -27,14 +27,14 @@ jobs:

27 pull-requests: read

28 security-events: write

29 steps:

30− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

30+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

31 with:

32 persist-credentials: false

33− - uses: github/codeql-action/init@54f647b7e1bb85c95cddabcd46b0c578ec92bc1a # v4.36.3

33+ - uses: github/codeql-action/init@19b2f06db2b6f5108140aeb04014ef02b648f789 # v4.31.11

34 with:

35 languages: ${{ matrix.language }}

36− - uses: github/codeql-action/autobuild@54f647b7e1bb85c95cddabcd46b0c578ec92bc1a # v4.36.3

36+ - uses: github/codeql-action/autobuild@19b2f06db2b6f5108140aeb04014ef02b648f789 # v4.31.11

37− - uses: github/codeql-action/analyze@54f647b7e1bb85c95cddabcd46b0c578ec92bc1a # v4.36.3

37+ - uses: github/codeql-action/analyze@19b2f06db2b6f5108140aeb04014ef02b648f789 # v4.31.11

39 grype:

40 runs-on: ubuntu-latest

@@ -43,16 +43,16 @@ jobs:

43 actions: read

44 contents: read

45 steps:

46− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

46+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

47 with:

48 persist-credentials: false

49− - uses: anchore/scan-action@e1165082ffb1fe366ebaf02d8526e7c4989ea9d2 # v7.4.0

49+ - uses: anchore/scan-action@0d444ed77d83ee2ba7f5ced0d90d640a1281d762 # v7.3.0

50 id: scan

51 with:

52 path: "."

53 fail-build: true

54 severity-cutoff: critical

55− - uses: github/codeql-action/upload-sarif@54f647b7e1bb85c95cddabcd46b0c578ec92bc1a # v4.36.3

55+ - uses: github/codeql-action/upload-sarif@19b2f06db2b6f5108140aeb04014ef02b648f789 # v4.31.11

56 with:

57 sarif_file: ${{ steps.scan.outputs.sarif }}

@@ -62,18 +62,18 @@ jobs:

62 security-events: write

63 contents: read

64 steps:

65− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

65+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

66 with:

67 persist-credentials: false

68− - uses: actions/setup-go@924ae3a1cded613372ab5595356fb5720e22ba16 # v6.5.0

68+ - uses: actions/setup-go@7a3fe6cf4cb3a834922a1244abfce67bcef6a0c5 # v6.2.0

69 with:

70− go-version: stable

70+ go-version: 1.26.0-rc.1 # change to "stable" once Go 1.26 is released

71 - name: Install govulncheck

72 run: go install golang.org/x/vuln/cmd/govulncheck@latest

73 - name: Run govulncheck

74 run: |

75 govulncheck -C . -format sarif ./... > results.sarif

76− - uses: github/codeql-action/upload-sarif@54f647b7e1bb85c95cddabcd46b0c578ec92bc1a # v4.36.3

76+ - uses: github/codeql-action/upload-sarif@19b2f06db2b6f5108140aeb04014ef02b648f789 # v4.31.11

77 with:

78 sarif_file: results.sarif

@@ -83,10 +83,10 @@ jobs:

83 permissions:

84 contents: read

85 steps:

86− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

86+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

87 with:

88 persist-credentials: false

89− - uses: actions/dependency-review-action@a1d282b36b6f3519aa1f3fc636f609c47dddb294 # v5.0.0

89+ - uses: actions/dependency-review-action@3c4e3dcb1aa7874d2c16be7d79418e9b7efd6261 # v4.8.2

90 with:

91 fail-on-severity: critical

92− allow-licenses: BSD-2-Clause, BSD-3-Clause, MIT, MIT-0, Apache-2.0, MPL-2.0, ISC, LicenseRef-scancode-google-patent-license-golang, Unlicense

92+ allow-licenses: BSD-2-Clause, BSD-3-Clause, MIT, Apache-2.0, MPL-2.0, ISC, LicenseRef-scancode-google-patent-license-golang

.github/workflows/snapshot.yml+3−3

@@ -18,14 +18,14 @@ jobs:

18 # Use our own large runners with more CPU and RAM for faster builds

19 group: releasers

20 steps:

21− - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

21+ - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

22 with:

23 fetch-depth: 0

24 persist-credentials: false

25− - uses: actions/setup-go@924ae3a1cded613372ab5595356fb5720e22ba16 # v6.5.0

25+ - uses: actions/setup-go@7a3fe6cf4cb3a834922a1244abfce67bcef6a0c5 # v6.2.0

26 with:

27 go-version-file: go.mod

28− - uses: goreleaser/goreleaser-action@f06c13b6b1a9625abc9e6e439d9c05a8f2190e94 # v7.2.3

28+ - uses: goreleaser/goreleaser-action@e435ccd777264be153ace6237001ef4d979d3a7a # v6.4.0

29 with:

30 version: "nightly"

31 distribution: goreleaser-pro

.gitignore−5

@@ -42,10 +42,6 @@ Thumbs.db

42.env

43.env.local

45−# Nix/direnv

46−.direnv/

47−/result

48−

49**/.crush/**

45**/.crush/**

51/crush

47/crush

@@ -54,4 +50,3 @@ Thumbs.db

54manpages/

50manpages/

55completions/crush.*sh

51completions/crush.*sh

56.prettierignore

52.prettierignore

57−.task

.golangci.yml+4−16

@@ -2,7 +2,6 @@ version: "2"

2linters:

3 enable:

4 - bodyclose

5− - errcheck

6 # - exhaustive

5 # - exhaustive

7 # - goconst

6 # - goconst

8 # - godot

7 # - godot

@@ -10,7 +9,6 @@ linters:

10 # - gomoddirectives

9 # - gomoddirectives

11 - goprintffuncname

10 - goprintffuncname

12 # - gosec

11 # - gosec

13− - ineffassign

14 - misspell

12 - misspell

15 # - nakedret

13 # - nakedret

16 # - nestif

14 # - nestif

@@ -25,30 +23,20 @@ linters:

25 - tparallel

23 - tparallel

26 # - unconvert

24 # - unconvert

27 # - unparam

25 # - unparam

28− - unused

29 - whitespace

26 - whitespace

30 # - wrapcheck

27 # - wrapcheck

28+ disable:

29+ - errcheck

30+ - ineffassign

31+ - unused

31 exclusions:

32 exclusions:

32 generated: lax

33 generated: lax

33 presets:

34 presets:

34 - common-false-positives

35 - common-false-positives

35− # Ignores the idiomatic-to-drop error family (*Close, *Flush,

36− # Fprint*, os.Remove/RemoveAll, os.Setenv, stdout/stderr writes) so

37− # errcheck only flags genuinely meaningful dropped errors.

38− - std-error-handling

39 rules:

36 rules:

40− # Dropped errors in tests (setup writes, httptest encoders) are

41− # low-signal — test failures surface them anyway.

42− - path: _test\.go

43− linters:

44− - errcheck

45 - text: '(slog|log)\.\w+'

37 - text: '(slog|log)\.\w+'

46 linters:

38 linters:

47 - noctx

39 - noctx

48− - path: internal/shell/run.go

49− text: "SA1019.*interp.ExecHandler"

50− linters:

51− - staticcheck

52issues:

40issues:

53 max-issues-per-linter: 0

41 max-issues-per-linter: 0

54 max-same-issues: 0

42 max-same-issues: 0

.goreleaser.yml−2

@@ -297,8 +297,6 @@ winget:

297 owner: microsoft

298 name: winget-pkgs

299 branch: master

300− body: |

301− /cc @andreynering

302

300

303changelog:

301changelog:

304 sort: asc

302 sort: asc

AGENTS.md+24−170

@@ -1,141 +1,38 @@

1# Crush Development Guide

3−## Project Overview

4−

5−Crush is a terminal-based AI coding assistant built in Go by

6−[Charm](https://charm.land). It connects to LLMs and gives them tools to read,

7−write, and execute code. It supports multiple providers (Anthropic, OpenAI,

8−Gemini, Bedrock, Copilot, Hyper, MiniMax, Vercel, and more), integrates with

9−LSPs for code intelligence, and supports extensibility via MCP servers and

10−agent skills.

11−

12−The module path is `github.com/charmbracelet/crush`.

13−

14−## Architecture

15−

16−```

17−main.go CLI entry point (cobra via internal/cmd)

18−internal/

19− app/app.go Top-level wiring: DB, config, agents, LSP, MCP, events

20− cmd/ CLI commands (root, run, login, models, stats, sessions)

21− config/

22− config.go Config struct, context file paths, agent definitions

23− load.go crush.json loading and validation

24− provider.go Provider configuration and model resolution

25− agent/

26− agent.go SessionAgent: runs LLM conversations per session

27− coordinator.go Coordinator: manages named agents ("coder", "task")

28− hooked_tool.go Decorator that runs PreToolUse hooks before tool execution

29− prompts.go Loads Go-template system prompts

30− templates/ System prompt templates (coder.md.tpl, task.md.tpl, etc.)

31− tools/ All built-in tools (bash, edit, view, grep, glob, etc.)

32− mcp/ MCP client integration

33− hooks/ Hook engine: runs user shell commands on hook events

34− hooks.go Decision types, aggregation logic, event constants

35− runner.go Parallel hook execution, timeout, dedup

36− input.go Stdin payload builder, env vars, stdout parsing (Crush + Claude Code compat)

37− session/session.go Session CRUD backed by SQLite

38− message/ Message model and content types

39− db/ SQLite via sqlc, with migrations

40− sql/ Raw SQL queries (consumed by sqlc)

41− migrations/ Schema migrations

42− lsp/ LSP client manager, auto-discovery, on-demand startup

43− ui/ Bubble Tea v2 TUI (see internal/ui/AGENTS.md)

44− permission/ Tool permission checking and allow-lists

45− skills/ Skill file discovery and loading

46− shell/ Bash command execution with background job support

47− event/ Telemetry (PostHog)

48− pubsub/ Internal pub/sub for cross-component messaging

49− filetracker/ Tracks files touched per session

50− history/ Prompt history

51−```

52−

53−### Key Dependency Roles

54−

55−- **`charm.land/fantasy`**: LLM provider abstraction layer. Handles protocol

56− differences between Anthropic, OpenAI, Gemini, etc. Used in `internal/app`

57− and `internal/agent`.

58−- **`charm.land/bubbletea/v2`**: TUI framework powering the interactive UI.

59−- **`charm.land/lipgloss/v2`**: Terminal styling.

60−- **`charm.land/glamour/v2`**: Markdown rendering in the terminal.

61−- **`charm.land/catwalk`**: Snapshot/golden-file testing for TUI components.

62−- **`sqlc`**: Generates Go code from SQL queries in `internal/db/sql/`.

63−

64−### Key Patterns

65−

66−- **Config is a Service**: accessed via `config.Service`, not global state.

67−- **Tools are self-documenting**: each tool has a `.go` implementation and a

68− `.md` description file in `internal/agent/tools/`.

69−- **System prompts are Go templates**: `internal/agent/templates/*.md.tpl`

70− with runtime data injected.

71−- **Context files**: Crush reads AGENTS.md, CRUSH.md, CLAUDE.md, GEMINI.md

72− (and `.local` variants) from the working directory for project-specific

73− instructions.

74−- **Persistence**: SQLite + sqlc. All queries live in `internal/db/sql/`,

75− generated code in `internal/db/`. Migrations in `internal/db/migrations/`.

76−- **Pub/sub**: `internal/pubsub` for decoupled communication between agent,

77− UI, and services.

78−- **Hooks**: User-defined shell commands in `crush.json` that fire before

79− tool execution. The engine (`internal/hooks/`) is independent of fantasy

80− and agent — it takes inputs, runs commands, returns decisions. The

81− `hookedTool` decorator in `internal/agent/hooked_tool.go` wraps tools at

82− the coordinator level. Hooks run before permission checks. See

83− `HOOKS.md` for the user-facing protocol.

84−- **CGO disabled**: builds with `CGO_ENABLED=0` and

85− `GOEXPERIMENT=greenteagc`.

86−

87## Build/Test/Lint Commands

3## Build/Test/Lint Commands

89- **Build**: `go build .` or `go run .`

5- **Build**: `go build .` or `go run .`

90−- **Test**: `task test` or `go test ./...` (run single test:

91− `go test ./internal/llm/prompt -run TestGetContextFromPaths`)

6+- **Test**: `task test` or `go test ./...` (run single test: `go test ./internal/llm/prompt -run TestGetContextFromPaths`)

92−- **Update Golden Files**: `go test ./... -update` (regenerates `.golden`

7+- **Update Golden Files**: `go test ./... -update` (regenerates .golden files when test output changes)

93− files when test output changes)

94− - Update specific package:

95− `go test ./internal/tui/components/core -update` (in this case,

8+ - Update specific package: `go test ./internal/tui/components/core -update` (in this case, we're updating "core")

96− we're updating "core")

97- **Lint**: `task lint:fix`

9- **Lint**: `task lint:fix`

98−- **Format**: `task fmt` (`gofumpt -w .`)

10+- **Format**: `task fmt` (gofumpt -w .)

99−- **Modernize**: `task modernize` (runs `modernize` which makes code

100− simplifications)

101- **Dev**: `task dev` (runs with profiling enabled)

11- **Dev**: `task dev` (runs with profiling enabled)

102

103## Code Style Guidelines

13## Code Style Guidelines

104

105−- **Imports**: Use `goimports` formatting, group stdlib, external, internal

15+- **Imports**: Use goimports formatting, group stdlib, external, internal packages

106− packages.

107−- **Formatting**: Use gofumpt (stricter than gofmt), enabled in

16+- **Formatting**: Use gofumpt (stricter than gofmt), enabled in golangci-lint

108− golangci-lint.

109−- **Naming**: Standard Go conventions — PascalCase for exported, camelCase

17+- **Naming**: Standard Go conventions - PascalCase for exported, camelCase for unexported

110− for unexported.

111−- **Types**: Prefer explicit types, use type aliases for clarity (e.g.,

18+- **Types**: Prefer explicit types, use type aliases for clarity (e.g., `type AgentName string`)

112− `type AgentName string`).

113−- **Error handling**: Return errors explicitly, use `fmt.Errorf` for

19+- **Error handling**: Return errors explicitly, use `fmt.Errorf` for wrapping

114− wrapping.

115−- **Context**: Always pass `context.Context` as first parameter for

20+- **Context**: Always pass context.Context as first parameter for operations

116− operations.

117−- **Interfaces**: Define interfaces in consuming packages, keep them small

21+- **Interfaces**: Define interfaces in consuming packages, keep them small and focused

118− and focused.

119−- **Structs**: Use struct embedding for composition, group related fields.

22+- **Structs**: Use struct embedding for composition, group related fields

120−- **Constants**: Use typed constants with iota for enums, group in const

23+- **Constants**: Use typed constants with iota for enums, group in const blocks

121− blocks.

122−- **Testing**: Use testify's `require` package, parallel tests with

24+- **Testing**: Use testify's `require` package, parallel tests with `t.Parallel()`,

123− `t.Parallel()`, `t.SetEnv()` to set environment variables. Always use

25+ `t.SetEnv()` to set environment variables. Always use `t.Tempdir()` when in

124− `t.Tempdir()` when in need of a temporary directory. This directory does

26+ need of a temporary directory. This directory does not need to be removed.

125− not need to be removed.

126−- **JSON tags**: Use snake_case for JSON field names.

27+- **JSON tags**: Use snake_case for JSON field names

127−- **File permissions**: Use octal notation (0o755, 0o644) for file

28+- **File permissions**: Use octal notation (0o755, 0o644) for file permissions

128− permissions.

129−- **Log messages**: Log messages must start with a capital letter (e.g.,

29+- **Log messages**: Log messages must start with a capital letter (e.g., "Failed to save session" not "failed to save session")

130− "Failed to save session" not "failed to save session").

131− - This is enforced by `task lint:log` which runs as part of `task lint`.

30+ - This is enforced by `task lint:log` which runs as part of `task lint`

132−- **Comments**: End comments in periods unless comments are at the end of the

31+- **Comments**: End comments in periods unless comments are at the end of the line.

133− line.

134

135## Testing with Mock Providers

33## Testing with Mock Providers

136

137−When writing tests that involve provider configurations, use the mock

35+When writing tests that involve provider configurations, use the mock providers to avoid API calls:

138−providers to avoid API calls:

139

140```go

37```go

141func TestYourFunction(t *testing.T) {

38func TestYourFunction(t *testing.T) {

@@ -172,52 +69,9 @@ func TestYourFunction(t *testing.T) {

172

173## Committing

70## Committing

174

175−- ALWAYS use semantic commits (`fix:`, `feat:`, `chore:`, `refactor:`,

72+- ALWAYS use semantic commits (`fix:`, `feat:`, `chore:`, `refactor:`, `docs:`, `sec:`, etc).

176− `docs:`, `sec:`, etc).

177- Try to keep commits to one line, not including your attribution. Only use

73- Try to keep commits to one line, not including your attribution. Only use

178 multi-line commits when additional context is truly necessary.

74 multi-line commits when additional context is truly necessary.

179

180## Working on the TUI (UI)

76## Working on the TUI (UI)

181−

182−Anytime you need to work on the TUI, read `internal/ui/AGENTS.md` before

77+Anytime you need to work on the tui before starting work read the internal/ui/AGENTS.md file

183−starting work.

184−

185−## Styling System

186−

187−The styling system lives in `internal/ui/styles/` and is organized into

188−three layers:

189−

190−- **`quickstyle.go`**: The stable base theme builder. `quickStyle(opts)`

191− constructs a `Styles` struct from `quickStyleOpts` — a palette of

192− design tokens (primary, secondary, fgBase, bgBase, success, error, etc.).

193− `quickStyle` must be fully token-driven: never hardcode specific

194− `charmtone.*` colors here (except Chroma syntax highlighting, which is

195− pending tokenization). This lets any theme reuse the base without

196− inheriting Charmtone-specific colors.

197−- **`themes.go`**: Defines concrete themes. Each theme function (e.g.

198− `CharmtonePantera`) calls `quickStyle` with its palette, then applies

199− theme-specific overrides as needed.

200−- **`styles.go`**: Defines the `Styles` struct and its documentation —

201− the shape of what `quickStyle` produces.

202−

203−**Adding theme-specific overrides**: When a style genuinely needs a

204−color that doesn't fit the token model (e.g. the bang prompt uses

205−Salt/Hazy/Larple), keep `quickStyle` on the closest semantic token and

206−override only the differing colors in the theme function:

207−

208−```go

209−func CharmtonePantera() Styles {

210− s := quickStyle(quickStyleOpts{ /* palette */ })

211−

212− // Override only the colors that differ from the token defaults.

213− s.Editor.PromptBangIconFocused = s.Editor.PromptBangIconFocused.

214− Foreground(charmtone.Salt).

215− Background(charmtone.Hazy)

216−

217− return s

218−}

219−```

220−

221−**Adding a new theme**: Add a function in `themes.go` that returns the

222−result of `quickStyle` with a `quickStyleOpts` palette (plus any needed

223−overrides), then wire it into `ThemeForProvider`.

LICENSE.md+1−1

@@ -6,7 +6,7 @@ FSL-1.1-MIT

7## Notice

11## Terms and Conditions

README.md+48−271

@@ -1,8 +1,5 @@

1# Crush

3−> [!NOTE]

4−> This is a personal fork of [charmbracelet/crush](https://github.com/charmbracelet/crush) with UI/UX tweaks and performance improvements.

5−

6

3

7 <a href="https://stuff.charm.sh/crush/charm-crush.png"><img width="450" alt="Charm Crush Logo" src="https://github.com/user-attachments/assets/cf8ca3ce-8b02-43f0-9d0f-5a331488da4b" /></a>

4 <a href="https://stuff.charm.sh/crush/charm-crush.png"><img width="450" alt="Charm Crush Logo" src="https://github.com/user-attachments/assets/cf8ca3ce-8b02-43f0-9d0f-5a331488da4b" /></a>

8 <a href="https://github.com/charmbracelet/crush/releases"><img src="https://img.shields.io/github/release/charmbracelet/crush" alt="Latest Release"></a>

5 <a href="https://github.com/charmbracelet/crush/releases"><img src="https://img.shields.io/github/release/charmbracelet/crush" alt="Latest Release"></a>

@@ -172,7 +169,7 @@ go install github.com/charmbracelet/crush@latest

172> [!WARNING]

169> [!WARNING]

173> Productivity may increase when using Crush and you may find yourself nerd

170> Productivity may increase when using Crush and you may find yourself nerd

174> sniped when first using the application. If the symptoms persist, join the

171> sniped when first using the application. If the symptoms persist, join the

175−> [Slack][slack] or [Discord][discord] and nerd snipe the rest of us.

172+> [Discord][discord] and nerd snipe the rest of us.

176

173

177## Getting Started

174## Getting Started

178

175

@@ -184,23 +181,16 @@ That said, you can also set environment variables for preferred providers.

184

181

185| Environment Variable | Provider |

182| Environment Variable | Provider |

186| --------------------------- | -------------------------------------------------- |

183| --------------------------- | -------------------------------------------------- |

187−| `HYPER_API_KEY` | Charm Hyper |

188| `ANTHROPIC_API_KEY` | Anthropic |

184| `ANTHROPIC_API_KEY` | Anthropic |

189| `OPENAI_API_KEY` | OpenAI |

185| `OPENAI_API_KEY` | OpenAI |

190| `VERCEL_API_KEY` | Vercel AI Gateway |

186| `VERCEL_API_KEY` | Vercel AI Gateway |

191| `GEMINI_API_KEY` | Google Gemini |

187| `GEMINI_API_KEY` | Google Gemini |

192| `SYNTHETIC_API_KEY` | Synthetic |

188| `SYNTHETIC_API_KEY` | Synthetic |

193| `ZAI_API_KEY` | Z.ai |

189| `ZAI_API_KEY` | Z.ai |

194−| `MINIMAX_API_KEY` | MiniMax |

195| `HF_TOKEN` | Hugging Face Inference |

190| `HF_TOKEN` | Hugging Face Inference |

196| `CEREBRAS_API_KEY` | Cerebras |

191| `CEREBRAS_API_KEY` | Cerebras |

197| `OPENROUTER_API_KEY` | OpenRouter |

192| `OPENROUTER_API_KEY` | OpenRouter |

198−| `IONET_API_KEY` | io.net |

199−| `ALIBABA_SINGAPORE_API_KEY` | Alibaba (Singapore) |

200−| `ALIBABA_US_API_KEY` | Alibaba (United States) |

201| `GROQ_API_KEY` | Groq |

193| `GROQ_API_KEY` | Groq |

202−| `AVIAN_API_KEY` | Avian |

203−| `OPENCODE_API_KEY` | OpenCode Zen & Go |

204| `VERTEXAI_PROJECT` | Google Cloud VertexAI (Gemini) |

194| `VERTEXAI_PROJECT` | Google Cloud VertexAI (Gemini) |

205| `VERTEXAI_LOCATION` | Google Cloud VertexAI (Gemini) |

195| `VERTEXAI_LOCATION` | Google Cloud VertexAI (Gemini) |

206| `AWS_ACCESS_KEY_ID` | Amazon Bedrock (Claude) |

196| `AWS_ACCESS_KEY_ID` | Amazon Bedrock (Claude) |

@@ -211,17 +201,6 @@ That said, you can also set environment variables for preferred providers.

211| `AZURE_OPENAI_API_ENDPOINT` | Azure OpenAI models |

201| `AZURE_OPENAI_API_ENDPOINT` | Azure OpenAI models |

212| `AZURE_OPENAI_API_KEY` | Azure OpenAI models (optional when using Entra ID) |

202| `AZURE_OPENAI_API_KEY` | Azure OpenAI models (optional when using Entra ID) |

213| `AZURE_OPENAI_API_VERSION` | Azure OpenAI models |

203| `AZURE_OPENAI_API_VERSION` | Azure OpenAI models |

214−| `MOONSHOT_API_KEY` | Moonshot |

215−

216−### Subscriptions

217−

218−If you prefer subscription-based usage, here are some plans that work well in

219−Crush:

220−

221−- [Synthetic](https://synthetic.new/pricing)

222−- [GLM Coding Plan](https://z.ai/subscribe)

223−- [Kimi Code](https://www.kimi.com/membership/pricing)

224−- [MiniMax Coding Plan](https://platform.minimax.io/subscribe/coding-plan)

225

204

226### By the Way

205### By the Way

227

206

@@ -233,10 +212,6 @@ Crush’s default model listing is managed in [Catwalk](https://github.com/charm

233

212

234## Configuration

213## Configuration

235

214

236−> [!TIP]

237−> Crush ships with a builtin `crush-config` skill for configuring itself. In

238−> many cases you can simply ask Crush to configure itself.

239−

240Crush runs great with no configuration. That said, if you do need or want to

215Crush runs great with no configuration. That said, if you do need or want to

241customize Crush, configuration can be added either local to the project itself,

216customize Crush, configuration can be added either local to the project itself,

242or globally, with the following priority:

217or globally, with the following priority:

@@ -254,8 +229,7 @@ Configuration itself is stored as a JSON object:

254}

229}

255```

230```

256

231

257−As an additional note, Crush also stores ephemeral data, such as application

232+As an additional note, Crush also stores ephemeral data, such as application state, in one additional location:

258−state, in one additional location:

259

233

260```bash

234```bash

261# Unix

235# Unix

@@ -267,9 +241,8 @@ $HOME/.local/share/crush/crush.json

267

241

268> [!TIP]

242> [!TIP]

269> You can override the user and data config locations by setting:

243> You can override the user and data config locations by setting:

270−>

271−> - `CRUSH_GLOBAL_CONFIG`

244+> * `CRUSH_GLOBAL_CONFIG`

272−> - `CRUSH_GLOBAL_DATA`

245+> * `CRUSH_GLOBAL_DATA`

273

246

274### LSPs

247### LSPs

275

248

@@ -299,36 +272,10 @@ like you would. LSPs can be added manually like so:

299

272

300### MCPs

273### MCPs

301

274

302−Crush also supports Model Context Protocol (MCP) servers through three transport

275+Crush also supports Model Context Protocol (MCP) servers through three

303−types: `stdio` for command-line servers, `http` for HTTP endpoints, and `sse`

276+transport types: `stdio` for command-line servers, `http` for HTTP endpoints,

304−for Server-Sent Events.

277+and `sse` for Server-Sent Events. Environment variable expansion is supported

305−

278+using `$(echo $VAR)` syntax.

306−Shell-style value expansion (`$VAR`, `${VAR:-default}`, `$(command)`, quoting,

307−nesting) works in `command`, `args`, `env`, `headers`, and `url`, so

308−file-based secrets work out of the box. You can use values like `"$TOKEN"`

309−or `"$(cat /path/to/secret/token)"`. Expansion runs through Crush's embedded

310−shell, so the same syntax works on every supported system, Windows included.

311−

312−Unset variables expand to the empty string by default, matching bash. For

313−required credentials, use `${VAR:?message}` so an unset variable fails loudly

314−at load time with `message` instead of silently resolving to empty:

315−

316−```json

317−{ "api_key": "${CODEBERG_TOKEN:?set CODEBERG_TOKEN}" }

318−```

319−

320−Headers (both MCP `headers` and provider `extra_headers`) whose value

321−resolves to the empty string are dropped from the outgoing request rather

322−than sent as `Header:`. That keeps optional env-gated headers like

323−`"OpenAI-Organization": "$OPENAI_ORG_ID"` clean when the variable is unset.

324−

325−Provider `extra_body` is a non-expanding JSON passthrough; put env-driven

326−values in `extra_headers` or the provider's `api_key` / `base_url`, all of

327−which do expand.

328−

329−> **Security note:** `crush.json` is trusted code. Any `$(...)` in it runs at

330−> load time with your shell's privileges, before the UI appears. Don't launch

331−> Crush in a directory whose `crush.json` you haven't reviewed.

332

279

333```json

280```json

334{

281{

@@ -368,72 +315,6 @@ which do expand.

368}

315}

369```

316```

370

317

371−### Hooks

372−

373−Crush has preliminary support for hooks. For details, see

374−[the hook guide](./docs/hooks/).

375−

376−### Sharing a workspace across clients

377−

378−When Crush is run against a shared backend (for example two TUIs talking to

379−the same `crush serve`), clients are grouped into **workspaces** keyed by

380−their resolved `--cwd`. Two clients with the same `--cwd` join the same

381−underlying workspace, so they share the session list, message history,

382−permission queue, LSP, and MCP state.

383−

384−Joining is implicit: pointing a second client at the same working directory

385−attaches it to the existing workspace. Each new invocation, however, starts

386−in its own fresh session by default. To pick up the conversation another

387−client already has open, use the session manager (the session picker) and

388−select it. Sessions surface two signals there:

389−

390−- `IsBusy` is set while an agent turn is in flight for that session.

391−- `AttachedClients` reports how many clients are currently viewing it.

392−

393−A non-zero `AttachedClients` (often combined with `IsBusy`) is the cue that a

394−session is "in progress" on another client and joining it will mirror that

395−view live.

396−

397−The first client to create a workspace fixes its process-wide flags. In

398−particular, `--yolo` and `--debug` follow a **first-wins** rule: later

399−clients that arrive at the same `--cwd` with different values for those

400−flags do not change the running workspace. A debug log line is emitted

401−recording the mismatch, and the workspace keeps the flags it was created

402−with.

403−

404−A workspace lives as long as at least one client has an SSE event stream

405−open against it. When the last stream disconnects, the workspace is torn

406−down. There is a short grace window right after `POST /v1/workspaces` so a

407−client that has created the workspace but not yet opened its event stream

408−does not get reaped before it can attach.

409−

410−### Global context files

411−

412−Crush automatically includes two files for cross-project instructions.

413−

414−- `~/.config/crush/CRUSH.md`: Crush-specific rules that would confuse other

415− agentic coding tools. If you only use Crush, this is the only one you need to

416− edit.

417−- `~/.config/AGENTS.md`: generic instructions that other coding tools might

418− read. Avoid referring to Crush-specific features or workflows here. You

419− probably only care about this if you use multiple agentic coding tools and

420− want to share instructions between them.

421−

422−You can customize these paths using the `global_context_paths` option in your

423−configuration:

424−

425−```jsonc

426−{

427− "$schema": "https://charm.land/crush.json",

428− "options": {

429− "global_context_paths": [

430− "~/path/to/custom/context/file.md",

431− "/full/path/to/folder/of/files/" // recursively load all .md files in folder

432− ]

433− }

434−}

435−```

436−

437### Ignoring Files

318### Ignoring Files

438

319

439Crush respects `.gitignore` files by default, but you can also create a

320Crush respects `.gitignore` files by default, but you can also create a

@@ -478,28 +359,16 @@ completely hidden from the agent.

478{

359{

479 "$schema": "https://charm.land/crush.json",

360 "$schema": "https://charm.land/crush.json",

480 "options": {

361 "options": {

481− "disabled_tools": ["bash", "sourcegraph"]

362+ "disabled_tools": [

363+ "bash",

364+ "sourcegraph"

365+ ]

482 }

366 }

483}

367}

484```

368```

485

369

486To disable tools from MCP servers, see the [MCP config section](#mcps).

370To disable tools from MCP servers, see the [MCP config section](#mcps).

487

371

488−### Disabling Skills

489−

490−If you'd like to prevent Crush from using certain skills entirely, you can

491−disable them via the `options.disabled_skills` list. Disabled skills are hidden

492−from the agent, including builtin skills and skills discovered from disk.

493−

494−```json

495−{

496− "$schema": "https://charm.land/crush.json",

497− "options": {

498− "disabled_skills": ["crush-config"]

499− }

500−}

501−```

502−

503### Agent Skills

372### Agent Skills

504

373

505Crush supports the [Agent Skills](https://agentskills.io) open standard for

374Crush supports the [Agent Skills](https://agentskills.io) open standard for

@@ -507,25 +376,11 @@ extending agent capabilities with reusable skill packages. Skills are folders

507containing a `SKILL.md` file with instructions that Crush can discover and

376containing a `SKILL.md` file with instructions that Crush can discover and

508activate on demand.

377activate on demand.

509

378

510−The global paths we looks for skills are:

379+Skills are discovered from:

511−

512−* `$CRUSH_SKILLS_DIR`

513−* `$XDG_CONFIG_HOME/agents/skills` or `~/.config/agents/skills/`

514−* `$XDG_CONFIG_HOME/crush/skills` or `~/.config/crush/skills/`

515−* `~/.agents/skills/`

516−* `~/.claude/skills/`

517−* On Windows, we _also_ look at

518− * `%LOCALAPPDATA%\agents\skills\` or `%USERPROFILE%\AppData\Local\agents\skills\`

519− * `%LOCALAPPDATA%\crush\skills\` or `%USERPROFILE%\AppData\Local\crush\skills\`

520−* Additional paths configured via `options.skills_paths`

521−

522−On top of that, we _also_ load skills in your project from the following

523−relative paths:

524

380

525−* `.agents/skills`

381+- `~/.config/crush/skills/` on Unix (default, can be overridden with `CRUSH_SKILLS_DIR`)

526−* `.crush/skills`

382+- `%LOCALAPPDATA%\crush\skills\` on Windows (default, can be overridden with `CRUSH_SKILLS_DIR`)

527−* `.claude/skills`

383+- Additional paths configured via `options.skills_paths`

528−* `.cursor/skills`

529

384

530```jsonc

385```jsonc

531{

386{

@@ -533,9 +388,9 @@ relative paths:

533 "options": {

388 "options": {

534 "skills_paths": [

389 "skills_paths": [

535 "~/.config/crush/skills", // Windows: "%LOCALAPPDATA%\\crush\\skills",

390 "~/.config/crush/skills", // Windows: "%LOCALAPPDATA%\\crush\\skills",

536− "./project-skills",

391+ "./project-skills"

537− ],

392+ ]

538− },

393+ }

539}

394}

540```

395```

541

396

@@ -557,56 +412,6 @@ git clone https://github.com/anthropics/skills.git _temp

557mv _temp/skills/* . ; rm -r -force _temp

412mv _temp/skills/* . ; rm -r -force _temp

558```

413```

559

414

560−#### User-Invocable Skills

561−

562−Skills can be made invocable as commands from the commands palette (Ctrl+P). Add `user-invocable: true` to the skill's YAML frontmatter:

563−

564−```yaml

565−---

566−name: my-skill

567−description: A skill that can be invoked as a command.

568−user-invocable: true

569−---

570−```

571−

572−User-invocable skills appear in the commands palette with a `user:` or `project:` prefix:

573−- Skills from global directories show as `user:skill-name`

574−- Skills from project directories show as `project:skill-name`

575−

576−When invoked, the skill's instructions are loaded into the conversation context.

577−

578−To prevent the model from auto-triggering a skill (while still allowing user invocation), add `disable-model-invocation: true`:

579−

580−```yaml

581−---

582−name: my-skill

583−description: Only invocable by users, not the model.

584−user-invocable: true

585−disable-model-invocation: true

586−---

587−```

588−

589−Skills with `disable-model-invocation` won't appear in the model's available skills list but can still be invoked manually by users.

590−

591−### Desktop notifications

592−

593−Crush sends desktop notifications when a tool call requires permission and when

594−the agent finishes its turn. They're only sent when the terminal window isn't

595−focused _and_ your terminal supports reporting the focus state.

596−

597−```jsonc

598−{

599− "$schema": "https://charm.land/crush.json",

600− "options": {

601− "disable_notifications": false, // default

602− },

603−}

604−```

605−

606−To disable desktop notifications, set `disable_notifications` to `true` in your

607−configuration. On macOS, notifications currently lack icons due to platform

608−limitations.

609−

610### Initialization

415### Initialization

611

416

612When you initialize a project, Crush analyzes your codebase and creates

417When you initialize a project, Crush analyzes your codebase and creates

@@ -648,9 +453,10 @@ it creates. You can customize this behavior with the `attribution` option:

648

453

649- `trailer_style`: Controls the attribution trailer added to commit messages

454- `trailer_style`: Controls the attribution trailer added to commit messages

650 (default: `assisted-by`)

455 (default: `assisted-by`)

651− - `assisted-by`: Adds `Assisted-by: Crush:[ModelID]` as specified in [the convention](https://docs.kernel.org/process/coding-assistants.html#attribution)

456+ - `assisted-by`: Adds `Assisted-by: [Model Name] via Crush <crush@charm.land>`

457+ (includes the model name)

652− - `co-authored-by`: Adds `Co-Authored-By: Crush <crush@charm.land>`

458+ - `co-authored-by`: Adds `Co-Authored-By: Crush <crush@charm.land>`

653− - `none`: No attribution trailer

459+ - `none`: No attribution trailer

654- `generated_with`: When true (default), adds `💘 Generated with Crush` line to

460- `generated_with`: When true (default), adds `💘 Generated with Crush` line to

655 commit messages and PR descriptions

461 commit messages and PR descriptions

656

462

@@ -662,9 +468,8 @@ Anthropic-compatible APIs.

662> [!NOTE]

468> [!NOTE]

663> Note that we support two "types" for OpenAI. Make sure to choose the right one

469> Note that we support two "types" for OpenAI. Make sure to choose the right one

664> to ensure the best experience!

470> to ensure the best experience!

665−>

666−> - `openai` should be used when proxying or routing requests through OpenAI.

471+> * `openai` should be used when proxying or routing requests through OpenAI.

667−> - `openai-compat` should be used when using non-OpenAI providers that have OpenAI-compatible APIs.

472+> * `openai-compat` should be used when using non-OpenAI providers that have OpenAI-compatible APIs.

668

473

669#### OpenAI-Compatible APIs

474#### OpenAI-Compatible APIs

670

475

@@ -775,10 +580,9 @@ To add specific models to the configuration, configure as such:

775

580

776### Local Models

581### Local Models

777

582

778−Crush can auto-discovers models from local providers. Add a custom provider

583+Local models can also be configured via OpenAI-compatible API. Here are two common examples:

779−with `type` set to `llamacpp`, `omlx`, `lmstudio`, `litellm`, or `ollama`

584+

780−and leave out the models list. Crush will populate the model list

585+#### Ollama

781−automatically.

782

586

783```json

587```json

784{

588{

@@ -786,50 +590,37 @@ automatically.

786 "ollama": {

590 "ollama": {

787 "name": "Ollama",

591 "name": "Ollama",

788 "base_url": "http://localhost:11434/v1/",

592 "base_url": "http://localhost:11434/v1/",

789− "type": "ollama"

593+ "type": "openai-compat",

790− }

594+ "models": [

791− }

595+ {

792−}

596+ "name": "Qwen 3 30B",

793−```

597+ "id": "qwen3:30b",

794−

598+ "context_window": 256000,

795−For llama.cpp (`llama-server`), point at the server's base URL:

599+ "default_max_tokens": 20000

796−

600+ }

797−```json

601+ ]

798−{

799− "providers": {

800− "llamacpp": {

801− "name": "llama.cpp",

802− "base_url": "http://localhost:2222",

803− "type": "llamacpp"

804 }

602 }

805 }

603 }

806}

604}

807```

605```

808

606

809−#### Manual Model Configuration

607+#### LM Studio

810−

811−You can still list models explicitly. User-defined models always take

812−precedence over discovered ones, and any fields you set won't be overwritten

813−by auto-discovery. Auto discovery will run if the model list is empty for any

814−`openai-compat` provider or if you pass `"discover_models": true` it will merge

815− the found models with your hand configured ones.

816

608

817```json

609```json

818{

610{

819 "providers": {

611 "providers": {

820− "ollama": {

612+ "lmstudio": {

821− "name": "Ollama",

613+ "name": "LM Studio",

822− "base_url": "http://localhost:11434/v1/",

614+ "base_url": "http://localhost:1234/v1/",

823− "type": "ollama",

615+ "type": "openai-compat",

824 "models": [

616 "models": [

825 {

617 {

826 "name": "Qwen 3 30B",

618 "name": "Qwen 3 30B",

827− "id": "qwen3:30b",

619+ "id": "qwen/qwen3-30b-a3b-2507",

828 "context_window": 256000,

620 "context_window": 256000,

829 "default_max_tokens": 20000

621 "default_max_tokens": 20000

830 }

622 }

831− ],

623+ ]

832− "discover_models": true

833 }

624 }

834 }

625 }

835}

626}

@@ -947,21 +738,8 @@ Or by setting the following in your config:

947}

738}

948```

739```

949

740

950−Crush also respects the [`DO_NOT_TRACK`](https://donottrack.sh/) convention

741+Crush also respects the `DO_NOT_TRACK` convention which can be enabled via

951−which can be enabled via `export DO_NOT_TRACK=1`.

742+`export DO_NOT_TRACK=1`.

952−

953−## Q&A

954−

955−### Why is clipboard copy and paste not working?

956−

957−Installing an extra tool might be needed on Unix-like environments.

958−

959−| Environment | Tool |

960−| ------------------- | ------------------------ |

961−| Windows | Native support |

962−| macOS | Native support |

963−| Linux/BSD + Wayland | `wl-copy` and `wl-paste` |

964−| Linux/BSD + X11 | `xclip` or `xsel` |

965

743

966## Contributing

744## Contributing

967

745

@@ -972,12 +750,11 @@ See the [contributing guide](https://github.com/charmbracelet/crush?tab=contribu

972We’d love to hear your thoughts on this project. Need help? We gotchu. You can find us on:

750We’d love to hear your thoughts on this project. Need help? We gotchu. You can find us on:

973

751

974- [Twitter](https://twitter.com/charmcli)

752- [Twitter](https://twitter.com/charmcli)

975−- [Slack][slack]

753+- [Slack](https://charm.land/slack)

976- [Discord][discord]

754- [Discord][discord]

977- [The Fediverse](https://mastodon.social/@charmcli)

755- [The Fediverse](https://mastodon.social/@charmcli)

978- [Bluesky](https://bsky.app/profile/charm.land)

756- [Bluesky](https://bsky.app/profile/charm.land)

979

757

980−[slack]: https://charm.land/slack

981[discord]: https://charm.land/discord

758[discord]: https://charm.land/discord

982

759

983## License

760## License

@@ -988,7 +765,7 @@ We’d love to hear your thoughts on this project. Need help? We gotchu. You can

988

765

989Part of [Charm](https://charm.land).

766Part of [Charm](https://charm.land).

990

767

991−<a href="https://charm.land/"><img alt="The Charm logo" width="400" src="https://stuff.charm.sh/charm-banner-softy.jpg" /></a>

768+<a href="https://charm.land/"><img alt="The Charm logo" width="400" src="https://stuff.charm.sh/charm-banner-next.jpg" /></a>

992

769

993

770

994Charm热爱开源 • Charm loves open source

771Charm热爱开源 • Charm loves open source

Taskfile.yaml+5−63

@@ -45,39 +45,15 @@ tasks:

45 vars:

46 LDFLAGS: '{{if .VERSION}}-ldflags="-X github.com/charmbracelet/crush/internal/version.Version={{.VERSION}}"{{end}}'

47 cmds:

48− - "go build -v {{if .RACE}}-race{{end}} {{.LDFLAGS}} ."

48+ - "go build {{if .RACE}}-race{{end}} {{.LDFLAGS}} ."

49− sources:

50− - ./**/*.go

51− - go.mod

52− - internal/agent/**/*.md

53− - internal/agent/**/*.md.tpl

54− - internal/agent/hyper/provider.json

55 generates:

49 generates:

56− - crush{{exeExt}}

50+ - crush

58 run:

52 run:

59 desc: Run build

53 desc: Run build

60 cmds:

54 cmds:

61 - task: build

55 - task: build

62− - "./crush{{exeExt}} {{.CLI_ARGS}} {{if .RACE}}2>race.log{{end}}"

56+ - "./crush {{.CLI_ARGS}} {{if .RACE}}2>race.log{{end}}"

63−

64− run:catwalk:

65− desc: Run build with local Catwalk

66− env:

67− CATWALK_URL: http://localhost:8080

68− cmds:

69− - task: build

70− - ./crush{{exeExt}} {{.CLI_ARGS}}

71−

72− run:onboarding:

73− desc: Run build with custom config to test onboarding

74− env:

75− CRUSH_GLOBAL_DATA: tmp/onboarding/data

76− CRUSH_GLOBAL_CONFIG: tmp/onboarding/config

77− cmds:

78− - task: build

79− - rm -rf tmp/onboarding

80− - ./crush{{exeExt}} {{.CLI_ARGS}}

82 test:

58 test:

83 desc: Run tests

59 desc: Run tests

@@ -101,11 +77,6 @@ tasks:

101 cmds:

77 cmds:

102 - prettier --write internal/cmd/stats/index.html internal/cmd/stats/index.css internal/cmd/stats/index.js

78 - prettier --write internal/cmd/stats/index.html internal/cmd/stats/index.css internal/cmd/stats/index.js

103

104− modernize:

105− desc: Run modernize

106− cmds:

107− - go run golang.org/x/tools/go/analysis/passes/modernize/cmd/modernize@latest -fix -test ./...

108−

109 dev:

80 dev:

110 desc: Run with profiling enabled

81 desc: Run with profiling enabled

111 env:

82 env:

@@ -120,12 +91,6 @@ tasks:

120 cmds:

91 cmds:

121 - task: fetch-tags

92 - task: fetch-tags

122 - go install {{.LDFLAGS}} -v .

93 - go install {{.LDFLAGS}} -v .

123− sources:

124− - ./**/*.go

125− - go.mod

126− - internal/agent/**/*.md

127− - internal/agent/**/*.md.tpl

128− - internal/agent/hyper/provider.json

129

130 profile:cpu:

95 profile:cpu:

131 desc: 10s CPU profile

96 desc: 10s CPU profile

@@ -186,30 +151,7 @@ tasks:

186

151

187 deps:

152 deps:

188 desc: Update Fantasy and Catwalk

153 desc: Update Fantasy and Catwalk

189− env:

190− # The Go proxy takes a bit of time to see the latest release. Setting

191− # these bypass it to ensure we can update to a release from a minute ago.

192− GOPROXY: direct

193− GONOSUMDB: charm.land/*

194 cmds:

154 cmds:

195− - go get charm.land/fantasy@latest

155+ - go get charm.land/fantasy

196− - go get charm.land/catwalk@latest

156+ - go get github.com/charmbracelet/catwalk

197 - go mod tidy

157 - go mod tidy

198−

199− swag:

200− desc: Generate OpenAPI spec from swag annotations

201− cmds:

202− - go run github.com/swaggo/swag/cmd/swag@v1.16.6 init --generalInfo main.go --dir . --output internal/swagger --packageName swagger --parseDependency --parseInternal --parseDepth 5

203− sources:

204− - internal/server/*.go

205− - internal/proto/*.go

206− - main.go

207− generates:

208− - internal/swagger/docs.go

209− - internal/swagger/swagger.json

210− - internal/swagger/swagger.yaml

211−

212− sqlc:

213− desc: Generate code using SQLC

214− cmds:

215− - sqlc generate

docs/hooks/FUTURE.mddeleted−260

@@ -1,260 +0,0 @@

1−# Hooks: Future Work

2−

3−This document tracks planned features and design notes for hooks that are not

4−yet implemented. Nothing here is part of the current contract. Treat it as a

5−scratchpad for what's next, not as documentation of current behavior.

6−

7−> [!NOTE]

8−> This document was largely LLM-generated.

9−

10−## `context_files`

11−

12−**Status:** planned, not implemented.

13−

14−### Motivation

15−

16−Today, a hook that wants to inject reference material into the agent's context

17−has exactly one knob: `context` (string or array of strings). Whatever the hook

18−puts there is concatenated into what the model sees. That's fine for short notes

19−("current branch: main", "scrubbed secrets") but it scales badly:

20−

21−- Dumping a whole `README.md` or `package.json` into `context` burns tokens on

22− every tool call where the hook fires.

23−- The model sees the file contents even if it doesn't need them.

24−- Large files can push the turn past the context window.

25−

26−`context_files` is the lazy alternative: the hook returns **paths**, not

27−contents. Crush tells the agent the files exist and are relevant, and the agent

28−decides whether to open them with its existing `view` tool.

29−

30−### Proposed shape

31−

32−Additive envelope field. Accepts a list of strings:

33−

34−```jsonc

35−{

36− "decision": "allow",

37− "context": "Scrubbed one secret",

38− "context_files": ["README.md", "docs/ARCHITECTURE.md"],

39−}

40−```

41−

42−Paths are resolved relative to `CRUSH_CWD`. Non-existent paths are dropped with

43−a debug log (don't fail the hook over a missing file).

44−

45−### How the agent sees it

46−

47−Crush appends a short note to the turn's context along the lines of:

48−

49−```

50−## Referenced files

51−- README.md

52−- docs/ARCHITECTURE.md

53−```

54−

55−No file contents are inlined. The agent opens them with `view` if it decides

56−they're relevant. This keeps cost proportional to need.

57−

58−### Aggregation

59−

60−Matches the existing rules for lists:

61−

62−- Concatenates across matching hooks in config order.

63−- Deduplicates paths (same file referenced by two hooks → listed once).

64−- Dropped entirely if the final decision is `deny` or `halt`.

65−

66−### Backwards compatibility

67−

68−Purely additive. Hooks that don't emit `context_files` are unaffected. Existing

69−envelopes keep working unchanged. No version bump required.

70−

71−### Open questions

72−

73−- Should `context_files` paths be constrained to `CRUSH_PROJECT_DIR`? Probably

74− yes, to avoid hooks smuggling in arbitrary filesystem reads.

75−- Do we want a per-file line range (`"README.md:1-40"`) or keep it dead simple

76− (whole-file references only)? Start simple; add ranges only if asked for.

77−- Should we annotate "why this file is relevant" per entry? An object form

78− (`{"path": "...", "reason": "..."}`) would allow that but complicates the

79− schema. Defer until there's a real user need.

80−

81−## Sub-agent opt-in

82−

83−**Status:** not implemented.

84−

85−### Background

86−

87−Today hooks fire **only** on the top-level agent's tool calls. Sub-agents

88−(`agent` task tool, `agentic_fetch`, future delegated loops) run without hook

89−interception so a single delegated turn doesn't trigger the user's hook N times.

90−

91−The outer sub-agent tool call itself is hooked, so blanket policy like "never

92−spawn sub-agents" or "rewrite prompts sent to the task agent" still works from

93−the coder's side. The sub-agent's inner loop is the part that's exempt.

94−

95−### Why users might want the escape hatch

96−

97−- Audit logging of every tool call, including delegated ones.

98−- Redaction hooks that want to apply uniformly regardless of who called the

99− tool.

100−- Policy that cares about the _tool_ not the _caller_: "never fetch from this

101− domain, even in `agentic_fetch`."

102−

103−Until someone actually asks, don't ship this. YAGNI.

104−

105−### Proposed shape

106−

107−Additive, per-hook. Zero-value matches current default (skip sub-agents):

108−

109−```jsonc

110−{

111− "hooks": {

112− "PreToolUse": [

113− {

114− "matcher": "^bash$",

115− "command": "./hooks/audit.sh",

116− "include_sub_agents": true, // default false

117− },

118− ],

119− },

120−}

121−```

122−

123−Implementation changes where `wrapToolsWithHooks` decides to skip. Instead of a

124−single `isSubAgent` bailout, the runner filters per-hook matches by the hook's

125−`include_sub_agents` flag. Hooks that opt in get wrapped into sub-agent tool

126−slices too; everything else stays skipped.

127−

128−### Backwards Compatibility

129−

130−Purely additive. Hooks that don't set `include_sub_agents` get the default

131−(`false` = skip sub-agents). No wire format change, no version bump. The initial

132−transition from "hooks fire everywhere" to "hooks skip sub-agents by default"

133−was a one-time behavior change; adding the opt-in is pure addition.

134−

135−### Side benefit: payload awareness

136−

137−Extend the stdin payload with `"is_sub_agent": true|false` so hook scripts that

138−opt in can branch on caller type ("audit top-level and sub-agent calls

139−differently"). Also purely additive — hooks that don't read the field are

140−unaffected.

141−

142−### Open questions

143−

144−- Per-hook flag (above) vs a global `hooks.include_sub_agents` default? A global

145− toggle is simpler but coarse-grained; per-hook is more flexible and

146− composable. Start per-hook; a global default can be layered on later with

147− explicit precedence ("per-hook overrides global").

148−- Does an opt-in hook see hooks from _nested_ sub-agents too (a sub-agent that

149− itself calls a sub-agent)? Probably yes — once you've opted in you want the

150− full tree. But call it out explicitly in docs so users aren't surprised by N²

151− explosions on pathological configs.

152−

153−## `UserPromptSubmit` event

154−

155−**Status:** not implemented.

156−

157−### Motivation

158−

159−Today Crush supports exactly one hook event, `PreToolUse`. That's enough to gate

160−and rewrite tool calls but nothing else. The next-most-useful event is

161−`UserPromptSubmit`: fires after the user hits Enter but before the turn hits the

162−LLM. Lets hooks inject context, rewrite prompts, or gate on content without the

163−mutation complexity of `PostToolUse` (output scrubbing, error coercion, size

164−limits — all rabbit holes).

165−

166−### Use cases

167−

168−- Prepend project context the user didn't think to include ("current branch:

169− `feat/x`; last commit: `<sha> <title>`").

170−- Point at reference files via `context_files` (when that lands) so the agent

171− knows where to look without being force-fed contents.

172−- Redact secrets out of the prompt before it leaves the machine.

173−- Refuse prompts matching a policy ("don't send anything mentioning

174− `production.env`") — with `deny` and a reason the user sees.

175−- Expand shorthand (`@TODO` → "please address the TODO in …").

176−

177−### Proposed shape

178−

179−Stdin payload extends the common envelope with the prompt:

180−

181−```jsonc

182−{

183− "event": "UserPromptSubmit",

184− "session_id": "…",

185− "cwd": "/home/user/project",

186− "prompt": "fix the login flow",

187− "attachments": ["screenshot.png"],

188−}

189−```

190−

191−Output envelope reuses common fields plus one new per-event field,

192−`updated_prompt`:

193−

194−```jsonc

195−{

196− "decision": "allow", // optional; deny blocks the submission entirely

197− "reason": "includes a production secret", // shown to the user when denying

198− "context": "Current branch: feat/login",

199− "updated_prompt": "fix the login flow\n\n(from @TODO on line 42)",

200−}

201−```

202−

203−`updated_prompt` is a **full replacement** — not a merge patch — because a

204−prompt is a single string with no natural key structure. If multiple hooks emit

205−`updated_prompt`, later hooks in config order win.

206−

207−### Aggregation

208−

209−Reuses the universal rules:

210−

211−- `halt` is sticky. Halts the whole turn before the LLM is called.

212−- `context` concatenates in config order.

213−- `updated_prompt`: last writer wins.

214−- `decision: "deny"` blocks the submission. The user sees `reason`; the turn

215− never reaches the LLM.

216−

217−### Differences from `PreToolUse`

218−

219−- No `updated_input`: there are no tool inputs at this point.

220−- No permission-prompt bypass: there's no permission prompt for a user prompt.

221−- `decision: "allow"` is functionally identical to silence. It exists only for

222− symmetry with `PreToolUse` and to give hook authors a consistent vocabulary.

223− (Could be argued both ways — consider dropping it here.)

224−- Fires on every user submission, including follow-ups in the same session.

225− Hooks should be fast; no subprocess-per-keystroke scenarios but the per-turn

226− overhead is real.

227−

228−### Implementation sketch

229−

230−- New event constant `EventUserPromptSubmit` in `internal/hooks/hooks.go`.

231−- `Runner.Run` already takes an event name; no interface change.

232−- A new call site in `sessionAgent.Run` (or the coordinator's Run path) that

233− fires hooks after creating the user message but before the first LLM call. If

234− the aggregate decision is `deny` or `halt`, abort the turn and surface

235− `reason` to the user.

236−- If hooks return `context`, prepend it to the prompt seen by the LLM (or attach

237− as a system-message-level note — decide based on how the prompt is threaded

238− through fantasy).

239−- If hooks return `updated_prompt`, replace the prompt body before the first LLM

240− call. The message row in the DB should still store the _original_ prompt so

241− the user sees what they typed; only the outbound version is rewritten. (Or:

242− store both, show the original, send the rewritten — mirror how `updated_input`

243− is handled today.)

244−

245−### Open questions

246−

247−- Store original vs rewritten prompt? Probably both, with UI showing original

248− and a subtle indicator that a hook modified it.

249−- Do hooks fire on queued prompts too, or only when actually dispatched? If the

250− user queues three prompts and the hook blocks the second, what happens to the

251− third? Simplest rule: fire when dispatched; denial skips to the next queued

252− prompt with a visible note.

253−- What about the `/commands` prefix? Does `UserPromptSubmit` fire for slash

254− commands, or are those intercepted earlier? Probably earlier — hooks see only

255− freeform prompts that would actually reach the LLM.

256−

257−## Cross-platform shell (Windows support)

258−

259−**Status:** implemented. See the [Execution model](README.md#execution-model)

260−section in `README.md` for the current behavior and contract.

docs/hooks/README.mddeleted−759

@@ -1,759 +0,0 @@

1−# Hooks

2−

3−> [!NOTE]

4−> This document was designed for both humans and agents.

5−

6−Hooks are user-defined shell scripts that run when various events happen during

7−the agent lifecycle, allowing you to both build on top of Crush, customize

8−its behavior, and exert deterministic control over an agent's wily behavior.

9−

10−Hooks are just shell commands, and were designed to be both simple and future

11−forward.

12−

13−### Hot Hook Facts

14−

15−- Hooks are just shell commands

16−- Hooks can be written in any language because they’re just executables: Bash, Python, Node, Rust, Haskell, whatever

17−- Hooks are Claude Code-compatible

18−- Crush ships with a builtin `crush-hook` skill write, edit, and configure

19− hooks; just tell Crush how to configure Crush

20−- Crush currently supports just one hook, `PreToolUse`, with plans to support

21− the full gamut; please let us know which hooks you'd like to see next

22−- Hooks run in parallel for speed, but their results compose in config order

23− for determinism

24−

25−### Some things you can do with hooks:

26−

27−- Block "dangerous" commands: no more `git push -f` or `cabal init`

28−- Rewrite tool input: turn `node` calls info `deno`, scrub secrets from

29− commands, rewrite all mentions of "Haskell" into "Haskell, The Best

30− Language", and so on

31−- Inject context: add notes to the model's context whenever certain tools are

32− called. For example: "remember to run gofumpt after editing Go files"

33−- Auto-approve tools: skip the permission prompt for bash commands that

34− you know are safe

35−- Log certain tool calls

36−

37−…And lots more. Show us what you're building!

38−

39−## Baby's First Hook

40−

41−Let's just dive into it and make a simple hook. This particular hook will

42−disallow the use of Haskell (but we love you, Simon Peyton Jones).

43−

44−### Config

45−

46−The first thing we need to do is hook up our hook. Let's add the following to

47−our **project-level** `crush.json`. Relative paths like `./no-haskell.sh` work

48−here because the project root is your working directory. If you're configuring

49−a global hook (`~/.config/crush/crush.json`), use an absolute path instead.

50−

51−```jsonc

52−{

53− // As expected, hooks go in a "hooks" object.

54− "hooks": {

55− // PreToolUse is an event that fires before a tool is used.

56− "PreToolUse": [

57− {

58− // What tool do we want to hook into? In this case, Bash, because it

59− // runs the stuff we wanna block.

60− "matcher": "^bash$",

61−

62− // The path to our actual hook script.

63− "command": "./no-haskell.sh",

64− },

65− ],

66− },

67−}

68−```

69−

70−Now, let's make our `no-haskell.sh` hook script.

71−

72−```bash

73−#!/usr/bin/env bash

74−

75−# Disallow ghc, cabal, and stack. Pipe the bash command output

76−# ($CRUSH_TOOL_INPUT_COMMAND) to grep and match on a regexp.

78−

79− # Someone is trying to use Haskell. Let's send a message back to the model

80− # and user explaining why we're blocking this. Note that we send all feedback

81− # like this to stderr.

82− echo "No Haskell allowed, kiddo." >&2

83−

84− # Now, block the tool call by exiting with code 2.

85− exit 2

86−fi

87−```

88−

89−That's basically it. For the full guide on how hooks work, however, read on.

90−

91−---

92−

93−## Execution model

94−

95−Hooks run through Crush's embedded POSIX shell (`mvdan.cc/sh`) — the same

96−interpreter the `bash` tool uses. Inline commands and shebang-less scripts

97−execute in-process; scripts with a `#!` shebang dispatch to the named

98−interpreter via `os/exec`. This contract is identical on macOS, Linux, and

99−Windows.

100−

101−What this means in practice:

102−

103−- **Windows without Unix tooling**: inline shell (`echo`, pipelines, `jq`,

104− `grep`), shebang-less `.sh` scripts, inline PowerShell

105− (`powershell -Command …`), and `.exe` invocations all work out of the box

106− with no WSL, Git Bash, Cygwin, or MSYS required.

107−- **PowerShell scripts** (`.ps1`) are not auto-dispatched by extension.

108− Invoke them explicitly: `powershell -File ./audit.ps1` (or

109− `pwsh -File ./audit.ps1`).

110−- **Shebang'd scripts** require the named interpreter on `PATH`. Git for

111− Windows ships `bash.exe`, which makes `#!/bin/bash` and

112− `#!/usr/bin/env bash` scripts work on Windows the same way they do on

113− Unix. CRLF line endings in the shebang line are tolerated.

114−- **Permissive shebang fallback**: if the absolute path in a shebang

115− doesn't exist (e.g. `#!/bin/bash` on Windows), Crush falls back to a

116− `PATH` lookup of the base name (`bash`) before giving up. A debug-level

117− log records the fallback. If the interpreter isn't on `PATH` either, the

118− hook fails cleanly as a non-blocking warning and the agent proceeds as

119− "no opinion".

120−- **Environment**: every hook sees `CRUSH=1`, `AGENT=crush`, and

121− `AI_AGENT=crush` on top of the `CRUSH_*` hook-specific variables. These

122− three markers are guaranteed and match what the `bash` tool sets, so

123− scripts that detect "am I being run by an AI agent?" behave the same in

124− both contexts.

125−- **Timeout behavior**: when a hook exceeds its timeout, Crush cancels the

126− context and waits a short grace period (~1s) for the interpreter to

127− yield. If the hook still hasn't returned, Crush abandons it, logs a

128− warning, and treats the result as "no opinion" so the agent can proceed.

129− Long-running work should honor context cancellation or run in a

130− subprocess via a shebang.

131−

132−## Configuration

133−

134−Hooks can be added to your `crush.json` (or `.crush.json`) at both the global

135−and project-level, with project level hooks taking precedence.

136−

137−```jsonc

138−{

139− "hooks": {

140− "PreToolUse": [

141− {

142− "name": "no-rm-rf", // friendly name shown in the TUI

143− "matcher": "bash", // regex tested against the tool name

144− "command": "./hooks/my-hot-hook.sh", // the path to the hook

145− "timeout": 10, // in seconds; default 30

146− },

147− ],

148− },

149−}

150−```

151−

152−> [!IMPORTANT]

153−> The `command` is resolved relative to your **current working directory** —

154−> not relative to the config file. Relative paths like `./hooks/whatever.sh`

155−> work fine in project-level `crush.json` because the project root is also

156−> your working directory. For **global** config (`~/.config/crush/`),

157−> however, you must use either an absolute path or an inline command:

158−>

159−> ```jsonc

160−> // Global ~/.config/crush/crush.json

161−> {

162−> "hooks": {

163−> "PreToolUse": [

164−> {

165−> "command": "/home/you/.config/crush/hooks/no-haskell.sh"

166−> // or use an inline command:

167−> // "command": "echo '{\"decision\":\"allow\"}'"

168−> }

169−> ]

170−> }

171−> }

172−> ```

173−

174−Remember, hooks will run in parallel but resolve in config order. Last hook

175−wins when rewriting input, but first deny wins when blocking.

176−

177−## Events

178−

179−Here are the events you can hook into (spoiler: there's currently just one):

180−

181−### PreToolUse

182−

183−This hook fires before every tool call. Use it to block dangerous commands,

184−enforce policies, rewrite tool input, inject context the model should see, log

185−stuff, and so on.

186−

187−**Matched against**: the tool name (e.g. `bash`, `edit`, `write`,

188−`mcp_github_create_pull_request`).

189−

190−> [!NOTE]

191−> Event names are case insensitive and snake-caseable, so `PreToolUse`,

192−> `pretooluse`, `PRETOOLUSE`, `pre_tool_use`, and `PRE_TOOL_USE` all work.

193−

194−**Scope**: `PreToolUse` only fires on the **top-level agent's** tool calls.

195−Sub-agents (the `agent` task tool, `agentic_fetch`, etc.) run without hook

196−interception so a single delegated turn doesn't trigger your hook N times. The

197−outer sub-agent tool call itself _is_ hooked, so policy like "never let the

198−agent spawn sub-agents" still works.

199−

200−Hooks are keyed by event name. Only `command` is required, and you can omit

201−`matcher` to match all tools.

202−

203−## Building Hooks

204−

205−When a hook fires, Crush:

206−

207−1. Filters hooks whose `matcher` regex matches the tool name (no matcher = match

208− all).

209−2. Deduplicates by `command` (identical commands run once).

210−3. Runs all matching hooks **in parallel** through Crush's embedded POSIX

211− shell (see [Execution model](#execution-model)).

212−4. Waits for all to finish (or time out), then aggregates results **in config

213− order**: deny wins over allow, allow wins over none; `updated_input` patches

214− shallow-merge in order.

215−5. Applies the result **before** permission checks. If the aggregated decision

216− is `deny`, the tool call is blocked and you never see a permission prompt

217− for it. If it's `allow`, Crush treats that as affirmative pre-approval and

218− also skips the prompt. Silence (no decision) falls through to the normal

219− permission flow.

220−

221−Note that you can omit `matcher` and match in your shell script instead,

222−however you'll incur some additional overhead as Crush will still parse and

223−run each hook.

224−

225−### Input

226−

227−Each hook receives data two ways: environment variables and stdin (as JSON).

228−Environment variables are typically easier to work with, with JSON being

229−available when input is more complex.

230−

231−#### Environment Variables

232−

233−The available environment variables are:

234−

235−| Variable | Description |

236−| ---------------------------- | ---------------------------------------------- |

237−| `CRUSH` | Always `1` when running under Crush. |

238−| `AGENT` | Always `crush`. |

239−| `AI_AGENT` | Always `crush`. |

240−| `CRUSH_EVENT` | The hook event name (e.g. `PreToolUse`). |

241−| `CRUSH_TOOL_NAME` | The tool being called (e.g. `bash`). |

242−| `CRUSH_SESSION_ID` | Current session ID. |

243−| `CRUSH_CWD` | Working directory. |

244−| `CRUSH_PROJECT_DIR` | Project root directory. |

245−| `CRUSH_TOOL_INPUT_COMMAND` | For `bash` calls: the shell command being run. |

246−| `CRUSH_TOOL_INPUT_FILE_PATH` | For file tools: the target file path. |

247−

248−The `CRUSH`, `AGENT`, and `AI_AGENT` markers are also set by the `bash`

249−tool, so a script can detect "am I running under Crush?" the same way in

250−either context.

251−

252−#### JSON

253−

254−Standard input provides the full context as JSON:

255−

256−```jsonc

257−{

258− "event": "PreToolUse", // Hook event name

259− "session_id": "313909e", // Current session ID

260− "cwd": "/home/user/project", // Working directory

261− "tool_name": "bash", // The tool being called

262− "tool_input": { "command": "rm -rf /" }, // The tool's input

263−}

264−```

265−

266−Note that `tool_input` field contains the raw JSON the model sent to the tool.

267−

268−To parse the stdin JSON in your hook script, read from stdin and use a tool like

269−`jq`:

270−

271−```bash

272−#!/usr/bin/env bash

273−read -r input

274−tool_name=$(echo "$input" | jq -r '.tool_name')

275−command=$(echo "$input" | jq -r '.tool_input.command // empty')

276−```

277−

278−You can also use tools like Python:

279−

280−```python

281−#!/usr/bin/env python3

282−import json, sys

283−

284−data = json.load(sys.stdin)

285−tool_name = data.get("tool_name", "")

286−command = data.get("tool_input", {}).get("command", "")

287−```

288−

289−### Output

290−

291−Hooks communicate back to Crush via **exit code** and `stdout`/`stderr`. The

292−simplest way to do this is to return an error code and print additional context

293−to stderr. For example:

294−

295−```bash

296−# Here, error code 2 blocks the tool, using stderr as the reason:

297−if some_bad_condition; then

298− echo "Blocked: reason here" >&2

299− exit 2

300−fi

301−```

302−

303−| Exit Code | Meaning |

304−| --------- | ---------------------------------------------------------------- |

305−| 0 | Success. Stdout is parsed as JSON (see fields below). |

306−| 2 | **Block the tool.** Stderr is used as the deny reason (no JSON). |

307−| 49 | **Halt the turn.** Stderr is used as the halt reason (no JSON). |

308−| Other | Non-blocking error. Logged and ignored — the tool call proceeds. |

309−

310−The difference between exit 2 and exit 49:

311−

312−- **Exit 2** blocks the current tool call. The agent sees the error and can try

313− something else.

314−- **Exit 49** halts the whole turn. The agent doesn't get to respond further;

315− the user takes over. Use this when something is wrong enough that the agent

316− shouldn't keep trying. 49 sits in an empty slice of the exit-code space —

317− between the generic-error range (1-30), the BSD `sysexits.h` range (64-78),

318− and the killed-by-signal range (128+) — so it can't be hit by accident.

319−

320−That said, if you need more control, or if you need to rewrite input, you can

321−use JSON on stdout. Exit 0 and print a JSON object to provide context, update

322−the input, or still deny/halt with a reason:

323−

324−```jsonc

325−{

326− "version": 1, // Output envelope version. Optional; defaults to 1.

327− "decision": "allow", // "allow", "deny", or null. Omit for no opinion.

328− "halt": false, // If true, halts the turn entirely.

329− "reason": "LGTM", // Shown when denying or halting.

330− "context": "Scrubbed secrets", // String or array of strings. Appended to what the model sees.

331− "updated_input": { "command": "…" }, // Shallow-merged into the tool's input before execution.

332−}

333−```

334−

335−`version` is an optional integer at the top of the envelope. It defaults to `1`

336−if omitted. Unknown higher versions are still parsed; the field exists so the

337−envelope can evolve without a compatibility shim.

338−

339−`decision: "allow"` is **affirmative**: it pre-approves the tool call and

340−bypasses the permission prompt entirely. Silence (no `decision`, or

341−`decision: null`) means "no opinion" — the tool still goes through the

342−normal permission flow. Use `"allow"` when you want to auto-approve; omit it

343−when you only want to inject context or rewrite input without also vouching

344−for the call.

345−

346−`updated_input` is a shallow-merge patch. Keys you include overwrite matching

347−keys in `tool_input`; keys you don't include are preserved. If the model called

348−`bash` with `{"command": "npm test", "timeout": 60000}` and your hook returns

349−`{"updated_input": {"command": "bun test"}}`, the tool runs with

350−`{"command": "bun test", "timeout": 60000}` — the timeout isn't dropped. The

351−merge is shallow: nested objects are replaced wholesale, not deep-merged.

352−

353−`halt: true` stops the turn entirely. The agent doesn't get to respond further;

354−the user takes over. The exit-code shorthand is `exit 49` with stderr as the

355−reason.

356−

357−`context` accepts either a string or an array of strings. Use the string form

358−for a single observation; use the array form when a hook produces multiple

359−distinct notes and you'd rather not concatenate them by hand. Empty strings and

360−empty array entries are dropped.

361−

362−Here's a full shell script that produces this JSON:

363−

364−```bash

365−#!/usr/bin/env bash

366−# Example: rewrite a bash command using RTK

367−

368−read -r input

369−original_cmd=$(echo "$input" | jq -r '.tool_input.command')

370−rewritten=$(secret-scrubber rewrite "$original_cmd")

371−

372−cat <<EOF

373−{

374− "decision": "allow",

375− "context": "Scrubbed secrets",

376− "updated_input": {"command": "$rewritten"}

377−}

378−EOF

379−```

380−

381−### Multiple Hooks

382−

383−Hooks run in parallel, but their results compose in config order. Whichever hook

384−finishes first doesn't get to "win" by virtue of timing; composition is

385−deterministic based on the order hooks appear in `crush.json`.

386−

387−When multiple hooks match the same tool call:

388−

389−- If **any** hook denies, the tool call is blocked. `reason` values are

390− concatenated in config order (newline-separated).

391−- If **any** hook halts, the turn ends after the tool call is blocked.

392−- If no hook denies or halts but at least one allows, the tool call proceeds

393− **and the permission prompt is skipped**.

394−- `context` values are concatenated in config order. Strings and arrays compose

395− uniformly — each string becomes one entry, and array entries are flattened in.

396−- `updated_input` patches shallow-merge in config order against the original

397− tool input. Later hooks override earlier ones on colliding keys. If denied or

398− halted, `updated_input` patches are ignored.

399−

400−### Timeouts

401−

402−If a hook exceeds its timeout, Crush cancels its context and treats the

403−result as a non-blocking error so the tool call proceeds. The default

404−timeout is 30 seconds. Shebang-dispatched subprocesses are killed through

405−`exec.CommandContext`; in-process hooks get a short grace period to yield

406−and are then abandoned (the agent moves on regardless). Long-running work

407−should honor context cancellation or run out-of-process via a shebang.

408−

409−## Examples

410−

411−### Block destructive commands

412−

413−Prevent the agent from running `rm -rf` in bash:

414−

415−```json

416−{

417− "hooks": {

418− "PreToolUse": [

419− {

420− "matcher": "^bash$",

421− "command": "./hooks/no-rm-rf.sh"

422− }

423− ]

424− }

425−}

426−```

427−

428−`hooks/no-rm-rf.sh`:

429−

430−```bash

431−#!/usr/bin/env bash

432−# Block rm -rf commands in the bash tool. Otherwise stay silent so the

433−# normal permission flow runs.

434−

435−if echo "$CRUSH_TOOL_INPUT_COMMAND" | grep -qE 'rm\s+-(rf|fr)\s+/'; then

436− echo "Refusing to run rm -rf against root" >&2

437− exit 2

438−fi

439−

440−exit 0

441−```

442−

443−### Auto-approve read-only tools

444−

445−Skip the permission prompt for tools that can't change anything. The hook

446−returns `decision: "allow"`, which tells Crush to pre-approve the call:

447−

448−```jsonc

449−{

450− "hooks": {

451− "PreToolUse": [

452− {

453− "matcher": "^(view|ls|grep|glob)$",

454− "command": "echo '{\"decision\":\"allow\"}'",

455− },

456− ],

457− },

458−}

459−```

460−

461−No script file needed — the command is inline. Every `view`/`ls`/`grep`/`glob`

462−call now runs without prompting. Add the `bash` tool to this list at your own

463−risk; consider a more targeted allowlist instead:

464−

465−```bash

466−#!/usr/bin/env bash

467−# hooks/safe-bash.sh — auto-approve read-only bash commands.

468−

469−case "$CRUSH_TOOL_INPUT_COMMAND" in

470− ls*|cat*|grep*|rg*|echo*|pwd*)

471− echo '{"decision":"allow"}'

472− ;;

473− *)

474− # Silent — fall through to the normal permission prompt.

475− exit 0

476− ;;

477−esac

478−```

479−

480−### Inject context into file writes

481−

482−Add a reminder to the model whenever it writes a Go file:

483−

484−```json

485−{

486− "hooks": {

487− "PreToolUse": [

488− {

489− "matcher": "^(edit|write|multiedit)$",

490− "command": "./hooks/go-context.sh"

491− }

492− ]

493− }

494−}

495−```

496−

497−`hooks/go-context.sh`:

498−

499−```bash

500−#!/usr/bin/env bash

501−# Remind the model about Go formatting when editing .go files.

502−# Emit context only; stay silent on `decision` so the normal permission

503−# prompt still runs for edits/writes.

504−

505−if [[ "$CRUSH_TOOL_INPUT_FILE_PATH" == *.go ]]; then

506− echo '{"context": "Remember: run gofumpt after editing Go files."}'

507−else

508− echo '{}'

509−fi

510−```

511−

512−### Block all MCP tools

513−

514−The `command` can be inline. This one-liner matches all MCP tools and blocks

515−them:

516−

517−```jsonc

518−{ "matcher": "^mcp_", "command": "echo 'MCP tools are disabled' >&2; exit 2" }

519−```

520−

521−### Log every tool call

522−

523−With no `matcher` this fires for every tool. It exits 0 with no stdout so the

524−tool call always proceeds.

525−

526−```jsonc

527−{ "command": "echo \"$(date -Iseconds) $CRUSH_TOOL_NAME\" >> ./tools.log" }

528−```

529−

530−### A real-world Example:

531−

532−For a more practical example, see [`rtk-rewrite.sh`](./examples/rtk-rewrite.sh),

533−which demonstrates how to rewrite tool input using

534−[RTK](https://github.com/rtk-ai/rtk) to save tokens.

535−

536−### Using other languages

537−

538−Hooks aren't limited to shell scripts: any executable works. Here's the same

539−"block rm -rf" example in some other languages.

540−

541−#### Lua

542−

543−`{"matcher": "^bash$", "command": "lua ./hooks/no-rm-rf.lua"}`

544−

545−```lua

546−local input = io.read("*a")

547−local tool_input = input:match('"command":"(.-)"') or ""

548−

549−if tool_input:match("rm%s+%-[rf][rf]%s+/") then

550− io.stderr:write("Refusing to run rm -rf against root\n")

551− os.exit(2)

552−end

553−```

554−

555−#### JavaScript

556−

557−`{"matcher": "^bash$", "command": "node ./hooks/no-rm-rf.js"}`

558−

559−```js

560−let input = "";

561−process.stdin.on("data", (chunk) => (input += chunk));

562−process.stdin.on("end", () => {

563− const { tool_input: toolInput } = JSON.parse(input);

564−

565− if (/rm\s+-[rf]{2}\s+\//.test(toolInput.command)) {

566− process.stderr.write("Refusing to run rm -rf against root\n");

567− process.exit(2);

568− }

569−});

570−```

571−

572−---

573−

574−## Claude Code compatibility

575−

576−Crush hooks are broadly compatible with [Claude Code

577−hooks](https://docs.claude.com/en/docs/claude-code/hooks): the config shape,

578−stdin payload, output envelope, and exit codes line up so most Claude Code

579−hooks run under Crush unchanged. This document covers the Crush-specific API

580−only — anything not documented here isn't guaranteed to work.

581−

582−One intentional divergence: Crush treats `updated_input` as a shallow-merge

583−patch against the original `tool_input` rather than a full replacement. Keys

584−you omit are preserved. See [Output](#output) for details.

585−

586−---

587−

588−## Reference

589−

590−This is the official reference of the narrative above. If prose and this section

591−disagree, the prose should be presumed canonical for intent, while this section

592−is canonical for shape.

593−

594−Both the stdin payload and the output envelope have **common fields** that apply

595−to every event and **per-event fields** that only some events recognize. When an

596−event doesn't understand a field, it's ignored.

597−

598−### Hook config

599−

600−Each entry under a `hooks.<EventName>` array:

601−

602−```jsonc

603−{

604− // string. Optional. Friendly display name shown in the TUI. Falls back to

605− // command when omitted.

606− "name": "no-rm-rf",

607−

608− // string. Optional. Regex tested against the tool name. Omit to match all.

609− "matcher": "^bash$",

610−

611− // string. Required. Shell command to run.

612− "command": "./hooks/my-hook.sh",

613−

614− // number. Optional. Seconds before the hook is killed. Defaults to 30.

615− "timeout": 10,

616−}

617−```

618−

619−### Stdin payload (common)

620−

621−Present in every hook event:

622−

623−```jsonc

624−{

625− // string. Hook event name.

626− "event": "PreToolUse",

627−

628− // string. Current session ID.

629− "session_id": "313909e",

630−

631− // string. Working directory when invoked.

632− "cwd": "/home/user/project",

633−}

634−```

635−

636−### Stdin payload — PreToolUse

637−

638−Extends the common payload:

639−

640−```jsonc

641−{

642− // ...common fields...

643−

644− // string. The tool being called.

645− "tool_name": "bash",

646−

647− // object. Raw JSON input the model sent to the tool. Shape is per-tool.

648− "tool_input": {

649− "command": "npm test",

650− },

651−}

652−```

653−

654−### Output envelope (common)

655−

656−Fields a hook may print to stdout on exit 0. All are optional and apply to every

657−event:

658−

659−```jsonc

660−{

661− // number. Defaults to 1. Unknown higher values still parse; exists for

662− // forward-compat.

663− "version": 1,

664−

665− // boolean. If true, ends the turn entirely. User takes over.

666− "halt": false,

667−

668− // string. Shown when denying (to the model) or halting (to the model and

669− // user).

670− "reason": "not allowed",

671−

672− // string | string[]. Appended to what the model sees. Empty entries are

673− // dropped.

674− "context": "Rewrote with RTK",

675−}

676−```

677−

678−### Output envelope — PreToolUse

679−

680−Extends the common envelope:

681−

682−```jsonc

683−{

684− // ...common fields...

685−

686− // "allow" | "deny" | null. null/omitted = no opinion, the tool still goes

687− // through the normal permission prompt. "allow" is affirmative: pre-approves

688− // the tool call and bypasses the prompt. "deny" blocks the call; the model

689− // sees the error and may try something else.

690− "decision": "allow",

691−

692− // object. Shallow-merge patch against tool_input. Nested objects are

693− // replaced wholesale, not deep-merged.

694− "updated_input": {

695− "command": "bun test",

696− },

697−}

698−```

699−

700−### Exit codes

701−

702−| Code | Meaning |

703−| ----- | ------------------------------------------------------------------------ |

704−| `0` | Success. Stdout is parsed as the output envelope. |

705−| `2` | Block this tool call. Stderr becomes the deny reason. Stdout is ignored. |

706−| `49` | Halt the whole turn. Stderr becomes the halt reason. Stdout is ignored. |

707−| other | Non-blocking error. Logged and ignored; the tool call proceeds. |

708−

709−Exit `2` only applies to events that can block something. On events where

710−there's nothing to block, it's treated as a non-blocking error.

711−

712−### Aggregation

713−

714−When multiple hooks match the same event, results compose in **config order**.

715−

716−Universal rules:

717−

718−1. `halt` is sticky: if any hook halts, the turn ends.

719−2. `reason` values concatenate with `\n` in config order. Halt-only hooks

720− without a deny still contribute their reason.

721−3. `context` values concatenate with `\n` in config order. String entries and

722− array entries flatten uniformly.

723−

724−PreToolUse-specific rules:

725−

726−4. `decision` precedence: `deny` > `allow` > `null`. First deny determines the

727− outcome; subsequent allows don't override. If the final aggregated decision

728− is `allow`, Crush pre-approves the tool call and skips the permission

729− prompt. If it's `null` (no hook allowed), the tool goes through the normal

730− permission flow.

731−5. `updated_input` patches shallow-merge sequentially against the original

732− `tool_input`. Later patches override earlier ones on colliding keys. Patches

733− are **ignored** if the final decision is deny or halt.

734−

735−### Environment variables

736−

737−See [Environment Variables](#environment-variables) above for the full list.

738−

739−---

740−

741−## Whatcha think?

742−

743−We'd love to hear your thoughts on this project. Need help? We gotchu. You can

744−find us on:

745−

746−- [Twitter](https://twitter.com/charmcli)

747−- [Slack](https://charm.land/slack)

748−- [Discord](https://charm.land/discord)

749−- [The Fediverse](https://mastodon.social/@charmcli)

750−- [Bluesky](https://bsky.app/profile/charm.land)

751−

752−---

753−

754−Part of [Charm](https://charm.land).

755−

756−<a href="https://charm.land/"><img alt="The Charm logo" width="400" src="https://stuff.charm.sh/charm-banner-softy.jpg" /></a>

757−

758−

759−Charm热爱开源 • Charm loves open source

docs/hooks/examples/rtk-rewrite.shdeleted−60

@@ -1,60 +0,0 @@

1−#!/usr/bin/env bash

2−

3−# RTK hook for Crush

4−# This hook rewrites bash commands to use rtk for token savings.

5−#

6−# https://github.com/rtk-ai/rtk

7−#

8−# Requires: rtk >= 0.23.0, jq

9−#

10−# All rewrite logic lives in `rtk rewrite`, which is the single source of

11−# truth. To add or change rewrite rules, edit the Rust registry, not this

12−# file.

13−#

14−# Exit code protocol for `rtk rewrite`:

15−# 0 + stdout Rewrite found → allow with updated input

16−# 1 No RTK equivalent → pass through unchanged

17−# 2 Deny rule matched → pass through

18−

19−set -euo pipefail

20−

21−if ! command -v jq &>/dev/null; then

22− echo "[rtk] WARNING: jq is not installed. Install: https://jqlang.github.io/jq/download/" >&2

23− exit 0

24−fi

25−

26−if ! command -v rtk &>/dev/null; then

27− echo "[rtk] WARNING: rtk is not installed. Install: https://github.com/rtk-ai/rtk#installation" >&2

28− exit 0

29−fi

30−

31−# Version guard: rtk rewrite requires >= 0.23.0.

32−RTK_VERSION=$(rtk --version 2>/dev/null | grep -oE '[0-9]+\.[0-9]+\.[0-9]+' | head -1)

33−if [ -n "$RTK_VERSION" ]; then

34− MAJOR=$(echo "$RTK_VERSION" | cut -d. -f1)

35− MINOR=$(echo "$RTK_VERSION" | cut -d. -f2)

36− if [ "$MAJOR" -eq 0 ] && [ "$MINOR" -lt 23 ]; then

37− echo "[rtk] WARNING: rtk $RTK_VERSION is too old (need >= 0.23.0). Upgrade: cargo install rtk" >&2

38− exit 0

39− fi

40−fi

41−

42−CMD="${CRUSH_TOOL_INPUT_COMMAND:-}"

43−if [ -z "$CMD" ]; then

44− exit 0

45−fi

46−

47−REWRITTEN=$(rtk rewrite "$CMD" 2>/dev/null) && EXIT_CODE=0 || EXIT_CODE=$?

48−

49−case $EXIT_CODE in

50−0 | 3)

51− # Rewrite found. If identical, the command already uses rtk.

52− [ "$CMD" = "$REWRITTEN" ] && exit 0

53− jq -n --arg cmd "$REWRITTEN" \

54− '{decision: "allow", updated_input: ({command: $cmd} | tostring)}'

55− ;;

56−*)

57− # No rewrite (1), deny (2), or unexpected — pass through.

58− exit 0

59− ;;

60−esac

flake.lockdeleted−60

@@ -1,60 +0,0 @@

1−{

2− "nodes": {

3− "flake-utils": {

4− "inputs": {

5− "systems": "systems"

6− },

7− "locked": {

8− "lastModified": 1731533236,

9− "narHash": "sha256-l0KFg5HjrsfsO/JpG+r7fRrqm12kzFHyUHqHCVpMMbI=",

10− "owner": "numtide",

11− "repo": "flake-utils",

12− "rev": "11707dc2f618dd54ca8739b309ec4fc024de578b",

13− "type": "github"

14− },

15− "original": {

16− "owner": "numtide",

17− "repo": "flake-utils",

18− "type": "github"

19− }

20− },

21− "nixpkgs": {

22− "locked": {

23− "lastModified": 1782964654,

24− "narHash": "sha256-oqyac50OyRdni5ZK9H7ZuAJoR3LzgsOdUcbBlUsk9lg=",

25− "owner": "NixOS",

26− "repo": "nixpkgs",

27− "rev": "3130a1cf8f7c794ea5a4760c3fa4c745d64ec269",

28− "type": "github"

29− },

30− "original": {

31− "owner": "NixOS",

32− "repo": "nixpkgs",

33− "type": "github"

34− }

35− },

36− "root": {

37− "inputs": {

38− "flake-utils": "flake-utils",

39− "nixpkgs": "nixpkgs"

40− }

41− },

42− "systems": {

43− "locked": {

44− "lastModified": 1681028828,

45− "narHash": "sha256-Vy1rq5AaRuLzOxct8nz4T6wlgyUR7zLU309k9mBC768=",

46− "owner": "nix-systems",

47− "repo": "default",

48− "rev": "da67096a3b9bf56a91d16901293e51ba5b49a27e",

49− "type": "github"

50− },

51− "original": {

52− "owner": "nix-systems",

53− "repo": "default",

54− "type": "github"

55− }

56− }

57− },

58− "root": "root",

59− "version": 7

60−}

flake.nixdeleted−75

@@ -1,75 +0,0 @@

1−{

2− description = "Crush development environment";

3−

4− inputs = {

5− nixpkgs.url = "github:NixOS/nixpkgs";

6− flake-utils.url = "github:numtide/flake-utils";

7− };

8−

9− outputs = {

10− self,

11− nixpkgs,

12− flake-utils,

13− }:

14− flake-utils.lib.eachDefaultSystem (

15− system: let

16− pkgs = nixpkgs.legacyPackages.${system};

17− version = "devel";

18− in {

19− packages.default = pkgs.buildGoModule {

20− pname = "crush";

21− inherit version;

22−

23− src = ./.;

24−

25− vendorHash = null;

26−

27− # Match the Go toolchain used in the dev shell.

28− go = pkgs.go_1_26;

29−

30− env.CGO_ENABLED = 0;

31−

32− ldflags = [

33− "-s"

34− "-w"

35− "-X github.com/charmbracelet/crush/internal/version.Version=${version}"

36− ];

37−

38− # The test suite requires network access and external tools.

39− doCheck = false;

40−

41− meta = {

42− description = "Terminal-based AI coding assistant";

43− homepage = "https://github.com/charmbracelet/crush";

44− license = pkgs.lib.licenses.mit;

45− mainProgram = "crush";

46− };

47− };

48−

49− devShells.default = pkgs.mkShell {

50− buildInputs = with pkgs; [

51− # Go toolchain

52− go_1_26

53−

54− # Development tools

55− gopls # Go language server

56− golangci-lint # Linter

57− gofumpt # Formatter (stricter than gofmt)

58− go-task # Task runner

59− delve # Go debugger

60−

61− # Additional tools

62− git # Version control

63− gh # GitHub CLI

64− svu # Semantic version utility

65− sqlc # SQL code generator

66− ];

67−

68− shellHook = ''

69− # Set Go environment variables

70− export CGO_ENABLED=0

71− '';

72− };

73− }

74− );

75−}

go.mod+111−136

@@ -1,159 +1,146 @@

1module github.com/charmbracelet/crush

3−go 1.26.4

3+go 1.25.5

5require (

6− charm.land/bubbles/v2 v2.1.1

6+ charm.land/bubbles/v2 v2.0.0-rc.1.0.20260109112849-ae99f46cec66

7− charm.land/bubbletea/v2 v2.0.8

7+ charm.land/bubbletea/v2 v2.0.0-rc.2.0.20251216153312-819e2e89c62e

8− charm.land/catwalk v0.49.0

9− charm.land/fang/v2 v2.0.1

10− charm.land/fantasy v0.35.0

8+ charm.land/fantasy v0.6.1

11− charm.land/glamour/v2 v2.0.1

9+ charm.land/glamour/v2 v2.0.0-20260123212943-6014aa153a9b

12− charm.land/lipgloss/v2 v2.0.5

10+ charm.land/lipgloss/v2 v2.0.0-beta.3.0.20251205162909-7869489d8971

13− charm.land/log/v2 v2.0.0

11+ charm.land/log/v2 v2.0.0-20251110204020-529bb77f35da

14 charm.land/x/vcr v0.1.1

12 charm.land/x/vcr v0.1.1

15 github.com/JohannesKaufmann/html-to-markdown v1.6.0

13 github.com/JohannesKaufmann/html-to-markdown v1.6.0

16 github.com/MakeNowJust/heredoc v1.0.0

14 github.com/MakeNowJust/heredoc v1.0.0

17− github.com/Microsoft/go-winio v0.6.2

18− github.com/PuerkitoBio/goquery v1.12.0

15+ github.com/PuerkitoBio/goquery v1.11.0

19− github.com/alecthomas/chroma/v2 v2.27.0

16+ github.com/alecthomas/chroma/v2 v2.23.1

20− github.com/aymanbagabas/go-udiff v0.4.1

17+ github.com/atotto/clipboard v0.1.4

18+ github.com/aymanbagabas/go-nativeclipboard v0.1.2

19+ github.com/aymanbagabas/go-udiff v0.3.1

21 github.com/bmatcuk/doublestar/v4 v4.10.0

20 github.com/bmatcuk/doublestar/v4 v4.10.0

22 github.com/charlievieth/fastwalk v1.0.14

21 github.com/charlievieth/fastwalk v1.0.14

23− github.com/charmbracelet/colorprofile v0.4.3

22+ github.com/charmbracelet/catwalk v0.15.0

23+ github.com/charmbracelet/colorprofile v0.4.1

24− github.com/charmbracelet/openai-go v0.0.0-20260617131321-5e4b9c18c4be

24+ github.com/charmbracelet/fang v0.4.4

25− github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7

25+ github.com/charmbracelet/ultraviolet v0.0.0-20251212194010-b927aa605560

26− github.com/charmbracelet/x/ansi v0.11.7

26+ github.com/charmbracelet/x/ansi v0.11.4

27 github.com/charmbracelet/x/editor v0.2.0

28 github.com/charmbracelet/x/etag v0.2.0

29− github.com/charmbracelet/x/exp/charmtone v0.0.0-20260527151214-009e6338d40d

29+ github.com/charmbracelet/x/exp/charmtone v0.0.0-20260109001716-2fbdffcb221f

30 github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f

31 github.com/charmbracelet/x/exp/ordered v0.1.0

32− github.com/charmbracelet/x/exp/slice v0.0.0-20260622092256-25656177ba8e

32+ github.com/charmbracelet/x/exp/slice v0.0.0-20251201173703-9f73bfd934ff

33 github.com/charmbracelet/x/exp/strings v0.1.0

34− github.com/charmbracelet/x/powernap v0.1.6

34+ github.com/charmbracelet/x/powernap v0.0.0-20260127155452-b72a9a918687

35 github.com/charmbracelet/x/term v0.2.2

36− github.com/clipperhouse/displaywidth v0.11.0

37− github.com/clipperhouse/uax29/v2 v2.7.0

38 github.com/denisbrodbeck/machineid v1.0.1

36 github.com/denisbrodbeck/machineid v1.0.1

37+ github.com/disintegration/imageorient v0.0.0-20180920195336-8147d86e83ec

39 github.com/disintegration/imaging v1.6.2

38 github.com/disintegration/imaging v1.6.2

40 github.com/dustin/go-humanize v1.0.1

39 github.com/dustin/go-humanize v1.0.1

41− github.com/gen2brain/beeep v0.11.2

42− github.com/go-git/go-git/v5 v5.19.1

43 github.com/google/uuid v1.6.0

40 github.com/google/uuid v1.6.0

44− github.com/invopop/jsonschema v0.14.0

41+ github.com/invopop/jsonschema v0.13.0

45− github.com/itchyny/gojq v0.12.19

46 github.com/joho/godotenv v1.5.1

42 github.com/joho/godotenv v1.5.1

47 github.com/jordanella/go-ansi-paintbrush v0.0.0-20240728195301-b7ad996ecf3d

43 github.com/jordanella/go-ansi-paintbrush v0.0.0-20240728195301-b7ad996ecf3d

48− github.com/lucasb-eyer/go-colorful v1.4.0

44+ github.com/lucasb-eyer/go-colorful v1.3.0

49− github.com/mattn/go-isatty v0.0.22

45+ github.com/mattn/go-isatty v0.0.20

50− github.com/modelcontextprotocol/go-sdk v1.6.1

46+ github.com/modelcontextprotocol/go-sdk v1.2.0

47+ github.com/muesli/termenv v0.16.0

51− github.com/ncruces/go-sqlite3 v0.35.1

48+ github.com/ncruces/go-sqlite3 v0.30.5

49+ github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646

52 github.com/nxadm/tail v1.4.11

50 github.com/nxadm/tail v1.4.11

51+ github.com/openai/openai-go/v2 v2.7.1

53 github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c

52 github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c

54− github.com/posthog/posthog-go v1.17.5

53+ github.com/posthog/posthog-go v1.9.1

55− github.com/pressly/goose/v3 v3.27.2

54+ github.com/pressly/goose/v3 v3.26.0

56− github.com/qjebbs/go-jsons v1.0.0-alpha.6

55+ github.com/qjebbs/go-jsons v1.0.0-alpha.4

57 github.com/rivo/uniseg v0.4.7

56 github.com/rivo/uniseg v0.4.7

57+ github.com/sabhiram/go-gitignore v0.0.0-20210923224102-525f6e181f06

58− github.com/sahilm/fuzzy v0.1.3

58+ github.com/sahilm/fuzzy v0.1.1

59− github.com/sourcegraph/jsonrpc2 v0.2.1

60 github.com/spf13/cobra v1.10.2

59 github.com/spf13/cobra v1.10.2

60+ github.com/srwiley/oksvg v0.0.0-20221011165216-be6e8873101c

61+ github.com/srwiley/rasterx v0.0.0-20220730225603-2ab79fcdd4ef

61 github.com/stretchr/testify v1.11.1

62 github.com/stretchr/testify v1.11.1

62− github.com/swaggo/http-swagger/v2 v2.0.2

63− github.com/swaggo/swag v1.16.6

64− github.com/tidwall/gjson v1.19.0

63+ github.com/tidwall/gjson v1.18.0

65 github.com/tidwall/sjson v1.2.5

64 github.com/tidwall/sjson v1.2.5

66 github.com/zeebo/xxh3 v1.1.0

65 github.com/zeebo/xxh3 v1.1.0

67− go.uber.org/goleak v1.3.0

68− golang.design/x/clipboard v0.8.0

66+ golang.org/x/mod v0.32.0

69− golang.org/x/net v0.56.0

67+ golang.org/x/net v0.49.0

70− golang.org/x/sync v0.21.0

68+ golang.org/x/sync v0.19.0

71− golang.org/x/sys v0.46.0

72− golang.org/x/text v0.38.0

69+ golang.org/x/text v0.33.0

73 gopkg.in/natefinch/lumberjack.v2 v2.2.1

70 gopkg.in/natefinch/lumberjack.v2 v2.2.1

74 gopkg.in/yaml.v3 v3.0.1

71 gopkg.in/yaml.v3 v3.0.1

75− modernc.org/sqlite v1.53.0

72+ modernc.org/sqlite v1.44.3

76 mvdan.cc/sh/moreinterp v0.0.0-20250902163504-3cf4fd5717a5

73 mvdan.cc/sh/moreinterp v0.0.0-20250902163504-3cf4fd5717a5

77− mvdan.cc/sh/v3 v3.13.1

74+ mvdan.cc/sh/v3 v3.12.1-0.20250902163504-3cf4fd5717a5

78)

75)

80require (

77require (

81− cloud.google.com/go v0.123.0 // indirect

78+ cloud.google.com/go v0.116.0 // indirect

82− cloud.google.com/go/auth v0.20.0 // indirect

79+ cloud.google.com/go/auth v0.18.0 // indirect

83 cloud.google.com/go/auth/oauth2adapt v0.2.8 // indirect

80 cloud.google.com/go/auth/oauth2adapt v0.2.8 // indirect

84 cloud.google.com/go/compute/metadata v0.9.0 // indirect

81 cloud.google.com/go/compute/metadata v0.9.0 // indirect

85− git.sr.ht/~jackmordaunt/go-toast v1.1.2 // indirect

86 github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0 // indirect

82 github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0 // indirect

87 github.com/Azure/azure-sdk-for-go/sdk/internal v1.10.0 // indirect

83 github.com/Azure/azure-sdk-for-go/sdk/internal v1.10.0 // indirect

88− github.com/KyleBanks/depth v1.2.1 // indirect

84+ github.com/RealAlexandreAI/json-repair v0.0.14 // indirect

89− github.com/andybalholm/brotli v1.2.1 // indirect

90 github.com/andybalholm/cascadia v1.3.3 // indirect

85 github.com/andybalholm/cascadia v1.3.3 // indirect

91− github.com/atotto/clipboard v0.1.4 // indirect

92− github.com/aws/aws-sdk-go-v2 v1.42.0 // indirect

86+ github.com/aws/aws-sdk-go-v2 v1.41.1 // indirect

93− github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.13 // indirect

87+ github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3 // indirect

94− github.com/aws/aws-sdk-go-v2/config v1.32.25 // indirect

88+ github.com/aws/aws-sdk-go-v2/config v1.32.7 // indirect

95− github.com/aws/aws-sdk-go-v2/credentials v1.19.24 // indirect

89+ github.com/aws/aws-sdk-go-v2/credentials v1.19.7 // indirect

96− github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.29 // indirect

90+ github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.17 // indirect

97− github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.29 // indirect

91+ github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.17 // indirect

98− github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.29 // indirect

92+ github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.17 // indirect

99− github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.30 // indirect

93+ github.com/aws/aws-sdk-go-v2/internal/ini v1.8.4 // indirect

100− github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.12 // indirect

94+ github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.4 // indirect

101− github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.29 // indirect

95+ github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.17 // indirect

102− github.com/aws/aws-sdk-go-v2/service/signin v1.2.0 // indirect

96+ github.com/aws/aws-sdk-go-v2/service/signin v1.0.5 // indirect

103− github.com/aws/aws-sdk-go-v2/service/sso v1.31.3 // indirect

97+ github.com/aws/aws-sdk-go-v2/service/sso v1.30.9 // indirect

104− github.com/aws/aws-sdk-go-v2/service/ssooidc v1.36.6 // indirect

98+ github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.13 // indirect

105− github.com/aws/aws-sdk-go-v2/service/sts v1.43.3 // indirect

99+ github.com/aws/aws-sdk-go-v2/service/sts v1.41.6 // indirect

106− github.com/aws/smithy-go v1.27.3 // indirect

100+ github.com/aws/smithy-go v1.24.0 // indirect

101+ github.com/aymanbagabas/go-osc52/v2 v2.0.1 // indirect

107 github.com/aymerick/douceur v0.2.0 // indirect

102 github.com/aymerick/douceur v0.2.0 // indirect

108 github.com/bahlo/generic-list-go v0.2.0 // indirect

103 github.com/bahlo/generic-list-go v0.2.0 // indirect

109− github.com/buger/jsonparser v1.1.2 // indirect

104+ github.com/buger/jsonparser v1.1.1 // indirect

110− github.com/cespare/xxhash/v2 v2.3.0 // indirect

111− github.com/charmbracelet/anthropic-sdk-go v0.0.0-20260223140439-63879b0b8dab // indirect

105+ github.com/charmbracelet/anthropic-sdk-go v0.0.0-20251024181547-21d6f3d9a904 // indirect

106+ github.com/charmbracelet/x/json v0.2.0 // indirect

112 github.com/charmbracelet/x/termios v0.1.1 // indirect

107 github.com/charmbracelet/x/termios v0.1.1 // indirect

113 github.com/charmbracelet/x/windows v0.2.2 // indirect

108 github.com/charmbracelet/x/windows v0.2.2 // indirect

114− github.com/cyphar/filepath-securejoin v0.6.1 // indirect

109+ github.com/clipperhouse/displaywidth v0.7.0 // indirect

115− github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc // indirect

110+ github.com/clipperhouse/stringish v0.1.1 // indirect

111+ github.com/clipperhouse/uax29/v2 v2.3.0 // indirect

112+ github.com/davecgh/go-spew v1.1.1 // indirect

113+ github.com/disintegration/gift v1.1.2 // indirect

116− github.com/dlclark/regexp2/v2 v2.2.1 // indirect

114+ github.com/dlclark/regexp2 v1.11.5 // indirect

117− github.com/ebitengine/purego v0.10.1 // indirect

115+ github.com/ebitengine/purego v0.10.0-alpha.3.0.20260102153238-200df6041cff // indirect

118− github.com/esiqveland/notify v0.13.3 // indirect

119− github.com/felixge/httpsnoop v1.1.0 // indirect

116+ github.com/felixge/httpsnoop v1.0.4 // indirect

120 github.com/fsnotify/fsnotify v1.9.0 // indirect

117 github.com/fsnotify/fsnotify v1.9.0 // indirect

121− github.com/go-git/gcfg v1.5.1-0.20230307220236-3a3c6141e376 // indirect

122− github.com/go-git/go-billy/v5 v5.9.0 // indirect

123− github.com/go-json-experiment/json v0.0.0-20260623181947-01eb4420fa68 // indirect

118+ github.com/go-json-experiment/json v0.0.0-20251027170946-4849db3c2f7e // indirect

124 github.com/go-logfmt/logfmt v0.6.0 // indirect

119 github.com/go-logfmt/logfmt v0.6.0 // indirect

125 github.com/go-logr/logr v1.4.3 // indirect

120 github.com/go-logr/logr v1.4.3 // indirect

126 github.com/go-logr/stdr v1.2.2 // indirect

121 github.com/go-logr/stdr v1.2.2 // indirect

127− github.com/go-ole/go-ole v1.3.0 // indirect

128− github.com/go-openapi/jsonpointer v0.19.5 // indirect

129− github.com/go-openapi/jsonreference v0.20.0 // indirect

130− github.com/go-openapi/spec v0.20.6 // indirect

131− github.com/go-openapi/swag v0.19.15 // indirect

132− github.com/go-viper/mapstructure/v2 v2.5.0 // indirect

122+ github.com/go-viper/mapstructure/v2 v2.4.0 // indirect

133 github.com/goccy/go-json v0.10.5 // indirect

123 github.com/goccy/go-json v0.10.5 // indirect

134− github.com/goccy/go-yaml v1.19.2 // indirect

124+ github.com/goccy/go-yaml v1.19.0 // indirect

135− github.com/godbus/dbus/v5 v5.2.2 // indirect

136 github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0 // indirect

125 github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0 // indirect

137 github.com/google/go-cmp v0.7.0 // indirect

126 github.com/google/go-cmp v0.7.0 // indirect

138− github.com/google/jsonschema-go v0.4.3 // indirect

127+ github.com/google/jsonschema-go v0.3.0 // indirect

139 github.com/google/s2a-go v0.1.9 // indirect

128 github.com/google/s2a-go v0.1.9 // indirect

140− github.com/googleapis/enterprise-certificate-proxy v0.3.17 // indirect

129+ github.com/googleapis/enterprise-certificate-proxy v0.3.7 // indirect

141− github.com/googleapis/gax-go/v2 v2.22.0 // indirect

130+ github.com/googleapis/gax-go/v2 v2.15.0 // indirect

142 github.com/gorilla/css v1.0.1 // indirect

131 github.com/gorilla/css v1.0.1 // indirect

143 github.com/gorilla/websocket v1.5.3 // indirect

132 github.com/gorilla/websocket v1.5.3 // indirect

144 github.com/hashicorp/golang-lru/v2 v2.0.7 // indirect

133 github.com/hashicorp/golang-lru/v2 v2.0.7 // indirect

145 github.com/inconshreveable/mousetrap v1.1.0 // indirect

134 github.com/inconshreveable/mousetrap v1.1.0 // indirect

146− github.com/itchyny/timefmt-go v0.1.8 // indirect

135+ github.com/kaptinlin/go-i18n v0.2.2 // indirect

147− github.com/jackmordaunt/icns/v3 v3.0.1 // indirect

136+ github.com/kaptinlin/jsonpointer v0.4.8 // indirect

148− github.com/jbenet/go-context v0.0.0-20150711004518-d14ea06fba99 // indirect

149− github.com/josharian/intern v1.0.0 // indirect

150− github.com/kaptinlin/jsonpointer v0.4.26 // indirect

151− github.com/kaptinlin/jsonschema v0.9.2 // indirect

137+ github.com/kaptinlin/jsonschema v0.6.6 // indirect

138+ github.com/kaptinlin/messageformat-go v0.4.7 // indirect

152− github.com/klauspost/compress v1.18.6 // indirect

139+ github.com/klauspost/compress v1.18.0 // indirect

153− github.com/klauspost/cpuid/v2 v2.3.0 // indirect

140+ github.com/klauspost/cpuid/v2 v2.2.10 // indirect

154 github.com/klauspost/pgzip v1.2.6 // indirect

141 github.com/klauspost/pgzip v1.2.6 // indirect

155 github.com/mailru/easyjson v0.7.7 // indirect

142 github.com/mailru/easyjson v0.7.7 // indirect

156− github.com/mattn/go-runewidth v0.0.24 // indirect

143+ github.com/mattn/go-runewidth v0.0.19 // indirect

157 github.com/mfridman/interpolate v0.0.2 // indirect

144 github.com/mfridman/interpolate v0.0.2 // indirect

158 github.com/microcosm-cc/bluemonday v1.0.27 // indirect

145 github.com/microcosm-cc/bluemonday v1.0.27 // indirect

159 github.com/mitchellh/mapstructure v1.5.0 // indirect

146 github.com/mitchellh/mapstructure v1.5.0 // indirect

@@ -162,60 +149,48 @@ require (

162 github.com/muesli/mango-cobra v1.2.0 // indirect

149 github.com/muesli/mango-cobra v1.2.0 // indirect

163 github.com/muesli/mango-pflag v0.1.0 // indirect

150 github.com/muesli/mango-pflag v0.1.0 // indirect

164 github.com/muesli/roff v0.1.0 // indirect

151 github.com/muesli/roff v0.1.0 // indirect

165− github.com/ncruces/go-sqlite3-wasm/v3 v3.1.35302 // indirect

166 github.com/ncruces/go-strftime v1.0.0 // indirect

152 github.com/ncruces/go-strftime v1.0.0 // indirect

167 github.com/ncruces/julianday v1.0.0 // indirect

153 github.com/ncruces/julianday v1.0.0 // indirect

168− github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646 // indirect

169− github.com/pb33f/ordered-map/v2 v2.3.1 // indirect

170− github.com/pierrec/lz4/v4 v4.1.26 // indirect

154+ github.com/pierrec/lz4/v4 v4.1.22 // indirect

171− github.com/pjbgf/sha1cd v0.6.0 // indirect

172− github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 // indirect

155+ github.com/pmezard/go-difflib v1.0.0 // indirect

173 github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect

156 github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect

174− github.com/segmentio/asm v1.2.1 // indirect

157+ github.com/sergi/go-diff v1.3.2-0.20230802210424-5b0b94c5c0d3 // indirect

175− github.com/segmentio/encoding v0.5.4 // indirect

176− github.com/sergeymakinen/go-bmp v1.0.0 // indirect

177− github.com/sergeymakinen/go-ico v1.0.0-beta.0 // indirect

178 github.com/sethvargo/go-retry v0.3.0 // indirect

158 github.com/sethvargo/go-retry v0.3.0 // indirect

159+ github.com/sourcegraph/jsonrpc2 v0.2.1 // indirect

179 github.com/spf13/pflag v1.0.9 // indirect

160 github.com/spf13/pflag v1.0.9 // indirect

180− github.com/swaggo/files/v2 v2.0.0 // indirect

161+ github.com/tetratelabs/wazero v1.11.0 // indirect

181− github.com/tadvi/systray v0.0.0-20190226123456-11a2b8fa57af // indirect

182 github.com/tidwall/match v1.1.1 // indirect

162 github.com/tidwall/match v1.1.1 // indirect

183 github.com/tidwall/pretty v1.2.1 // indirect

163 github.com/tidwall/pretty v1.2.1 // indirect

184 github.com/u-root/u-root v0.14.1-0.20250807200646-5e7721023dc7 // indirect

164 github.com/u-root/u-root v0.14.1-0.20250807200646-5e7721023dc7 // indirect

185 github.com/u-root/uio v0.0.0-20240224005618-d2acac8f3701 // indirect

165 github.com/u-root/uio v0.0.0-20240224005618-d2acac8f3701 // indirect

166+ github.com/wk8/go-ordered-map/v2 v2.1.8 // indirect

186 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect

167 github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e // indirect

187 github.com/yosida95/uritemplate/v3 v3.0.2 // indirect

168 github.com/yosida95/uritemplate/v3 v3.0.2 // indirect

188 github.com/yuin/goldmark v1.7.8 // indirect

169 github.com/yuin/goldmark v1.7.8 // indirect

189 github.com/yuin/goldmark-emoji v1.0.5 // indirect

170 github.com/yuin/goldmark-emoji v1.0.5 // indirect

190− go.opentelemetry.io/auto/sdk v1.2.1 // indirect

171+ go.opentelemetry.io/auto/sdk v1.1.0 // indirect

191− go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.69.0 // indirect

172+ go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.61.0 // indirect

192− go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.69.0 // indirect

173+ go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.61.0 // indirect

193− go.opentelemetry.io/otel v1.44.0 // indirect

174+ go.opentelemetry.io/otel v1.37.0 // indirect

194− go.opentelemetry.io/otel/metric v1.44.0 // indirect

175+ go.opentelemetry.io/otel/metric v1.37.0 // indirect

195− go.opentelemetry.io/otel/trace v1.44.0 // indirect

176+ go.opentelemetry.io/otel/trace v1.37.0 // indirect

196 go.uber.org/multierr v1.11.0 // indirect

177 go.uber.org/multierr v1.11.0 // indirect

197 go.yaml.in/yaml/v4 v4.0.0-rc.3 // indirect

178 go.yaml.in/yaml/v4 v4.0.0-rc.3 // indirect

198− golang.design/x/x11 v0.2.0 // indirect

199− golang.org/x/crypto v0.53.0 // indirect

179+ golang.org/x/crypto v0.47.0 // indirect

200− golang.org/x/exp v0.0.0-20260410095643-746e56fc9e2f // indirect

201− golang.org/x/exp/shiny v0.0.0-20250606033433-dcc06ee1d476 // indirect

180+ golang.org/x/exp v0.0.0-20251023183803-a4bb9ffd2546 // indirect

202− golang.org/x/image v0.43.0 // indirect

181+ golang.org/x/image v0.34.0 // indirect

203− golang.org/x/mobile v0.0.0-20250606033058-a2a15c67f36f // indirect

204− golang.org/x/mod v0.36.0 // indirect

205− golang.org/x/oauth2 v0.36.0 // indirect

182+ golang.org/x/oauth2 v0.34.0 // indirect

183+ golang.org/x/sys v0.40.0 // indirect

206− golang.org/x/term v0.44.0 // indirect

184+ golang.org/x/term v0.39.0 // indirect

207− golang.org/x/time v0.15.0 // indirect

185+ golang.org/x/time v0.14.0 // indirect

208− golang.org/x/tools v0.45.0 // indirect

209− google.golang.org/api v0.286.0 // indirect

186+ google.golang.org/api v0.239.0 // indirect

210− google.golang.org/genai v1.62.0 // indirect

187+ google.golang.org/genai v1.41.0 // indirect

211− google.golang.org/genproto/googleapis/rpc v0.0.0-20260622175928-b703f567277d // indirect

188+ google.golang.org/genproto/googleapis/rpc v0.0.0-20250804133106-a7a43d27e69b // indirect

212− google.golang.org/grpc v1.81.1 // indirect

189+ google.golang.org/grpc v1.76.0 // indirect

213− google.golang.org/protobuf v1.36.11 // indirect

190+ google.golang.org/protobuf v1.36.10 // indirect

214 gopkg.in/dnaeon/go-vcr.v4 v4.0.6-0.20251110073552-01de4eb40290 // indirect

191 gopkg.in/dnaeon/go-vcr.v4 v4.0.6-0.20251110073552-01de4eb40290 // indirect

215 gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7 // indirect

192 gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7 // indirect

216− gopkg.in/warnings.v0 v0.1.2 // indirect

217− gopkg.in/yaml.v2 v2.4.0 // indirect

218− modernc.org/libc v1.73.4 // indirect

193+ modernc.org/libc v1.67.6 // indirect

219 modernc.org/mathutil v1.7.1 // indirect

194 modernc.org/mathutil v1.7.1 // indirect

220 modernc.org/memory v1.11.0 // indirect

195 modernc.org/memory v1.11.0 // indirect

221)

196)

go.sum+240−314

@@ -1,31 +1,25 @@

1−charm.land/bubbles/v2 v2.1.1 h1:7r55WzBxpo/R3z98hGmY7KKPd3ET6vsf0Fb9sDHOV60=

1+charm.land/bubbles/v2 v2.0.0-rc.1.0.20260109112849-ae99f46cec66 h1:2BdJynsAW+8rv9xq6ZS+x0mtacfxpxjIK1KUIeTqBOs=

2−charm.land/bubbles/v2 v2.1.1/go.mod h1:GE6M31gaWZVXzGw73OeuTTgy4lX+OtkH0E5ymnNsHxo=

2+charm.land/bubbles/v2 v2.0.0-rc.1.0.20260109112849-ae99f46cec66/go.mod h1:5AbN6cEd/47gkEf8TgiQ2O3RZ5QxMS14l9W+7F9fPC4=

3−charm.land/bubbletea/v2 v2.0.8 h1:SxTJMhCAI3lbPmy4SgX5LWZ24AdINr4I6UEqzZvYJuY=

3+charm.land/bubbletea/v2 v2.0.0-rc.2.0.20251216153312-819e2e89c62e h1:tXwTmgGpwZT7ParKF5xbEQBVjM2e1uKhKi/GpfU3mYQ=

4−charm.land/bubbletea/v2 v2.0.8/go.mod h1:2SkdgoTXluXJHOUwAoRlRXF/28vklb1rFl6GcgV1/ss=

4+charm.land/bubbletea/v2 v2.0.0-rc.2.0.20251216153312-819e2e89c62e/go.mod h1:pDM18flq3Z4njKZPA3zCvyVSSIJbMcoqlE82BdGUtL8=

5−charm.land/catwalk v0.49.0 h1:pXcHh9H76VvwGqPYMZhJe5vRqMLpzGsv0anJArKw70g=

5+charm.land/fantasy v0.6.1 h1:v3pavSHpZ5xTw98TpNYoj6DRq4ksCBWwJiZeiG/mVIc=

6−charm.land/catwalk v0.49.0/go.mod h1:bw6/oiChsa9Fkr8LgYn84KrbbJY2z1+m7Jmlmmv3B2A=

6+charm.land/fantasy v0.6.1/go.mod h1:Ifj41bNnIXJ1aF6sLKcS9y3MzWbDnObmcHrCaaHfpZ0=

7−charm.land/fang/v2 v2.0.1 h1:zQCM8JQJ1JnQX/66B5jlCYBUxL2as5JXQZ2KJ6EL0mY=

7+charm.land/glamour/v2 v2.0.0-20260123212943-6014aa153a9b h1:A6IUUyChZDWP16RUdRJCfmYISAKWQGyIcfhZJUCViQ0=

8−charm.land/fang/v2 v2.0.1/go.mod h1:S1GmkpcvK+OB5w9caywUnJcsMew45Ot8FXqoz8ALrII=

8+charm.land/glamour/v2 v2.0.0-20260123212943-6014aa153a9b/go.mod h1:J3kVhY6oHXZq5f+8vC3hmDO95fEvbqj3z7xDwxrfzU8=

9−charm.land/fantasy v0.35.0 h1:iFES7cnt3/o20T01+y2PDkHj5dHEWRFQH0vSXYzV8xo=

9+charm.land/lipgloss/v2 v2.0.0-beta.3.0.20251205162909-7869489d8971 h1:xZFcNsJMiIDbFtWRyDmkKNk1sjojfaom4Zoe0cyH/8c=

10−charm.land/fantasy v0.35.0/go.mod h1:rUqUhunz+c6k66L0hHd51CrXB3SSzPNKkwBrzIe6XnY=

10+charm.land/lipgloss/v2 v2.0.0-beta.3.0.20251205162909-7869489d8971/go.mod h1:i61Y3FmdbcBNSKa+pKB3DaE4uVQmBLMs/xlvRyHcXAE=

11−charm.land/glamour/v2 v2.0.1 h1:xl+r00A4aJWU0z8fgwKd9fQQ4rsphqGUzuEiXZP5n+c=

11+charm.land/log/v2 v2.0.0-20251110204020-529bb77f35da h1:vZa/Ow0uLclpfaDY0ubjzE+B0eLQqi2zanmpeALanow=

12−charm.land/glamour/v2 v2.0.1/go.mod h1:jo9z8XqVKPeEFMVdvCRLGk++RyJ3CdUwgNr7EvXLw3k=

12+charm.land/log/v2 v2.0.0-20251110204020-529bb77f35da/go.mod h1:Tj12StbPc4GwksDF6XwhC9wdXouinIVxRGKKmmmzdSU=

13−charm.land/lipgloss/v2 v2.0.5 h1:kbNxgeeUOYv5J0YdpxFjfvf3dFvqH8Aci4zB6xqFtrY=

14−charm.land/lipgloss/v2 v2.0.5/go.mod h1:9oqhxt4yxIMe6q5A4kHr44DremZk7J9UNh74GlWa5nc=

15−charm.land/log/v2 v2.0.0 h1:SY3Cey7ipx86/MBXQHwsguOT6X1exT94mmJRdzTNs+s=

16−charm.land/log/v2 v2.0.0/go.mod h1:c3cZSRqm20qUVVAR1WmS/7ab8bgha3C6G7DjPcaVZz0=

17charm.land/x/vcr v0.1.1 h1:PXCFMUG0rPtyk35rhfzYCJEduOzWXCIbrXTFq4OF/9Q=

13charm.land/x/vcr v0.1.1 h1:PXCFMUG0rPtyk35rhfzYCJEduOzWXCIbrXTFq4OF/9Q=

18charm.land/x/vcr v0.1.1/go.mod h1:eByq2gqzWvcct/8XE2XO5KznoWEBiXH56+y2gphbltM=

14charm.land/x/vcr v0.1.1/go.mod h1:eByq2gqzWvcct/8XE2XO5KznoWEBiXH56+y2gphbltM=

19−cloud.google.com/go v0.123.0 h1:2NAUJwPR47q+E35uaJeYoNhuNEM9kM8SjgRgdeOJUSE=

15+cloud.google.com/go v0.116.0 h1:B3fRrSDkLRt5qSHWe40ERJvhvnQwdZiHu0bJOpldweE=

20−cloud.google.com/go v0.123.0/go.mod h1:xBoMV08QcqUGuPW65Qfm1o9Y4zKZBpGS+7bImXLTAZU=

16+cloud.google.com/go v0.116.0/go.mod h1:cEPSRWPzZEswwdr9BxE6ChEn01dWlTaF05LiC2Xs70U=

21−cloud.google.com/go/auth v0.20.0 h1:kXTssoVb4azsVDoUiF8KvxAqrsQcQtB53DcSgta74CA=

17+cloud.google.com/go/auth v0.18.0 h1:wnqy5hrv7p3k7cShwAU/Br3nzod7fxoqG+k0VZ+/Pk0=

22−cloud.google.com/go/auth v0.20.0/go.mod h1:942/yi/itH1SsmpyrbnTMDgGfdy2BUqIKyd0cyYLc5Q=

18+cloud.google.com/go/auth v0.18.0/go.mod h1:wwkPM1AgE1f2u6dG443MiWoD8C3BtOywNsUMcUTVDRo=

23cloud.google.com/go/auth/oauth2adapt v0.2.8 h1:keo8NaayQZ6wimpNSmW5OPc283g65QNIiLpZnkHRbnc=

19cloud.google.com/go/auth/oauth2adapt v0.2.8 h1:keo8NaayQZ6wimpNSmW5OPc283g65QNIiLpZnkHRbnc=

24cloud.google.com/go/auth/oauth2adapt v0.2.8/go.mod h1:XQ9y31RkqZCcwJWNSx2Xvric3RrU88hAYYbjDWYDL+c=

20cloud.google.com/go/auth/oauth2adapt v0.2.8/go.mod h1:XQ9y31RkqZCcwJWNSx2Xvric3RrU88hAYYbjDWYDL+c=

25cloud.google.com/go/compute/metadata v0.9.0 h1:pDUj4QMoPejqq20dK0Pg2N4yG9zIkYGdBtwLoEkH9Zs=

21cloud.google.com/go/compute/metadata v0.9.0 h1:pDUj4QMoPejqq20dK0Pg2N4yG9zIkYGdBtwLoEkH9Zs=

26cloud.google.com/go/compute/metadata v0.9.0/go.mod h1:E0bWwX5wTnLPedCKqk3pJmVgCBSM6qQI1yTBdEb3C10=

22cloud.google.com/go/compute/metadata v0.9.0/go.mod h1:E0bWwX5wTnLPedCKqk3pJmVgCBSM6qQI1yTBdEb3C10=

27−git.sr.ht/~jackmordaunt/go-toast v1.1.2 h1:/yrfI55LRt1M7H1vkaw+NaH1+L1CDxrqDltwm5euVuE=

28−git.sr.ht/~jackmordaunt/go-toast v1.1.2/go.mod h1:jA4OqHKTQ4AFBdwrSnwnskUIIS3HYzlJSgdzCKqfavo=

29github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0 h1:g0EZJwz7xkXQiZAI5xi9f3WWFYBlX1CPTrR+NDToRkQ=

23github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0 h1:g0EZJwz7xkXQiZAI5xi9f3WWFYBlX1CPTrR+NDToRkQ=

30github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0/go.mod h1:XCW7KnZet0Opnr7HccfUw1PLc4CjHqpcaxW8DHklNkQ=

24github.com/Azure/azure-sdk-for-go/sdk/azcore v1.17.0/go.mod h1:XCW7KnZet0Opnr7HccfUw1PLc4CjHqpcaxW8DHklNkQ=

31github.com/Azure/azure-sdk-for-go/sdk/azidentity v1.7.0 h1:tfLQ34V6F7tVSwoTf/4lH5sE0o6eCJuNDTmH09nDpbc=

25github.com/Azure/azure-sdk-for-go/sdk/azidentity v1.7.0 h1:tfLQ34V6F7tVSwoTf/4lH5sE0o6eCJuNDTmH09nDpbc=

@@ -36,154 +30,146 @@ github.com/AzureAD/microsoft-authentication-library-for-go v1.2.2 h1:XHOnouVk1mx

36github.com/AzureAD/microsoft-authentication-library-for-go v1.2.2/go.mod h1:wP83P5OoQ5p6ip3ScPr0BAq0BvuPAvacpEuSzyouqAI=

30github.com/AzureAD/microsoft-authentication-library-for-go v1.2.2/go.mod h1:wP83P5OoQ5p6ip3ScPr0BAq0BvuPAvacpEuSzyouqAI=

37github.com/JohannesKaufmann/html-to-markdown v1.6.0 h1:04VXMiE50YYfCfLboJCLcgqF5x+rHJnb1ssNmqpLH/k=

31github.com/JohannesKaufmann/html-to-markdown v1.6.0 h1:04VXMiE50YYfCfLboJCLcgqF5x+rHJnb1ssNmqpLH/k=

38github.com/JohannesKaufmann/html-to-markdown v1.6.0/go.mod h1:NUI78lGg/a7vpEJTz/0uOcYMaibytE4BUOQS8k78yPQ=

32github.com/JohannesKaufmann/html-to-markdown v1.6.0/go.mod h1:NUI78lGg/a7vpEJTz/0uOcYMaibytE4BUOQS8k78yPQ=

39−github.com/KyleBanks/depth v1.2.1 h1:5h8fQADFrWtarTdtDudMmGsC7GPbOAu6RVB3ffsVFHc=

40−github.com/KyleBanks/depth v1.2.1/go.mod h1:jzSb9d0L43HxTQfT+oSA1EEp2q+ne2uh6XgeJcm8brE=

41github.com/MakeNowJust/heredoc v1.0.0 h1:cXCdzVdstXyiTqTvfqk9SDHpKNjxuom+DOlyEeQ4pzQ=

33github.com/MakeNowJust/heredoc v1.0.0 h1:cXCdzVdstXyiTqTvfqk9SDHpKNjxuom+DOlyEeQ4pzQ=

42github.com/MakeNowJust/heredoc v1.0.0/go.mod h1:mG5amYoWBHf8vpLOuehzbGGw0EHxpZZ6lCpQ4fNJ8LE=

34github.com/MakeNowJust/heredoc v1.0.0/go.mod h1:mG5amYoWBHf8vpLOuehzbGGw0EHxpZZ6lCpQ4fNJ8LE=

43−github.com/Microsoft/go-winio v0.6.2 h1:F2VQgta7ecxGYO8k3ZZz3RS8fVIXVxONVUPlNERoyfY=

44−github.com/Microsoft/go-winio v0.6.2/go.mod h1:yd8OoFMLzJbo9gZq8j5qaps8bJ9aShtEA8Ipt1oGCvU=

45github.com/PuerkitoBio/goquery v1.9.2/go.mod h1:GHPCaP0ODyyxqcNoFGYlAprUFH81NuRPd0GX3Zu2Mvk=

35github.com/PuerkitoBio/goquery v1.9.2/go.mod h1:GHPCaP0ODyyxqcNoFGYlAprUFH81NuRPd0GX3Zu2Mvk=

46−github.com/PuerkitoBio/goquery v1.12.0 h1:pAcL4g3WRXekcB9AU/y1mbKez2dbY2AajVhtkO8RIBo=

36+github.com/PuerkitoBio/goquery v1.11.0 h1:jZ7pwMQXIITcUXNH83LLk+txlaEy6NVOfTuP43xxfqw=

47−github.com/PuerkitoBio/goquery v1.12.0/go.mod h1:802ej+gV2y7bbIhOIoPY5sT183ZW0YFofScC4q/hIpQ=

37+github.com/PuerkitoBio/goquery v1.11.0/go.mod h1:wQHgxUOU3JGuj3oD/QFfxUdlzW6xPHfqyHre6VMY4DQ=

38+github.com/RealAlexandreAI/json-repair v0.0.14 h1:4kTqotVonDVTio5n2yweRUELVcNe2x518wl0bCsw0t0=

39+github.com/RealAlexandreAI/json-repair v0.0.14/go.mod h1:GKJi5borR78O8c7HCVbgqjhoiVibZ6hJldxbc6dGrAI=

48github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=

40github.com/alecthomas/assert/v2 v2.11.0 h1:2Q9r3ki8+JYXvGsDyBXwH3LcJ+WK5D0gc5E8vS6K3D0=

49github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=

41github.com/alecthomas/assert/v2 v2.11.0/go.mod h1:Bze95FyfUr7x34QZrjL+XP+0qgp/zg8yS+TtBj1WA3k=

50−github.com/alecthomas/chroma/v2 v2.27.0 h1:FodwmyOBgJULFYmDqibcp9pvfDLWdtPRh9v/r5BXYZs=

42+github.com/alecthomas/chroma/v2 v2.23.1 h1:nv2AVZdTyClGbVQkIzlDm/rnhk1E9bU9nXwmZ/Vk/iY=

51−github.com/alecthomas/chroma/v2 v2.27.0/go.mod h1:NjJ3ciIgrqBNeIkWZ4e46nseoLDslxU1LmfCoL+wcY8=

43+github.com/alecthomas/chroma/v2 v2.23.1/go.mod h1:NqVhfBR0lte5Ouh3DcthuUCTUpDC9cxBOfyMbMQPs3o=

52github.com/alecthomas/repr v0.5.2 h1:SU73FTI9D1P5UNtvseffFSGmdNci/O6RsqzeXJtP0Qs=

44github.com/alecthomas/repr v0.5.2 h1:SU73FTI9D1P5UNtvseffFSGmdNci/O6RsqzeXJtP0Qs=

53github.com/alecthomas/repr v0.5.2/go.mod h1:Fr0507jx4eOXV7AlPV6AVZLYrLIuIeSOWtW57eE/O/4=

45github.com/alecthomas/repr v0.5.2/go.mod h1:Fr0507jx4eOXV7AlPV6AVZLYrLIuIeSOWtW57eE/O/4=

54−github.com/andybalholm/brotli v1.2.1 h1:R+f5xP285VArJDRgowrfb9DqL18yVK0gKAW/F+eTWro=

55−github.com/andybalholm/brotli v1.2.1/go.mod h1:rzTDkvFWvIrjDXZHkuS16NPggd91W3kUSvPlQ1pLaKY=

56github.com/andybalholm/cascadia v1.3.2/go.mod h1:7gtRlve5FxPPgIgX36uWBX58OdBsSS6lUvCFb+h7KvU=

46github.com/andybalholm/cascadia v1.3.2/go.mod h1:7gtRlve5FxPPgIgX36uWBX58OdBsSS6lUvCFb+h7KvU=

57github.com/andybalholm/cascadia v1.3.3 h1:AG2YHrzJIm4BZ19iwJ/DAua6Btl3IwJX+VI4kktS1LM=

47github.com/andybalholm/cascadia v1.3.3 h1:AG2YHrzJIm4BZ19iwJ/DAua6Btl3IwJX+VI4kktS1LM=

58github.com/andybalholm/cascadia v1.3.3/go.mod h1:xNd9bqTn98Ln4DwST8/nG+H0yuB8Hmgu1YHNnWw0GeA=

48github.com/andybalholm/cascadia v1.3.3/go.mod h1:xNd9bqTn98Ln4DwST8/nG+H0yuB8Hmgu1YHNnWw0GeA=

59github.com/atotto/clipboard v0.1.4 h1:EH0zSVneZPSuFR11BlR9YppQTVDbh5+16AmcJi4g1z4=

49github.com/atotto/clipboard v0.1.4 h1:EH0zSVneZPSuFR11BlR9YppQTVDbh5+16AmcJi4g1z4=

60github.com/atotto/clipboard v0.1.4/go.mod h1:ZY9tmq7sm5xIbd9bOK4onWV4S6X0u6GY7Vn0Yu86PYI=

50github.com/atotto/clipboard v0.1.4/go.mod h1:ZY9tmq7sm5xIbd9bOK4onWV4S6X0u6GY7Vn0Yu86PYI=

61−github.com/aws/aws-sdk-go-v2 v1.42.0 h1:XvXMJTkFQtpBKIWZnmr9ZEOc2InWM2yldjXEJ/bymhA=

51+github.com/aws/aws-sdk-go-v2 v1.41.1 h1:ABlyEARCDLN034NhxlRUSZr4l71mh+T5KAeGh6cerhU=

62−github.com/aws/aws-sdk-go-v2 v1.42.0/go.mod h1:27+ACypSLljLAEKsCYOmrjKh83vuTRkuAe9Uv/3A4bg=

52+github.com/aws/aws-sdk-go-v2 v1.41.1/go.mod h1:MayyLB8y+buD9hZqkCW3kX1AKq07Y5pXxtgB+rRFhz0=

63−github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.13 h1:p1BBrg/Hhp6uK7zpejeI8QFXHJeC/mynzi04Sl03k9g=

53+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3 h1:tW1/Rkad38LA15X4UQtjXZXNKsCgkshC3EbmcUmghTg=

64−github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.13/go.mod h1:8cIfkE9MDhkRZGpQ22aV6/lkYeYSozpz16Smrs5x4Ls=

54+github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3/go.mod h1:UbnqO+zjqk3uIt9yCACHJ9IVNhyhOCnYk8yA19SAWrM=

65−github.com/aws/aws-sdk-go-v2/config v1.32.25 h1:ACCejvStYoilgwrfegSt5ZntCbPrk52qfwyNcnl3omM=

55+github.com/aws/aws-sdk-go-v2/config v1.32.7 h1:vxUyWGUwmkQ2g19n7JY/9YL8MfAIl7bTesIUykECXmY=

66−github.com/aws/aws-sdk-go-v2/config v1.32.25/go.mod h1:LJyU8sDRbXUxFn8xMJIGP+v9QYYwveNLI8a/giAOiAs=

56+github.com/aws/aws-sdk-go-v2/config v1.32.7/go.mod h1:2/Qm5vKUU/r7Y+zUk/Ptt2MDAEKAfUtKc1+3U1Mo3oY=

67−github.com/aws/aws-sdk-go-v2/credentials v1.19.24 h1:2hQqYCV9yqyePQ9o6dCrZc/zO8U3TwPr9mIKlZnPu/I=

57+github.com/aws/aws-sdk-go-v2/credentials v1.19.7 h1:tHK47VqqtJxOymRrNtUXN5SP/zUTvZKeLx4tH6PGQc8=

68−github.com/aws/aws-sdk-go-v2/credentials v1.19.24/go.mod h1:IDwpACtwqHLISdzfwUUNq4P9DsB/h5BLg4FwJPNfqFY=

58+github.com/aws/aws-sdk-go-v2/credentials v1.19.7/go.mod h1:qOZk8sPDrxhf+4Wf4oT2urYJrYt3RejHSzgAquYeppw=

69−github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.29 h1:r6qZHbT+wxgWO/e9vYNUEtg7lv5+UN3pRqKhLXvnArg=

59+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.17 h1:I0GyV8wiYrP8XpA70g1HBcQO1JlQxCMTW9npl5UbDHY=

70−github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.29/go.mod h1:QRnaRcTVGKPGRy8w78HMQtKUGRYcnMZAANATkeVA6Mo=

60+github.com/aws/aws-sdk-go-v2/feature/ec2/imds v1.18.17/go.mod h1:tyw7BOl5bBe/oqvoIeECFJjMdzXoa/dfVz3QQ5lgHGA=

71−github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.29 h1:f3vKqSo13fhTYb+JEcXwXefZQE26I1FB5eTSniU67ko=

61+github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.17 h1:xOLELNKGp2vsiteLsvLPwxC+mYmO6OZ8PYgiuPJzF8U=

72−github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.29/go.mod h1:MzoLFUArKGpGD+ukmPiTPG1X5x4o6M2kq4v2dr1FiEc=

62+github.com/aws/aws-sdk-go-v2/internal/configsources v1.4.17/go.mod h1:5M5CI3D12dNOtH3/mk6minaRwI2/37ifCURZISxA/IQ=

73−github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.29 h1:RdwIf/CuUsvJX3RgJagbOyotl/cxoLY4xviKuE7p2GY=

63+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.17 h1:WWLqlh79iO48yLkj1v3ISRNiv+3KdQoZ6JWyfcsyQik=

74−github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.29/go.mod h1:71wt8W2EgswdZy9Mf9KNnzxZ3TiZlv4caKghPktDOkA=

64+github.com/aws/aws-sdk-go-v2/internal/endpoints/v2 v2.7.17/go.mod h1:EhG22vHRrvF8oXSTYStZhJc1aUgKtnJe+aOiFEV90cM=

75−github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.30 h1:VTGy885W5DKBxWRUJbym9hytNaYzsyaPkCHGRRMAOhU=

65+github.com/aws/aws-sdk-go-v2/internal/ini v1.8.4 h1:WKuaxf++XKWlHWu9ECbMlha8WOEGm0OUEZqm4K/Gcfk=

76−github.com/aws/aws-sdk-go-v2/internal/v4a v1.4.30/go.mod h1:AS0HycUvJRFvTt613AYDOgO2jzw+00cVSMny8XB3yMY=

66+github.com/aws/aws-sdk-go-v2/internal/ini v1.8.4/go.mod h1:ZWy7j6v1vWGmPReu0iSGvRiise4YI5SkR3OHKTZ6Wuc=

77−github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.12 h1:ZD2+BSw9vFsNlKYIasSNt3uDbjqqXIBcM13UJv/Lx2k=

67+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.4 h1:0ryTNEdJbzUCEWkVXEXoqlXV72J5keC1GvILMOuD00E=

78−github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.12/go.mod h1:Ms4zlcVBbXbiP7EVLhl+lgjvA/a7YphqQ3Ih3174EmI=

68+github.com/aws/aws-sdk-go-v2/service/internal/accept-encoding v1.13.4/go.mod h1:HQ4qwNZh32C3CBeO6iJLQlgtMzqeG17ziAA/3KDJFow=

79−github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.29 h1:DRebniUGZ2MqiiIVmQJ04vIXr918hubdHMnarSLEWyU=

69+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.17 h1:RuNSMoozM8oXlgLG/n6WLaFGoea7/CddrCfIiSA+xdY=

80−github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.29/go.mod h1:LfRkPCD8YHDM2E5eTkos2UpwYeZnBcVarTa8L59bJHA=

70+github.com/aws/aws-sdk-go-v2/service/internal/presigned-url v1.13.17/go.mod h1:F2xxQ9TZz5gDWsclCtPQscGpP0VUOc8RqgFM3vDENmU=

81−github.com/aws/aws-sdk-go-v2/service/signin v1.2.0 h1:3nXpRcFwRCW8n7HgO2QGy0Dc20eQNfBuUemGQhpF8m8=

71+github.com/aws/aws-sdk-go-v2/service/signin v1.0.5 h1:VrhDvQib/i0lxvr3zqlUwLwJP4fpmpyD9wYG1vfSu+Y=

82−github.com/aws/aws-sdk-go-v2/service/signin v1.2.0/go.mod h1:LxYujSTLPRlp2vTtcUO/+1ilrew8ytt6SvQyOgejzFQ=

72+github.com/aws/aws-sdk-go-v2/service/signin v1.0.5/go.mod h1:k029+U8SY30/3/ras4G/Fnv/b88N4mAfliNn08Dem4M=

83−github.com/aws/aws-sdk-go-v2/service/sso v1.31.3 h1:ey1XLTYXb9PcLt4535632o5kCGXNXEhNb620Dqwuylo=

73+github.com/aws/aws-sdk-go-v2/service/sso v1.30.9 h1:v6EiMvhEYBoHABfbGB4alOYmCIrcgyPPiBE1wZAEbqk=

84−github.com/aws/aws-sdk-go-v2/service/sso v1.31.3/go.mod h1:Lk7PlmoTYryQmyBG0EXqj5BcUbj3whXdU2s3yGI3EAc=

74+github.com/aws/aws-sdk-go-v2/service/sso v1.30.9/go.mod h1:yifAsgBxgJWn3ggx70A3urX2AN49Y5sJTD1UQFlfqBw=

85−github.com/aws/aws-sdk-go-v2/service/ssooidc v1.36.6 h1:yLr03zQE/5Eu5l3QU0Si+xMbLMbSDF2YXsigqXngs6g=

75+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.13 h1:gd84Omyu9JLriJVCbGApcLzVR3XtmC4ZDPcAI6Ftvds=

86−github.com/aws/aws-sdk-go-v2/service/ssooidc v1.36.6/go.mod h1:Q5N6icH+KJZDLh+ESNwzdv6cZ6vLFF/egy3IOxWhmz4=

76+github.com/aws/aws-sdk-go-v2/service/ssooidc v1.35.13/go.mod h1:sTGThjphYE4Ohw8vJiRStAcu3rbjtXRsdNB0TvZ5wwo=

87−github.com/aws/aws-sdk-go-v2/service/sts v1.43.3 h1:VrIhKRCSK1umelSgB9RghvA9RTUYeQffyAS5ApXehNI=

77+github.com/aws/aws-sdk-go-v2/service/sts v1.41.6 h1:5fFjR/ToSOzB2OQ/XqWpZBmNvmP/pJ1jOWYlFDJTjRQ=

88−github.com/aws/aws-sdk-go-v2/service/sts v1.43.3/go.mod h1:r8wkDOuLaaMFqFiYAb8dGY2A3gJCOujMc6CFOVC4Zhc=

78+github.com/aws/aws-sdk-go-v2/service/sts v1.41.6/go.mod h1:qgFDZQSD/Kys7nJnVqYlWKnh0SSdMjAi0uSwON4wgYQ=

89−github.com/aws/smithy-go v1.27.3 h1:F3Zb497UhhskkfpJmfkXswyo+t0sh9OTBnIHjogWbVY=

79+github.com/aws/smithy-go v1.24.0 h1:LpilSUItNPFr1eY85RYgTIg5eIEPtvFbskaFcmmIUnk=

90−github.com/aws/smithy-go v1.27.3/go.mod h1:YE2RhdIuDbA5E5bTdciG9KrW3+TiEONeUWCqxX9i1Fc=

80+github.com/aws/smithy-go v1.24.0/go.mod h1:LEj2LM3rBRQJxPZTB4KuzZkaZYnZPnvgIhb4pu07mx0=

81+github.com/aymanbagabas/go-nativeclipboard v0.1.2 h1:Z2iVRWQ4IynMLWM6a+lWH2Nk5gPyEtPRMuBIyZ2dECM=

82+github.com/aymanbagabas/go-nativeclipboard v0.1.2/go.mod h1:BVJhN7hs5DieCzUB2Atf4Yk9Y9kFe62E95+gOjpJq6Q=

83+github.com/aymanbagabas/go-osc52/v2 v2.0.1 h1:HwpRHbFMcZLEVr42D4p7XBqjyuxQH5SMiErDT4WkJ2k=

84+github.com/aymanbagabas/go-osc52/v2 v2.0.1/go.mod h1:uYgXzlJ7ZpABp8OJ+exZzJJhRNQ2ASbcXHWsFqH8hp8=

91−github.com/aymanbagabas/go-udiff v0.4.1 h1:OEIrQ8maEeDBXQDoGCbbTTXYJMYRCRO1fnodZ12Gv5o=

85+github.com/aymanbagabas/go-udiff v0.3.1 h1:LV+qyBQ2pqe0u42ZsUEtPiCaUoqgA9gYRDs3vj1nolY=

92−github.com/aymanbagabas/go-udiff v0.4.1/go.mod h1:0L9PGwj20lrtmEMeyw4WKJ/TMyDtvAoK9bf2u/mNo3w=

86+github.com/aymanbagabas/go-udiff v0.3.1/go.mod h1:G0fsKmG+P6ylD0r6N/KgQD/nWzgfnl8ZBcNLgcbrw8E=

93github.com/aymerick/douceur v0.2.0 h1:Mv+mAeH1Q+n9Fr+oyamOlAkUNPWPlA8PPGR0QAaYuPk=

87github.com/aymerick/douceur v0.2.0 h1:Mv+mAeH1Q+n9Fr+oyamOlAkUNPWPlA8PPGR0QAaYuPk=

94github.com/aymerick/douceur v0.2.0/go.mod h1:wlT5vV2O3h55X9m7iVYN0TBM0NH/MmbLnd30/FjWUq4=

88github.com/aymerick/douceur v0.2.0/go.mod h1:wlT5vV2O3h55X9m7iVYN0TBM0NH/MmbLnd30/FjWUq4=

95github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=

89github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=

96github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=

90github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=

97github.com/bmatcuk/doublestar/v4 v4.10.0 h1:zU9WiOla1YA122oLM6i4EXvGW62DvKZVxIe6TYWexEs=

91github.com/bmatcuk/doublestar/v4 v4.10.0 h1:zU9WiOla1YA122oLM6i4EXvGW62DvKZVxIe6TYWexEs=

98github.com/bmatcuk/doublestar/v4 v4.10.0/go.mod h1:xBQ8jztBU6kakFMg+8WGxn0c6z1fTSPVIjEY1Wr7jzc=

92github.com/bmatcuk/doublestar/v4 v4.10.0/go.mod h1:xBQ8jztBU6kakFMg+8WGxn0c6z1fTSPVIjEY1Wr7jzc=

99−github.com/buger/jsonparser v1.1.2 h1:frqHqw7otoVbk5M8LlE/L7HTnIq2v9RX6EJ48i9AxJk=

93+github.com/buger/jsonparser v1.1.1 h1:2PnMjfWD7wBILjqQbt530v576A/cAbQvEW9gGIpYMUs=

100−github.com/buger/jsonparser v1.1.2/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=

94+github.com/buger/jsonparser v1.1.1/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=

101−github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=

102−github.com/cespare/xxhash/v2 v2.3.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=

103github.com/charlievieth/fastwalk v1.0.14 h1:3Eh5uaFGwHZd8EGwTjJnSpBkfwfsak9h6ICgnWlhAyg=

95github.com/charlievieth/fastwalk v1.0.14 h1:3Eh5uaFGwHZd8EGwTjJnSpBkfwfsak9h6ICgnWlhAyg=

104github.com/charlievieth/fastwalk v1.0.14/go.mod h1:diVcUreiU1aQ4/Wu3NbxxH4/KYdKpLDojrQ1Bb2KgNY=

96github.com/charlievieth/fastwalk v1.0.14/go.mod h1:diVcUreiU1aQ4/Wu3NbxxH4/KYdKpLDojrQ1Bb2KgNY=

105−github.com/charmbracelet/anthropic-sdk-go v0.0.0-20260223140439-63879b0b8dab h1:J7XQLgl9sefgTnTGrmX3xqvp5o6MCiBzEjGv5igAlc4=

97+github.com/charmbracelet/anthropic-sdk-go v0.0.0-20251024181547-21d6f3d9a904 h1:rwLdEpG9wE6kL69KkEKDiWprO8pQOZHZXeod6+9K+mw=

106−github.com/charmbracelet/anthropic-sdk-go v0.0.0-20260223140439-63879b0b8dab/go.mod h1:hqlYqR7uPKOKfnNeicUbZp0Ps0GeYFlKYtwh5HGDCx8=

98+github.com/charmbracelet/anthropic-sdk-go v0.0.0-20251024181547-21d6f3d9a904/go.mod h1:8TIYxZxsuCqqeJ0lga/b91tBwrbjoHDC66Sq5t8N2R4=

107−github.com/charmbracelet/colorprofile v0.4.3 h1:QPa1IWkYI+AOB+fE+mg/5/4HRMZcaXex9t5KX76i20Q=

99+github.com/charmbracelet/catwalk v0.15.0 h1:5oWJdvchTPfF7855A0n40+XbZQz4+vouZ/NhQ661JKI=

100+github.com/charmbracelet/catwalk v0.15.0/go.mod h1:qg+Yl9oaZTkTvRscqbxfttzOFQ4v0pOT5XwC7b5O0NQ=

101+github.com/charmbracelet/colorprofile v0.4.1 h1:a1lO03qTrSIRaK8c3JRxJDZOvhvIeSco3ej+ngLk1kk=

108−github.com/charmbracelet/colorprofile v0.4.3/go.mod h1:/zT4BhpD5aGFpqQQqw7a+VtHCzu+zrQtt1zhMt9mR4Q=

102+github.com/charmbracelet/colorprofile v0.4.1/go.mod h1:U1d9Dljmdf9DLegaJ0nGZNJvoXAhayhmidOdcBwAvKk=

109−github.com/charmbracelet/openai-go v0.0.0-20260617131321-5e4b9c18c4be h1:pg+OWlIkk9HOe/8P5J95aKe2wGDzFUiiyFOUpwR30B4=

103+github.com/charmbracelet/fang v0.4.4 h1:G4qKxF6or/eTPgmAolwPuRNyuci3hTUGGX1rj1YkHJY=

110−github.com/charmbracelet/openai-go v0.0.0-20260617131321-5e4b9c18c4be/go.mod h1:1DahUaExbUZx/jD+FNT2PKP4L9rLE5+ZBRuI8mZjd/E=

104+github.com/charmbracelet/fang v0.4.4/go.mod h1:P5/DNb9DddQ0Z0dbc0P3ol4/ix5Po7Ofr2KMBfAqoCo=

111−github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7 h1:3FmWoGNWK4STvqg0O0Aeav2T7rodWJAPeF0QpH+8gFw=

105+github.com/charmbracelet/ultraviolet v0.0.0-20251212194010-b927aa605560 h1:j3PW2hypGoPKBy3ooKzW0TFxaxhyHK3NbkLLn4KeRFc=

112−github.com/charmbracelet/ultraviolet v0.0.0-20260703014108-f5a850f9c2b7/go.mod h1:f/jRa757WUmaOZrbPspXymbg/GnbF+rwe4OLsG7aXYo=

106+github.com/charmbracelet/ultraviolet v0.0.0-20251212194010-b927aa605560/go.mod h1:VWATWLRwYP06VYCEur7FsNR2B1xAo7Y+xl1PTbd1ePc=

113−github.com/charmbracelet/x/ansi v0.11.7 h1:kzv1kJvjg2S3r9KHo8hDdHFQLEqn4RBCb39dAYC84jI=

107+github.com/charmbracelet/x/ansi v0.11.4 h1:6G65PLu6HjmE858CnTUQY1LXT3ZUWwfvqEROLF8vqHI=

114−github.com/charmbracelet/x/ansi v0.11.7/go.mod h1:9qGpnAVYz+8ACONkZBUWPtL7lulP9No6p1epAihUZwQ=

108+github.com/charmbracelet/x/ansi v0.11.4/go.mod h1:/5AZ+UfWExW3int5H5ugnsG/PWjNcSQcwYsHBlPFQN4=

115github.com/charmbracelet/x/editor v0.2.0 h1:7XLUKtaRaB8jN7bWU2p2UChiySyaAuIfYiIRg8gGWwk=

109github.com/charmbracelet/x/editor v0.2.0 h1:7XLUKtaRaB8jN7bWU2p2UChiySyaAuIfYiIRg8gGWwk=

116github.com/charmbracelet/x/editor v0.2.0/go.mod h1:p3oQ28TSL3YPd+GKJ1fHWcp+7bVGpedHpXmo0D6t1dY=

110github.com/charmbracelet/x/editor v0.2.0/go.mod h1:p3oQ28TSL3YPd+GKJ1fHWcp+7bVGpedHpXmo0D6t1dY=

117github.com/charmbracelet/x/etag v0.2.0 h1:Euj1VkheoHfTYA9y+TCwkeXF/hN8Fb9l4LqZl79pt04=

111github.com/charmbracelet/x/etag v0.2.0 h1:Euj1VkheoHfTYA9y+TCwkeXF/hN8Fb9l4LqZl79pt04=

118github.com/charmbracelet/x/etag v0.2.0/go.mod h1:C1B7/bsgvzzxpfu0Rabbd+rTHJa5TmC/qgTseCf6DF0=

112github.com/charmbracelet/x/etag v0.2.0/go.mod h1:C1B7/bsgvzzxpfu0Rabbd+rTHJa5TmC/qgTseCf6DF0=

119−github.com/charmbracelet/x/exp/charmtone v0.0.0-20260527151214-009e6338d40d h1:sMilwx1YIYTrQva6jsB522AoRYAerNaDIKP4ZPtUq0A=

113+github.com/charmbracelet/x/exp/charmtone v0.0.0-20260109001716-2fbdffcb221f h1:OKFNbG2sSmgpQW9EC3gYNG+QrcQ4+wWYjzfmJvWkkDo=

120−github.com/charmbracelet/x/exp/charmtone v0.0.0-20260527151214-009e6338d40d/go.mod h1:nsExn0DGyX0lh9LwLHTn2Gg+hafdzfSXnC+QmEJTZFY=

114+github.com/charmbracelet/x/exp/charmtone v0.0.0-20260109001716-2fbdffcb221f/go.mod h1:nsExn0DGyX0lh9LwLHTn2Gg+hafdzfSXnC+QmEJTZFY=

121github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f h1:pk6gmGpCE7F3FcjaOEKYriCvpmIN4+6OS/RD0vm4uIA=

115github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f h1:pk6gmGpCE7F3FcjaOEKYriCvpmIN4+6OS/RD0vm4uIA=

122github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f/go.mod h1:IfZAMTHB6XkZSeXUqriemErjAWCCzT0LwjKFYCZyw0I=

116github.com/charmbracelet/x/exp/golden v0.0.0-20250806222409-83e3a29d542f/go.mod h1:IfZAMTHB6XkZSeXUqriemErjAWCCzT0LwjKFYCZyw0I=

123github.com/charmbracelet/x/exp/ordered v0.1.0 h1:55/qLwjIh0gL0Vni+QAWk7T/qRVP6sBf+2agPBgnOFE=

117github.com/charmbracelet/x/exp/ordered v0.1.0 h1:55/qLwjIh0gL0Vni+QAWk7T/qRVP6sBf+2agPBgnOFE=

124github.com/charmbracelet/x/exp/ordered v0.1.0/go.mod h1:5UHwmG+is5THxMyCJHNPCn2/ecI07aKNrW+LcResjJ8=

118github.com/charmbracelet/x/exp/ordered v0.1.0/go.mod h1:5UHwmG+is5THxMyCJHNPCn2/ecI07aKNrW+LcResjJ8=

125−github.com/charmbracelet/x/exp/slice v0.0.0-20260622092256-25656177ba8e h1:TerxTCW/3BrnpqKxeouW0cWSjZb7PsVAmtNftRAAFRE=

119+github.com/charmbracelet/x/exp/slice v0.0.0-20251201173703-9f73bfd934ff h1:Uwr+/JS+qnRcO/++xjYEDtW7x+P5E4+4cBiOHTt2Xfk=

126−github.com/charmbracelet/x/exp/slice v0.0.0-20260622092256-25656177ba8e/go.mod h1:vqEfX6xzqW1pKKZUUiFOKg0OQ7bCh54Q2vR/tserrRA=

120+github.com/charmbracelet/x/exp/slice v0.0.0-20251201173703-9f73bfd934ff/go.mod h1:vqEfX6xzqW1pKKZUUiFOKg0OQ7bCh54Q2vR/tserrRA=

127github.com/charmbracelet/x/exp/strings v0.1.0 h1:i69S2XI7uG1u4NLGeJPSYU++Nmjvpo9nwd6aoEm7gkA=

121github.com/charmbracelet/x/exp/strings v0.1.0 h1:i69S2XI7uG1u4NLGeJPSYU++Nmjvpo9nwd6aoEm7gkA=

128github.com/charmbracelet/x/exp/strings v0.1.0/go.mod h1:/ehtMPNh9K4odGFkqYJKpIYyePhdp1hLBRvyY4bWkH8=

122github.com/charmbracelet/x/exp/strings v0.1.0/go.mod h1:/ehtMPNh9K4odGFkqYJKpIYyePhdp1hLBRvyY4bWkH8=

129−github.com/charmbracelet/x/powernap v0.1.6 h1:tsnQWQVl9arYm5S/tQVXuMJcvnkUNNkXhBWlEBDRNqM=

123+github.com/charmbracelet/x/json v0.2.0 h1:DqB+ZGx2h+Z+1s98HOuOyli+i97wsFQIxP2ZQANTPrQ=

124+github.com/charmbracelet/x/json v0.2.0/go.mod h1:opFIflx2YgXgi49xVUu8gEQ21teFAxyMwvOiZhIvWNM=

125+github.com/charmbracelet/x/powernap v0.0.0-20260127155452-b72a9a918687 h1:h1XMgTkpBt9kEJ+9DkARNBXEgaigUQ0cI2Bot7Awnt8=

130−github.com/charmbracelet/x/powernap v0.1.6/go.mod h1:cmdl5zlP5mR8TF2Y68UKc7hdGUDiSJ2+4hk0h04Hsx4=

126+github.com/charmbracelet/x/powernap v0.0.0-20260127155452-b72a9a918687/go.mod h1:cmdl5zlP5mR8TF2Y68UKc7hdGUDiSJ2+4hk0h04Hsx4=

131github.com/charmbracelet/x/term v0.2.2 h1:xVRT/S2ZcKdhhOuSP4t5cLi5o+JxklsoEObBSgfgZRk=

127github.com/charmbracelet/x/term v0.2.2 h1:xVRT/S2ZcKdhhOuSP4t5cLi5o+JxklsoEObBSgfgZRk=

132github.com/charmbracelet/x/term v0.2.2/go.mod h1:kF8CY5RddLWrsgVwpw4kAa6TESp6EB5y3uxGLeCqzAI=

128github.com/charmbracelet/x/term v0.2.2/go.mod h1:kF8CY5RddLWrsgVwpw4kAa6TESp6EB5y3uxGLeCqzAI=

133github.com/charmbracelet/x/termios v0.1.1 h1:o3Q2bT8eqzGnGPOYheoYS8eEleT5ZVNYNy8JawjaNZY=

129github.com/charmbracelet/x/termios v0.1.1 h1:o3Q2bT8eqzGnGPOYheoYS8eEleT5ZVNYNy8JawjaNZY=

134github.com/charmbracelet/x/termios v0.1.1/go.mod h1:rB7fnv1TgOPOyyKRJ9o+AsTU/vK5WHJ2ivHeut/Pcwo=

130github.com/charmbracelet/x/termios v0.1.1/go.mod h1:rB7fnv1TgOPOyyKRJ9o+AsTU/vK5WHJ2ivHeut/Pcwo=

135github.com/charmbracelet/x/windows v0.2.2 h1:IofanmuvaxnKHuV04sC0eBy/smG6kIKrWG2/jYn2GuM=

131github.com/charmbracelet/x/windows v0.2.2 h1:IofanmuvaxnKHuV04sC0eBy/smG6kIKrWG2/jYn2GuM=

136github.com/charmbracelet/x/windows v0.2.2/go.mod h1:/8XtdKZzedat74NQFn0NGlGL4soHB0YQZrETF96h75k=

132github.com/charmbracelet/x/windows v0.2.2/go.mod h1:/8XtdKZzedat74NQFn0NGlGL4soHB0YQZrETF96h75k=

137−github.com/clipperhouse/displaywidth v0.11.0 h1:lBc6kY44VFw+TDx4I8opi/EtL9m20WSEFgwIwO+UVM8=

133+github.com/clipperhouse/displaywidth v0.7.0 h1:QNv1GYsnLX9QBrcWUtMlogpTXuM5FVnBwKWp1O5NwmE=

138−github.com/clipperhouse/displaywidth v0.11.0/go.mod h1:bkrFNkf81G8HyVqmKGxsPufD3JhNl3dSqnGhOoSD/o0=

134+github.com/clipperhouse/displaywidth v0.7.0/go.mod h1:R+kHuzaYWFkTm7xoMmK1lFydbci4X2CicfbGstSGg0o=

139−github.com/clipperhouse/uax29/v2 v2.7.0 h1:+gs4oBZ2gPfVrKPthwbMzWZDaAFPGYK72F0NJv2v7Vk=

135+github.com/clipperhouse/stringish v0.1.1 h1:+NSqMOr3GR6k1FdRhhnXrLfztGzuG+VuFDfatpWHKCs=

136+github.com/clipperhouse/stringish v0.1.1/go.mod h1:v/WhFtE1q0ovMta2+m+UbpZ+2/HEXNWYXQgCt4hdOzA=

137+github.com/clipperhouse/uax29/v2 v2.3.0 h1:SNdx9DVUqMoBuBoW3iLOj4FQv3dN5mDtuqwuhIGpJy4=

140−github.com/clipperhouse/uax29/v2 v2.7.0/go.mod h1:EFJ2TJMRUaplDxHKj1qAEhCtQPW2tJSwu5BF98AuoVM=

138+github.com/clipperhouse/uax29/v2 v2.3.0/go.mod h1:Wn1g7MK6OoeDT0vL+Q0SQLDz/KpfsVRgg6W7ihQeh4g=

141−github.com/cncf/xds/go v0.0.0-20260202195803-dba9d589def2 h1:aBangftG7EVZoUb69Os8IaYg++6uMOdKK83QtkkvJik=

139+github.com/cncf/xds/go v0.0.0-20250501225837-2ac532fd4443 h1:aQ3y1lwWyqYPiWZThqv1aFbZMiM9vblcSArJRf2Irls=

142−github.com/cncf/xds/go v0.0.0-20260202195803-dba9d589def2/go.mod h1:qwXFYgsP6T7XnJtbKlf1HP8AjxZZyzxMmc+Lq5GjlU4=

140+github.com/cncf/xds/go v0.0.0-20250501225837-2ac532fd4443/go.mod h1:W+zGtBO5Y1IgJhy4+A9GOqVhqLpfZi+vwmdNXUehLA8=

143github.com/cpuguy83/go-md2man/v2 v2.0.6/go.mod h1:oOW0eioCTA6cOiMLiUPZOpcVxMig6NIQQ7OS05n1F4g=

141github.com/cpuguy83/go-md2man/v2 v2.0.6/go.mod h1:oOW0eioCTA6cOiMLiUPZOpcVxMig6NIQQ7OS05n1F4g=

144−github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=

145github.com/creack/pty v1.1.24 h1:bJrF4RRfyJnbTJqzRLHzcGaZK1NeM5kTC9jGgovnR1s=

142github.com/creack/pty v1.1.24 h1:bJrF4RRfyJnbTJqzRLHzcGaZK1NeM5kTC9jGgovnR1s=

146github.com/creack/pty v1.1.24/go.mod h1:08sCNb52WyoAwi2QDyzUCTgcvVFhUzewun7wtTfvcwE=

143github.com/creack/pty v1.1.24/go.mod h1:08sCNb52WyoAwi2QDyzUCTgcvVFhUzewun7wtTfvcwE=

147−github.com/cyphar/filepath-securejoin v0.6.1 h1:5CeZ1jPXEiYt3+Z6zqprSAgSWiggmpVyciv8syjIpVE=

148−github.com/cyphar/filepath-securejoin v0.6.1/go.mod h1:A8hd4EnAeyujCJRrICiOWqjS1AX0a9kM5XL+NwKoYSc=

149github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=

144github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=

145+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=

150github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=

146github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=

151−github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc h1:U9qPSI2PIWSS1VwoXQT9A3Wy9MM3WgvqSxFWenqJduM=

152−github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=

153github.com/denisbrodbeck/machineid v1.0.1 h1:geKr9qtkB876mXguW2X6TU4ZynleN6ezuMSRhl4D7AQ=

147github.com/denisbrodbeck/machineid v1.0.1 h1:geKr9qtkB876mXguW2X6TU4ZynleN6ezuMSRhl4D7AQ=

154github.com/denisbrodbeck/machineid v1.0.1/go.mod h1:dJUwb7PTidGDeYyUBmXZ2GphQBbjJCrnectwCyxcUSI=

148github.com/denisbrodbeck/machineid v1.0.1/go.mod h1:dJUwb7PTidGDeYyUBmXZ2GphQBbjJCrnectwCyxcUSI=

149+github.com/disintegration/gift v1.1.2 h1:9ZyHJr+kPamiH10FX3Pynt1AxFUob812bU9Wt4GMzhs=

150+github.com/disintegration/gift v1.1.2/go.mod h1:Jh2i7f7Q2BM7Ezno3PhfezbR1xpUg9dUg3/RlKGr4HI=

151+github.com/disintegration/imageorient v0.0.0-20180920195336-8147d86e83ec h1:YrB6aVr9touOt75I9O1SiancmR2GMg45U9UYf0gtgWg=

152+github.com/disintegration/imageorient v0.0.0-20180920195336-8147d86e83ec/go.mod h1:K0KBFIr1gWu/C1Gp10nFAcAE4hsB7JxE6OgLijrJ8Sk=

155github.com/disintegration/imaging v1.6.2 h1:w1LecBlG2Lnp8B3jk5zSuNqd7b4DXhcjwek1ei82L+c=

153github.com/disintegration/imaging v1.6.2 h1:w1LecBlG2Lnp8B3jk5zSuNqd7b4DXhcjwek1ei82L+c=

156github.com/disintegration/imaging v1.6.2/go.mod h1:44/5580QXChDfwIclfc/PCwrr44amcmDAg8hxG0Ewe4=

154github.com/disintegration/imaging v1.6.2/go.mod h1:44/5580QXChDfwIclfc/PCwrr44amcmDAg8hxG0Ewe4=

157−github.com/dlclark/regexp2/v2 v2.2.1 h1:mf4KkFUj0gJuarK8P+LgiS+Lit7m9N1yAwEfPbee7R0=

155+github.com/dlclark/regexp2 v1.11.5 h1:Q/sSnsKerHeCkc/jSTNq1oCm7KiVgUMZRDUoRu0JQZQ=

158−github.com/dlclark/regexp2/v2 v2.2.1/go.mod h1:avUrQvPaLz2DrFNHJF0taWAFFX2C1GMSSoeiqFjcBmU=

156+github.com/dlclark/regexp2 v1.11.5/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=

159−github.com/dnaeon/go-vcr v1.2.0 h1:zHCHvJYTMh1N7xnV7zf1m1GPBF9Ad0Jk/whtQ1663qI=

160−github.com/dnaeon/go-vcr v1.2.0/go.mod h1:R4UdLID7HZT3taECzJs4YgbbH6PIGXB6W/sc5OLb6RQ=

161github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=

157github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=

162github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=

158github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=

163−github.com/ebitengine/purego v0.10.1 h1:dewVBCBT2GaMu1SrNTYxQhgQBethzfhiwvZiLGP/qyY=

159+github.com/ebitengine/purego v0.10.0-alpha.3.0.20260102153238-200df6041cff h1:vAcU1VsCRstZ9ty11yD/L0WDyT73S/gVfmuWvcWX5DA=

164−github.com/ebitengine/purego v0.10.1/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=

160+github.com/ebitengine/purego v0.10.0-alpha.3.0.20260102153238-200df6041cff/go.mod h1:iIjxzd6CiRiOG0UyXP+V1+jWqUXVjPKLAI0mRfJZTmQ=

165−github.com/envoyproxy/go-control-plane v0.14.0 h1:hbG2kr4RuFj222B6+7T83thSPqLjwBIfQawTkC++2HA=

161+github.com/envoyproxy/go-control-plane v0.13.4 h1:zEqyPVyku6IvWCFwux4x9RxkLOMUL+1vC9xUFv5l2/M=

166−github.com/envoyproxy/go-control-plane/envoy v1.37.0 h1:u3riX6BoYRfF4Dr7dwSOroNfdSbEPe9Yyl09/B6wBrQ=

162+github.com/envoyproxy/go-control-plane/envoy v1.32.4 h1:jb83lalDRZSpPWW2Z7Mck/8kXZ5CQAFYVjQcdVIr83A=

167−github.com/envoyproxy/go-control-plane/envoy v1.37.0/go.mod h1:DReE9MMrmecPy+YvQOAOHNYMALuowAnbjjEMkkWOi6A=

163+github.com/envoyproxy/go-control-plane/envoy v1.32.4/go.mod h1:Gzjc5k8JcJswLjAx1Zm+wSYE20UrLtt7JZMWiWQXQEw=

168−github.com/envoyproxy/protoc-gen-validate v1.3.3 h1:MVQghNeW+LZcmXe7SY1V36Z+WFMDjpqGAGacLe2T0ds=

164+github.com/envoyproxy/protoc-gen-validate v1.2.1 h1:DEo3O99U8j4hBFwbJfrz9VtgcDfUKS7KJ7spH3d86P8=

169−github.com/envoyproxy/protoc-gen-validate v1.3.3/go.mod h1:TsndJ/ngyIdQRhMcVVGDDHINPLWB7C82oDArY51KfB0=

165+github.com/envoyproxy/protoc-gen-validate v1.2.1/go.mod h1:d/C80l/jxXLdfEIhX1W2TmLfsJ31lvEjwamM4DxlWXU=

170−github.com/esiqveland/notify v0.13.3 h1:QCMw6o1n+6rl+oLUfg8P1IIDSFsDEb2WlXvVvIJbI/o=

166+github.com/felixge/httpsnoop v1.0.4 h1:NFTV2Zj1bL4mc9sqWACXbQFVBBg2W3GPvqp8/ESS2Wg=

171−github.com/esiqveland/notify v0.13.3/go.mod h1:hesw/IRYTO0x99u1JPweAl4+5mwXJibQVUcP0Iu5ORE=

172−github.com/felixge/httpsnoop v1.1.0 h1:3YtUj32ZZkqZtt3sZZsClsymw/QDuVfpNhoA31zeORc=

173−github.com/felixge/httpsnoop v1.1.0/go.mod h1:Zqxgdd+1Rkcz8euOqdr7lqgCRJztwr5hp9vDSi5UZCE=

167+github.com/felixge/httpsnoop v1.0.4/go.mod h1:m8KPJKqk1gH5J9DgRY2ASl2lWCfGKXixSwevea8zH2U=

174github.com/fsnotify/fsnotify v1.6.0/go.mod h1:sl3t1tCWJFWoRz9R8WJCbQihKKwmorjAbSClcnxKAGw=

168github.com/fsnotify/fsnotify v1.6.0/go.mod h1:sl3t1tCWJFWoRz9R8WJCbQihKKwmorjAbSClcnxKAGw=

175github.com/fsnotify/fsnotify v1.9.0 h1:2Ml+OJNzbYCTzsxtv8vKSFD9PbJjmhYF14k/jKC7S9k=

169github.com/fsnotify/fsnotify v1.9.0 h1:2Ml+OJNzbYCTzsxtv8vKSFD9PbJjmhYF14k/jKC7S9k=

176github.com/fsnotify/fsnotify v1.9.0/go.mod h1:8jBTzvmWwFyi3Pb8djgCCO5IBqzKJ/Jwo8TRcHyHii0=

170github.com/fsnotify/fsnotify v1.9.0/go.mod h1:8jBTzvmWwFyi3Pb8djgCCO5IBqzKJ/Jwo8TRcHyHii0=

177−github.com/gen2brain/beeep v0.11.2 h1:+KfiKQBbQCuhfJFPANZuJ+oxsSKAYNe88hIpJuyKWDA=

171+github.com/go-json-experiment/json v0.0.0-20251027170946-4849db3c2f7e h1:Lf/gRkoycfOBPa42vU2bbgPurFong6zXeFtPoxholzU=

178−github.com/gen2brain/beeep v0.11.2/go.mod h1:jQVvuwnLuwOcdctHn/uyh8horSBNJ8uGb9Cn2W4tvoc=

172+github.com/go-json-experiment/json v0.0.0-20251027170946-4849db3c2f7e/go.mod h1:uNVvRXArCGbZ508SxYYTC5v1JWoz2voff5pm25jU1Ok=

179−github.com/go-git/gcfg v1.5.1-0.20230307220236-3a3c6141e376 h1:+zs/tPmkDkHx3U66DAb0lQFJrpS6731Oaa12ikc+DiI=

180−github.com/go-git/gcfg v1.5.1-0.20230307220236-3a3c6141e376/go.mod h1:an3vInlBmSxCcxctByoQdvwPiA7DTK7jaaFDBTtu0ic=

181−github.com/go-git/go-billy/v5 v5.9.0 h1:jItGXszUDRtR/AlferWPTMN4j38BQ88XnXKbilmmBPA=

182−github.com/go-git/go-billy/v5 v5.9.0/go.mod h1:jCnQMLj9eUgGU7+ludSTYoZL/GGmii14RxKFj7ROgHw=

183−github.com/go-git/go-git/v5 v5.19.1 h1:nX27AnaU43/K5bKktKwgBmR9lawoYVe1Ckg0rgzzN00=

184−github.com/go-git/go-git/v5 v5.19.1/go.mod h1:Pb1v0c7/g8aGQJwx9Us09W85yGoyvSwuhEGMH7zjDKQ=

185−github.com/go-json-experiment/json v0.0.0-20260623181947-01eb4420fa68 h1:KZaTBSyshWX3MP5jukJcNSuXDQTO+rNpt0J564dX/eg=

186−github.com/go-json-experiment/json v0.0.0-20260623181947-01eb4420fa68/go.mod h1:tphK2c80bpPhMOI4v6bIc2xWywPfbqi1Z06+RcrMkDg=

187github.com/go-logfmt/logfmt v0.6.0 h1:wGYYu3uicYdqXVgoYbvnkrPVXkuLM1p1ifugDMEdRi4=

173github.com/go-logfmt/logfmt v0.6.0 h1:wGYYu3uicYdqXVgoYbvnkrPVXkuLM1p1ifugDMEdRi4=

188github.com/go-logfmt/logfmt v0.6.0/go.mod h1:WYhtIu8zTZfxdn5+rREduYbwxfcBr/Vr6KEVveWlfTs=

174github.com/go-logfmt/logfmt v0.6.0/go.mod h1:WYhtIu8zTZfxdn5+rREduYbwxfcBr/Vr6KEVveWlfTs=

189github.com/go-logr/logr v1.2.2/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=

175github.com/go-logr/logr v1.2.2/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=

@@ -191,31 +177,16 @@ github.com/go-logr/logr v1.4.3 h1:CjnDlHq8ikf6E492q6eKboGOC0T8CDaOvkHCIg8idEI=

191github.com/go-logr/logr v1.4.3/go.mod h1:9T104GzyrTigFIr8wt5mBrctHMim0Nb2HLGrmQ40KvY=

177github.com/go-logr/logr v1.4.3/go.mod h1:9T104GzyrTigFIr8wt5mBrctHMim0Nb2HLGrmQ40KvY=

192github.com/go-logr/stdr v1.2.2 h1:hSWxHoqTgW2S2qGc0LTAI563KZ5YKYRhT3MFKZMbjag=

178github.com/go-logr/stdr v1.2.2 h1:hSWxHoqTgW2S2qGc0LTAI563KZ5YKYRhT3MFKZMbjag=

193github.com/go-logr/stdr v1.2.2/go.mod h1:mMo/vtBO5dYbehREoey6XUKy/eSumjCCveDpRre4VKE=

179github.com/go-logr/stdr v1.2.2/go.mod h1:mMo/vtBO5dYbehREoey6XUKy/eSumjCCveDpRre4VKE=

194−github.com/go-ole/go-ole v1.3.0 h1:Dt6ye7+vXGIKZ7Xtk4s6/xVdGDQynvom7xCFEdWr6uE=

195−github.com/go-ole/go-ole v1.3.0/go.mod h1:5LS6F96DhAwUc7C+1HLexzMXY1xGRSryjyPPKW6zv78=

196−github.com/go-openapi/jsonpointer v0.19.3/go.mod h1:Pl9vOtqEWErmShwVjC8pYs9cog34VGT37dQOVbmoatg=

197−github.com/go-openapi/jsonpointer v0.19.5 h1:gZr+CIYByUqjcgeLXnQu2gHYQC9o73G2XUeOFYEICuY=

198−github.com/go-openapi/jsonpointer v0.19.5/go.mod h1:Pl9vOtqEWErmShwVjC8pYs9cog34VGT37dQOVbmoatg=

199−github.com/go-openapi/jsonreference v0.20.0 h1:MYlu0sBgChmCfJxxUKZ8g1cPWFOB37YSZqewK7OKeyA=

200−github.com/go-openapi/jsonreference v0.20.0/go.mod h1:Ag74Ico3lPc+zR+qjn4XBUmXymS4zJbYVCZmcgkasdo=

201−github.com/go-openapi/spec v0.20.6 h1:ich1RQ3WDbfoeTqTAb+5EIxNmpKVJZWBNah9RAT0jIQ=

202−github.com/go-openapi/spec v0.20.6/go.mod h1:2OpW+JddWPrpXSCIX8eOx7lZ5iyuWj3RYR6VaaBKcWA=

203−github.com/go-openapi/swag v0.19.5/go.mod h1:POnQmlKehdgb5mhVOsnJFsivZCEZ/vjK9gh66Z9tfKk=

204−github.com/go-openapi/swag v0.19.15 h1:D2NRCBzS9/pEY3gP9Nl8aDqGUcPFrwG2p+CNFrLyrCM=

205−github.com/go-openapi/swag v0.19.15/go.mod h1:QYRuS/SOXUCsnplDa677K7+DxSOj6IPNl/eQntq43wQ=

206github.com/go-quicktest/qt v1.101.0 h1:O1K29Txy5P2OK0dGo59b7b0LR6wKfIhttaAhHUyn7eI=

180github.com/go-quicktest/qt v1.101.0 h1:O1K29Txy5P2OK0dGo59b7b0LR6wKfIhttaAhHUyn7eI=

207github.com/go-quicktest/qt v1.101.0/go.mod h1:14Bz/f7NwaXPtdYEgzsx46kqSxVwTbzVZsDC26tQJow=

181github.com/go-quicktest/qt v1.101.0/go.mod h1:14Bz/f7NwaXPtdYEgzsx46kqSxVwTbzVZsDC26tQJow=

208−github.com/go-viper/mapstructure/v2 v2.5.0 h1:vM5IJoUAy3d7zRSVtIwQgBj7BiWtMPfmPEgAXnvj1Ro=

182+github.com/go-viper/mapstructure/v2 v2.4.0 h1:EBsztssimR/CONLSZZ04E8qAkxNYq4Qp9LvH92wZUgs=

209−github.com/go-viper/mapstructure/v2 v2.5.0/go.mod h1:oJDH3BJKyqBA2TXFhDsKDGDTlndYOZ6rGS0BRZIxGhM=

183+github.com/go-viper/mapstructure/v2 v2.4.0/go.mod h1:oJDH3BJKyqBA2TXFhDsKDGDTlndYOZ6rGS0BRZIxGhM=

210github.com/goccy/go-json v0.10.5 h1:Fq85nIqj+gXn/S5ahsiTlK3TmC85qgirsdTP/+DeaC4=

184github.com/goccy/go-json v0.10.5 h1:Fq85nIqj+gXn/S5ahsiTlK3TmC85qgirsdTP/+DeaC4=

211github.com/goccy/go-json v0.10.5/go.mod h1:oq7eo15ShAhp70Anwd5lgX2pLfOS3QCiwU/PULtXL6M=

185github.com/goccy/go-json v0.10.5/go.mod h1:oq7eo15ShAhp70Anwd5lgX2pLfOS3QCiwU/PULtXL6M=

212−github.com/goccy/go-yaml v1.19.2 h1:PmFC1S6h8ljIz6gMRBopkjP1TVT7xuwrButHID66PoM=

186+github.com/goccy/go-yaml v1.19.0 h1:EmkZ9RIsX+Uq4DYFowegAuJo8+xdX3T/2dwNPXbxEYE=

213−github.com/goccy/go-yaml v1.19.2/go.mod h1:XBurs7gK8ATbW4ZPGKgcbrY1Br56PdM69F7LkFRi1kA=

187+github.com/goccy/go-yaml v1.19.0/go.mod h1:XBurs7gK8ATbW4ZPGKgcbrY1Br56PdM69F7LkFRi1kA=

214−github.com/godbus/dbus/v5 v5.1.0/go.mod h1:xhWf0FNVPg57R7Z0UbKHbJfkEywrmjJnf7w5xrFpKfA=

188+github.com/golang-jwt/jwt/v5 v5.2.2 h1:Rl4B7itRWVtYIHFrSNd7vhTiz9UpLdi6gZhZ3wEeDy8=

215−github.com/godbus/dbus/v5 v5.2.2 h1:TUR3TgtSVDmjiXOgAAyaZbYmIeP3DPkld3jgKGV8mXQ=

216−github.com/godbus/dbus/v5 v5.2.2/go.mod h1:3AAv2+hPq5rdnr5txxxRwiGjPXamgoIHgz9FPBfOp3c=

217−github.com/golang-jwt/jwt/v5 v5.3.1 h1:kYf81DTWFe7t+1VvL7eS+jKFVWaUnK9cB1qbwn63YCY=

218−github.com/golang-jwt/jwt/v5 v5.3.1/go.mod h1:fxCRLWMO43lRc8nhHWY6LGqRcf+1gQWArsqaEUEa5bE=

189+github.com/golang-jwt/jwt/v5 v5.2.2/go.mod h1:pqrtFR0X4osieyHYxtmOUWsAWrfe1Q5UVIyoH402zdk=

219github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0 h1:DACJavvAHhabrF08vX0COfcOBJRhZ8lUbR+ZWIs0Y5g=

190github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0 h1:DACJavvAHhabrF08vX0COfcOBJRhZ8lUbR+ZWIs0Y5g=

220github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0/go.mod h1:E/TSTwGwJL78qG/PmXZO1EjYhfJinVAhrmmHX6Z8B9k=

191github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0/go.mod h1:E/TSTwGwJL78qG/PmXZO1EjYhfJinVAhrmmHX6Z8B9k=

221github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=

192github.com/golang/protobuf v1.5.4 h1:i7eJL8qZTpSEXOPTxNKhASYpMn+8e5Q6AdndVa1dWek=

@@ -223,18 +194,18 @@ github.com/golang/protobuf v1.5.4/go.mod h1:lnTiLA8Wa4RWRcIUkrtSVa5nRhsEGBg48fD6

223github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=

194github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=

224github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=

195github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=

225github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=

196github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=

226−github.com/google/jsonschema-go v0.4.3 h1:/DBOLZTfDow7pe2GmaJNhltueGTtDKICi8V8p+DQPd0=

197+github.com/google/jsonschema-go v0.3.0 h1:6AH2TxVNtk3IlvkkhjrtbUc4S8AvO0Xii0DxIygDg+Q=

227−github.com/google/jsonschema-go v0.4.3/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=

198+github.com/google/jsonschema-go v0.3.0/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=

228github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17kjQEVQ1XRhq2/JR1M3sGqeJoxs=

199github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e h1:ijClszYn+mADRFY17kjQEVQ1XRhq2/JR1M3sGqeJoxs=

229github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=

200github.com/google/pprof v0.0.0-20250317173921-a4b03ec1a45e/go.mod h1:boTsfXsheKC2y+lKOCMpSfarhxDeIzfZG1jqGcPl3cA=

230github.com/google/s2a-go v0.1.9 h1:LGD7gtMgezd8a/Xak7mEWL0PjoTQFvpRudN895yqKW0=

201github.com/google/s2a-go v0.1.9 h1:LGD7gtMgezd8a/Xak7mEWL0PjoTQFvpRudN895yqKW0=

231github.com/google/s2a-go v0.1.9/go.mod h1:YA0Ei2ZQL3acow2O62kdp9UlnvMmU7kA6Eutn0dXayM=

202github.com/google/s2a-go v0.1.9/go.mod h1:YA0Ei2ZQL3acow2O62kdp9UlnvMmU7kA6Eutn0dXayM=

232github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=

203github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=

233github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=

204github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=

234−github.com/googleapis/enterprise-certificate-proxy v0.3.17 h1:73NfMHdiqo9JFU9+7a5ExpVa10/R29pXfZIaW559nrg=

205+github.com/googleapis/enterprise-certificate-proxy v0.3.7 h1:zrn2Ee/nWmHulBx5sAVrGgAa0f2/R35S4DJwfFaUPFQ=

235−github.com/googleapis/enterprise-certificate-proxy v0.3.17/go.mod h1:rSEsBUemEBZEexP2y6jPp16LUmUbjmSbcPMQizR0o4k=

206+github.com/googleapis/enterprise-certificate-proxy v0.3.7/go.mod h1:MkHOF77EYAE7qfSuSS9PU6g4Nt4e11cnsDUowfwewLA=

236−github.com/googleapis/gax-go/v2 v2.22.0 h1:PjIWBpgGIVKGoCXuiCoP64altEJCj3/Ei+kSU5vlZD4=

207+github.com/googleapis/gax-go/v2 v2.15.0 h1:SyjDc1mGgZU5LncH8gimWo9lW1DtIfPibOG81vgd/bo=

237−github.com/googleapis/gax-go/v2 v2.22.0/go.mod h1:irWBbALSr0Sk3qlqb9SyJ1h68WjgeFuiOzI4Rqw5+aY=

208+github.com/googleapis/gax-go/v2 v2.15.0/go.mod h1:zVVkkxAQHa1RQpg9z2AUCMnKhi0Qld9rcmyfL1OZhoc=

238github.com/gorilla/css v1.0.1 h1:ntNaBIghp6JmvWnxbZKANoLyuXTPZ4cAMlo6RyhlbO8=

209github.com/gorilla/css v1.0.1 h1:ntNaBIghp6JmvWnxbZKANoLyuXTPZ4cAMlo6RyhlbO8=

239github.com/gorilla/css v1.0.1/go.mod h1:BvnYkspnSzMmwRK+b8/xgNPLiIuNZr6vbZBTPQ2A3b0=

210github.com/gorilla/css v1.0.1/go.mod h1:BvnYkspnSzMmwRK+b8/xgNPLiIuNZr6vbZBTPQ2A3b0=

240github.com/gorilla/websocket v1.4.1/go.mod h1:YR8l580nyteQvAITg2hZ9XVh4b55+EU/adAjf1fMHhE=

211github.com/gorilla/websocket v1.4.1/go.mod h1:YR8l580nyteQvAITg2hZ9XVh4b55+EU/adAjf1fMHhE=

@@ -246,30 +217,25 @@ github.com/hexops/gotextdiff v1.0.3 h1:gitA9+qJrrTCsiCl7+kh75nPqQt1cx4ZkudSTLoUq

246github.com/hexops/gotextdiff v1.0.3/go.mod h1:pSWU5MAI3yDq+fZBTazCSJysOMbxWL1BSow5/V2vxeg=

217github.com/hexops/gotextdiff v1.0.3/go.mod h1:pSWU5MAI3yDq+fZBTazCSJysOMbxWL1BSow5/V2vxeg=

247github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=

218github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=

248github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=

219github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=

249−github.com/invopop/jsonschema v0.14.0 h1:MHQqLhvpNUZfw+hM3AZDYK7jxO8FZoQeQM77g8iyZjg=

220+github.com/invopop/jsonschema v0.13.0 h1:KvpoAJWEjR3uD9Kbm2HWJmqsEaHt8lBUpd0qHcIi21E=

250−github.com/invopop/jsonschema v0.14.0/go.mod h1:ygm6C2EaVNMBDPpaPlnOA2pFAxBnxGjFlMZABxm9n2I=

221+github.com/invopop/jsonschema v0.13.0/go.mod h1:ffZ5Km5SWWRAIN6wbDXItl95euhFz2uON45H2qjYt+0=

251−github.com/itchyny/gojq v0.12.19 h1:ttXA0XCLEMoaLOz5lSeFOZ6u6Q3QxmG46vfgI4O0DEs=

252−github.com/itchyny/gojq v0.12.19/go.mod h1:5galtVPDywX8SPSOrqjGxkBeDhSxEW1gSxoy7tn1iZY=

253−github.com/itchyny/timefmt-go v0.1.8 h1:1YEo1JvfXeAHKdjelbYr/uCuhkybaHCeTkH8Bo791OI=

254−github.com/itchyny/timefmt-go v0.1.8/go.mod h1:5E46Q+zj7vbTgWY8o5YkMeYb4I6GeWLFnetPy5oBrAI=

255−github.com/jackmordaunt/icns/v3 v3.0.1 h1:xxot6aNuGrU+lNgxz5I5H0qSeCjNKp8uTXB1j8D4S3o=

256−github.com/jackmordaunt/icns/v3 v3.0.1/go.mod h1:5sHL59nqTd2ynTnowxB/MDQFhKNqkK8X687uKNygaSQ=

257−github.com/jbenet/go-context v0.0.0-20150711004518-d14ea06fba99 h1:BQSFePA1RWJOlocH6Fxy8MmwDt+yVQYULKfN0RoTN8A=

258−github.com/jbenet/go-context v0.0.0-20150711004518-d14ea06fba99/go.mod h1:1lJo3i6rXxKeerYnT8Nvf0QmHCRC1n8sfWVwXF2Frvo=

259github.com/joho/godotenv v1.5.1 h1:7eLL/+HRGLY0ldzfGMeQkb7vMd0as4CfYvUVzLqw0N0=

222github.com/joho/godotenv v1.5.1 h1:7eLL/+HRGLY0ldzfGMeQkb7vMd0as4CfYvUVzLqw0N0=

260github.com/joho/godotenv v1.5.1/go.mod h1:f4LDr5Voq0i2e/R5DDNOoa2zzDfwtkZa6DnEwAbqwq4=

223github.com/joho/godotenv v1.5.1/go.mod h1:f4LDr5Voq0i2e/R5DDNOoa2zzDfwtkZa6DnEwAbqwq4=

261github.com/jordanella/go-ansi-paintbrush v0.0.0-20240728195301-b7ad996ecf3d h1:on25kP+Sx7sxUMRQiA8gdcToAGet4DK/EIA30mXre+4=

224github.com/jordanella/go-ansi-paintbrush v0.0.0-20240728195301-b7ad996ecf3d h1:on25kP+Sx7sxUMRQiA8gdcToAGet4DK/EIA30mXre+4=

262github.com/jordanella/go-ansi-paintbrush v0.0.0-20240728195301-b7ad996ecf3d/go.mod h1:SV0W0APWP9MZ1/gfDQ/NzzTlWdIgYZ/ZbpN4d/UXRYw=

225github.com/jordanella/go-ansi-paintbrush v0.0.0-20240728195301-b7ad996ecf3d/go.mod h1:SV0W0APWP9MZ1/gfDQ/NzzTlWdIgYZ/ZbpN4d/UXRYw=

263−github.com/josharian/intern v1.0.0 h1:vlS4z54oSdjm0bgjRigI+G1HpF+tI+9rE5LLzOg8HmY=

264github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=

226github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=

265−github.com/kaptinlin/jsonpointer v0.4.26 h1:tw616yszHek+B3/GtDSia+uzBa3sLXGpmo4tYeMhBZw=

227+github.com/kaptinlin/go-i18n v0.2.2 h1:kebVCZme/BrCTqonh/J+VYCl1+Of5C18bvyn3DRPl5M=

228+github.com/kaptinlin/go-i18n v0.2.2/go.mod h1:MiwkeHryBopAhC/M3zEwIM/2IN8TvTqJQswPw6kceqM=

229+github.com/kaptinlin/jsonpointer v0.4.8 h1:HocHcXrOBfP/nUJw0YYjed/TlQvuCAY6uRs3Qok7F6g=

266−github.com/kaptinlin/jsonpointer v0.4.26/go.mod h1:wVOBaXGGnP42YsMb6zev/3W5POTvspdNfh8DXzf8XS8=

230+github.com/kaptinlin/jsonpointer v0.4.8/go.mod h1:9y0LgXavlmVE5FSHShY5LRlURJJVhbyVJSRWkilrTqA=

231+github.com/kaptinlin/jsonschema v0.6.6 h1:UmIF1amA5ijCGSk4tl4ViNlgYL4jzHHvY+Nd5cnkfDI=

267−github.com/kaptinlin/jsonschema v0.9.2 h1:zqowDnBkpg80OeIt4Bl1bOyfKY1/KVVzst1m3MCsxd4=

232+github.com/kaptinlin/jsonschema v0.6.6/go.mod h1:EbhSbdxZ4QjzIORdMWOrRXJeCHrLTJqXDA8JzNaeFc8=

268−github.com/kaptinlin/jsonschema v0.9.2/go.mod h1:mCH2W5lXd29tdDjvoFfY32nedPORnlk7pCVrrcs/NkQ=

233+github.com/kaptinlin/messageformat-go v0.4.7 h1:HQ/OvFUSU7+fAHWkZnP2ug9y+A/ZyTE8j33jfWr8O3Q=

234+github.com/kaptinlin/messageformat-go v0.4.7/go.mod h1:DusKpv8CIybczGvwIVn3j13hbR3psr5mOwhFudkiq1c=

269−github.com/klauspost/compress v1.18.6 h1:2jupLlAwFm95+YDR+NwD2MEfFO9d4z4Prjl1XXDjuao=

235+github.com/klauspost/compress v1.18.0 h1:c/Cqfb0r+Yi+JtIEq73FWXVkRonBlf0CRNYc8Zttxdo=

270−github.com/klauspost/compress v1.18.6/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=

236+github.com/klauspost/compress v1.18.0/go.mod h1:2Pp+KzxcywXVXMr50+X0Q/Lsb43OQHYWRCY2AiWywWQ=

271−github.com/klauspost/cpuid/v2 v2.3.0 h1:S4CRMLnYUhGeDFDqkGriYKdfoFlDnMtqTiI/sFzhA9Y=

237+github.com/klauspost/cpuid/v2 v2.2.10 h1:tBs3QSyvjDyFTq3uoc/9xFpCuOsJQFNPiAhYdw2skhE=

272−github.com/klauspost/cpuid/v2 v2.3.0/go.mod h1:hqwkgyIinND0mEev00jJYCxPNVRVXFQeu1XKlok6oO0=

238+github.com/klauspost/cpuid/v2 v2.2.10/go.mod h1:hqwkgyIinND0mEev00jJYCxPNVRVXFQeu1XKlok6oO0=

273github.com/klauspost/pgzip v1.2.6 h1:8RXeL5crjEUFnR2/Sn6GJNWtSQ3Dk8pq4CL3jvdDyjU=

239github.com/klauspost/pgzip v1.2.6 h1:8RXeL5crjEUFnR2/Sn6GJNWtSQ3Dk8pq4CL3jvdDyjU=

274github.com/klauspost/pgzip v1.2.6/go.mod h1:Ch1tH69qFZu15pkjo5kYi6mth2Zzwzt50oCQKQE9RUs=

240github.com/klauspost/pgzip v1.2.6/go.mod h1:Ch1tH69qFZu15pkjo5kYi6mth2Zzwzt50oCQKQE9RUs=

275github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=

241github.com/kr/pretty v0.1.0/go.mod h1:dAy3ld7l9f0ibDNOQOHHMYYIIbhfbHSm3C4ZsoJORNo=

@@ -281,25 +247,22 @@ github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=

281github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=

247github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=

282github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=

248github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc=

283github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=

249github.com/kylelemons/godebug v1.1.0/go.mod h1:9/0rRGxNHcop5bhtWyNeEfOS8JIWk580+fNqagV/RAw=

284−github.com/lucasb-eyer/go-colorful v1.4.0 h1:UtrWVfLdarDgc44HcS7pYloGHJUjHV/4FwW4TvVgFr4=

250+github.com/lucasb-eyer/go-colorful v1.3.0 h1:2/yBRLdWBZKrf7gB40FoiKfAWYQ0lqNcbuQwVHXptag=

285−github.com/lucasb-eyer/go-colorful v1.4.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=

251+github.com/lucasb-eyer/go-colorful v1.3.0/go.mod h1:R4dSotOR9KMtayYi1e77YzuveK+i7ruzyGqttikkLy0=

286−github.com/mailru/easyjson v0.0.0-20190614124828-94de47d64c63/go.mod h1:C1wdFJiN94OJF2b5HbByQZoLdCWB1Yqtg26g4irojpc=

287−github.com/mailru/easyjson v0.0.0-20190626092158-b2ccc519800e/go.mod h1:C1wdFJiN94OJF2b5HbByQZoLdCWB1Yqtg26g4irojpc=

288−github.com/mailru/easyjson v0.7.6/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=

289github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=

252github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=

290github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=

253github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=

291−github.com/mattn/go-isatty v0.0.22 h1:j8l17JJ9i6VGPUFUYoTUKPSgKe/83EYU2zBC7YNKMw4=

254+github.com/mattn/go-isatty v0.0.20 h1:xfD0iDuEKnDkl03q4limB+vH+GxLEtL/jb4xVJSWWEY=

292−github.com/mattn/go-isatty v0.0.22/go.mod h1:ZXfXG4SQHsB/w3ZeOYbR0PrPwLy+n6xiMrJlRFqopa4=

255+github.com/mattn/go-isatty v0.0.20/go.mod h1:W+V8PltTTMOvKvAeJH7IuucS94S2C6jfK/D7dTCTo3Y=

293−github.com/mattn/go-runewidth v0.0.24 h1:cpokDiIn0MGnhdHwuWnJBITySJ20QyNGnY2kR/ay2DU=

256+github.com/mattn/go-runewidth v0.0.19 h1:v++JhqYnZuu5jSKrk9RbgF5v4CGUjqRfBm05byFGLdw=

294−github.com/mattn/go-runewidth v0.0.24/go.mod h1:XBkDxAl56ILZc9knddidhrOlY5R/pDhgLpndooCuJAs=

257+github.com/mattn/go-runewidth v0.0.19/go.mod h1:XBkDxAl56ILZc9knddidhrOlY5R/pDhgLpndooCuJAs=

295github.com/mfridman/interpolate v0.0.2 h1:pnuTK7MQIxxFz1Gr+rjSIx9u7qVjf5VOoM/u6BbAxPY=

258github.com/mfridman/interpolate v0.0.2 h1:pnuTK7MQIxxFz1Gr+rjSIx9u7qVjf5VOoM/u6BbAxPY=

296github.com/mfridman/interpolate v0.0.2/go.mod h1:p+7uk6oE07mpE/Ik1b8EckO0O4ZXiGAfshKBWLUM9Xg=

259github.com/mfridman/interpolate v0.0.2/go.mod h1:p+7uk6oE07mpE/Ik1b8EckO0O4ZXiGAfshKBWLUM9Xg=

297github.com/microcosm-cc/bluemonday v1.0.27 h1:MpEUotklkwCSLeH+Qdx1VJgNqLlpY2KXwXFM08ygZfk=

260github.com/microcosm-cc/bluemonday v1.0.27 h1:MpEUotklkwCSLeH+Qdx1VJgNqLlpY2KXwXFM08ygZfk=

298github.com/microcosm-cc/bluemonday v1.0.27/go.mod h1:jFi9vgW+H7c3V0lb6nR74Ib/DIB5OBs92Dimizgw2cA=

261github.com/microcosm-cc/bluemonday v1.0.27/go.mod h1:jFi9vgW+H7c3V0lb6nR74Ib/DIB5OBs92Dimizgw2cA=

299github.com/mitchellh/mapstructure v1.5.0 h1:jeMsZIYE/09sWLaz43PL7Gy6RuMjD2eJVyuac5Z2hdY=

262github.com/mitchellh/mapstructure v1.5.0 h1:jeMsZIYE/09sWLaz43PL7Gy6RuMjD2eJVyuac5Z2hdY=

300github.com/mitchellh/mapstructure v1.5.0/go.mod h1:bFUtVrKA4DC2yAKiSyO/QUcy7e+RRV2QTWOzhPopBRo=

263github.com/mitchellh/mapstructure v1.5.0/go.mod h1:bFUtVrKA4DC2yAKiSyO/QUcy7e+RRV2QTWOzhPopBRo=

301−github.com/modelcontextprotocol/go-sdk v1.6.1 h1:0zOSupjKUxPKSocPT1Wtago+mUHU2/uZ4xSOY0FGReU=

264+github.com/modelcontextprotocol/go-sdk v1.2.0 h1:Y23co09300CEk8iZ/tMxIX1dVmKZkzoSBZOpJwUnc/s=

302−github.com/modelcontextprotocol/go-sdk v1.6.1/go.mod h1:kzm3kzFL1/+AziGOE0nUs3gvPoNxMCvkxokMkuFapXQ=

265+github.com/modelcontextprotocol/go-sdk v1.2.0/go.mod h1:6fM3LCm3yV7pAs8isnKLn07oKtB0MP9LHd3DfAcKw10=

303github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELUXHmA=

266github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELUXHmA=

304github.com/muesli/cancelreader v0.2.2/go.mod h1:3XuTXfFS2VjM+HTLZY9Ak0l6eUKfijIfMUZ4EgX0QYo=

267github.com/muesli/cancelreader v0.2.2/go.mod h1:3XuTXfFS2VjM+HTLZY9Ak0l6eUKfijIfMUZ4EgX0QYo=

305github.com/muesli/mango v0.1.0 h1:DZQK45d2gGbql1arsYA4vfg4d7I9Hfx5rX/GCmzsAvI=

268github.com/muesli/mango v0.1.0 h1:DZQK45d2gGbql1arsYA4vfg4d7I9Hfx5rX/GCmzsAvI=

@@ -310,43 +273,37 @@ github.com/muesli/mango-pflag v0.1.0 h1:UADqbYgpUyRoBja3g6LUL+3LErjpsOwaC9ywvBWe

310github.com/muesli/mango-pflag v0.1.0/go.mod h1:YEQomTxaCUp8PrbhFh10UfbhbQrM/xJ4i2PB8VTLLW0=

273github.com/muesli/mango-pflag v0.1.0/go.mod h1:YEQomTxaCUp8PrbhFh10UfbhbQrM/xJ4i2PB8VTLLW0=

311github.com/muesli/roff v0.1.0 h1:YD0lalCotmYuF5HhZliKWlIx7IEhiXeSfq7hNjFqGF8=

274github.com/muesli/roff v0.1.0 h1:YD0lalCotmYuF5HhZliKWlIx7IEhiXeSfq7hNjFqGF8=

312github.com/muesli/roff v0.1.0/go.mod h1:pjAHQM9hdUUwm/krAfrLGgJkXJ+YuhtsfZ42kieB2Ig=

275github.com/muesli/roff v0.1.0/go.mod h1:pjAHQM9hdUUwm/krAfrLGgJkXJ+YuhtsfZ42kieB2Ig=

313−github.com/ncruces/go-sqlite3 v0.35.1 h1:h/LaVyQwIvBBT0+2JmVe2tbYyWjUQ093/pYhpBqdxJo=

276+github.com/muesli/termenv v0.16.0 h1:S5AlUN9dENB57rsbnkPyfdGuWIlkmzJjbFf0Tf5FWUc=

314−github.com/ncruces/go-sqlite3 v0.35.1/go.mod h1:fXOSIkWwN5NXgbJk+7Zls8QIW4xOflmgh11OFvcY+J0=

277+github.com/muesli/termenv v0.16.0/go.mod h1:ZRfOIKPFDYQoDFF4Olj7/QJbW60Ol/kL1pU3VfY/Cnk=

315−github.com/ncruces/go-sqlite3-wasm/v3 v3.1.35302 h1:Cew7/eNAMd1zhpXYBjofBua/63pFvbvB2h4PM/p6gKU=

278+github.com/ncruces/go-sqlite3 v0.30.5 h1:6usmTQ6khriL8oWilkAZSJM/AIpAlVL2zFrlcpDldCE=

316−github.com/ncruces/go-sqlite3-wasm/v3 v3.1.35302/go.mod h1:xe0CfafDUxfh+fSVKjHHMiAxoG9KALt5nFtbGNb/jRs=

279+github.com/ncruces/go-sqlite3 v0.30.5/go.mod h1:0I0JFflTKzfs3Ogfv8erP7CCoV/Z8uxigVDNOR0AQ5E=

317github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=

280github.com/ncruces/go-strftime v1.0.0 h1:HMFp8mLCTPp341M/ZnA4qaf7ZlsbTc+miZjCLOFAw7w=

318github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=

281github.com/ncruces/go-strftime v1.0.0/go.mod h1:Fwc5htZGVVkseilnfgOVb9mKy6w1naJmn9CehxcKcls=

319github.com/ncruces/julianday v1.0.0 h1:fH0OKwa7NWvniGQtxdJRxAgkBMolni2BjDHaWTxqt7M=

282github.com/ncruces/julianday v1.0.0 h1:fH0OKwa7NWvniGQtxdJRxAgkBMolni2BjDHaWTxqt7M=

320github.com/ncruces/julianday v1.0.0/go.mod h1:Dusn2KvZrrovOMJuOt0TNXL6tB7U2E8kvza5fFc9G7g=

283github.com/ncruces/julianday v1.0.0/go.mod h1:Dusn2KvZrrovOMJuOt0TNXL6tB7U2E8kvza5fFc9G7g=

321github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646 h1:zYyBkD/k9seD2A7fsi6Oo2LfFZAehjjQMERAvZLEDnQ=

284github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646 h1:zYyBkD/k9seD2A7fsi6Oo2LfFZAehjjQMERAvZLEDnQ=

322github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646/go.mod h1:jpp1/29i3P1S/RLdc7JQKbRpFeM1dOBd8T9ki5s+AY8=

285github.com/nfnt/resize v0.0.0-20180221191011-83c6a9932646/go.mod h1:jpp1/29i3P1S/RLdc7JQKbRpFeM1dOBd8T9ki5s+AY8=

323−github.com/niemeyer/pretty v0.0.0-20200227124842-a10e7caefd8e/go.mod h1:zD1mROLANZcx1PVRCS0qkT7pwLkGfwJo4zjcN/Tysno=

324github.com/nxadm/tail v1.4.11 h1:8feyoE3OzPrcshW5/MJ4sGESc5cqmGkGCWlco4l0bqY=

286github.com/nxadm/tail v1.4.11 h1:8feyoE3OzPrcshW5/MJ4sGESc5cqmGkGCWlco4l0bqY=

325github.com/nxadm/tail v1.4.11/go.mod h1:OTaG3NK980DZzxbRq6lEuzgU+mug70nY11sMd4JXXHc=

287github.com/nxadm/tail v1.4.11/go.mod h1:OTaG3NK980DZzxbRq6lEuzgU+mug70nY11sMd4JXXHc=

326−github.com/onsi/gomega v1.34.1 h1:EUMJIKUjM8sKjYbtxQI9A4z2o+rruxnzNvpknOXie6k=

288+github.com/openai/openai-go/v2 v2.7.1 h1:/tfvTJhfv7hTSL8mWwc5VL4WLLSDL5yn9VqVykdu9r8=

327−github.com/onsi/gomega v1.34.1/go.mod h1:kU1QgUvBDLXBJq618Xvm2LUX6rSAfRaFRTcdOeDLwwY=

289+github.com/openai/openai-go/v2 v2.7.1/go.mod h1:jrJs23apqJKKbT+pqtFgNKpRju/KP9zpUTZhz3GElQE=

328−github.com/pb33f/ordered-map/v2 v2.3.1 h1:5319HDO0aw4DA4gzi+zv4FXU9UlSs3xGZ40wcP1nBjY=

290+github.com/pelletier/go-toml/v2 v2.2.4 h1:mye9XuhQ6gvn5h28+VilKrrPoQVanw5PMw/TB0t5Ec4=

329−github.com/pb33f/ordered-map/v2 v2.3.1/go.mod h1:qxFQgd0PkVUtOMCkTapqotNgzRhMPL7VvaHKbd1HnmQ=

291+github.com/pelletier/go-toml/v2 v2.2.4/go.mod h1:2gIqNv+qfxSVS7cM2xJQKtLSTLUE9V8t9Stt+h56mCY=

330−github.com/pierrec/lz4/v4 v4.1.26 h1:GrpZw1gZttORinvzBdXPUXATeqlJjqUG/D87TKMnhjY=

292+github.com/pierrec/lz4/v4 v4.1.22 h1:cKFw6uJDK+/gfw5BcDL0JL5aBsAFdsIT18eRtLj7VIU=

331−github.com/pierrec/lz4/v4 v4.1.26/go.mod h1:EoQMVJgeeEOMsCqCzqFm2O0cJvljX2nGZjcRIPL34O4=

293+github.com/pierrec/lz4/v4 v4.1.22/go.mod h1:gZWDp/Ze/IJXGXf23ltt2EXimqmTUXEy0GFuRQyBid4=

332−github.com/pjbgf/sha1cd v0.6.0 h1:3WJ8Wz8gvDz29quX1OcEmkAlUg9diU4GxJHqs0/XiwU=

333−github.com/pjbgf/sha1cd v0.6.0/go.mod h1:lhpGlyHLpQZoxMv8HcgXvZEhcGs0PG/vsZnEJ7H0iCM=

334github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c h1:+mdjkGKdHQG3305AYmdv1U2eRNDiU2ErMBj1gwrq8eQ=

294github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c h1:+mdjkGKdHQG3305AYmdv1U2eRNDiU2ErMBj1gwrq8eQ=

335github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c/go.mod h1:7rwL4CYBLnjLxUqIJNnCWiEdr3bn6IUYi15bNlnbCCU=

295github.com/pkg/browser v0.0.0-20240102092130-5ac0b6a4141c/go.mod h1:7rwL4CYBLnjLxUqIJNnCWiEdr3bn6IUYi15bNlnbCCU=

336github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=

296github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=

337−github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=

338−github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=

339github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10 h1:GFCKgmp0tecUJ0sJuv4pzYCqS9+RGSn52M3FUwPs+uo=

297github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10 h1:GFCKgmp0tecUJ0sJuv4pzYCqS9+RGSn52M3FUwPs+uo=

340github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10/go.mod h1:t/avpk3KcrXxUnYOhZhMXJlSEyie6gQbtLq5NM3loB8=

298github.com/planetscale/vtprotobuf v0.6.1-0.20240319094008-0393e58bdf10/go.mod h1:t/avpk3KcrXxUnYOhZhMXJlSEyie6gQbtLq5NM3loB8=

299+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=

341github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=

300github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=

342−github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 h1:Jamvg5psRIccs7FGNTlIRMkT8wgtp5eCXdBlqhYGL6U=

301+github.com/posthog/posthog-go v1.9.1 h1:9bkcRnYSvcgMxL2s9QlCnd1DVnm2qWXxWu5o0HSF0xM=

343−github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=

302+github.com/posthog/posthog-go v1.9.1/go.mod h1:wB3/9Q7d9gGb1P/yf/Wri9VBlbP8oA8z++prRzL5OcY=

344−github.com/posthog/posthog-go v1.17.5 h1:mrLiAdyiQpl8Yeyg23iShyAztJMaUrYzsBjKeH52Aak=

303+github.com/pressly/goose/v3 v3.26.0 h1:KJakav68jdH0WDvoAcj8+n61WqOIaPGgH0bJWS6jpmM=

345−github.com/posthog/posthog-go v1.17.5/go.mod h1://M430hNH3e8CDv4i8SJesb26816Mpa6GIZaiP4pNQU=

346−github.com/pressly/goose/v3 v3.27.2 h1:FjKNzcmMdGrQlSIu5alMSmakQtJFBgtw+A0bb1p/LC8=

347−github.com/pressly/goose/v3 v3.27.2/go.mod h1:qWW+/8dkVtJYjJrbIpwD5xxnEJTUKvxkQ9JKQp9LaIM=

304+github.com/pressly/goose/v3 v3.26.0/go.mod h1:4hC1KrritdCxtuFsqgs1R4AU5bWtTAf+cnWvfhf2DNY=

348−github.com/qjebbs/go-jsons v1.0.0-alpha.6 h1:9pMjdrCT7rj9CDIM0FXf9Ilie8PEy9g1b4qMkCnedjQ=

305+github.com/qjebbs/go-jsons v1.0.0-alpha.4 h1:Qsb4ohRUHQODIUAsJKdKJ/SIDbsO7oGOzsfy+h1yQZs=

349−github.com/qjebbs/go-jsons v1.0.0-alpha.6/go.mod h1:wNJrtinHyC3YSf6giEh4FJN8+yZV7nXBjvmfjhBIcw4=

306+github.com/qjebbs/go-jsons v1.0.0-alpha.4/go.mod h1:wNJrtinHyC3YSf6giEh4FJN8+yZV7nXBjvmfjhBIcw4=

350github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec h1:W09IVJc94icq4NjY3clb7Lk8O1qJ8BdBEF8z0ibU0rE=

307github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec h1:W09IVJc94icq4NjY3clb7Lk8O1qJ8BdBEF8z0ibU0rE=

351github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo=

308github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo=

352github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=

309github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=

@@ -354,18 +311,12 @@ github.com/rivo/uniseg v0.4.7/go.mod h1:FN3SvrM+Zdj16jyLfmOkMNblXMcoc8DfTHruCPUc

354github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=

311github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=

355github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=

312github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=

356github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=

313github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=

357−github.com/sahilm/fuzzy v0.1.3 h1:juByESSS32nVD81vr6tHmKmA/8zde7gE+x5CLxrzXPU=

314+github.com/sabhiram/go-gitignore v0.0.0-20210923224102-525f6e181f06 h1:OkMGxebDjyw0ULyrTYWeN0UNCCkmCWfjPnIA2W6oviI=

315+github.com/sabhiram/go-gitignore v0.0.0-20210923224102-525f6e181f06/go.mod h1:+ePHsJ1keEjQtpvf9HHw0f4ZeJ0TLRsxhunSI2hYJSs=

316+github.com/sahilm/fuzzy v0.1.1 h1:ceu5RHF8DGgoi+/dR5PsECjCDH1BE3Fnmpo7aVXOdRA=

358−github.com/sahilm/fuzzy v0.1.3/go.mod h1:au6//VbVSqu6DFrkL2CfjlJ5iURpNCPeE+1GwY3XsT8=

317+github.com/sahilm/fuzzy v0.1.1/go.mod h1:VFvziUEIMCrT6A6tw2RFIXPXXmzXbOsSHF0DOI8ZK9Y=

359github.com/sebdah/goldie/v2 v2.5.3 h1:9ES/mNN+HNUbNWpVAlrzuZ7jE+Nrczbj8uFRjM7624Y=

318github.com/sebdah/goldie/v2 v2.5.3 h1:9ES/mNN+HNUbNWpVAlrzuZ7jE+Nrczbj8uFRjM7624Y=

360github.com/sebdah/goldie/v2 v2.5.3/go.mod h1:oZ9fp0+se1eapSRjfYbsV/0Hqhbuu3bJVvKI/NNtssI=

319github.com/sebdah/goldie/v2 v2.5.3/go.mod h1:oZ9fp0+se1eapSRjfYbsV/0Hqhbuu3bJVvKI/NNtssI=

361−github.com/segmentio/asm v1.2.1 h1:DTNbBqs57ioxAD4PrArqftgypG4/qNpXoJx8TVXxPR0=

362−github.com/segmentio/asm v1.2.1/go.mod h1:BqMnlJP91P8d+4ibuonYZw9mfnzI9HfxselHZr5aAcs=

363−github.com/segmentio/encoding v0.5.4 h1:OW1VRern8Nw6ITAtwSZ7Idrl3MXCFwXHPgqESYfvNt0=

364−github.com/segmentio/encoding v0.5.4/go.mod h1:HS1ZKa3kSN32ZHVZ7ZLPLXWvOVIiZtyJnO1gPH1sKt0=

365−github.com/sergeymakinen/go-bmp v1.0.0 h1:SdGTzp9WvCV0A1V0mBeaS7kQAwNLdVJbmHlqNWq0R+M=

366−github.com/sergeymakinen/go-bmp v1.0.0/go.mod h1:/mxlAQZRLxSvJFNIEGGLBE/m40f3ZnUifpgVDlcUIEY=

367−github.com/sergeymakinen/go-ico v1.0.0-beta.0 h1:m5qKH7uPKLdrygMWxbamVn+tl2HfiA3K6MFJw4GfZvQ=

368−github.com/sergeymakinen/go-ico v1.0.0-beta.0/go.mod h1:wQ47mTczswBO5F0NoDt7O0IXgnV4Xy3ojrroMQzyhUk=

369github.com/sergi/go-diff v1.0.0/go.mod h1:0CfEIISq7TuYL3j771MWULgwwjU+GofnZX9QAmXWZgo=

320github.com/sergi/go-diff v1.0.0/go.mod h1:0CfEIISq7TuYL3j771MWULgwwjU+GofnZX9QAmXWZgo=

370github.com/sergi/go-diff v1.3.1/go.mod h1:aMJSSKb2lpPvRNec0+w3fl7LP9IOFzdc9Pa4NFbPK1I=

321github.com/sergi/go-diff v1.3.1/go.mod h1:aMJSSKb2lpPvRNec0+w3fl7LP9IOFzdc9Pa4NFbPK1I=

371github.com/sergi/go-diff v1.3.2-0.20230802210424-5b0b94c5c0d3 h1:n661drycOFuPLCN3Uc8sB6B/s6Z4t2xvBgU1htSHuq8=

322github.com/sergi/go-diff v1.3.2-0.20230802210424-5b0b94c5c0d3 h1:n661drycOFuPLCN3Uc8sB6B/s6Z4t2xvBgU1htSHuq8=

@@ -378,28 +329,21 @@ github.com/spf13/cobra v1.10.2 h1:DMTTonx5m65Ic0GOoRY2c16WCbHxOOw6xxezuLaBpcU=

378github.com/spf13/cobra v1.10.2/go.mod h1:7C1pvHqHw5A4vrJfjNwvOdzYu0Gml16OCs2GRiTUUS4=

329github.com/spf13/cobra v1.10.2/go.mod h1:7C1pvHqHw5A4vrJfjNwvOdzYu0Gml16OCs2GRiTUUS4=

379github.com/spf13/pflag v1.0.9 h1:9exaQaMOCwffKiiiYk6/BndUBv+iRViNW+4lEMi0PvY=

330github.com/spf13/pflag v1.0.9 h1:9exaQaMOCwffKiiiYk6/BndUBv+iRViNW+4lEMi0PvY=

380github.com/spf13/pflag v1.0.9/go.mod h1:McXfInJRrz4CZXVZOBLb0bTZqETkiAhM9Iw0y3An2Bg=

331github.com/spf13/pflag v1.0.9/go.mod h1:McXfInJRrz4CZXVZOBLb0bTZqETkiAhM9Iw0y3An2Bg=

332+github.com/srwiley/oksvg v0.0.0-20221011165216-be6e8873101c h1:km8GpoQut05eY3GiYWEedbTT0qnSxrCjsVbb7yKY1KE=

333+github.com/srwiley/oksvg v0.0.0-20221011165216-be6e8873101c/go.mod h1:cNQ3dwVJtS5Hmnjxy6AgTPd0Inb3pW05ftPSX7NZO7Q=

334+github.com/srwiley/rasterx v0.0.0-20220730225603-2ab79fcdd4ef h1:Ch6Q+AZUxDBCVqdkI8FSpFyZDtCVBc2VmejdNrm5rRQ=

335+github.com/srwiley/rasterx v0.0.0-20220730225603-2ab79fcdd4ef/go.mod h1:nXTWP6+gD5+LUJ8krVhhoeHjvHTutPxMYl5SvkcnJNE=

381github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=

336github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=

382−github.com/stretchr/objx v0.4.0/go.mod h1:YvHI0jy2hoMjB+UWwv71VJQ9isScKT/TqJzVSSt89Yw=

383−github.com/stretchr/objx v0.5.0/go.mod h1:Yh+to48EsGEfYuaHDzXPcE3xhTkx73EhmCGUpEOglKo=

384github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=

337github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=

385github.com/stretchr/testify v1.4.0/go.mod h1:j7eGeouHqKxXV5pUuKE4zz7dFj8WfuZ+81PSLYec5m4=

338github.com/stretchr/testify v1.4.0/go.mod h1:j7eGeouHqKxXV5pUuKE4zz7dFj8WfuZ+81PSLYec5m4=

386github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=

339github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=

387−github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=

388−github.com/stretchr/testify v1.8.0/go.mod h1:yNjHg4UonilssWZ8iaSj1OCr/vHnekPRkoO+kdMU+MU=

389−github.com/stretchr/testify v1.8.4/go.mod h1:sz/lmYIOXD/1dqDmKjjqLyZ2RngseejIcXlSw2iwfAo=

390github.com/stretchr/testify v1.11.1 h1:7s2iGBzp5EwR7/aIZr8ao5+dra3wiQyKjjFuvgVKu7U=

340github.com/stretchr/testify v1.11.1 h1:7s2iGBzp5EwR7/aIZr8ao5+dra3wiQyKjjFuvgVKu7U=

391github.com/stretchr/testify v1.11.1/go.mod h1:wZwfW3scLgRK+23gO65QZefKpKQRnfz6sD981Nm4B6U=

341github.com/stretchr/testify v1.11.1/go.mod h1:wZwfW3scLgRK+23gO65QZefKpKQRnfz6sD981Nm4B6U=

392−github.com/swaggo/files/v2 v2.0.0 h1:hmAt8Dkynw7Ssz46F6pn8ok6YmGZqHSVLZ+HQM7i0kw=

342+github.com/tetratelabs/wazero v1.11.0 h1:+gKemEuKCTevU4d7ZTzlsvgd1uaToIDtlQlmNbwqYhA=

393−github.com/swaggo/files/v2 v2.0.0/go.mod h1:24kk2Y9NYEJ5lHuCra6iVwkMjIekMCaFq/0JQj66kyM=

343+github.com/tetratelabs/wazero v1.11.0/go.mod h1:eV28rsN8Q+xwjogd7f4/Pp4xFxO7uOGbLcD/LzB1wiU=

394−github.com/swaggo/http-swagger/v2 v2.0.2 h1:FKCdLsl+sFCx60KFsyM0rDarwiUSZ8DqbfSyIKC9OBg=

395−github.com/swaggo/http-swagger/v2 v2.0.2/go.mod h1:r7/GBkAWIfK6E/OLnE8fXnviHiDeAHmgIyooa4xm3AQ=

396−github.com/swaggo/swag v1.16.6 h1:qBNcx53ZaX+M5dxVyTrgQ0PJ/ACK+NzhwcbieTt+9yI=

397−github.com/swaggo/swag v1.16.6/go.mod h1:ngP2etMK5a0P3QBizic5MEwpRmluJZPHjXcMoj4Xesg=

398−github.com/tadvi/systray v0.0.0-20190226123456-11a2b8fa57af h1:6yITBqGTE2lEeTPG04SN9W+iWHCRyHqlVYILiSXziwk=

399−github.com/tadvi/systray v0.0.0-20190226123456-11a2b8fa57af/go.mod h1:4F09kP5F+am0jAwlQLddpoMDM+iewkxxt6nxUQ5nq5o=

400github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=

344github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=

401−github.com/tidwall/gjson v1.19.0 h1:xwxm7n691Uf3u5OFjzngavjGTh55KX5q/9w9xHW88JU=

345+github.com/tidwall/gjson v1.18.0 h1:FIDeeyB800efLX89e5a8Y0BNH+LOngJyGrIWxG2FKQY=

402−github.com/tidwall/gjson v1.19.0/go.mod h1:V37/opeE/JbLUOfH0QTXiNez2l0RUjYUhpT4szFQAfc=

346+github.com/tidwall/gjson v1.18.0/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=

403github.com/tidwall/match v1.1.1 h1:+Ho715JplO36QYgwN9PGYNhgZvoUSc9X2c80KVTi+GA=

347github.com/tidwall/match v1.1.1 h1:+Ho715JplO36QYgwN9PGYNhgZvoUSc9X2c80KVTi+GA=

404github.com/tidwall/match v1.1.1/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM=

348github.com/tidwall/match v1.1.1/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM=

405github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=

349github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=

@@ -411,10 +355,10 @@ github.com/u-root/u-root v0.14.1-0.20250807200646-5e7721023dc7 h1:ax+jBy7xFhh+Ka

411github.com/u-root/u-root v0.14.1-0.20250807200646-5e7721023dc7/go.mod h1:/0Qr7qJeDwWxoKku2xKQ4Szc+SwBE3g9VE8jNiamsmc=

355github.com/u-root/u-root v0.14.1-0.20250807200646-5e7721023dc7/go.mod h1:/0Qr7qJeDwWxoKku2xKQ4Szc+SwBE3g9VE8jNiamsmc=

412github.com/u-root/uio v0.0.0-20240224005618-d2acac8f3701 h1:pyC9PaHYZFgEKFdlp3G8RaCKgVpHZnecvArXvPXcFkM=

356github.com/u-root/uio v0.0.0-20240224005618-d2acac8f3701 h1:pyC9PaHYZFgEKFdlp3G8RaCKgVpHZnecvArXvPXcFkM=

413github.com/u-root/uio v0.0.0-20240224005618-d2acac8f3701/go.mod h1:P3a5rG4X7tI17Nn3aOIAYr5HbIMukwXG0urG0WuL8OA=

357github.com/u-root/uio v0.0.0-20240224005618-d2acac8f3701/go.mod h1:P3a5rG4X7tI17Nn3aOIAYr5HbIMukwXG0urG0WuL8OA=

358+github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc=

359+github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=

414github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavMF/ppJZNG9ZpyihvCd0w101no=

360github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e h1:JVG44RsyaB9T2KIHavMF/ppJZNG9ZpyihvCd0w101no=

415github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=

361github.com/xo/terminfo v0.0.0-20220910002029-abceb7e1c41e/go.mod h1:RbqR21r5mrJuqunuUZ/Dhy/avygyECGrLceyNeo4LiM=

416−github.com/xyproto/randomstring v1.0.5 h1:YtlWPoRdgMu3NZtP45drfy1GKoojuR7hmRcnhZqKjWU=

417−github.com/xyproto/randomstring v1.0.5/go.mod h1:rgmS5DeNXLivK7YprL0pY+lTuhNQW3iGxZ18UQApw/E=

418github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=

362github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=

419github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT040gmv5BQTMR2HP4=

363github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT040gmv5BQTMR2HP4=

420github.com/yuin/goldmark v1.4.13/go.mod h1:6yULJ656Px+3vBD8DxQVa3kxgyrAnzto9xy5taEt/CY=

364github.com/yuin/goldmark v1.4.13/go.mod h1:6yULJ656Px+3vBD8DxQVa3kxgyrAnzto9xy5taEt/CY=

@@ -427,33 +371,27 @@ github.com/zeebo/assert v1.3.0 h1:g7C04CbJuIDKNPFHmsk4hwZDO5O+kntRxzaUoNXj+IQ=

427github.com/zeebo/assert v1.3.0/go.mod h1:Pq9JiuJQpG8JLJdtkwrJESF0Foym2/D9XMU5ciN/wJ0=

371github.com/zeebo/assert v1.3.0/go.mod h1:Pq9JiuJQpG8JLJdtkwrJESF0Foym2/D9XMU5ciN/wJ0=

428github.com/zeebo/xxh3 v1.1.0 h1:s7DLGDK45Dyfg7++yxI0khrfwq9661w9EN78eP/UZVs=

372github.com/zeebo/xxh3 v1.1.0 h1:s7DLGDK45Dyfg7++yxI0khrfwq9661w9EN78eP/UZVs=

429github.com/zeebo/xxh3 v1.1.0/go.mod h1:IisAie1LELR4xhVinxWS5+zf1lA4p0MW4T+w+W07F5s=

373github.com/zeebo/xxh3 v1.1.0/go.mod h1:IisAie1LELR4xhVinxWS5+zf1lA4p0MW4T+w+W07F5s=

430−go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=

374+go.opentelemetry.io/auto/sdk v1.1.0 h1:cH53jehLUN6UFLY71z+NDOiNJqDdPRaXzTel0sJySYA=

431−go.opentelemetry.io/auto/sdk v1.2.1/go.mod h1:KRTj+aOaElaLi+wW1kO/DZRXwkF4C5xPbEe3ZiIhN7Y=

375+go.opentelemetry.io/auto/sdk v1.1.0/go.mod h1:3wSPjt5PWp2RhlCcmmOial7AvC4DQqZb7a7wCow3W8A=

432−go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.69.0 h1:2yEATaop1/a1I4psnSLgWVPLWwCzkqWakgJy7xTDVy0=

376+go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.61.0 h1:q4XOmH/0opmeuJtPsbFNivyl7bCt7yRBbeEm2sC/XtQ=

433−go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.69.0/go.mod h1:D7J12YRapIekYyPWgGPlA/23pRmpSEZC5xJC/TTLI9U=

377+go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.61.0/go.mod h1:snMWehoOh2wsEwnvvwtDyFCxVeDAODenXHtn5vzrKjo=

434−go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.69.0 h1:8tvICD4vSTOOsNrsI4Ljf6C+6UKvpTEH5XY3JMoyPoo=

378+go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.61.0 h1:F7Jx+6hwnZ41NSFTO5q4LYDtJRXBf2PD0rNBkeB/lus=

435−go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.69.0/go.mod h1:z9+yiacE0IHRqM4qFfkbt/JYlmYXgss8GY/jXoNuPJI=

379+go.opentelemetry.io/contrib/instrumentation/net/http/otelhttp v0.61.0/go.mod h1:UHB22Z8QsdRDrnAtX4PntOl36ajSxcdUMt1sF7Y6E7Q=

436−go.opentelemetry.io/otel v1.44.0 h1:JjwHmHpA4iZ3wBxluu2fbbE7j4kqlE8jXyAyPXH7HqU=

380+go.opentelemetry.io/otel v1.37.0 h1:9zhNfelUvx0KBfu/gb+ZgeAfAgtWrfHJZcAqFC228wQ=

437−go.opentelemetry.io/otel v1.44.0/go.mod h1:BMgjTHL9WPRlRjL2oZCBTL4whCGtXch2H4BhOPIAyYc=

381+go.opentelemetry.io/otel v1.37.0/go.mod h1:ehE/umFRLnuLa/vSccNq9oS1ErUlkkK71gMcN34UG8I=

438−go.opentelemetry.io/otel/metric v1.44.0 h1:1w0gILTcHdr3YI+ixLyjemwrVnsMURbTZFrSYCdDdmc=

382+go.opentelemetry.io/otel/metric v1.37.0 h1:mvwbQS5m0tbmqML4NqK+e3aDiO02vsf/WgbsdpcPoZE=

439−go.opentelemetry.io/otel/metric v1.44.0/go.mod h1:8O7hanEPBNgEMmybD3s2VBKcgWOCsA6tzHBPODAiquo=

383+go.opentelemetry.io/otel/metric v1.37.0/go.mod h1:04wGrZurHYKOc+RKeye86GwKiTb9FKm1WHtO+4EVr2E=

440−go.opentelemetry.io/otel/sdk v1.44.0 h1:nHYwb9lK+fJPU/dnT6s7W7Z8itMWyqrnVfbheVYrZ58=

384+go.opentelemetry.io/otel/sdk v1.37.0 h1:ItB0QUqnjesGRvNcmAcU0LyvkVyGJ2xftD29bWdDvKI=

441−go.opentelemetry.io/otel/sdk v1.44.0/go.mod h1:Osuydd3Se74nqjAKxid74N5eC+jfEqfTegHRnq58oK0=

385+go.opentelemetry.io/otel/sdk v1.37.0/go.mod h1:VredYzxUvuo2q3WRcDnKDjbdvmO0sCzOvVAiY+yUkAg=

442−go.opentelemetry.io/otel/sdk/metric v1.44.0 h1:3LlKgI+VjbVsjNRFZJZAJ30WjXC5VkNRks6si09iEfI=

386+go.opentelemetry.io/otel/sdk/metric v1.37.0 h1:90lI228XrB9jCMuSdA0673aubgRobVZFhbjxHHspCPc=

443−go.opentelemetry.io/otel/sdk/metric v1.44.0/go.mod h1:5B5pMARnXxKhltooO4xUuCBorl65a4EpnTalObqOigA=

387+go.opentelemetry.io/otel/sdk/metric v1.37.0/go.mod h1:cNen4ZWfiD37l5NhS+Keb5RXVWZWpRE+9WyVCpbo5ps=

444−go.opentelemetry.io/otel/trace v1.44.0 h1:jxF5CsGYCe74MCRx2X4g7WsY/VBKRqqpNvXlX/6gtIk=

388+go.opentelemetry.io/otel/trace v1.37.0 h1:HLdcFNbRQBE2imdSEgm/kwqmQj1Or1l/7bW6mxVK7z4=

445−go.opentelemetry.io/otel/trace v1.44.0/go.mod h1:oLl1jrMQAVo6v3GAggN+1VH9VIz9iUSvW53sW1Q8PIE=

389+go.opentelemetry.io/otel/trace v1.37.0/go.mod h1:TlgrlQ+PtQO5XFerSPUYG0JSgGyryXewPGyayAWSBS0=

446−go.uber.org/goleak v1.3.0 h1:2K3zAYmnTNqV73imy9J1T3WC+gmCePx2hEGkimedGto=

447−go.uber.org/goleak v1.3.0/go.mod h1:CoHD4mav9JJNrW/WLlf7HGZPjdw8EucARQHekz1X6bE=

448go.uber.org/multierr v1.11.0 h1:blXXJkSxSSfBVBlC76pxqeO+LN3aDfLQo+309xJstO0=

390go.uber.org/multierr v1.11.0 h1:blXXJkSxSSfBVBlC76pxqeO+LN3aDfLQo+309xJstO0=

449go.uber.org/multierr v1.11.0/go.mod h1:20+QtiLqy0Nd6FdQB9TLXag12DsQkrbs3htMFfDN80Y=

391go.uber.org/multierr v1.11.0/go.mod h1:20+QtiLqy0Nd6FdQB9TLXag12DsQkrbs3htMFfDN80Y=

450go.yaml.in/yaml/v3 v3.0.4/go.mod h1:DhzuOOF2ATzADvBadXxruRBLzYTpT36CKvDb3+aBEFg=

392go.yaml.in/yaml/v3 v3.0.4/go.mod h1:DhzuOOF2ATzADvBadXxruRBLzYTpT36CKvDb3+aBEFg=

451go.yaml.in/yaml/v4 v4.0.0-rc.3 h1:3h1fjsh1CTAPjW7q/EMe+C8shx5d8ctzZTrLcs/j8Go=

393go.yaml.in/yaml/v4 v4.0.0-rc.3 h1:3h1fjsh1CTAPjW7q/EMe+C8shx5d8ctzZTrLcs/j8Go=

452go.yaml.in/yaml/v4 v4.0.0-rc.3/go.mod h1:aZqd9kCMsGL7AuUv/m/PvWLdg5sjJsZ4oHDEnfPPfY0=

394go.yaml.in/yaml/v4 v4.0.0-rc.3/go.mod h1:aZqd9kCMsGL7AuUv/m/PvWLdg5sjJsZ4oHDEnfPPfY0=

453−golang.design/x/clipboard v0.8.0 h1:6VEcH28wwcSgKc+vnxHHDWiRjrakTQIAJnPUrt3aOgg=

454−golang.design/x/clipboard v0.8.0/go.mod h1:s0pwrtA3Q9fgnVtGDmP5ZK/pp55cQKB23esKsjwWhWM=

455−golang.design/x/x11 v0.2.0 h1:Uiwu2guGihsJX/ZCzpoDPFz5gR/Qntm08mvoBCmRydo=

456−golang.design/x/x11 v0.2.0/go.mod h1:/5q1mFkdc1rL8mvB7DsQFi6as4tIkBv4FXjcP07mrkE=

457golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=

395golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w=

458golang.org/x/crypto v0.0.0-20210921155107-089bfa567519/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=

396golang.org/x/crypto v0.0.0-20210921155107-089bfa567519/go.mod h1:GvvjBRRGRdwPK5ydBHafDWAxML/pGHZbMvKqRZ5+Abc=

459golang.org/x/crypto v0.13.0/go.mod h1:y6Z2r+Rw4iayiXXAIxJIDAJ1zMW4yaTpebo8fPOliYc=

397golang.org/x/crypto v0.13.0/go.mod h1:y6Z2r+Rw4iayiXXAIxJIDAJ1zMW4yaTpebo8fPOliYc=

@@ -461,24 +399,20 @@ golang.org/x/crypto v0.19.0/go.mod h1:Iy9bg/ha4yyC70EfRS8jz+B6ybOBKMaSxLj6P6oBDf

461golang.org/x/crypto v0.22.0/go.mod h1:vr6Su+7cTlO45qkww3VDJlzDn0ctJvRgYbC2NvXHt+M=

399golang.org/x/crypto v0.22.0/go.mod h1:vr6Su+7cTlO45qkww3VDJlzDn0ctJvRgYbC2NvXHt+M=

462golang.org/x/crypto v0.23.0/go.mod h1:CKFgDieR+mRhux2Lsu27y0fO304Db0wZe70UKqHu0v8=

400golang.org/x/crypto v0.23.0/go.mod h1:CKFgDieR+mRhux2Lsu27y0fO304Db0wZe70UKqHu0v8=

463golang.org/x/crypto v0.31.0/go.mod h1:kDsLvtWBEx7MV9tJOj9bnXsPbxwJQ6csT/x4KIN4Ssk=

401golang.org/x/crypto v0.31.0/go.mod h1:kDsLvtWBEx7MV9tJOj9bnXsPbxwJQ6csT/x4KIN4Ssk=

464−golang.org/x/crypto v0.53.0 h1:QZ4Muo8THX6CizN2vPPd5fBGHyogrdK9fG4wLPFUsto=

402+golang.org/x/crypto v0.47.0 h1:V6e3FRj+n4dbpw86FJ8Fv7XVOql7TEwpHapKoMJ/GO8=

465−golang.org/x/crypto v0.53.0/go.mod h1:DNLU434OwVakk9PzuwV8w62mAJpRJL3vsgcfp4Qnsio=

403+golang.org/x/crypto v0.47.0/go.mod h1:ff3Y9VzzKbwSSEzWqJsJVBnWmRwRSHt/6Op5n9bQc4A=

466−golang.org/x/exp v0.0.0-20260410095643-746e56fc9e2f h1:W3F4c+6OLc6H2lb//N1q4WpJkhzJCK5J6kUi1NTVXfM=

404+golang.org/x/exp v0.0.0-20251023183803-a4bb9ffd2546 h1:mgKeJMpvi0yx/sU5GsxQ7p6s2wtOnGAHZWCHUM4KGzY=

467−golang.org/x/exp v0.0.0-20260410095643-746e56fc9e2f/go.mod h1:J1xhfL/vlindoeF/aINzNzt2Bket5bjo9sdOYzOsU80=

405+golang.org/x/exp v0.0.0-20251023183803-a4bb9ffd2546/go.mod h1:j/pmGrbnkbPtQfxEe5D0VQhZC6qKbfKifgD0oM7sR70=

468−golang.org/x/exp/shiny v0.0.0-20250606033433-dcc06ee1d476 h1:Wdx0vgH5Wgsw+lF//LJKmWOJBLWX6nprsMqnf99rYDE=

469−golang.org/x/exp/shiny v0.0.0-20250606033433-dcc06ee1d476/go.mod h1:ygj7T6vSGhhm/9yTpOQQNvuAUFziTH7RUiH74EoE2C8=

470golang.org/x/image v0.0.0-20191009234506-e7c1f5e7dbb8/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=

406golang.org/x/image v0.0.0-20191009234506-e7c1f5e7dbb8/go.mod h1:FeLwcggjj3mMvU+oOTbSwawSJRM1uh48EjtB4UJZlP0=

471−golang.org/x/image v0.43.0 h1:FLxcP4ec2350nTfOC8ysKtqYSIFbk/QGjw1ZHNP4tsY=

407+golang.org/x/image v0.34.0 h1:33gCkyw9hmwbZJeZkct8XyR11yH889EQt/QH4VmXMn8=

472−golang.org/x/image v0.43.0/go.mod h1:rrpelvGFt+kLPAjPM4HeWPgrl0FtafueU//e5N0qk/Q=

408+golang.org/x/image v0.34.0/go.mod h1:2RNFBZRB+vnwwFil8GkMdRvrJOFd1AzdZI6vOY+eJVU=

473−golang.org/x/mobile v0.0.0-20250606033058-a2a15c67f36f h1:/n+PL2HlfqeSiDCuhdBbRNlGS/g2fM4OHufalHaTVG8=

474−golang.org/x/mobile v0.0.0-20250606033058-a2a15c67f36f/go.mod h1:ESkJ836Z6LpG6mTVAhA48LpfW/8fNR0ifStlH2axyfg=

475golang.org/x/mod v0.6.0-dev.0.20220419223038-86c51ed26bb4/go.mod h1:jJ57K6gSWd91VN4djpZkiMVwK6gcyfeH4XE8wZrZaV4=

409golang.org/x/mod v0.6.0-dev.0.20220419223038-86c51ed26bb4/go.mod h1:jJ57K6gSWd91VN4djpZkiMVwK6gcyfeH4XE8wZrZaV4=

476golang.org/x/mod v0.8.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs=

410golang.org/x/mod v0.8.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs=

477golang.org/x/mod v0.12.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs=

411golang.org/x/mod v0.12.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs=

478golang.org/x/mod v0.15.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=

412golang.org/x/mod v0.15.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=

479golang.org/x/mod v0.17.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=

413golang.org/x/mod v0.17.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c=

480−golang.org/x/mod v0.36.0 h1:JJjpVx6myfUsUdAzZuOSTTmRE0PfZeNWzzvKrP7amb4=

414+golang.org/x/mod v0.32.0 h1:9F4d3PHLljb6x//jOyokMv3eX+YDeepZSEo3mFJy93c=

481−golang.org/x/mod v0.36.0/go.mod h1:moc6ELqsWcOw5Ef3xVprK5ul/MvtVvkIXLziUOICjUQ=

415+golang.org/x/mod v0.32.0/go.mod h1:SgipZ/3h2Ci89DlEtEXWUk/HteuRin+HHhN+WbNhguU=

482golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=

416golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s=

483golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=

417golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg=

484golang.org/x/net v0.0.0-20220722155237-a158d28d115b/go.mod h1:XRhObCWvk6IyKnWLug+ECip1KBveYUHfp+8e9klMJ9c=

418golang.org/x/net v0.0.0-20220722155237-a158d28d115b/go.mod h1:XRhObCWvk6IyKnWLug+ECip1KBveYUHfp+8e9klMJ9c=

@@ -490,10 +424,10 @@ golang.org/x/net v0.21.0/go.mod h1:bIjVDfnllIU7BJ2DNgfnXvpSvtn8VRwhlsaeUTyUS44=

490golang.org/x/net v0.24.0/go.mod h1:2Q7sJY5mzlzWjKtYUEXSlBWCdyaioyXzRB2RtU8KVE8=

424golang.org/x/net v0.24.0/go.mod h1:2Q7sJY5mzlzWjKtYUEXSlBWCdyaioyXzRB2RtU8KVE8=

491golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=

425golang.org/x/net v0.25.0/go.mod h1:JkAGAh7GEvH74S6FOH42FLoXpXbE/aqXSrIQjXgsiwM=

492golang.org/x/net v0.33.0/go.mod h1:HXLR5J+9DxmrqMwG9qjGCxZ+zKXxBru04zlTvWlWuN4=

426golang.org/x/net v0.33.0/go.mod h1:HXLR5J+9DxmrqMwG9qjGCxZ+zKXxBru04zlTvWlWuN4=

493−golang.org/x/net v0.56.0 h1:Rw8j/hFzGvJUZwNBXnAtf5sVDVt+65SK2C7IxCxZt5o=

427+golang.org/x/net v0.49.0 h1:eeHFmOGUTtaaPSGNmjBKpbng9MulQsJURQUAfUwY++o=

494−golang.org/x/net v0.56.0/go.mod h1:D3Ku6r+V6JROoZK144D2XfMHFcMq/0zSfLelVTCFKec=

428+golang.org/x/net v0.49.0/go.mod h1:/ysNB2EvaqvesRkuLAyjI1ycPZlQHM3q01F02UY/MV8=

495−golang.org/x/oauth2 v0.36.0 h1:peZ/1z27fi9hUOFCAZaHyrpWG5lwe0RJEEEeH0ThlIs=

429+golang.org/x/oauth2 v0.34.0 h1:hqK/t4AKgbqWkdkcAeI8XLmbK+4m4G5YeQRrmiotGlw=

496−golang.org/x/oauth2 v0.36.0/go.mod h1:YDBUJMTkDnJS+A4BP4eZBjCqtokkg1hODuPjwiGPO7Q=

430+golang.org/x/oauth2 v0.34.0/go.mod h1:lzm5WQJQwKZ3nwavOZ3IS5Aulzxi68dUSgRHujetwEA=

497golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=

431golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=

498golang.org/x/sync v0.0.0-20220722155255-886fb9371eb4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=

432golang.org/x/sync v0.0.0-20220722155255-886fb9371eb4/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=

499golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=

433golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=

@@ -501,8 +435,8 @@ golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y=

501golang.org/x/sync v0.6.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=

435golang.org/x/sync v0.6.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=

502golang.org/x/sync v0.7.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=

436golang.org/x/sync v0.7.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=

503golang.org/x/sync v0.10.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=

437golang.org/x/sync v0.10.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk=

504−golang.org/x/sync v0.21.0 h1:HLII4xRRTtCRkxYp4HNFF0Js/Og6q2i++KXbg0gHCwM=

438+golang.org/x/sync v0.19.0 h1:vV+1eWNmZ5geRlYjzm2adRgW2/mcpevXNg50YZtPCE4=

505−golang.org/x/sync v0.21.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=

439+golang.org/x/sync v0.19.0/go.mod h1:9KTHXmSnoGruLpwFjVSX0lNNA75CykiMECbovNTZqGI=

506golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=

440golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY=

507golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=

441golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs=

508golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

442golang.org/x/sys v0.0.0-20210615035016-665e8c7367d1/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

@@ -511,6 +445,7 @@ golang.org/x/sys v0.0.0-20220722155257-8c9f86f7a55f/go.mod h1:oPkhp1MJrh7nUepCBc

511golang.org/x/sys v0.0.0-20220908164124-27713097b956/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

445golang.org/x/sys v0.0.0-20220908164124-27713097b956/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

512golang.org/x/sys v0.1.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

446golang.org/x/sys v0.1.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

513golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

447golang.org/x/sys v0.5.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

448+golang.org/x/sys v0.6.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

514golang.org/x/sys v0.7.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

449golang.org/x/sys v0.7.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

515golang.org/x/sys v0.8.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

450golang.org/x/sys v0.8.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

516golang.org/x/sys v0.12.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

451golang.org/x/sys v0.12.0/go.mod h1:oPkhp1MJrh7nUepCBck5+mAzfO9JrbApNNgaTdGDITg=

@@ -518,8 +453,8 @@ golang.org/x/sys v0.17.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

518golang.org/x/sys v0.19.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

453golang.org/x/sys v0.19.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

519golang.org/x/sys v0.20.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

454golang.org/x/sys v0.20.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

520golang.org/x/sys v0.28.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

455golang.org/x/sys v0.28.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=

521−golang.org/x/sys v0.46.0 h1:noSf2Fq6F8DBgS+LysIkx7rIExoNHJsxOAtPp4rthXw=

456+golang.org/x/sys v0.40.0 h1:DBZZqJ2Rkml6QMQsZywtnjnnGvHza6BTfYFWY9kjEWQ=

522−golang.org/x/sys v0.46.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=

457+golang.org/x/sys v0.40.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=

523golang.org/x/telemetry v0.0.0-20240228155512-f48c80bd79b2/go.mod h1:TeRTkGYfJXctD9OcfyVLyj2J3IxLnKwHJR8f4D8a3YE=

458golang.org/x/telemetry v0.0.0-20240228155512-f48c80bd79b2/go.mod h1:TeRTkGYfJXctD9OcfyVLyj2J3IxLnKwHJR8f4D8a3YE=

524golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=

459golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=

525golang.org/x/term v0.0.0-20210927222741-03fcf44c2211/go.mod h1:jbD1KX2456YbFQfuXm/mYQcufACuNUgVhRMnK/tPxf8=

460golang.org/x/term v0.0.0-20210927222741-03fcf44c2211/go.mod h1:jbD1KX2456YbFQfuXm/mYQcufACuNUgVhRMnK/tPxf8=

@@ -531,8 +466,8 @@ golang.org/x/term v0.17.0/go.mod h1:lLRBjIVuehSbZlaOtGMbcMncT+aqLLLmKrsjNrUguwk=

531golang.org/x/term v0.19.0/go.mod h1:2CuTdWZ7KHSQwUzKva0cbMg6q2DMI3Mmxp+gKJbskEk=

466golang.org/x/term v0.19.0/go.mod h1:2CuTdWZ7KHSQwUzKva0cbMg6q2DMI3Mmxp+gKJbskEk=

532golang.org/x/term v0.20.0/go.mod h1:8UkIAJTvZgivsXaD6/pH6U9ecQzZ45awqEOzuCvwpFY=

467golang.org/x/term v0.20.0/go.mod h1:8UkIAJTvZgivsXaD6/pH6U9ecQzZ45awqEOzuCvwpFY=

533golang.org/x/term v0.27.0/go.mod h1:iMsnZpn0cago0GOrHO2+Y7u7JPn5AylBrcoWkElMTSM=

468golang.org/x/term v0.27.0/go.mod h1:iMsnZpn0cago0GOrHO2+Y7u7JPn5AylBrcoWkElMTSM=

534−golang.org/x/term v0.44.0 h1:0rLvDRCtNj0gZkyIXhCyOb2OAzEhLVqc4B+hrsBhrmc=

469+golang.org/x/term v0.39.0 h1:RclSuaJf32jOqZz74CkPA9qFuVTX7vhLlpfj/IGWlqY=

535−golang.org/x/term v0.44.0/go.mod h1:7ze4MdzUzLXpSAoFP1H0bOI9aXDqveSvatT5vKcFh2Y=

470+golang.org/x/term v0.39.0/go.mod h1:yxzUCTP/U+FzoxfdKmLaA0RV1WgE0VY7hXBwKtY/4ww=

536golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=

471golang.org/x/text v0.3.0/go.mod h1:NqM8EUOU14njkJ3fqMW+pc6Ldnwhi/IjpwHt7yyuwOQ=

537golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=

472golang.org/x/text v0.3.3/go.mod h1:5Zoc/QRtKVWzQhOtBMvqHzDpF6irO9z98xDceosuGiQ=

538golang.org/x/text v0.3.7/go.mod h1:u+2+/6zg+i71rQMx5EYifcz6MCKuco9NR6JIITiCfzQ=

473golang.org/x/text v0.3.7/go.mod h1:u+2+/6zg+i71rQMx5EYifcz6MCKuco9NR6JIITiCfzQ=

@@ -542,39 +477,33 @@ golang.org/x/text v0.13.0/go.mod h1:TvPlkZtksWOMsz7fbANvkp4WM8x/WCo/om8BMLbz+aE=

542golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=

477golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=

543golang.org/x/text v0.15.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=

478golang.org/x/text v0.15.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=

544golang.org/x/text v0.21.0/go.mod h1:4IBbMaMmOPCJ8SecivzSH54+73PCFmPWxNTLm+vZkEQ=

479golang.org/x/text v0.21.0/go.mod h1:4IBbMaMmOPCJ8SecivzSH54+73PCFmPWxNTLm+vZkEQ=

545−golang.org/x/text v0.38.0 h1:sXmwo9DwP3OK9EZ7PqAdaooSGozfl/3a6/xJcbzPRhE=

480+golang.org/x/text v0.33.0 h1:B3njUFyqtHDUI5jMn1YIr5B0IE2U0qck04r6d4KPAxE=

546−golang.org/x/text v0.38.0/go.mod h1:YXZt3QhHUKYT53r2lLKFIVi6Ao1jdzrTR/KQ09qyxF4=

481+golang.org/x/text v0.33.0/go.mod h1:LuMebE6+rBincTi9+xWTY8TztLzKHc/9C1uBCG27+q8=

547−golang.org/x/time v0.15.0 h1:bbrp8t3bGUeFOx08pvsMYRTCVSMk89u4tKbNOZbp88U=

482+golang.org/x/time v0.14.0 h1:MRx4UaLrDotUKUdCIqzPC48t1Y9hANFKIRpNx+Te8PI=

548−golang.org/x/time v0.15.0/go.mod h1:Y4YMaQmXwGQZoFaVFk4YpCt4FLQMYKZe9oeV/f4MSno=

483+golang.org/x/time v0.14.0/go.mod h1:eL/Oa2bBBK0TkX57Fyni+NgnyQQN4LitPmob2Hjnqw4=

549golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=

484golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ=

550golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo=

485golang.org/x/tools v0.0.0-20191119224855-298f0cb1881e/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo=

551golang.org/x/tools v0.1.12/go.mod h1:hNGJHUnrk76NpqgfD5Aqm5Crs+Hm0VOH/i9J2+nxYbc=

486golang.org/x/tools v0.1.12/go.mod h1:hNGJHUnrk76NpqgfD5Aqm5Crs+Hm0VOH/i9J2+nxYbc=

552golang.org/x/tools v0.6.0/go.mod h1:Xwgl3UAJ/d3gWutnCtw505GrjyAbvKui8lOU390QaIU=

487golang.org/x/tools v0.6.0/go.mod h1:Xwgl3UAJ/d3gWutnCtw505GrjyAbvKui8lOU390QaIU=

553golang.org/x/tools v0.13.0/go.mod h1:HvlwmtVNQAhOuCjW7xxvovg8wbNq7LwfXh/k7wXUl58=

488golang.org/x/tools v0.13.0/go.mod h1:HvlwmtVNQAhOuCjW7xxvovg8wbNq7LwfXh/k7wXUl58=

554golang.org/x/tools v0.21.1-0.20240508182429-e35e4ccd0d2d/go.mod h1:aiJjzUbINMkxbQROHiO6hDPo2LHcIPhhQsa9DLh0yGk=

489golang.org/x/tools v0.21.1-0.20240508182429-e35e4ccd0d2d/go.mod h1:aiJjzUbINMkxbQROHiO6hDPo2LHcIPhhQsa9DLh0yGk=

555−golang.org/x/tools v0.45.0 h1:18qN3FAooORvApf5XjCXgsuayZOEtXf6JK18I3+ONa8=

490+golang.org/x/tools v0.40.0 h1:yLkxfA+Qnul4cs9QA3KnlFu0lVmd8JJfoq+E41uSutA=

556−golang.org/x/tools v0.45.0/go.mod h1:LuUGqqaXcXMEFEruIVJVm5mgDD8vww/z/SR1gQ4uE/0=

491+golang.org/x/tools v0.40.0/go.mod h1:Ik/tzLRlbscWpqqMRjyWYDisX8bG13FrdXp3o4Sr9lc=

557golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

492golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0=

558−gonum.org/v1/gonum v0.17.0 h1:VbpOemQlsSMrYmn7T2OUvQ4dqxQXU+ouZFQsZOx50z4=

493+gonum.org/v1/gonum v0.16.0 h1:5+ul4Swaf3ESvrOnidPp4GZbzf0mxVQpDCYUQE7OJfk=

559−gonum.org/v1/gonum v0.17.0/go.mod h1:El3tOrEuMpv2UdMrbNlKEh9vd86bmQ6vqIcDwxEOc1E=

494+gonum.org/v1/gonum v0.16.0/go.mod h1:fef3am4MQ93R2HHpKnLk4/Tbh/s0+wqD5nfa6Pnwy4E=

560−google.golang.org/api v0.286.0 h1:TdTXMvzYKnWV1/lPbCdbXRqBrkDqjPto22H2xeZZ8LI=

495+google.golang.org/api v0.239.0 h1:2hZKUnFZEy81eugPs4e2XzIJ5SOwQg0G82bpXD65Puo=

561−google.golang.org/api v0.286.0/go.mod h1:NlOlUIr8MPoIhT9Bb/oUnRuHbJOLwxb6JSYJM8Yz+jQ=

496+google.golang.org/api v0.239.0/go.mod h1:cOVEm2TpdAGHL2z+UwyS+kmlGr3bVWQQ6sYEqkKje50=

562−google.golang.org/genai v1.62.0 h1:PaBju84orf4Vbcc6OfHe4vxhxhjwulKTgOpEc3iIc00=

497+google.golang.org/genai v1.41.0 h1:ayXl75LjTmqTu0y94yr96d17gIb4zF8gWVzX2TgioEY=

563−google.golang.org/genai v1.62.0/go.mod h1:mDdPDFXo1Ats7f1WXVyZgWb/CkMzFWTWJruIMy7hGIU=

498+google.golang.org/genai v1.41.0/go.mod h1:A3kkl0nyBjyFlNjgxIwKq70julKbIxpSxqKO5gw/gmk=

564−google.golang.org/genproto v0.0.0-20260622175928-b703f567277d h1:CP5omUq8AJTiWMrPKM1WRLJ7zZeXd9OPcQD3TbBNAyY=

499+google.golang.org/genproto/googleapis/rpc v0.0.0-20250804133106-a7a43d27e69b h1:zPKJod4w6F1+nRGDI9ubnXYhU9NSWoFAijkHkUXeTK8=

565−google.golang.org/genproto v0.0.0-20260622175928-b703f567277d/go.mod h1:DrwuGJgFSEVNpv3S5Q5VxhRTvdnjauw9GtvwVOEARfA=

566−google.golang.org/genproto/googleapis/api v0.0.0-20260622175928-b703f567277d h1:xr2lwHI91bn3UiXcnyzRMQjp2LRiM8wEHzwUaE0YhTs=

567−google.golang.org/genproto/googleapis/api v0.0.0-20260622175928-b703f567277d/go.mod h1:O0ZOWSrfWfJ+Z5HbwZ+wNtHsg/vk1k2C/w67eww8PfQ=

568−google.golang.org/genproto/googleapis/rpc v0.0.0-20260622175928-b703f567277d h1:mpAgMyM9vQHxycBlDq50y1VHpfSfVwzXvrQKtYbXuUY=

569−google.golang.org/genproto/googleapis/rpc v0.0.0-20260622175928-b703f567277d/go.mod h1:4Hqkh8ycfw05ld/3BWL7rJOSfebL2Q+DVDeRgYgxUU8=

500+google.golang.org/genproto/googleapis/rpc v0.0.0-20250804133106-a7a43d27e69b/go.mod h1:qQ0YXyHHx3XkvlzUtpXDkS29lDSafHMZBAZDc03LQ3A=

570−google.golang.org/grpc v1.81.1 h1:VnnIIZ88UzOOKLukQi+ImGz8O1Wdp8nAGGnvOfEIWQQ=

501+google.golang.org/grpc v1.76.0 h1:UnVkv1+uMLYXoIz6o7chp59WfQUYA2ex/BXQ9rHZu7A=

571−google.golang.org/grpc v1.81.1/go.mod h1:xGH9GfzOyMTGIOXBJmXt+BX/V0kcdQbdcuwQ/zNw42I=

502+google.golang.org/grpc v1.76.0/go.mod h1:Ju12QI8M6iQJtbcsV+awF5a4hfJMLi4X0JLo94ULZ6c=

572−google.golang.org/protobuf v1.36.11 h1:fV6ZwhNocDyBLK0dj+fg8ektcVegBBuEolpbTQyBNVE=

503+google.golang.org/protobuf v1.36.10 h1:AYd7cD/uASjIL6Q9LiTjz8JLcrh/88q5UObnmY3aOOE=

573−google.golang.org/protobuf v1.36.11/go.mod h1:HTf+CrKn2C3g5S8VImy6tdcUvCska2kB7j23XfzDpco=

504+google.golang.org/protobuf v1.36.10/go.mod h1:HTf+CrKn2C3g5S8VImy6tdcUvCska2kB7j23XfzDpco=

574gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

505gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

575−gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

576gopkg.in/check.v1 v1.0.0-20190902080502-41f04d3bba15/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

506gopkg.in/check.v1 v1.0.0-20190902080502-41f04d3bba15/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

577−gopkg.in/check.v1 v1.0.0-20200227125254-8fa46927fb4f/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=

578gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=

507gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk=

579gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=

508gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c/go.mod h1:JHkPIbrfpd72SG/EVd6muEfDQjcINNoR0C8j2r3qZ4Q=

580gopkg.in/dnaeon/go-vcr.v4 v4.0.6-0.20251110073552-01de4eb40290 h1:g3ah7zaWmw41EtOgBNXpx8zk4HYuH3OMwB+qh1Dt834=

509gopkg.in/dnaeon/go-vcr.v4 v4.0.6-0.20251110073552-01de4eb40290 h1:g3ah7zaWmw41EtOgBNXpx8zk4HYuH3OMwB+qh1Dt834=

@@ -583,44 +512,41 @@ gopkg.in/natefinch/lumberjack.v2 v2.2.1 h1:bBRl1b0OH9s/DuPhuXpNl+VtCaJXFZ5/uEFST

583gopkg.in/natefinch/lumberjack.v2 v2.2.1/go.mod h1:YD8tP3GAjkrDg1eZH7EGmyESg/lsYskCTPBJVb9jqSc=

512gopkg.in/natefinch/lumberjack.v2 v2.2.1/go.mod h1:YD8tP3GAjkrDg1eZH7EGmyESg/lsYskCTPBJVb9jqSc=

584gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7 h1:uRGJdciOHaEIrze2W8Q3AKkepLTh2hOroT7a+7czfdQ=

513gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7 h1:uRGJdciOHaEIrze2W8Q3AKkepLTh2hOroT7a+7czfdQ=

585gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7/go.mod h1:dt/ZhP58zS4L8KSrWDmTeBkI65Dw0HsyUHuEVlX15mw=

514gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7/go.mod h1:dt/ZhP58zS4L8KSrWDmTeBkI65Dw0HsyUHuEVlX15mw=

586−gopkg.in/warnings.v0 v0.1.2 h1:wFXVbFY8DY5/xOe1ECiWdKCzZlxgshcYVNkBHstARME=

587−gopkg.in/warnings.v0 v0.1.2/go.mod h1:jksf8JmL6Qr/oQM2OXTHunEvvTAsrWBLb6OOjuVWRNI=

588gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=

515gopkg.in/yaml.v2 v2.2.2/go.mod h1:hI93XBmqTisBFMUTm0b8Fm+jr3Dg1NNxqwp+5A1VGuI=

589gopkg.in/yaml.v2 v2.4.0 h1:D8xgwECY7CYvx+Y2n4sBz93Jn9JRvxdiyyo8CTfuKaY=

516gopkg.in/yaml.v2 v2.4.0 h1:D8xgwECY7CYvx+Y2n4sBz93Jn9JRvxdiyyo8CTfuKaY=

590gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=

517gopkg.in/yaml.v2 v2.4.0/go.mod h1:RDklbk79AGWmwhnvt/jBztapEOGDOx6ZbXqjP6csGnQ=

591gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

518gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

592−gopkg.in/yaml.v3 v3.0.0-20200615113413-eeeca48fe776/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

593gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=

519gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=

594gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

520gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

595−modernc.org/cc/v4 v4.28.4 h1:Hd/4Es+MBj+/7hSdZaisNyu6bv3V0Dp2MdllyfqaH+c=

521+modernc.org/cc/v4 v4.27.1 h1:9W30zRlYrefrDV2JE2O8VDtJ1yPGownxciz5rrbQZis=

596−modernc.org/cc/v4 v4.28.4/go.mod h1:OnovgIhbbMXMu1aISnJ0wvVD1KnW+cAUJkIrAWh+kVI=

522+modernc.org/cc/v4 v4.27.1/go.mod h1:uVtb5OGqUKpoLWhqwNQo/8LwvoiEBLvZXIQ/SmO6mL0=

597−modernc.org/ccgo/v4 v4.34.4 h1:OVnSOWQjVKOYkFxoHYB+qQmSHK5gqMqARM+K9DpR/Ws=

523+modernc.org/ccgo/v4 v4.30.1 h1:4r4U1J6Fhj98NKfSjnPUN7Ze2c6MnAdL0hWw6+LrJpc=

598−modernc.org/ccgo/v4 v4.34.4/go.mod h1:qdKqE8FNIYyysougB1RX9MxCzp5oJOcQXSobANJ4TuE=

524+modernc.org/ccgo/v4 v4.30.1/go.mod h1:bIOeI1JL54Utlxn+LwrFyjCx2n2RDiYEaJVSrgdrRfM=

599−modernc.org/fileutil v1.4.0 h1:j6ZzNTftVS054gi281TyLjHPp6CPHr2KCxEXjEbD6SM=

525+modernc.org/fileutil v1.3.40 h1:ZGMswMNc9JOCrcrakF1HrvmergNLAmxOPjizirpfqBA=

600−modernc.org/fileutil v1.4.0/go.mod h1:EqdKFDxiByqxLk8ozOxObDSfcVOv/54xDs/DUHdvCUU=

526+modernc.org/fileutil v1.3.40/go.mod h1:HxmghZSZVAz/LXcMNwZPA/DRrQZEVP9VX0V4LQGQFOc=

601modernc.org/gc/v2 v2.6.5 h1:nyqdV8q46KvTpZlsw66kWqwXRHdjIlJOhG6kxiV/9xI=

527modernc.org/gc/v2 v2.6.5 h1:nyqdV8q46KvTpZlsw66kWqwXRHdjIlJOhG6kxiV/9xI=

602modernc.org/gc/v2 v2.6.5/go.mod h1:YgIahr1ypgfe7chRuJi2gD7DBQiKSLMPgBQe9oIiito=

528modernc.org/gc/v2 v2.6.5/go.mod h1:YgIahr1ypgfe7chRuJi2gD7DBQiKSLMPgBQe9oIiito=

603−modernc.org/gc/v3 v3.1.3 h1:6QAplYyVO+KdPW3pGnqmJDUxtkec8ooEWvks/hhU3lc=

529+modernc.org/gc/v3 v3.1.1 h1:k8T3gkXWY9sEiytKhcgyiZ2L0DTyCQ/nvX+LoCljoRE=

604−modernc.org/gc/v3 v3.1.3/go.mod h1:HFK/6AGESC7Ex+EZJhJ2Gni6cTaYpSMmU/cT9RmlfYY=

530+modernc.org/gc/v3 v3.1.1/go.mod h1:HFK/6AGESC7Ex+EZJhJ2Gni6cTaYpSMmU/cT9RmlfYY=

605modernc.org/goabi0 v0.2.0 h1:HvEowk7LxcPd0eq6mVOAEMai46V+i7Jrj13t4AzuNks=

531modernc.org/goabi0 v0.2.0 h1:HvEowk7LxcPd0eq6mVOAEMai46V+i7Jrj13t4AzuNks=

606modernc.org/goabi0 v0.2.0/go.mod h1:CEFRnnJhKvWT1c1JTI3Avm+tgOWbkOu5oPA8eH8LnMI=

532modernc.org/goabi0 v0.2.0/go.mod h1:CEFRnnJhKvWT1c1JTI3Avm+tgOWbkOu5oPA8eH8LnMI=

607−modernc.org/libc v1.73.4 h1:+ra4Ui8ngyt8HDcO1FTDPWlkAh6yOdaO2yAoh8MddQA=

533+modernc.org/libc v1.67.6 h1:eVOQvpModVLKOdT+LvBPjdQqfrZq+pC39BygcT+E7OI=

608−modernc.org/libc v1.73.4/go.mod h1:DXZ3eO8qMCNn2SnmTNCiC71nJ9Rcq3PsnpU6Vc4rWK8=

534+modernc.org/libc v1.67.6/go.mod h1:JAhxUVlolfYDErnwiqaLvUqc8nfb2r6S6slAgZOnaiE=

609modernc.org/mathutil v1.7.1 h1:GCZVGXdaN8gTqB1Mf/usp1Y/hSqgI2vAGGP4jZMCxOU=

535modernc.org/mathutil v1.7.1 h1:GCZVGXdaN8gTqB1Mf/usp1Y/hSqgI2vAGGP4jZMCxOU=

610modernc.org/mathutil v1.7.1/go.mod h1:4p5IwJITfppl0G4sUEDtCr4DthTaT47/N3aT6MhfgJg=

536modernc.org/mathutil v1.7.1/go.mod h1:4p5IwJITfppl0G4sUEDtCr4DthTaT47/N3aT6MhfgJg=

611modernc.org/memory v1.11.0 h1:o4QC8aMQzmcwCK3t3Ux/ZHmwFPzE6hf2Y5LbkRs+hbI=

537modernc.org/memory v1.11.0 h1:o4QC8aMQzmcwCK3t3Ux/ZHmwFPzE6hf2Y5LbkRs+hbI=

612modernc.org/memory v1.11.0/go.mod h1:/JP4VbVC+K5sU2wZi9bHoq2MAkCnrt2r98UGeSK7Mjw=

538modernc.org/memory v1.11.0/go.mod h1:/JP4VbVC+K5sU2wZi9bHoq2MAkCnrt2r98UGeSK7Mjw=

613−modernc.org/opt v0.2.0 h1:tGyef5ApycA7FSEOMraay9SaTk5zmbx7Tu+cJs4QKZg=

539+modernc.org/opt v0.1.4 h1:2kNGMRiUjrp4LcaPuLY2PzUfqM/w9N23quVwhKt5Qm8=

614−modernc.org/opt v0.2.0/go.mod h1:03fq9lsNfvkYSfxrfUhZCWPk1lm4cq4N+Bh//bEtgns=

540+modernc.org/opt v0.1.4/go.mod h1:03fq9lsNfvkYSfxrfUhZCWPk1lm4cq4N+Bh//bEtgns=

615modernc.org/sortutil v1.2.1 h1:+xyoGf15mM3NMlPDnFqrteY07klSFxLElE2PVuWIJ7w=

541modernc.org/sortutil v1.2.1 h1:+xyoGf15mM3NMlPDnFqrteY07klSFxLElE2PVuWIJ7w=

616modernc.org/sortutil v1.2.1/go.mod h1:7ZI3a3REbai7gzCLcotuw9AC4VZVpYMjDzETGsSMqJE=

542modernc.org/sortutil v1.2.1/go.mod h1:7ZI3a3REbai7gzCLcotuw9AC4VZVpYMjDzETGsSMqJE=

617−modernc.org/sqlite v1.53.0 h1:20WG8N9q4ji/dEqGk4uiI0c6OPjSeLTNYGFCc3+7c1M=

543+modernc.org/sqlite v1.44.3 h1:+39JvV/HWMcYslAwRxHb8067w+2zowvFOUrOWIy9PjY=

618−modernc.org/sqlite v1.53.0/go.mod h1:xoEpOIpGrgT48H5iiyt/YXPCZPEzlfmfFwtk8Lklw8s=

544+modernc.org/sqlite v1.44.3/go.mod h1:CzbrU2lSB1DKUusvwGz7rqEKIq+NUd8GWuBBZDs9/nA=

619modernc.org/strutil v1.2.1 h1:UneZBkQA+DX2Rp35KcM69cSsNES9ly8mQWD71HKlOA0=

545modernc.org/strutil v1.2.1 h1:UneZBkQA+DX2Rp35KcM69cSsNES9ly8mQWD71HKlOA0=

620modernc.org/strutil v1.2.1/go.mod h1:EHkiggD70koQxjVdSBM3JKM7k6L0FbGE5eymy9i3B9A=

546modernc.org/strutil v1.2.1/go.mod h1:EHkiggD70koQxjVdSBM3JKM7k6L0FbGE5eymy9i3B9A=

621modernc.org/token v1.1.0 h1:Xl7Ap9dKaEs5kLoOQeQmPWevfnk/DM5qcLcYlA8ys6Y=

547modernc.org/token v1.1.0 h1:Xl7Ap9dKaEs5kLoOQeQmPWevfnk/DM5qcLcYlA8ys6Y=

622modernc.org/token v1.1.0/go.mod h1:UGzOrNV1mAFSEB63lOFHIpNRUVMvYTc6yu1SMY/XTDM=

548modernc.org/token v1.1.0/go.mod h1:UGzOrNV1mAFSEB63lOFHIpNRUVMvYTc6yu1SMY/XTDM=

623mvdan.cc/sh/moreinterp v0.0.0-20250902163504-3cf4fd5717a5 h1:mO2lyKtGwu4mGQ+Qqjx0+fd5UU5BXhX/rslFmxd5aco=

549mvdan.cc/sh/moreinterp v0.0.0-20250902163504-3cf4fd5717a5 h1:mO2lyKtGwu4mGQ+Qqjx0+fd5UU5BXhX/rslFmxd5aco=

624mvdan.cc/sh/moreinterp v0.0.0-20250902163504-3cf4fd5717a5/go.mod h1:Of9PCedbLDYT8b3EyiYG64rNnx5nOp27OLCVdDrjJyo=

550mvdan.cc/sh/moreinterp v0.0.0-20250902163504-3cf4fd5717a5/go.mod h1:Of9PCedbLDYT8b3EyiYG64rNnx5nOp27OLCVdDrjJyo=

625−mvdan.cc/sh/v3 v3.13.1 h1:DP3TfgZhDkT7lerUdnp6PTGKyxxzz6T+cOlY/xEvfWk=

551+mvdan.cc/sh/v3 v3.12.1-0.20250902163504-3cf4fd5717a5 h1:e7Z/Lgw/zMijvQBVrfh/vUDZ+9FpuSLrJDVGBuoJtuo=

626−mvdan.cc/sh/v3 v3.13.1/go.mod h1:lXJ8SexMvEVcHCoDvAGLZgFJ9Wsm2sulmoNEXGhYZD0=

552+mvdan.cc/sh/v3 v3.12.1-0.20250902163504-3cf4fd5717a5/go.mod h1:P21wo2gLLe3426sP+CmANLBaixSEbRtPl35w3YlM6dg=

internal/agent/.env.sample−1

@@ -1,5 +1,4 @@

1CRUSH_ANTHROPIC_API_KEY=

2−CRUSH_HYPER_API_KEY=

3CRUSH_OPENAI_API_KEY=

2CRUSH_OPENAI_API_KEY=

4CRUSH_OPENROUTER_API_KEY=

3CRUSH_OPENROUTER_API_KEY=

5CRUSH_ZAI_API_KEY=

4CRUSH_ZAI_API_KEY=

internal/agent/accepted_run_test.godeleted−231

@@ -1,231 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "testing"

6−

7− "github.com/charmbracelet/crush/internal/message"

8− "github.com/stretchr/testify/assert"

9− "github.com/stretchr/testify/require"

10−)

11−

12−// newCancelTestAgent builds a DB-backed sessionAgent with no model. The

13−// tests here exercise the dispatch/cancel/persist paths, none of which

14−// reach agent.Stream, so a model is unnecessary.

15−func newCancelTestAgent(t *testing.T) (*sessionAgent, fakeEnv) {

16− t.Helper()

17− env := testEnv(t)

18− sa := NewSessionAgent(SessionAgentOptions{

19− Sessions: env.sessions,

20− Messages: env.messages,

21− }).(*sessionAgent)

22− return sa, env

23−}

24−

25−func (a *sessionAgent) acceptedCount(sessionID string) int {

26− c, _ := a.acceptedRuns.Get(sessionID)

27− return c

28−}

29−

30−func (a *sessionAgent) hasPendingCancel(sessionID string) bool {

31− mark, ok := a.cancelMark.Get(sessionID)

32− return ok && mark > 0

33−}

34−

35−func (a *sessionAgent) pendingCancelMark(sessionID string) uint64 {

36− mark, _ := a.cancelMark.Get(sessionID)

37− return mark

38−}

39−

40−func TestAcceptedRun_CloseIsIdempotent(t *testing.T) {

41− t.Parallel()

42− sa, _ := newCancelTestAgent(t)

43−

44− accept := sa.BeginAccepted("sid")

45− require.Equal(t, "sid", accept.SessionID())

46− require.Equal(t, 1, sa.acceptedCount("sid"))

47−

48− accept.Close()

49− require.Equal(t, 0, sa.acceptedCount("sid"))

50−

51− // Repeated Close must not underflow the counter.

52− accept.Close()

53− accept.Close()

54− require.Equal(t, 0, sa.acceptedCount("sid"))

55−}

56−

57−func TestAcceptedRun_MultipleReservations(t *testing.T) {

58− t.Parallel()

59− sa, _ := newCancelTestAgent(t)

60−

61− a1 := sa.BeginAccepted("sid")

62− a2 := sa.BeginAccepted("sid")

63− require.Equal(t, 2, sa.acceptedCount("sid"))

64−

65− a1.Close()

66− require.Equal(t, 1, sa.acceptedCount("sid"))

67−

68− a2.Close()

69− require.Equal(t, 0, sa.acceptedCount("sid"))

70−}

71−

72−func TestAcceptedRun_NilSafe(t *testing.T) {

73− t.Parallel()

74− var accept *AcceptedRun

75− require.Equal(t, "", accept.SessionID())

76− // Must not panic.

77− accept.Close()

78−}

79−

80−func TestCancel_IdleDoesNotRecordPendingCancel(t *testing.T) {

81− t.Parallel()

82− sa, _ := newCancelTestAgent(t)

83−

84− // No accepted run, no active request: a true no-op.

85− sa.Cancel("sid")

86− require.False(t, sa.hasPendingCancel("sid"))

87−}

88−

89−func TestCancel_AcceptedRecordsPendingCancel(t *testing.T) {

90− t.Parallel()

91− sa, _ := newCancelTestAgent(t)

92−

93− accept := sa.BeginAccepted("sid")

94− defer accept.Close()

95−

96− sa.Cancel("sid")

97− require.True(t, sa.hasPendingCancel("sid"))

98−}

99−

100−func TestCancel_SecondCancelWhilePendingIsNoOp(t *testing.T) {

101− t.Parallel()

102− sa, _ := newCancelTestAgent(t)

103−

104− accept := sa.BeginAccepted("sid")

105− defer accept.Close()

106−

107− sa.Cancel("sid")

108− require.True(t, sa.hasPendingCancel("sid"))

109−

110− // A second cancel while a pending cancel is already recorded must

111− // remain a single pending cancel; one Run consumes exactly one.

112− sa.Cancel("sid")

113− require.True(t, sa.hasPendingCancel("sid"))

114−}

115−

116−func TestRun_CancelOnEntryPersistsCanceledTurn(t *testing.T) {

117− t.Parallel()

118− sa, env := newCancelTestAgent(t)

119−

120− sess, err := env.sessions.Create(t.Context(), "session")

121− require.NoError(t, err)

122−

123− accept := sa.BeginAccepted(sess.ID)

124− // A cancel arrives in the accepted-but-not-yet-active window.

125− sa.Cancel(sess.ID)

126− require.True(t, sa.hasPendingCancel(sess.ID))

127−

128− result, err := sa.Run(t.Context(), SessionAgentCall{

129− SessionID: sess.ID,

130− Prompt: "hello",

131− Accepted: accept,

132− })

133− require.NoError(t, err)

134− require.Nil(t, result)

135−

136− // The pending cancel was consumed and the accept released.

137− require.False(t, sa.hasPendingCancel(sess.ID))

138− require.Equal(t, 0, sa.acceptedCount(sess.ID))

139−

140− msgs, err := env.messages.List(t.Context(), sess.ID)

141− require.NoError(t, err)

142− require.Len(t, msgs, 2)

143− assert.Equal(t, message.User, msgs[0].Role)

144− assert.Equal(t, message.Assistant, msgs[1].Role)

145− assert.Equal(t, message.FinishReasonCanceled, msgs[1].FinishReason())

146−}

147−

148−func TestPersistCanceledTurn_WritesBothWhenUserMissing(t *testing.T) {

149− t.Parallel()

150− sa, env := newCancelTestAgent(t)

151−

152− sess, err := env.sessions.Create(t.Context(), "session")

153− require.NoError(t, err)

154−

155− err = sa.persistCanceledTurn(t.Context(), SessionAgentCall{

156− SessionID: sess.ID,

157− Prompt: "hello",

158− }, false)

159− require.NoError(t, err)

160−

161− msgs, err := env.messages.List(t.Context(), sess.ID)

162− require.NoError(t, err)

163− require.Len(t, msgs, 2)

164− assert.Equal(t, message.User, msgs[0].Role)

165− assert.Equal(t, message.Assistant, msgs[1].Role)

166− assert.Equal(t, message.FinishReasonCanceled, msgs[1].FinishReason())

167−}

168−

169−func TestPersistCanceledTurn_WritesAssistantOnlyWhenUserCreated(t *testing.T) {

170− t.Parallel()

171− sa, env := newCancelTestAgent(t)

172−

173− sess, err := env.sessions.Create(t.Context(), "session")

174− require.NoError(t, err)

175−

176− // Simulate PrepareStep having already created the user message.

177− _, err = sa.createUserMessage(t.Context(), SessionAgentCall{

178− SessionID: sess.ID,

179− Prompt: "hello",

180− })

181− require.NoError(t, err)

182−

183− err = sa.persistCanceledTurn(t.Context(), SessionAgentCall{

184− SessionID: sess.ID,

185− Prompt: "hello",

186− }, true)

187− require.NoError(t, err)

188−

189− msgs, err := env.messages.List(t.Context(), sess.ID)

190− require.NoError(t, err)

191− require.Len(t, msgs, 2)

192− assert.Equal(t, message.User, msgs[0].Role)

193− assert.Equal(t, message.Assistant, msgs[1].Role)

194− assert.Equal(t, message.FinishReasonCanceled, msgs[1].FinishReason())

195−}

196−

197−func TestPersistCanceledTurn_SucceedsWithCanceledContext(t *testing.T) {

198− t.Parallel()

199− sa, env := newCancelTestAgent(t)

200−

201− sess, err := env.sessions.Create(t.Context(), "session")

202− require.NoError(t, err)

203−

204− // Simulate workspace shutdown having already canceled the run

205− // context. WithoutCancel must let the writes through.

206− ctx, cancel := context.WithCancel(t.Context())

207− cancel()

208−

209− err = sa.persistCanceledTurn(ctx, SessionAgentCall{

210− SessionID: sess.ID,

211− Prompt: "hello",

212− }, false)

213− require.NoError(t, err)

214−

215− msgs, err := env.messages.List(t.Context(), sess.ID)

216− require.NoError(t, err)

217− require.Len(t, msgs, 2)

218−}

219−

220−func TestClearPendingCancel(t *testing.T) {

221− t.Parallel()

222− sa, _ := newCancelTestAgent(t)

223−

224− accept := sa.BeginAccepted("sid")

225− defer accept.Close()

226− sa.Cancel("sid")

227− require.True(t, sa.hasPendingCancel("sid"))

228−

229− sa.clearPendingCancel("sid")

230− require.False(t, sa.hasPendingCancel("sid"))

231−}

internal/agent/agent.go+143−1237

@@ -12,46 +12,38 @@ import (

12 "context"

13 _ "embed"

14 "encoding/base64"

15− "encoding/json"

16 "errors"

15 "errors"

17 "fmt"

16 "fmt"

18 "log/slog"

17 "log/slog"

19− "math"

20− "net/http"

21 "os"

18 "os"

22 "regexp"

19 "regexp"

23 "strconv"

20 "strconv"

24 "strings"

21 "strings"

25 "sync"

22 "sync"

26− "sync/atomic"

27 "time"

23 "time"

29− "charm.land/catwalk/pkg/catwalk"

30 "charm.land/fantasy"

25 "charm.land/fantasy"

31 "charm.land/fantasy/providers/anthropic"

26 "charm.land/fantasy/providers/anthropic"

32 "charm.land/fantasy/providers/bedrock"

27 "charm.land/fantasy/providers/bedrock"

33 "charm.land/fantasy/providers/google"

28 "charm.land/fantasy/providers/google"

34 "charm.land/fantasy/providers/openai"

29 "charm.land/fantasy/providers/openai"

35 "charm.land/fantasy/providers/openrouter"

30 "charm.land/fantasy/providers/openrouter"

36− "charm.land/fantasy/providers/vercel"

37 "charm.land/lipgloss/v2"

31 "charm.land/lipgloss/v2"

32+ "github.com/charmbracelet/catwalk/pkg/catwalk"

38 "github.com/charmbracelet/crush/internal/agent/hyper"

33 "github.com/charmbracelet/crush/internal/agent/hyper"

39− "github.com/charmbracelet/crush/internal/agent/notify"

40 "github.com/charmbracelet/crush/internal/agent/tools"

34 "github.com/charmbracelet/crush/internal/agent/tools"

41 "github.com/charmbracelet/crush/internal/agent/tools/mcp"

35 "github.com/charmbracelet/crush/internal/agent/tools/mcp"

42 "github.com/charmbracelet/crush/internal/config"

36 "github.com/charmbracelet/crush/internal/config"

43 "github.com/charmbracelet/crush/internal/csync"

37 "github.com/charmbracelet/crush/internal/csync"

44 "github.com/charmbracelet/crush/internal/message"

38 "github.com/charmbracelet/crush/internal/message"

45− "github.com/charmbracelet/crush/internal/pubsub"

39+ "github.com/charmbracelet/crush/internal/permission"

46 "github.com/charmbracelet/crush/internal/session"

40 "github.com/charmbracelet/crush/internal/session"

47 "github.com/charmbracelet/crush/internal/stringext"

41 "github.com/charmbracelet/crush/internal/stringext"

48− "github.com/charmbracelet/crush/internal/version"

49− "github.com/charmbracelet/x/ansi"

50 "github.com/charmbracelet/x/exp/charmtone"

42 "github.com/charmbracelet/x/exp/charmtone"

51)

43)

53const (

45const (

54− DefaultSessionName = "Untitled Session"

46+ defaultSessionName = "Untitled Session"

56 // Constants for auto-summarization thresholds

48 // Constants for auto-summarization thresholds

57 largeContextWindowThreshold = 200_000

49 largeContextWindowThreshold = 200_000

@@ -59,8 +51,6 @@ const (

59 smallContextWindowRatio = 0.2

51 smallContextWindowRatio = 0.2

60)

52)

62−var userAgent = fmt.Sprintf("Charm-Crush/%s (https://charm.land/crush)", version.Version)

63−

64//go:embed templates/title.md

54//go:embed templates/title.md

65var titlePrompt []byte

55var titlePrompt []byte

@@ -68,22 +58,10 @@ var titlePrompt []byte

68var summaryPrompt []byte

58var summaryPrompt []byte

70// Used to remove <think> tags from generated titles.

60// Used to remove <think> tags from generated titles.

71−var (

72− thinkTagRegex = regexp.MustCompile(`(?s)<think>.*?</think>`)

61+var thinkTagRegex = regexp.MustCompile(`<think>.*?</think>`)

73− orphanThinkTagRegex = regexp.MustCompile(`</?think>`)

74−)

76type SessionAgentCall struct {

63type SessionAgentCall struct {

77− SessionID string

64+ SessionID string

78− // RunID, when non-empty, is the caller-supplied correlator that

79− // gets echoed back on the notify.RunComplete event emitted for

80− // this turn. It is preserved when the call is enqueued behind a

81− // busy session so the queued turn's terminal event is still

82− // recognisable to the original caller. Callers that need a

83− // reliable completion contract (e.g. `crush run` against a

84− // session that may be busy) MUST set it; SessionID alone is

85− // ambiguous when concurrent turns share the same session.

86− RunID string

87 Prompt string

65 Prompt string

88 ProviderOptions fantasy.ProviderOptions

66 ProviderOptions fantasy.ProviderOptions

89 Attachments []message.Attachment

67 Attachments []message.Attachment

@@ -93,42 +71,10 @@ type SessionAgentCall struct {

93 TopK *int64

71 TopK *int64

94 FrequencyPenalty *float64

72 FrequencyPenalty *float64

95 PresencePenalty *float64

73 PresencePenalty *float64

96− NonInteractive bool

97− // OnComplete, when non-nil, replaces the default RunComplete

98− // publish path: the inner Run hands the terminal payload to this

99− // callback instead of emitting it on the RunComplete broker. The

100− // coordinator uses this hook to coalesce the unauthorized →

101− // re-auth → retry chain into a single user-visible terminal

102− // event, so non-interactive clients (e.g. `crush run`) don't

103− // exit on a stale failed-attempt RunComplete before the

104− // successful retry. It is intentionally stripped when queueing

105− // a busy-session call (see Run): the originating

106− // coordinator.Run has long returned by the time the queued

107− // recursion drains, so falling back to the default broker

108− // publish keeps the event visible to subscribers.

109− OnComplete func(notify.RunComplete)

110− // Accepted, when non-nil, is the accept reservation taken by

111− // BeginAccepted before the call was dispatched onto a goroutine

112− // (the client/server fire-and-forget path). Run consumes it under

113− // dispatchMu[SessionID] once the accepted -> (cancel-on-entry |

114− // queued | active) transition has been chosen. When nil

115− // (in-process / local callers like AppWorkspace), behavior is

116− // unchanged and no accept tracking applies.

117− Accepted *AcceptedRun

118− // acceptSeq carries the accept sequence of the handle that produced

119− // this call after it has been enqueued and its Accepted handle

120− // stripped. The queue-drain paths compare it against a session's

121− // cancel mark so a follow-up queued before a cancel is dropped while

122− // one queued after the cancel survives. 0 means untracked (an

123− // in-process enqueue with no accept reservation), which the drain

124− // paths treat as covered by any present mark, preserving the

125− // pre-sequence behavior.

126− acceptSeq uint64

127}

74}

128

129type SessionAgent interface {

76type SessionAgent interface {

130 Run(context.Context, SessionAgentCall) (*fantasy.AgentResult, error)

77 Run(context.Context, SessionAgentCall) (*fantasy.AgentResult, error)

131− BeginAccepted(sessionID string) *AcceptedRun

132 SetModels(large Model, small Model)

78 SetModels(large Model, small Model)

133 SetTools(tools []fantasy.AgentTool)

79 SetTools(tools []fantasy.AgentTool)

134 SetSystemPrompt(systemPrompt string)

80 SetSystemPrompt(systemPrompt string)

@@ -141,14 +87,12 @@ type SessionAgent interface {

141 ClearQueue(sessionID string)

87 ClearQueue(sessionID string)

142 Summarize(context.Context, string, fantasy.ProviderOptions) error

88 Summarize(context.Context, string, fantasy.ProviderOptions) error

143 Model() Model

89 Model() Model

144− GenerateTitle(ctx context.Context, sessionID, userPrompt string)

145}

90}

146

147type Model struct {

92type Model struct {

148 Model fantasy.LanguageModel

93 Model fantasy.LanguageModel

149 CatwalkCfg catwalk.Model

94 CatwalkCfg catwalk.Model

150 ModelCfg config.SelectedModel

95 ModelCfg config.SelectedModel

151− FlatRate bool

152}

96}

153

154type sessionAgent struct {

98type sessionAgent struct {

@@ -157,55 +101,15 @@ type sessionAgent struct {

157 systemPromptPrefix *csync.Value[string]

101 systemPromptPrefix *csync.Value[string]

158 systemPrompt *csync.Value[string]

102 systemPrompt *csync.Value[string]

159 tools *csync.Slice[fantasy.AgentTool]

103 tools *csync.Slice[fantasy.AgentTool]

160− toolsGen atomic.Uint64 // bumped by SetTools; lets PrepareStep skip re-copying unchanged tools

161

104

162 isSubAgent bool

105 isSubAgent bool

163 sessions session.Service

106 sessions session.Service

164 messages message.Service

107 messages message.Service

165 disableAutoSummarize bool

108 disableAutoSummarize bool

166 isYolo bool

109 isYolo bool

167− notify pubsub.Publisher[notify.Notification]

168− runComplete pubsub.Publisher[notify.RunComplete]

169

110

170 messageQueue *csync.Map[string, []SessionAgentCall]

111 messageQueue *csync.Map[string, []SessionAgentCall]

171 activeRequests *csync.Map[string, context.CancelFunc]

112 activeRequests *csync.Map[string, context.CancelFunc]

172−

173− // dispatchMu holds a per-session mutex that serializes the

174− // accepted -> (cancel-on-entry | queued | active) transition in

175− // Run against a concurrent Cancel. The lock is held only during

176− // the brief handoff (no DB or LLM I/O under the lock).

177− dispatchMu *csync.Map[string, *sync.Mutex]

178− // acceptedRuns counts dispatched-but-not-yet-active runs per

179− // session. A counter > 0 means a dispatched prompt is in flight

180− // and has not yet completed the dispatch handoff in Run. Only

181− // BeginAccepted increments it; only AcceptedRun.Close decrements

182− // it.

183− acceptedRuns *csync.Map[string, int]

184− // cancelMark records, per session, a high-water accept sequence: an

185− // accepted handle is canceled by it iff the handle's sequence is at

186− // or below the mark. Cancel raises the mark to the latest sequence

187− // assigned at cancel time, so a single Cancel covers every prompt

188− // accepted-but-not-yet-active then, while a prompt accepted later

189− // (higher sequence) is never poisoned. Absent or 0 means no pending

190− // cancel. It is only raised by Cancel when acceptedRuns > 0, so an

191− // idle Escape never records a mark.

192− cancelMark *csync.Map[string, uint64]

193− // dispatchMuCreate guards lazy creation of per-session entries in

194− // dispatchMu so two goroutines can't race to lock different mutex

195− // instances for the same session.

196− dispatchMuCreate sync.Mutex

197− // acceptedMu serializes increments/decrements of acceptedRuns and

198− // the assignment of accept sequence numbers from acceptSeqGen. It

199− // is separate from dispatchMu so AcceptedRun.Close (which may run

200− // while Run holds dispatchMu for the same session) does not

201− // deadlock by re-entering the dispatch lock.

202− acceptedMu sync.Mutex

203− // acceptSeqGen is the monotonic source of accept sequence numbers.

204− // Each BeginAccepted increments it under acceptedMu and stamps the

205− // returned handle, so sequences strictly increase in accept order

206− // across the agent. Cancel uses its current value as the per-session

207− // high-water mark.

208− acceptSeqGen uint64

209}

113}

210

114

211type SessionAgentOptions struct {

115type SessionAgentOptions struct {

@@ -219,8 +123,6 @@ type SessionAgentOptions struct {

219 Sessions session.Service

123 Sessions session.Service

220 Messages message.Service

124 Messages message.Service

221 Tools []fantasy.AgentTool

125 Tools []fantasy.AgentTool

222− Notify pubsub.Publisher[notify.Notification]

223− RunComplete pubsub.Publisher[notify.RunComplete]

224}

126}

225

127

226func NewSessionAgent(

128func NewSessionAgent(

@@ -237,419 +139,32 @@ func NewSessionAgent(

237 disableAutoSummarize: opts.DisableAutoSummarize,

139 disableAutoSummarize: opts.DisableAutoSummarize,

238 tools: csync.NewSliceFrom(opts.Tools),

140 tools: csync.NewSliceFrom(opts.Tools),

239 isYolo: opts.IsYolo,

141 isYolo: opts.IsYolo,

240− notify: opts.Notify,

241− runComplete: opts.RunComplete,

242 messageQueue: csync.NewMap[string, []SessionAgentCall](),

142 messageQueue: csync.NewMap[string, []SessionAgentCall](),

243 activeRequests: csync.NewMap[string, context.CancelFunc](),

143 activeRequests: csync.NewMap[string, context.CancelFunc](),

244− dispatchMu: csync.NewMap[string, *sync.Mutex](),

245− acceptedRuns: csync.NewMap[string, int](),

246− cancelMark: csync.NewMap[string, uint64](),

247− }

248−}

249−

250−// AcceptedRun owns exactly one accept reservation taken by

251−// BeginAccepted. It is the only carrier of accept-state across the

252−// backend.runAgent / Coordinator.Run / sessionAgent.Run layers: a

253−// counter > 0 means a dispatched prompt is in flight and has not yet

254−// completed the dispatch handoff in Run. Close is the only way to

255−// release the reservation and is idempotent.

256−type AcceptedRun struct {

257− agent *sessionAgent

258− sessionID string

259− // seq is the monotonic accept sequence stamped by BeginAccepted. A

260− // cancel covers this handle iff seq is at or below the session's

261− // cancel mark, so a handle accepted after a cancel (higher seq) is

262− // never poisoned by it.

263− seq uint64

264− done atomic.Bool

265−}

266−

267−// Close decrements the accept counter for this reservation. It is safe

268−// to call multiple times; only the first call has effect.

269−func (r *AcceptedRun) Close() {

270− if r == nil {

271− return

272− }

273− if !r.done.CompareAndSwap(false, true) {

274− return

275− }

276− r.agent.endAccepted(r.sessionID)

277−}

278−

279−// SessionID exposes the session this reservation is for so the run path

280−// can use it without an extra parameter.

281−func (r *AcceptedRun) SessionID() string {

282− if r == nil {

283− return ""

284− }

285− return r.sessionID

286−}

287−

288−// BeginAccepted increments the accept counter for sessionID and returns

289−// a handle whose Close is the only way to decrement it. It is the only

290−// entry point that mutates acceptedRuns.

291−func (a *sessionAgent) BeginAccepted(sessionID string) *AcceptedRun {

292− a.acceptedMu.Lock()

293− defer a.acceptedMu.Unlock()

294− count, _ := a.acceptedRuns.Get(sessionID)

295− a.acceptedRuns.Set(sessionID, count+1)

296− a.acceptSeqGen++

297− return &AcceptedRun{agent: a, sessionID: sessionID, seq: a.acceptSeqGen}

298−}

299−

300−// endAccepted decrements the accept counter for sessionID. It is only

301−// called via AcceptedRun.Close. It uses a dedicated lock (not the

302−// per-session dispatch mutex) so it can run while Run holds dispatchMu

303−// for the same session without deadlocking.

304−//

305−// When the count reaches zero the session's cancel mark is dropped: no

306−// accepted handle remains for it to cover, and any handle accepted later

307−// gets a strictly higher sequence that the mark would not match anyway.

308−// Handles canceled on entry never reach RunComplete, so this is the only

309−// place that clears the mark for an all-canceled batch. Sibling handles

310−// covered by the same mark are serialized on the per-session dispatch

311−// mutex and read the mark before they Close, so this never clears it out

312−// from under a covered handle still waiting to enter Run.

313−func (a *sessionAgent) endAccepted(sessionID string) {

314− a.acceptedMu.Lock()

315− defer a.acceptedMu.Unlock()

316− count, ok := a.acceptedRuns.Get(sessionID)

317− if !ok || count <= 1 {

318− a.acceptedRuns.Del(sessionID)

319− a.cancelMark.Del(sessionID)

320− return

321− }

322− a.acceptedRuns.Set(sessionID, count-1)

323−}

324−

325−// sessionMu returns the per-session dispatch mutex, creating it on first

326−// use. Creation is guarded so concurrent callers always observe the same

327−// mutex instance for a given session.

328−func (a *sessionAgent) sessionMu(sessionID string) *sync.Mutex {

329− if mu, ok := a.dispatchMu.Get(sessionID); ok {

330− return mu

331− }

332− a.dispatchMuCreate.Lock()

333− defer a.dispatchMuCreate.Unlock()

334− if mu, ok := a.dispatchMu.Get(sessionID); ok {

335− return mu

336− }

337− mu := &sync.Mutex{}

338− a.dispatchMu.Set(sessionID, mu)

339− return mu

340−}

341−

342−// enqueueCall appends call to the session's message queue. The

343−// OnComplete hook is stripped: the caller that supplied it (typically

344−// coordinator.Run) has its own retry/coalesce scope that ends when it

345−// returns, so by the time the queue drains nobody is left to consume the

346−// buffered terminal event. The recursive Run falls back to the default

347−// broker publish, which is what existing subscribers expect for queued

348−// turns.

349−func (a *sessionAgent) enqueueCall(call SessionAgentCall) {

350− existing, ok := a.messageQueue.Get(call.SessionID)

351− if !ok {

352− existing = make([]SessionAgentCall, 0, 4)

353− }

354− queued := call

355− if call.Accepted != nil {

356− // Preserve the accept sequence after the handle is stripped so

357− // the queue-drain paths can tell a follow-up queued before a

358− // cancel (covered by the mark) from one queued after it.

359− queued.acceptSeq = call.Accepted.seq

360 }

144 }

361− queued.OnComplete = nil

362− queued.Accepted = nil

363− existing = append(existing, queued)

364− a.messageQueue.Set(call.SessionID, existing)

365}

145}

366

146

367−// drainQueueForStep partitions the session's queued calls for the current

368−// streaming step under the per-session dispatch mutex so the filtering is

369−// atomic against a concurrent Cancel: canceledBySeq requires the caller to

370−// hold that mutex, and evaluating it here (rather than after unlocking)

371−// prevents a cancel recorded between the drain and the check from being

372−// observed inconsistently.

373−//

374−// Calls covered by a pending cancel are dropped; the dropped ones that

375−// carry a RunID are returned in canceledWithRunID so the caller can

376−// publish their terminal cancelled RunComplete (a caller waiting on that

377−// RunID, e.g. `crush run`, would otherwise hang). Uncanceled calls without

378−// a RunID are returned in fold to be folded into the active turn,

379−// preserving the existing follow-up behavior. Uncanceled calls that carry

380−// a RunID are left in the queue so each runs as its own turn via the

381−// recursive run path and publishes its own RunComplete, giving every

382−// RunID-bearing prompt an explicit lifecycle instead of being silently

383−// absorbed into another turn. fold is processed by the caller without the

384−// lock held.

385−func (a *sessionAgent) drainQueueForStep(sessionID string) (fold, canceledWithRunID []SessionAgentCall) {

386− dispatchLock := a.sessionMu(sessionID)

387− dispatchLock.Lock()

388− defer dispatchLock.Unlock()

389− queuedCalls, _ := a.messageQueue.Get(sessionID)

390− var keep []SessionAgentCall

391− for _, queued := range queuedCalls {

392− if a.canceledBySeq(sessionID, queued.acceptSeq) {

393− if queued.RunID != "" {

394− canceledWithRunID = append(canceledWithRunID, queued)

395− }

396− continue

397− }

398− if queued.RunID != "" {

399− keep = append(keep, queued)

400− continue

401− }

402− fold = append(fold, queued)

403− }

404− if len(keep) == 0 {

405− a.messageQueue.Del(sessionID)

406− } else {

407− a.messageQueue.Set(sessionID, keep)

408− }

409− return fold, canceledWithRunID

410−}

411−

412−// publishCanceledQueueDrops emits a terminal cancelled RunComplete for

413−// every dropped queued call that carries a RunID. A queued prompt removed

414−// from the queue without ever running — covered by a pending cancel, or

415−// cleared by Cancel/ClearQueue — would otherwise leave a caller blocked on

416−// that RunID: `crush run` ignores live message events and exits only on a

417−// RunComplete whose RunID matches. Calls without a RunID had no such waiter

418−// and are dropped silently as before. A detached, bounded context keeps the

419−// must-deliver publish alive even when the run context that triggered the

420−// drop is already canceled.

421−func (a *sessionAgent) publishCanceledQueueDrops(drops []SessionAgentCall) {

422− var hasRunID bool

423− for _, d := range drops {

424− if d.RunID != "" {

425− hasRunID = true

426− break

427− }

428− }

429− if !hasRunID {

430− return

431− }

432− ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)

433− defer cancel()

434− for _, d := range drops {

435− if d.RunID == "" {

436− continue

437− }

438− a.publishRunComplete(ctx, d, notify.RunComplete{

439− SessionID: d.SessionID,

440− RunID: d.RunID,

441− Cancelled: true,

442− })

443− }

444−}

445−

446−// clearQueueAndNotify removes all queued prompts for the session and

447−// publishes a terminal cancelled RunComplete for any that carried a RunID,

448−// so callers waiting on those RunIDs (e.g. `crush run`) are not left

449−// hanging when their queued prompt is discarded without running.

450−func (a *sessionAgent) clearQueueAndNotify(sessionID string) {

451− queued, ok := a.messageQueue.Get(sessionID)

452− a.messageQueue.Del(sessionID)

453− if !ok {

454− return

455− }

456− a.publishCanceledQueueDrops(queued)

457−}

458−

459−// clearPendingCancel removes any pending-cancel mark for sessionID. It

460−// takes the per-session dispatch lock so it is ordered against Cancel

461−// and the dispatch handoff.

462−func (a *sessionAgent) clearPendingCancel(sessionID string) {

463− mu := a.sessionMu(sessionID)

464− mu.Lock()

465− defer mu.Unlock()

466− a.cancelMark.Del(sessionID)

467−}

468−

469−// canceledBySeq reports whether an accepted handle or queued call with

470−// the given accept sequence is covered by a pending cancel for the

471−// session. Callers must hold the session's dispatch mutex. A tracked

472−// sequence (seq > 0) is covered only when it is at or below the cancel

473−// high-water mark, so a prompt accepted after the cancel (higher seq) is

474−// never poisoned. An untracked sequence (seq == 0, an in-process enqueue

475−// with no accept reservation) is covered whenever any mark is present,

476−// preserving the pre-sequence behavior. The mark is not consumed: it

477−// stays so every sibling handle it covers observes the same cancel, and

478−// a later handle (higher seq) ignores it regardless.

479−func (a *sessionAgent) canceledBySeq(sessionID string, seq uint64) bool {

480− mark, ok := a.cancelMark.Get(sessionID)

481− if !ok || mark == 0 {

482− return false

483− }

484− return seq == 0 || seq <= mark

485−}

486−

487−// persistCanceledTurn writes the user/assistant records for a turn that

488−// was canceled before (or just as) streaming would have produced them.

489−// It creates the user message only when it was not already created by an

490−// earlier createUserMessage call (userMsgCreated), then writes an

491−// assistant message with FinishReasonCanceled. Both writes use

492−// context.WithoutCancel(ctx) so workspace shutdown (which cancels the run

493−// context) can't drop them.

494−func (a *sessionAgent) persistCanceledTurn(ctx context.Context, call SessionAgentCall, userMsgCreated bool) error {

147+func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (*fantasy.AgentResult, error) {

495− writeCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Second)

496− defer cancel()

497− if !userMsgCreated {

498− if _, err := a.createUserMessage(writeCtx, call); err != nil {

499− return err

500− }

501− }

502− largeModel := a.largeModel.Get()

503− assistant, err := a.messages.Create(writeCtx, call.SessionID, message.CreateMessageParams{

504− Role: message.Assistant,

505− Parts: []message.ContentPart{},

506− Model: largeModel.ModelCfg.Model,

507− Provider: largeModel.ModelCfg.Provider,

508− })

509− if err != nil {

510− return err

511− }

512− assistant.AddFinish(message.FinishReasonCanceled, "User canceled request", "")

513− return a.messages.Update(writeCtx, assistant)

514−}

515−

516−// publishRunComplete emits the authoritative terminal event for a turn.

517−// It honors the per-call OnComplete hook when set (so the coordinator can

518−// coalesce retries) and otherwise falls back to the RunComplete broker.

519−// ctx is used only for the bounded-blocking must-deliver publish; the

520−// terminal payload is supplied by the caller. This is the single emit path

521−// shared by the streaming defer and the cancel-on-entry early return so a

522−// caller waiting on RunComplete (e.g. `crush run` with a RunID) always

523−// observes exactly one terminal event regardless of which Run branch ends

524−// the turn.

525−func (a *sessionAgent) publishRunComplete(ctx context.Context, call SessionAgentCall, complete notify.RunComplete) {

526− if call.OnComplete != nil {

527− call.OnComplete(complete)

528− return

529− }

530− if a.runComplete == nil {

531− return

532− }

533− a.runComplete.PublishMustDeliver(ctx, pubsub.UpdatedEvent, complete)

534−}

535−

536−// ValidateCall performs the cheap structural validation that

537−// sessionAgent.Run requires before a call can be dispatched: a call must

538−// carry either a non-empty prompt or a text attachment, and it must name a

539−// session. It is exported so callers that accept a run before dispatching it

540−// (e.g. backend.SendMessage) can apply the same checks and keep the error

541−// contract consistent.

542−func ValidateCall(call SessionAgentCall) error {

543 if call.Prompt == "" && !message.ContainsTextAttachment(call.Attachments) {

148 if call.Prompt == "" && !message.ContainsTextAttachment(call.Attachments) {

544− return ErrEmptyPrompt

149+ return nil, ErrEmptyPrompt

545 }

150 }

546 if call.SessionID == "" {

151 if call.SessionID == "" {

547− return ErrSessionMissing

152+ return nil, ErrSessionMissing

548 }

153 }

549− return nil

550−}

551

154

552−func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *fantasy.AgentResult, retErr error) {

155+ // Queue the message if busy

553− if err := ValidateCall(call); err != nil {

156+ if a.IsSessionBusy(call.SessionID) {

554− return nil, err

555− }

556−

557− // genCtx/cancel are the run context and its cancel func. For the

558− // accepted (fire-and-forget) dispatch path they are created under

559− // dispatchMu below so a concurrent Cancel can observe the

560− // activeRequests entry before the assistant message exists. For

561− // the in-process path they stay nil here and are created later,

562− // preserving the original ordering.

563− var (

564− genCtx context.Context

565− cancel context.CancelFunc

566− activeRegistered bool

567− userMsgCreated bool

568− )

569−

570− if call.Accepted != nil {

571− // Serialize the accepted -> (cancel-on-entry | queued |

572− // active) transition against a concurrent Cancel. Cancel takes

573− // the same per-session lock, so every cancel observes at least

574− // one of: a cancel mark, an activeRequests entry, or a

575− // messageQueue entry it then clears.

576− mu := a.sessionMu(call.SessionID)

157+ existing, ok := a.messageQueue.Get(call.SessionID)

577− mu.Lock()

158+ if !ok {

578−

159+ existing = []SessionAgentCall{}

579− if a.canceledBySeq(call.SessionID, call.Accepted.seq) {

580− // Cancel-on-entry: a cancel arrived while this run was

581− // dispatched but not yet active, and this handle's accept

582− // sequence is at or below the session's cancel mark. The

583− // mark is left in place so sibling handles it also covers

584− // observe the same cancel; release the accept reservation,

585− // drop the lock, and persist a canceled turn without

586− // entering Stream.

587− //

588− // This path returns before the streaming defer that

589− // publishes RunComplete is installed, so emit the terminal

590− // event explicitly. Without it, a caller waiting on

591− // RunComplete for this RunID (e.g. `crush run`, which

592− // ignores message events and blocks on RunComplete) would

593− // hang on an immediately-canceled accepted run.

594− call.Accepted.Close()

595− mu.Unlock()

596− complete := notify.RunComplete{

597− SessionID: call.SessionID,

598− RunID: call.RunID,

599− Cancelled: true,

600− }

601− if err := a.persistCanceledTurn(ctx, call, false); err != nil {

602− complete.Error = err.Error()

603− a.publishRunComplete(ctx, call, complete)

604− return nil, err

605− }

606− a.publishRunComplete(ctx, call, complete)

607− return nil, nil

608− }

609−

610− if a.IsSessionBusy(call.SessionID) {

611− // Busy: an earlier prompt is active. Queue this call and

612− // release the accept reservation. A Cancel arriving after

613− // this point sees the active entry and clears the queue.

614− a.enqueueCall(call)

615− call.Accepted.Close()

616− mu.Unlock()

617− return nil, nil

618 }

160 }

619−

161+ existing = append(existing, call)

620− // Idle: become the active run. Register the cancel func before

621− // dropping the lock so a Cancel that arrives between here and

622− // assistant creation is not lost.

623− runCtx := context.WithValue(ctx, tools.SessionIDContextKey, call.SessionID)

624− genCtx, cancel = context.WithCancel(runCtx)

625− a.activeRequests.Set(call.SessionID, cancel)

162+ a.messageQueue.Set(call.SessionID, existing)

626− activeRegistered = true

627− call.Accepted.Close()

628− mu.Unlock()

629−

630− defer cancel()

631− defer a.activeRequests.Del(call.SessionID)

632− } else if a.IsSessionBusy(call.SessionID) {

633− // Queue the message if busy. Strip OnComplete: the caller that

634− // supplied the hook (typically coordinator.Run) has its own

635− // retry/coalesce scope that ends when it returns, so by the time

636− // the queue drains nobody is left to consume the buffered

637− // terminal event. The recursive Run will fall back to the

638− // default broker publish, which is what existing subscribers

639− // expect for queued turns.

640− a.enqueueCall(call)

641 return nil, nil

163 return nil, nil

642 }

164 }

643

165

644 // Copy mutable fields under lock to avoid races with SetTools/SetModels.

166 // Copy mutable fields under lock to avoid races with SetTools/SetModels.

645 agentTools := a.tools.Copy()

167 agentTools := a.tools.Copy()

646− // Per-step tool snapshot reused across PrepareStep calls; only

647− // re-copied when SetTools bumps toolsGen (see PrepareStep below).

648− var (

649− stepTools []fantasy.AgentTool

650− stepToolsGen uint64

651− stepToolsInit bool

652− )

653 largeModel := a.largeModel.Get()

168 largeModel := a.largeModel.Get()

654 systemPrompt := a.systemPrompt.Get()

169 systemPrompt := a.systemPrompt.Get()

655 promptPrefix := a.systemPromptPrefix.Get()

170 promptPrefix := a.systemPromptPrefix.Get()

@@ -678,7 +193,6 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

678 largeModel.Model,

193 largeModel.Model,

679 fantasy.WithSystemPrompt(systemPrompt),

194 fantasy.WithSystemPrompt(systemPrompt),

680 fantasy.WithTools(agentTools...),

195 fantasy.WithTools(agentTools...),

681− fantasy.WithUserAgent(userAgent),

682 )

196 )

683

197

684 sessionLock := sync.Mutex{}

198 sessionLock := sync.Mutex{}

@@ -693,11 +207,11 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

693 }

207 }

694

208

695 var wg sync.WaitGroup

209 var wg sync.WaitGroup

696− // Generate title from the first real (non-shell) user prompt.

210+ // Generate title if first message.

697− if !hasUserTextMessage(msgs) {

211+ if len(msgs) == 0 {

698 titleCtx := ctx // Copy to avoid race with ctx reassignment below.

212 titleCtx := ctx // Copy to avoid race with ctx reassignment below.

699 wg.Go(func() {

213 wg.Go(func() {

700− a.GenerateTitle(titleCtx, call.SessionID, call.Prompt)

214+ a.generateTitle(titleCtx, call.SessionID, call.Prompt)

701 })

215 })

702 }

216 }

703 defer wg.Wait()

217 defer wg.Wait()

@@ -707,101 +221,29 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

707 if err != nil {

221 if err != nil {

708 return nil, err

222 return nil, err

709 }

223 }

710− userMsgCreated = true

711

224

712 // Add the session to the context.

225 // Add the session to the context.

713 ctx = context.WithValue(ctx, tools.SessionIDContextKey, call.SessionID)

226 ctx = context.WithValue(ctx, tools.SessionIDContextKey, call.SessionID)

714

227

715− // For the accepted dispatch path the run context and cancel func

716− // were already created and registered under dispatchMu above; reuse

717− // them. For the in-process path create them here, preserving the

718− // original ordering.

719− if !activeRegistered {

720− genCtx, cancel = context.WithCancel(ctx)

228+ genCtx, cancel := context.WithCancel(ctx)

721− a.activeRequests.Set(call.SessionID, cancel)

229+ a.activeRequests.Set(call.SessionID, cancel)

722

230

723− defer cancel()

231+ defer cancel()

724− defer a.activeRequests.Del(call.SessionID)

232+ defer a.activeRequests.Del(call.SessionID)

725− }

233+

726− // skipRunComplete is set just before the queued-recursion path so

727− // the outer Run doesn't publish a RunComplete that would race

728− // with — and be superseded by — the recursive call's own

729− // RunComplete (each queued user prompt is its own turn and

730− // publishes exactly one terminal event).

731− var skipRunComplete bool

732− // currentAssistant is declared here so the deferred RunComplete

733− // publish below can capture the pointer that PrepareStep will

734− // later (re)assign for each streaming step. The final assistant

735− // message of the turn is the value reachable through this

736− // pointer when the defer runs.

737− var currentAssistant *message.Message

738− // Drain any debounced message updates before returning. message.Service

739− // already flushes synchronously on terminal updates, but a defer here

740− // guarantees the contract at every Run exit (success, error, panic

741− // recovery upstream) without callers needing to know.

742− //

743− // After the flush completes — meaning all per-message

744− // Publish(UpdatedEvent) calls have fired and been buffered into

745− // every subscriber's channel — publish the authoritative

746− // RunComplete event for this turn. The flush-then-publish order

747− // gives well-behaved clients the best chance of seeing the final

748− // message event before RunComplete; the embedded Text field

749− // reconciles for clients that observe the events out of order

750− // (the pubsub broker fan-in does not serialize publishes from

751− // different upstream brokers).

752− defer func() {

753− // Use a context detached from the run context: workspace

754− // shutdown cancels ctx before this goroutine returns, but the

755− // buffered streaming deltas must still land before the DB is

756− // closed. A short timeout bounds the flush.

757− flushCtx, flushCancel := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Second)

758− defer flushCancel()

759− if flushErr := a.messages.FlushAll(flushCtx); flushErr != nil {

760− slog.Error("Failed to flush pending message updates after run", "error", flushErr)

761− }

762− if skipRunComplete {

763− return

764− }

765− complete := notify.RunComplete{SessionID: call.SessionID, RunID: call.RunID}

766− if currentAssistant != nil {

767− complete.MessageID = currentAssistant.ID

768− complete.Text = currentAssistant.Content().String()

769− }

770− if retErr != nil {

771− complete.Error = retErr.Error()

772− complete.Cancelled = errors.Is(retErr, context.Canceled)

773− } else if ctx.Err() != nil {

774− complete.Cancelled = true

775− }

776− // Prefer the per-call hook when supplied so the coordinator

777− // can coalesce retries (e.g. unauthorized → re-auth → retry)

778− // into a single user-visible terminal event. The fallback

779− // must-deliver publish applies bounded-blocking semantics to

780− // the authoritative terminal event so a momentarily-full

781− // subscriber channel can't silently drop it and hang

782− // non-interactive clients waiting on RunComplete.

783− a.publishRunComplete(ctx, call, complete)

784− }()

785−

786− history, files := a.preparePrompt(msgs, largeModel.CatwalkCfg.SupportsImages, call.Attachments...)

234+ history, files := a.preparePrompt(msgs, call.Attachments...)

787

235

788 startTime := time.Now()

236 startTime := time.Now()

789 a.eventPromptSent(call.SessionID)

237 a.eventPromptSent(call.SessionID)

790

238

791− var stepMessages []fantasy.Message

239+ var currentAssistant *message.Message

792 var shouldSummarize bool

240 var shouldSummarize bool

793− sanitizedToolCalls := make(map[string]bool)

794− // Don't send MaxOutputTokens if 0 — some providers (e.g. LM Studio) reject it

795− var maxOutputTokens *int64

796− if call.MaxOutputTokens > 0 {

797− maxOutputTokens = &call.MaxOutputTokens

798− }

799− result, err = agent.Stream(genCtx, fantasy.AgentStreamCall{

241+ result, err := agent.Stream(genCtx, fantasy.AgentStreamCall{

800 Prompt: message.PromptWithTextAttachments(call.Prompt, call.Attachments),

242 Prompt: message.PromptWithTextAttachments(call.Prompt, call.Attachments),

801 Files: files,

243 Files: files,

802 Messages: history,

244 Messages: history,

803 ProviderOptions: call.ProviderOptions,

245 ProviderOptions: call.ProviderOptions,

804− MaxOutputTokens: maxOutputTokens,

246+ MaxOutputTokens: &call.MaxOutputTokens,

805 TopP: call.TopP,

247 TopP: call.TopP,

806 Temperature: call.Temperature,

248 Temperature: call.Temperature,

807 PresencePenalty: call.PresencePenalty,

249 PresencePenalty: call.PresencePenalty,

@@ -813,30 +255,9 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

813 prepared.Messages[i].ProviderOptions = nil

255 prepared.Messages[i].ProviderOptions = nil

814 }

256 }

815

257

816− // Use latest tools (updated by SetTools when MCP tools change).

258+ queuedCalls, _ := a.messageQueue.Get(call.SessionID)

817− // Reuse the previous step's copy unless SetTools has bumped the

259+ a.messageQueue.Del(call.SessionID)

818− // generation, so an unchanged tool set isn't re-copied every step.

819− if gen := a.toolsGen.Load(); !stepToolsInit || gen != stepToolsGen {

820− stepTools = a.tools.Copy()

821− stepToolsGen = gen

822− stepToolsInit = true

823− }

824− prepared.Tools = stepTools

825−

826− // Drain queued follow-up prompts for this step. Calls covered

827− // by a cancel recorded while they sat in the queue are dropped:

828− // a cancel that arrived after a prompt was queued must not let

829− // it run as part of this step. Coverage is per-call by accept

830− // sequence so a follow-up queued after the cancel (higher seq)

831− // is not dropped. A dropped prompt carrying a RunID still gets

832− // its terminal cancelled RunComplete so a caller waiting on it

833− // does not hang. Uncanceled prompts without a RunID are folded

834− // into this turn; uncanceled prompts with a RunID are left

835− // queued so each runs as its own turn (with its own

836− // RunComplete) via the recursive run path below.

837− fold, canceledRunIDs := a.drainQueueForStep(call.SessionID)

838− a.publishCanceledQueueDrops(canceledRunIDs)

839− for _, queued := range fold {

260+ for _, queued := range queuedCalls {

840 userMessage, createErr := a.createUserMessage(callContext, queued)

261 userMessage, createErr := a.createUserMessage(callContext, queued)

841 if createErr != nil {

262 if createErr != nil {

842 return callContext, prepared, createErr

263 return callContext, prepared, createErr

@@ -866,13 +287,6 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

866 prepared.Messages = append([]fantasy.Message{fantasy.NewSystemMessage(promptPrefix)}, prepared.Messages...)

287 prepared.Messages = append([]fantasy.Message{fantasy.NewSystemMessage(promptPrefix)}, prepared.Messages...)

867 }

288 }

868

289

869− // Save a reference to the prepared messages for token

870− // estimation in OnStepFinish. No clone needed here — the

871− // messages are not modified between PrepareStep and

872− // OnStepFinish (which run sequentially), and the clone is

873− // only needed when the provider returns zero usage.

874− stepMessages = prepared.Messages

875−

876 var assistantMsg message.Message

290 var assistantMsg message.Message

877 assistantMsg, err = a.messages.Create(callContext, call.SessionID, message.CreateMessageParams{

291 assistantMsg, err = a.messages.Create(callContext, call.SessionID, message.CreateMessageParams{

878 Role: message.Assistant,

292 Role: message.Assistant,

@@ -936,39 +350,25 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

936 Finished: false,

350 Finished: false,

937 }

351 }

938 currentAssistant.AddToolCall(toolCall)

352 currentAssistant.AddToolCall(toolCall)

939− // Use parent ctx instead of genCtx to ensure the update succeeds

940− // even if the request is canceled mid-stream

941− return a.messages.Update(ctx, *currentAssistant)

353+ return a.messages.Update(genCtx, *currentAssistant)

942 },

354 },

943 OnRetry: func(err *fantasy.ProviderError, delay time.Duration) {

355 OnRetry: func(err *fantasy.ProviderError, delay time.Duration) {

944− slog.Warn("Provider request failed, retrying", providerRetryLogFields(err, delay)...)

356+ // TODO: implement

945 },

357 },

946 OnToolCall: func(tc fantasy.ToolCallContent) error {

358 OnToolCall: func(tc fantasy.ToolCallContent) error {

947− input, wasSanitized := sanitizeToolInput(tc.ToolName, tc.ToolCallID, tc.Input)

948− if wasSanitized {

949− sanitizedToolCalls[tc.ToolCallID] = true

950− }

951 toolCall := message.ToolCall{

359 toolCall := message.ToolCall{

952 ID: tc.ToolCallID,

360 ID: tc.ToolCallID,

953 Name: tc.ToolName,

361 Name: tc.ToolName,

954− Input: input,

362+ Input: tc.Input,

955 ProviderExecuted: false,

363 ProviderExecuted: false,

956 Finished: true,

364 Finished: true,

957 }

365 }

958 currentAssistant.AddToolCall(toolCall)

366 currentAssistant.AddToolCall(toolCall)

959− // Use parent ctx instead of genCtx to ensure the update succeeds

960− // even if the request is canceled mid-stream

961− return a.messages.Update(ctx, *currentAssistant)

367+ return a.messages.Update(genCtx, *currentAssistant)

962 },

368 },

963 OnToolResult: func(result fantasy.ToolResultContent) error {

369 OnToolResult: func(result fantasy.ToolResultContent) error {

964 toolResult := a.convertToToolResult(result)

370 toolResult := a.convertToToolResult(result)

965− if sanitizedToolCalls[result.ToolCallID] {

966− toolResult.Content = "Tool call failed: arguments were not valid JSON. Please check your tool call format and try again."

967− toolResult.IsError = true

968− }

969− // Use parent ctx instead of genCtx to ensure the message is created

970− // even if the request is canceled mid-stream

971− _, createMsgErr := a.messages.Create(ctx, currentAssistant.SessionID, message.CreateMessageParams{

371+ _, createMsgErr := a.messages.Create(genCtx, currentAssistant.SessionID, message.CreateMessageParams{

972 Role: message.Tool,

372 Role: message.Tool,

973 Parts: []message.ContentPart{

373 Parts: []message.ContentPart{

974 toolResult,

374 toolResult,

@@ -977,9 +377,6 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

977 return createMsgErr

377 return createMsgErr

978 },

378 },

979 OnStepFinish: func(stepResult fantasy.StepResult) error {

379 OnStepFinish: func(stepResult fantasy.StepResult) error {

980− for _, w := range stepResult.Warnings {

981− slog.Warn("Provider warning", "type", w.Type, "message", w.Message)

982− }

983 finishReason := message.FinishReasonUnknown

380 finishReason := message.FinishReasonUnknown

984 switch stepResult.FinishReason {

381 switch stepResult.FinishReason {

985 case fantasy.FinishReasonLength:

382 case fantasy.FinishReasonLength:

@@ -989,18 +386,6 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

989 case fantasy.FinishReasonToolCalls:

386 case fantasy.FinishReasonToolCalls:

990 finishReason = message.FinishReasonToolUse

387 finishReason = message.FinishReasonToolUse

991 }

388 }

992− // If a tool result halted the turn (e.g. a hook halt or a

993− // permission denial), the step ends on FinishReasonToolCalls but

994− // the model will not be called again. Treat it as the end of the

995− // turn so the UI can render the assistant footer.

996− if finishReason == message.FinishReasonToolUse {

997− for _, tr := range stepResult.Content.ToolResults() {

998− if tr.StopTurn {

999− finishReason = message.FinishReasonEndTurn

1000− break

1001− }

1002− }

1003− }

1004 currentAssistant.AddFinish(finishReason, "", "")

389 currentAssistant.AddFinish(finishReason, "", "")

1005 sessionLock.Lock()

390 sessionLock.Lock()

1006 defer sessionLock.Unlock()

391 defer sessionLock.Unlock()

@@ -1009,9 +394,7 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1009 if getSessionErr != nil {

394 if getSessionErr != nil {

1010 return getSessionErr

395 return getSessionErr

1011 }

396 }

1012− usage, estimated := fallbackStepUsage(stepMessages, stepResult)

1013− a.updateSessionUsage(largeModel, &updatedSession, usage, a.openrouterCost(stepResult.ProviderMetadata), estimated)

397+ a.updateSessionUsage(largeModel, &updatedSession, stepResult.Usage, a.openrouterCost(stepResult.ProviderMetadata))

1014− extractHyperCredits(stepResult.ProviderMetadata)

1015 _, sessionErr := a.sessions.Save(ctx, updatedSession)

398 _, sessionErr := a.sessions.Save(ctx, updatedSession)

1016 if sessionErr != nil {

399 if sessionErr != nil {

1017 return sessionErr

400 return sessionErr

@@ -1022,11 +405,6 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1022 StopWhen: []fantasy.StopCondition{

405 StopWhen: []fantasy.StopCondition{

1023 func(_ []fantasy.StepResult) bool {

406 func(_ []fantasy.StepResult) bool {

1024 cw := int64(largeModel.CatwalkCfg.ContextWindow)

407 cw := int64(largeModel.CatwalkCfg.ContextWindow)

1025− // If context window is unknown (0), skip auto-summarize

1026− // to avoid immediately truncating custom/local models.

1027− if cw == 0 {

1028− return false

1029− }

1030 tokens := currentSession.CompletionTokens + currentSession.PromptTokens

408 tokens := currentSession.CompletionTokens + currentSession.PromptTokens

1031 remaining := cw - tokens

409 remaining := cw - tokens

1032 var threshold int64

410 var threshold int64

@@ -1041,48 +419,22 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1041 }

419 }

1042 return false

420 return false

1043 },

421 },

1044− func(steps []fantasy.StepResult) bool {

1045− return hasRepeatedToolCalls(steps, loopDetectionWindowSize, loopDetectionMaxRepeats)

1046− },

1047 },

422 },

1048 })

423 })

1049

424

1050 a.eventPromptResponded(call.SessionID, time.Since(startTime).Truncate(time.Second))

425 a.eventPromptResponded(call.SessionID, time.Since(startTime).Truncate(time.Second))

1051

426

1052 if err != nil {

427 if err != nil {

1053− isHyper := largeModel.ModelCfg.Provider == hyper.Name

1054 isCancelErr := errors.Is(err, context.Canceled)

428 isCancelErr := errors.Is(err, context.Canceled)

429+ isPermissionErr := errors.Is(err, permission.ErrorPermissionDenied)

1055 if currentAssistant == nil {

430 if currentAssistant == nil {

1056− // Cancel-before-assistant-creation window: the run was

1057− // canceled after activeRequests.Set but before PrepareStep

1058− // created the assistant message. Without this, the turn

1059− // would return with no FinishReasonCanceled marker and no

1060− // user-visible record. The user message was already created

1061− // above, so persistCanceledTurn only writes the assistant

1062− // record.

1063− if isCancelErr {

1064− if persistErr := a.persistCanceledTurn(ctx, call, userMsgCreated); persistErr != nil {

1065− return nil, persistErr

1066− }

1067− }

1068 return result, err

431 return result, err

1069 }

432 }

1070− // Persist final state with a context detached from the run

1071− // context. The run context (ctx) is derived from the

1072− // workspace context, which workspace shutdown cancels before

1073− // agent goroutines finish; using ctx here would drop the

1074− // final assistant state. WithoutCancel keeps the values

1075− // (e.g. session ID) while ignoring cancellation, and a short

1076− // timeout bounds the cleanup writes.

1077− cleanupCtx, cleanupCancel := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Second)

1078− defer cleanupCancel()

1079 // Ensure we finish thinking on error to close the reasoning state.

433 // Ensure we finish thinking on error to close the reasoning state.

1080 currentAssistant.FinishThinking()

434 currentAssistant.FinishThinking()

1081 toolCalls := currentAssistant.ToolCalls()

435 toolCalls := currentAssistant.ToolCalls()

1082− // INFO: we use the cleanup context here because the genCtx has been cancelled.

436+ // INFO: we use the parent context here because the genCtx has been cancelled.

1083− // Only load messages from the current assistant message onward — tool

1084− // results for this turn's tool calls can only exist after it.

1085− msgs, createErr := a.messages.ListAfter(cleanupCtx, currentAssistant.SessionID, currentAssistant.ID)

437+ msgs, createErr := a.messages.List(ctx, currentAssistant.SessionID)

1086 if createErr != nil {

438 if createErr != nil {

1087 return nil, createErr

439 return nil, createErr

1088 }

440 }

@@ -1091,7 +443,7 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1091 tc.Finished = true

443 tc.Finished = true

1092 tc.Input = "{}"

444 tc.Input = "{}"

1093 currentAssistant.AddToolCall(tc)

445 currentAssistant.AddToolCall(tc)

1094− updateErr := a.messages.Update(cleanupCtx, *currentAssistant)

446+ updateErr := a.messages.Update(ctx, *currentAssistant)

1095 if updateErr != nil {

447 if updateErr != nil {

1096 return nil, updateErr

448 return nil, updateErr

1097 }

449 }

@@ -1116,7 +468,9 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1116 }

468 }

1117 content := "There was an error while executing the tool"

469 content := "There was an error while executing the tool"

1118 if isCancelErr {

470 if isCancelErr {

1119− content = "Error: user cancelled assistant tool calling"

471+ content = "Tool execution canceled by user"

472+ } else if isPermissionErr {

473+ content = "User denied permission"

1120 }

474 }

1121 toolResult := message.ToolResult{

475 toolResult := message.ToolResult{

1122 ToolCallID: tc.ID,

476 ToolCallID: tc.ID,

@@ -1124,7 +478,7 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1124 Content: content,

478 Content: content,

1125 IsError: true,

479 IsError: true,

1126 }

480 }

1127− _, createErr = a.messages.Create(cleanupCtx, currentAssistant.SessionID, message.CreateMessageParams{

481+ _, createErr = a.messages.Create(ctx, currentAssistant.SessionID, message.CreateMessageParams{

1128 Role: message.Tool,

482 Role: message.Tool,

1129 Parts: []message.ContentPart{

483 Parts: []message.ContentPart{

1130 toolResult,

484 toolResult,

@@ -1140,9 +494,9 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1140 linkStyle := lipgloss.NewStyle().Foreground(charmtone.Guac).Underline(true)

494 linkStyle := lipgloss.NewStyle().Foreground(charmtone.Guac).Underline(true)

1141 if isCancelErr {

495 if isCancelErr {

1142 currentAssistant.AddFinish(message.FinishReasonCanceled, "User canceled request", "")

496 currentAssistant.AddFinish(message.FinishReasonCanceled, "User canceled request", "")

1143− } else if isHyper && errors.As(err, &providerErr) && providerErr.StatusCode == http.StatusUnauthorized {

497+ } else if isPermissionErr {

1144− currentAssistant.AddFinish(message.FinishReasonError, "Unauthorized", `Please re-authenticate with Hyper. You can also run "crush auth" to re-authenticate.`)

498+ currentAssistant.AddFinish(message.FinishReasonPermissionDenied, "User denied permission", "")

1145− } else if isHyper && errors.As(err, &providerErr) && providerErr.StatusCode == http.StatusPaymentRequired {

499+ } else if errors.Is(err, hyper.ErrNoCredits) {

1146 url := hyper.BaseURL()

500 url := hyper.BaseURL()

1147 link := linkStyle.Hyperlink(url, "id=hyper").Render(url)

501 link := linkStyle.Hyperlink(url, "id=hyper").Render(url)

1148 currentAssistant.AddFinish(message.FinishReasonError, "No credits", "You're out of credits. Add more at "+link)

502 currentAssistant.AddFinish(message.FinishReasonError, "No credits", "You're out of credits. Add more at "+link)

@@ -1163,9 +517,9 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1163 } else {

517 } else {

1164 currentAssistant.AddFinish(message.FinishReasonError, defaultTitle, err.Error())

518 currentAssistant.AddFinish(message.FinishReasonError, defaultTitle, err.Error())

1165 }

519 }

1166− // Note: we use the cleanup context here because the genCtx has been

520+ // Note: we use the parent context here because the genCtx has been

1167 // cancelled.

521 // cancelled.

1168− updateErr := a.messages.Update(cleanupCtx, *currentAssistant)

522+ updateErr := a.messages.Update(ctx, *currentAssistant)

1169 if updateErr != nil {

523 if updateErr != nil {

1170 return nil, updateErr

524 return nil, updateErr

1171 }

525 }

@@ -1189,123 +543,17 @@ func (a *sessionAgent) Run(ctx context.Context, call SessionAgentCall) (result *

1189 }

543 }

1190 }

544 }

1191

545

1192− // Release active request before publishing the notification.

546+ // Release active request before processing queued messages.

1193− // TUI handlers poll IsSessionBusy() and only re-evaluate when a

1194− // tea.Msg arrives, so the cleanup must precede the notify or

1195− // subscribers see stale busy state at the moment of receipt.

1196 a.activeRequests.Del(call.SessionID)

547 a.activeRequests.Del(call.SessionID)

1197 cancel()

548 cancel()

1198

549

1199− // Send notification that agent has finished its turn (skip for

550+ queuedMessages, ok := a.messageQueue.Get(call.SessionID)

1200− // nested/non-interactive sessions).

1201− if !call.NonInteractive && a.notify != nil {

1202− a.notify.Publish(pubsub.CreatedEvent, notify.Notification{

1203− SessionID: call.SessionID,

1204− SessionTitle: currentSession.Title,

1205− Type: notify.TypeAgentFinished,

1206− })

1207− }

1208−

1209− // Hand off to the next queued prompt (if any) under dispatchMu so

1210− // the transition from this finished run to the queued run is atomic

1211− // against a concurrent Cancel. activeRequests for this session was

1212− // just deleted above, so without the lock there is a window in

1213− // which the session looks idle and a cancel becomes a no-op that

1214− // fails to stop the queued prompt. Holding the lock lets us observe

1215− // a pending cancel recorded against the session and drop the queue

1216− // instead of running it, and (for the recursion) hand a fresh

1217− // accept reservation to the dequeued call so acceptedRuns stays > 0

1218− // across the recursive Run's own dispatch handoff — keeping the

1219− // session observable to Cancel for the entire transition and

1220− // closing the dequeue -> re-register window.

1221− mu := a.sessionMu(call.SessionID)

1222− mu.Lock()

1223− queuedMessages, _ := a.messageQueue.Get(call.SessionID)

1224− if mark, ok := a.cancelMark.Get(call.SessionID); ok && mark > 0 && len(queuedMessages) > 0 {

1225− // A cancel was recorded for this session (e.g. it arrived while

1226− // this run was active and follow-ups had been queued). Drop the

1227− // queued prompts it covers (accept sequence at or below the

1228− // mark, or untracked); keep any queued after the cancel (higher

1229− // sequence) so they still run.

1230− var kept []SessionAgentCall

1231− var canceledRunIDDrops []SessionAgentCall

1232− for _, q := range queuedMessages {

1233− if q.acceptSeq == 0 || q.acceptSeq <= mark {

1234− if q.RunID != "" {

1235− canceledRunIDDrops = append(canceledRunIDDrops, q)

1236− }

1237− continue

1238− }

1239− kept = append(kept, q)

1240− }

1241− queuedMessages = kept

1242− a.messageQueue.Set(call.SessionID, kept)

1243− // A dropped prompt carrying a RunID must still publish its

1244− // terminal cancelled RunComplete so a caller waiting on that

1245− // RunID does not hang.

1246− a.publishCanceledQueueDrops(canceledRunIDDrops)

1247− }

1248− if len(queuedMessages) == 0 {

551+ if !ok || len(queuedMessages) == 0 {

1249− // No queued work. Clear the cancel mark only when no accepted

1250− // run remains in flight that it might still cover; otherwise a

1251− // sibling prompt (sequence at or below the mark) waiting to

1252− // enter Run would lose its cancellation. When accepted runs are

1253− // gone, this also clears a stale mark so it can't catch a

1254− // future run.

1255− a.messageQueue.Del(call.SessionID)

1256− a.acceptedMu.Lock()

1257− inFlight, _ := a.acceptedRuns.Get(call.SessionID)

1258− a.acceptedMu.Unlock()

1259− if inFlight == 0 {

1260− a.cancelMark.Del(call.SessionID)

1261− }

1262− mu.Unlock()

1263 return result, err

552 return result, err

1264 }

553 }

1265− // There are queued messages, restart the loop. Suppress the outer

554+ // There are queued messages restart the loop.

1266− // defer's emit: it would otherwise observe the recursive Run's retErr

1267− // (named-return clobbering through the return below) against this

1268− // turn's MessageID/Text and publish a mixed, racing event.

1269− skipRunComplete = true

1270− // Decide whether this turn still owes its own terminal RunComplete.

1271− // Each submitted prompt with a RunID has its own lifecycle, so a turn

1272− // that is finished and handing off to a *different* queued prompt must

1273− // publish its own RunComplete here — leaving it to the recursive turn

1274− // (which carries a different RunID) would hang a caller waiting on

1275− // this turn's RunID. The exception is the summarize-continuation path,

1276− // which re-queues this same call (same RunID) to resume after a

1277− // summary; in that case the eventual terminal turn for this RunID

1278− // publishes, so publishing now would double-emit.

1279− outerOwesRunComplete := call.RunID != ""

1280− if outerOwesRunComplete {

1281− for _, q := range queuedMessages {

1282− if q.RunID == call.RunID {

1283− outerOwesRunComplete = false

1284− break

1285− }

1286− }

1287− }

1288 firstQueuedMessage := queuedMessages[0]

555 firstQueuedMessage := queuedMessages[0]

1289 a.messageQueue.Set(call.SessionID, queuedMessages[1:])

556 a.messageQueue.Set(call.SessionID, queuedMessages[1:])

1290− // Reserve a fresh accept for the dequeued prompt before dropping the

1291− // lock so acceptedRuns > 0 across the handoff into the recursive

1292− // Run. This closes the window between this dequeue and the recursive

1293− // Run registering its activeRequests entry: a cancel arriving in

1294− // that window now records a pending cancel (acceptedRuns > 0) that

1295− // the recursive Run's accepted path observes as cancel-on-entry.

1296− firstQueuedMessage.Accepted = a.BeginAccepted(call.SessionID)

1297− mu.Unlock()

1298− if outerOwesRunComplete {

1299− complete := notify.RunComplete{SessionID: call.SessionID, RunID: call.RunID}

1300− if currentAssistant != nil {

1301− complete.MessageID = currentAssistant.ID

1302− complete.Text = currentAssistant.Content().String()

1303− }

1304− if ctx.Err() != nil {

1305− complete.Cancelled = true

1306− }

1307− a.publishRunComplete(ctx, call, complete)

1308− }

1309 return a.Run(ctx, firstQueuedMessage)

557 return a.Run(ctx, firstQueuedMessage)

1310}

558}

1311

559

@@ -1331,27 +579,20 @@ func (a *sessionAgent) Summarize(ctx context.Context, sessionID string, opts fan

1331 return nil

579 return nil

1332 }

580 }

1333

581

1334− aiMsgs, _ := a.preparePrompt(msgs, largeModel.CatwalkCfg.SupportsImages)

582+ aiMsgs, _ := a.preparePrompt(msgs)

1335

583

1336 genCtx, cancel := context.WithCancel(ctx)

584 genCtx, cancel := context.WithCancel(ctx)

1337 a.activeRequests.Set(sessionID, cancel)

585 a.activeRequests.Set(sessionID, cancel)

1338 defer a.activeRequests.Del(sessionID)

586 defer a.activeRequests.Del(sessionID)

1339 defer cancel()

587 defer cancel()

1340− defer func() {

1341− if flushErr := a.messages.FlushAll(ctx); flushErr != nil {

1342− slog.Error("Failed to flush pending message updates after summarize", "error", flushErr)

1343− }

1344− }()

1345

588

1346− agent := fantasy.NewAgent(

589+ agent := fantasy.NewAgent(largeModel.Model,

1347− largeModel.Model,

1348 fantasy.WithSystemPrompt(string(summaryPrompt)),

590 fantasy.WithSystemPrompt(string(summaryPrompt)),

1349− fantasy.WithUserAgent(userAgent),

1350 )

591 )

1351 summaryMessage, err := a.messages.Create(ctx, sessionID, message.CreateMessageParams{

592 summaryMessage, err := a.messages.Create(ctx, sessionID, message.CreateMessageParams{

1352 Role: message.Assistant,

593 Role: message.Assistant,

1353− Model: largeModel.ModelCfg.Model,

594+ Model: largeModel.Model.Model(),

1354− Provider: largeModel.ModelCfg.Provider,

595+ Provider: largeModel.Model.Provider(),

1355 IsSummaryMessage: true,

596 IsSummaryMessage: true,

1356 })

597 })

1357 if err != nil {

598 if err != nil {

@@ -1397,12 +638,6 @@ func (a *sessionAgent) Summarize(ctx context.Context, sessionID string, opts fan

1397 deleteErr := a.messages.Delete(ctx, summaryMessage.ID)

638 deleteErr := a.messages.Delete(ctx, summaryMessage.ID)

1398 return deleteErr

639 return deleteErr

1399 }

640 }

1400− // Mark the summary message as finished with an error so the UI

1401− // stops spinning.

1402− summaryMessage.AddFinish(message.FinishReasonError, "Summarization Error", err.Error())

1403− if updateErr := a.messages.Update(ctx, summaryMessage); updateErr != nil {

1404− return updateErr

1405− }

1406 return err

641 return err

1407 }

642 }

1408

643

@@ -1422,42 +657,20 @@ func (a *sessionAgent) Summarize(ctx context.Context, sessionID string, opts fan

1422 }

657 }

1423 openrouterCost = &newCost

658 openrouterCost = &newCost

1424 }

659 }

1425− extractHyperCredits(step.ProviderMetadata)

1426 }

660 }

1427

661

1428− a.updateSessionUsage(largeModel, &currentSession, resp.TotalUsage, openrouterCost, false)

662+ a.updateSessionUsage(largeModel, &currentSession, resp.TotalUsage, openrouterCost)

1429

663

1430 // Just in case, get just the last usage info.

664 // Just in case, get just the last usage info.

1431 usage := resp.Response.Usage

665 usage := resp.Response.Usage

1432 currentSession.SummaryMessageID = summaryMessage.ID

666 currentSession.SummaryMessageID = summaryMessage.ID

1433− currentSession.CompletionTokens = summaryCompletionTokens(usage, summaryMessage)

667+ currentSession.CompletionTokens = usage.OutputTokens

1434 currentSession.PromptTokens = 0

668 currentSession.PromptTokens = 0

1435− currentSession.EstimatedUsage = usageIsZero(usage)

1436 _, err = a.sessions.Save(genCtx, currentSession)

669 _, err = a.sessions.Save(genCtx, currentSession)

1437− if err != nil {

1438− return err

670+ return err

1439− }

1440−

1441− // Release the active request before processing queued messages so that

1442− // Run() does not see the session as busy.

1443− a.activeRequests.Del(sessionID)

1444− cancel()

1445−

1446− // Process any messages that were queued while summarizing.

1447− queuedMessages, ok := a.messageQueue.Get(sessionID)

1448− if !ok || len(queuedMessages) == 0 {

1449− return nil

1450− }

1451− firstQueuedMessage := queuedMessages[0]

1452− a.messageQueue.Set(sessionID, queuedMessages[1:])

1453− _, qErr := a.Run(ctx, firstQueuedMessage)

1454− return qErr

1455}

671}

1456

672

1457−// cacheControlOptions is computed once for the process lifetime: the env check

1458−// is constant and callers only read (never mutate) the returned map, so the

1459−// same shared value can be reused across the ~3 calls per step.

1460−var cacheControlOptions = sync.OnceValue(func() fantasy.ProviderOptions {

673+func (a *sessionAgent) getCacheControlOptions() fantasy.ProviderOptions {

1461 if t, _ := strconv.ParseBool(os.Getenv("CRUSH_DISABLE_ANTHROPIC_CACHE")); t {

674 if t, _ := strconv.ParseBool(os.Getenv("CRUSH_DISABLE_ANTHROPIC_CACHE")); t {

1462 return fantasy.ProviderOptions{}

675 return fantasy.ProviderOptions{}

1463 }

676 }

@@ -1468,14 +681,7 @@ var cacheControlOptions = sync.OnceValue(func() fantasy.ProviderOptions {

1468 bedrock.Name: &anthropic.ProviderCacheControlOptions{

681 bedrock.Name: &anthropic.ProviderCacheControlOptions{

1469 CacheControl: anthropic.CacheControl{Type: "ephemeral"},

682 CacheControl: anthropic.CacheControl{Type: "ephemeral"},

1470 },

683 },

1471− vercel.Name: &anthropic.ProviderCacheControlOptions{

1472− CacheControl: anthropic.CacheControl{Type: "ephemeral"},

1473− },

1474 }

684 }

1475−})

1476−

1477−func (a *sessionAgent) getCacheControlOptions() fantasy.ProviderOptions {

1478− return cacheControlOptions()

1479}

685}

1480

686

1481func (a *sessionAgent) createUserMessage(ctx context.Context, call SessionAgentCall) (message.Message, error) {

687func (a *sessionAgent) createUserMessage(ctx context.Context, call SessionAgentCall) (message.Message, error) {

@@ -1495,66 +701,27 @@ func (a *sessionAgent) createUserMessage(ctx context.Context, call SessionAgentC

1495 return msg, nil

701 return msg, nil

1496}

702}

1497

703

1498−func (a *sessionAgent) preparePrompt(msgs []message.Message, supportsImages bool, attachments ...message.Attachment) ([]fantasy.Message, []fantasy.FilePart) {

704+func (a *sessionAgent) preparePrompt(msgs []message.Message, attachments ...message.Attachment) ([]fantasy.Message, []fantasy.FilePart) {

1499− history := make([]fantasy.Message, 0, len(msgs)+1)

705+ var history []fantasy.Message

1500 if !a.isSubAgent {

706 if !a.isSubAgent {

1501 history = append(history, fantasy.NewUserMessage(

707 history = append(history, fantasy.NewUserMessage(

1502− fmt.Sprintf(

1503− "<system_reminder>%s</system_reminder>",

708+ fmt.Sprintf("<system_reminder>%s</system_reminder>",

1504 `This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.

709 `This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.

1505If you are working on tasks that would benefit from a todo list please use the "todos" tool to create one.

710If you are working on tasks that would benefit from a todo list please use the "todos" tool to create one.

1506If not, please feel free to ignore. Again do not mention this message to the user.`,

711If not, please feel free to ignore. Again do not mention this message to the user.`,

1507 ),

712 ),

1508 ))

713 ))

1509 }

714 }

1510− // Collect all tool call IDs present in assistant messages and all tool

1511− // result IDs present in tool messages. This lets us detect both orphaned

1512− // tool results (result without a call) and orphaned tool calls (call

1513− // without a result).

1514− knownToolCallIDs := make(map[string]struct{})

1515− knownToolResultIDs := make(map[string]struct{})

1516− for _, m := range msgs {

1517− switch m.Role {

1518− case message.Assistant:

1519− for _, tc := range m.ToolCalls() {

1520− knownToolCallIDs[tc.ID] = struct{}{}

1521− }

1522− case message.Tool:

1523− for _, tr := range m.ToolResults() {

1524− knownToolResultIDs[tr.ToolCallID] = struct{}{}

1525− }

1526− }

1527− }

1528−

1529 for _, m := range msgs {

715 for _, m := range msgs {

1530 if len(m.Parts) == 0 {

716 if len(m.Parts) == 0 {

1531 continue

717 continue

1532 }

718 }

1533− // Assistant message without content or tool calls (cancelled before it returned anything).

719+ // Assistant message without content or tool calls (cancelled before it

720+ // returned anything).

1534 if m.Role == message.Assistant && len(m.ToolCalls()) == 0 && m.Content().Text == "" && m.ReasoningContent().String() == "" {

721 if m.Role == message.Assistant && len(m.ToolCalls()) == 0 && m.Content().Text == "" && m.ReasoningContent().String() == "" {

1535 continue

722 continue

1536 }

723 }

1537− if m.Role == message.Tool {

1538− if msg, ok := filterOrphanedToolResults(m, knownToolCallIDs); ok {

1539− history = append(history, msg)

724+ history = append(history, m.ToAIMessage()...)

1540− }

1541− continue

1542− }

1543− aiMsgs := m.ToAIMessage()

1544− if !supportsImages {

1545− for i := range aiMsgs {

1546− if aiMsgs[i].Role == fantasy.MessageRoleUser {

1547− aiMsgs[i].Content = filterFileParts(aiMsgs[i].Content)

1548− }

1549− }

1550− }

1551− history = append(history, aiMsgs...)

1552−

1553− if m.Role == message.Assistant {

1554− if msg, ok := syntheticToolResultsForOrphanedCalls(m, knownToolResultIDs); ok {

1555− history = append(history, msg)

1556− }

1557− }

1558 }

725 }

1559

726

1560 var files []fantasy.FilePart

727 var files []fantasy.FilePart

@@ -1572,152 +739,47 @@ If not, please feel free to ignore. Again do not mention this message to the use

1572 return history, files

739 return history, files

1573}

740}

1574

741

1575−// filterFileParts removes fantasy.FilePart entries from a slice of message

1576−// parts. Used to strip image attachments from historical user messages when

1577−// the current model does not support them.

1578−func filterFileParts(parts []fantasy.MessagePart) []fantasy.MessagePart {

1579− filtered := make([]fantasy.MessagePart, 0, len(parts))

1580− for _, part := range parts {

1581− if _, ok := fantasy.AsMessagePart[fantasy.FilePart](part); ok {

1582− continue

1583− }

1584− filtered = append(filtered, part)

1585− }

1586− return filtered

1587−}

1588−

1589−// filterOrphanedToolResults converts a tool message to a fantasy.Message,

1590−// dropping any tool result parts whose tool_call_id has no matching tool call

1591−// in the known set. An orphaned result causes API validation to fail on every

1592−// subsequent turn, permanently locking the session. Returns the filtered

1593−// message and true if at least one valid part remains.

1594−func filterOrphanedToolResults(m message.Message, knownToolCallIDs map[string]struct{}) (fantasy.Message, bool) {

1595− aiMsgs := m.ToAIMessage()

1596− if len(aiMsgs) == 0 {

1597− return fantasy.Message{}, false

1598− }

1599− validParts := make([]fantasy.MessagePart, 0, len(aiMsgs[0].Content))

1600− for _, part := range aiMsgs[0].Content {

1601− tr, ok := fantasy.AsMessagePart[fantasy.ToolResultPart](part)

1602− if !ok {

1603− validParts = append(validParts, part)

1604− continue

1605− }

1606− if _, known := knownToolCallIDs[tr.ToolCallID]; known {

1607− validParts = append(validParts, part)

1608− } else {

1609− slog.Warn(

1610− "Dropping orphaned tool result with no matching tool call",

1611− "tool_call_id", tr.ToolCallID,

1612− )

1613− }

1614− }

1615− if len(validParts) == 0 {

1616− return fantasy.Message{}, false

1617− }

1618− msg := aiMsgs[0]

1619− msg.Content = validParts

1620− return msg, true

1621−}

1622−

1623−// syntheticToolResultsForOrphanedCalls returns a tool message containing

1624−// synthetic tool results for any tool calls in the assistant message that

1625−// have no matching result in knownToolResultIDs. LLM APIs require every

1626−// tool_use to be immediately followed by a tool_result; an interrupted

1627−// session can leave orphaned tool_use blocks that permanently lock the

1628−// conversation. Returns the message and true if any synthetic results were

1629−// produced.

1630−func syntheticToolResultsForOrphanedCalls(m message.Message, knownToolResultIDs map[string]struct{}) (fantasy.Message, bool) {

1631− syntheticParts := make([]fantasy.MessagePart, 0, len(m.ToolCalls()))

1632− for _, tc := range m.ToolCalls() {

1633− if _, hasResult := knownToolResultIDs[tc.ID]; hasResult {

1634− continue

1635− }

1636− slog.Warn(

1637− "Injecting synthetic tool result for orphaned tool call",

1638− "tool_call_id", tc.ID,

1639− "tool_name", tc.Name,

1640− )

1641− syntheticParts = append(syntheticParts, fantasy.ToolResultPart{

1642− ToolCallID: tc.ID,

1643− Output: fantasy.ToolResultOutputContentError{

1644− Error: errors.New("tool call was interrupted and did not produce a result, you may retry this call if the result is still needed"),

1645− },

1646− })

1647− }

1648− if len(syntheticParts) == 0 {

1649− return fantasy.Message{}, false

1650− }

1651− return fantasy.Message{

1652− Role: fantasy.MessageRoleTool,

1653− Content: syntheticParts,

1654− }, true

1655−}

1656−

1657func (a *sessionAgent) getSessionMessages(ctx context.Context, session session.Session) ([]message.Message, error) {

742func (a *sessionAgent) getSessionMessages(ctx context.Context, session session.Session) ([]message.Message, error) {

1658− if session.SummaryMessageID != "" {

1659− msgs, err := a.messages.ListAfter(ctx, session.ID, session.SummaryMessageID)

1660− if err != nil {

1661− return nil, fmt.Errorf("failed to list messages after summary: %w", err)

1662− }

1663− if len(msgs) > 0 {

1664− msgs[0].Role = message.User

1665− }

1666− return msgs, nil

1667− }

1668−

1669 msgs, err := a.messages.List(ctx, session.ID)

743 msgs, err := a.messages.List(ctx, session.ID)

1670 if err != nil {

744 if err != nil {

1671 return nil, fmt.Errorf("failed to list messages: %w", err)

745 return nil, fmt.Errorf("failed to list messages: %w", err)

1672 }

746 }

1673− return msgs, nil

1674−}

1675

747

1676−// hasUserTextMessage reports whether any user message in msgs contains

748+ if session.SummaryMessageID != "" {

1677−// text content (as opposed to only shell commands or other non-text parts).

749+ summaryMsgIndex := -1

1678−func hasUserTextMessage(msgs []message.Message) bool {

1679− for _, msg := range msgs {

1680− if msg.Role != message.User {

1681− continue

1682− }

1683− for _, part := range msg.Parts {

750+ for i, msg := range msgs {

1684− if tc, ok := part.(message.TextContent); ok && tc.Text != "" {

751+ if msg.ID == session.SummaryMessageID {

1685− return true

752+ summaryMsgIndex = i

753+ break

1686 }

754 }

1687 }

755 }

756+ if summaryMsgIndex != -1 {

757+ msgs = msgs[summaryMsgIndex:]

758+ msgs[0].Role = message.User

759+ }

1688 }

760 }

1689− return false

761+ return msgs, nil

1690}

762}

1691

763

1692−// GenerateTitle generates a session title based on the initial prompt.

764+// generateTitle generates a session titled based on the initial prompt.

1693−func (a *sessionAgent) GenerateTitle(ctx context.Context, sessionID string, userPrompt string) {

765+func (a *sessionAgent) generateTitle(ctx context.Context, sessionID string, userPrompt string) {

1694 if userPrompt == "" {

766 if userPrompt == "" {

1695 return

767 return

1696 }

768 }

1697

769

1698− // Ensure the session always gets a title even if every path below

1699− // fails or the context is cancelled before we finish.

1700− var titleSaved bool

1701− defer func() {

1702− if !titleSaved {

1703− fallbackCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Second)

1704− defer cancel()

1705− if err := a.sessions.Rename(fallbackCtx, sessionID, DefaultSessionName); err != nil {

1706− slog.Error("Failed to save fallback session title", "error", err)

1707− }

1708− }

1709− }()

1710−

1711 smallModel := a.smallModel.Get()

770 smallModel := a.smallModel.Get()

1712 largeModel := a.largeModel.Get()

771 largeModel := a.largeModel.Get()

1713 systemPromptPrefix := a.systemPromptPrefix.Get()

772 systemPromptPrefix := a.systemPromptPrefix.Get()

1714

773

774+ var maxOutputTokens int64 = 40

775+ if smallModel.CatwalkCfg.CanReason {

776+ maxOutputTokens = smallModel.CatwalkCfg.DefaultMaxTokens

777+ }

778+

1715 newAgent := func(m fantasy.LanguageModel, p []byte, tok int64) fantasy.Agent {

779 newAgent := func(m fantasy.LanguageModel, p []byte, tok int64) fantasy.Agent {

1716− return fantasy.NewAgent(

780+ return fantasy.NewAgent(m,

1717− m,

1718 fantasy.WithSystemPrompt(string(p)+"\n /no_think"),

781 fantasy.WithSystemPrompt(string(p)+"\n /no_think"),

1719 fantasy.WithMaxOutputTokens(tok),

782 fantasy.WithMaxOutputTokens(tok),

1720− fantasy.WithUserAgent(userAgent),

1721 )

783 )

1722 }

784 }

1723

785

@@ -1734,40 +796,41 @@ func (a *sessionAgent) GenerateTitle(ctx context.Context, sessionID string, user

1734 },

796 },

1735 }

797 }

1736

798

1737− type modelAttempt struct {

799+ // Use the small model to generate the title.

1738− name string

1739− model Model

1740− }

1741− attempts := []modelAttempt{

1742− {"small", smallModel},

800+ model := smallModel

1743− {"large", largeModel},

801+ agent := newAgent(model.Model, titlePrompt, maxOutputTokens)

1744− }

802+ resp, err := agent.Stream(ctx, streamCall)

1745−

803+ if err == nil {

1746− var resp *fantasy.AgentResult

804+ // We successfully generated a title with the small model.

1747− var err error

805+ slog.Debug("Generated title with small model")

1748− var model Model

806+ } else {

1749− var success bool

807+ // It didn't work. Let's try with the big model.

1750− for _, attempt := range attempts {

808+ slog.Error("Error generating title with small model; trying big model", "err", err)

1751− tok := int64(40)

809+ model = largeModel

1752− if attempt.model.CatwalkCfg.CanReason {

1753− tok = attempt.model.CatwalkCfg.DefaultMaxTokens

1754− }

1755− agent := newAgent(attempt.model.Model, titlePrompt, tok)

810+ agent = newAgent(model.Model, titlePrompt, maxOutputTokens)

1756 resp, err = agent.Stream(ctx, streamCall)

811 resp, err = agent.Stream(ctx, streamCall)

1757− if err == nil && resp.Response.FinishReason != fantasy.FinishReasonLength {

812+ if err == nil {

1758− model = attempt.model

1759− slog.Debug("Generated title with " + attempt.name + " model")

813+ slog.Debug("Generated title with large model")

1760− success = true

1761− break

1762− }

1763− if err != nil {

1764− slog.Error("Error generating title with "+attempt.name+" model; trying next", "err", err)

1765 } else {

814 } else {

1766− slog.Error("Title generation hit token limit with " + attempt.name + " model; trying next")

815+ // Welp, the large model didn't work either. Use the default

816+ // session name and return.

817+ slog.Error("Error generating title with large model", "err", err)

818+ saveErr := a.sessions.UpdateTitleAndUsage(ctx, sessionID, defaultSessionName, 0, 0, 0)

819+ if saveErr != nil {

820+ slog.Error("Failed to save session title and usage", "error", saveErr)

821+ }

822+ return

1767 }

823 }

1768 }

824 }

1769− if !success {

825+

1770− // The deferred fallback will save the default session name.

826+ if resp == nil {

827+ // Actually, we didn't get a response so we can't. Use the default

828+ // session name and return.

829+ slog.Error("Response is nil; can't generate title")

830+ saveErr := a.sessions.UpdateTitleAndUsage(ctx, sessionID, defaultSessionName, 0, 0, 0)

831+ if saveErr != nil {

832+ slog.Error("Failed to save session title and usage", "error", saveErr)

833+ }

1771 return

834 return

1772 }

835 }

1773

836

@@ -1777,19 +840,11 @@ func (a *sessionAgent) GenerateTitle(ctx context.Context, sessionID string, user

1777

840

1778 // Remove thinking tags if present.

841 // Remove thinking tags if present.

1779 title = thinkTagRegex.ReplaceAllString(title, "")

842 title = thinkTagRegex.ReplaceAllString(title, "")

1780− title = orphanThinkTagRegex.ReplaceAllString(title, "")

1781

843

1782 title = strings.TrimSpace(title)

844 title = strings.TrimSpace(title)

1783 if title == "" {

845 if title == "" {

1784− // LLM returned empty content. Use the prompt itself as a

846+ slog.Debug("Empty title; using fallback")

1785− // fallback title, truncated to 50 chars, before resorting to

847+ title = defaultSessionName

1786− // the generic default.

1787− fallback := strings.ReplaceAll(userPrompt, "\n", " ")

1788− fallback = strings.TrimSpace(fallback)

1789− if len(fallback) > 50 {

1790− fallback = ansi.Truncate(fallback, 50, "…")

1791− }

1792− title = cmp.Or(fallback, DefaultSessionName)

1793 }

848 }

1794

849

1795 // Calculate usage and cost.

850 // Calculate usage and cost.

@@ -1803,7 +858,6 @@ func (a *sessionAgent) GenerateTitle(ctx context.Context, sessionID string, user

1803 }

858 }

1804 openrouterCost = &newCost

859 openrouterCost = &newCost

1805 }

860 }

1806− extractHyperCredits(step.ProviderMetadata)

1807 }

861 }

1808

862

1809 modelConfig := model.CatwalkCfg

863 modelConfig := model.CatwalkCfg

@@ -1817,11 +871,6 @@ func (a *sessionAgent) GenerateTitle(ctx context.Context, sessionID string, user

1817 cost = *openrouterCost

871 cost = *openrouterCost

1818 }

872 }

1819

873

1820− // Skip cost accumulation

1821− if model.FlatRate {

1822− cost = 0

1823− }

1824−

1825 promptTokens := resp.TotalUsage.InputTokens + resp.TotalUsage.CacheCreationTokens

874 promptTokens := resp.TotalUsage.InputTokens + resp.TotalUsage.CacheCreationTokens

1826 completionTokens := resp.TotalUsage.OutputTokens

875 completionTokens := resp.TotalUsage.OutputTokens

1827

876

@@ -1832,7 +881,6 @@ func (a *sessionAgent) GenerateTitle(ctx context.Context, sessionID string, user

1832 slog.Error("Failed to save session title and usage", "error", saveErr)

881 slog.Error("Failed to save session title and usage", "error", saveErr)

1833 return

882 return

1834 }

883 }

1835− titleSaved = true

1836}

884}

1837

885

1838func (a *sessionAgent) openrouterCost(metadata fantasy.ProviderMetadata) *float64 {

886func (a *sessionAgent) openrouterCost(metadata fantasy.ProviderMetadata) *float64 {

@@ -1848,85 +896,26 @@ func (a *sessionAgent) openrouterCost(metadata fantasy.ProviderMetadata) *float6

1848 return &opts.Usage.Cost

896 return &opts.Usage.Cost

1849}

897}

1850

898

1851−// extractHyperCredits reads usage.remaining.hypercredits from OpenAI

1852−// provider metadata and stores it for the next FetchCredits call.

1853−func extractHyperCredits(metadata fantasy.ProviderMetadata) {

1854− openaiMeta, ok := metadata[openai.Name]

1855− if !ok {

1856− return

1857− }

1858− pm, ok := openaiMeta.(*openai.ProviderMetadata)

1859− if !ok {

1860− return

1861− }

1862− var remaining struct {

1863− Hypercredits float64 `json:"hypercredits"`

1864− }

1865− if pm.ExtraField("remaining", &remaining) && remaining.Hypercredits > 0 {

1866− hyper.SetBalance(int(math.Round(remaining.Hypercredits)))

1867− }

1868−}

1869−

1870−func (a *sessionAgent) updateSessionUsage(model Model, session *session.Session, usage fantasy.Usage, overrideCost *float64, estimated bool) {

899+func (a *sessionAgent) updateSessionUsage(model Model, session *session.Session, usage fantasy.Usage, overrideCost *float64) {

1871− if !usageIsZero(usage) {

1872− session.EstimatedUsage = estimated

1873− }

1874−

1875 modelConfig := model.CatwalkCfg

900 modelConfig := model.CatwalkCfg

1876 cost := modelConfig.CostPer1MInCached/1e6*float64(usage.CacheCreationTokens) +

901 cost := modelConfig.CostPer1MInCached/1e6*float64(usage.CacheCreationTokens) +

1877 modelConfig.CostPer1MOutCached/1e6*float64(usage.CacheReadTokens) +

902 modelConfig.CostPer1MOutCached/1e6*float64(usage.CacheReadTokens) +

1878 modelConfig.CostPer1MIn/1e6*float64(usage.InputTokens) +

903 modelConfig.CostPer1MIn/1e6*float64(usage.InputTokens) +

1879 modelConfig.CostPer1MOut/1e6*float64(usage.OutputTokens)

904 modelConfig.CostPer1MOut/1e6*float64(usage.OutputTokens)

1880

905

1881− if !estimated {

1882− a.eventTokensUsed(session.ID, model, usage, cost)

906+ a.eventTokensUsed(session.ID, model, usage, cost)

1883− }

1884

907

1885− if estimated {

908+ if overrideCost != nil {

1886− cost = 0

909+ session.Cost += *overrideCost

1887 } else {

910 } else {

1888− // Use override cost if available (e.g., from OpenRouter).

911+ session.Cost += cost

1889− if overrideCost != nil {

1890− cost = *overrideCost

1891− }

1892−

1893− // Skip cost accumulation

1894− if model.FlatRate {

1895− cost = 0

1896− }

1897 }

912 }

1898

913

1899− session.Cost += cost

1900− updateSessionTokenCounters(session, usage)

1901−}

1902−

1903−func updateSessionTokenCounters(session *session.Session, usage fantasy.Usage) {

1904− if usage.OutputTokens != 0 {

1905− session.CompletionTokens = usage.OutputTokens

914+ session.CompletionTokens = usage.OutputTokens

1906− }

1907− if promptTokens := usage.InputTokens + usage.CacheReadTokens; promptTokens != 0 {

915+ session.PromptTokens = usage.InputTokens + usage.CacheReadTokens

1908− session.PromptTokens = promptTokens

1909− }

1910−}

1911−

1912−func summaryCompletionTokens(usage fantasy.Usage, summaryMessage message.Message) int64 {

1913− if usage.OutputTokens != 0 {

1914− return usage.OutputTokens

1915− }

1916− return approxTokenCount(summaryMessage.Content().Text) + approxTokenCount(summaryMessage.ReasoningContent().String())

1917}

916}

1918

917

1919func (a *sessionAgent) Cancel(sessionID string) {

918func (a *sessionAgent) Cancel(sessionID string) {

1920− // Serialize against the dispatch handoff in Run so the accepted ->

1921− // (cancel-on-entry | queued | active) transition is atomic against

1922− // this cancel. Every cancel observes at least one of: an active

1923− // request, an accepted run (recorded as a pending cancel), or a

1924− // queue entry it then clears. If none of those hold, an idle Escape

1925− // is a true no-op and must not poison the next prompt.

1926− mu := a.sessionMu(sessionID)

1927− mu.Lock()

1928− defer mu.Unlock()

1929−

1930 // Cancel regular requests. Don't use Take() here - we need the entry to

919 // Cancel regular requests. Don't use Take() here - we need the entry to

1931 // remain in activeRequests so IsBusy() returns true until the goroutine

920 // remain in activeRequests so IsBusy() returns true until the goroutine

1932 // fully completes (including error handling that may access the DB).

921 // fully completes (including error handling that may access the DB).

@@ -1942,40 +931,16 @@ func (a *sessionAgent) Cancel(sessionID string) {

1942 cancel()

931 cancel()

1943 }

932 }

1944

933

1945− // Record a pending cancel only when a dispatched-but-not-yet-active

1946− // run exists. This catches runs still in the goroutine scheduler or

1947− // about to enter Run's busy-queue branch, while leaving an idle

1948− // session untouched. Active and accepted are not mutually exclusive:

1949− // when a run is active and a follow-up has been accepted, both the

1950− // cancel above and this pending record fire.

1951− //

1952− // Raise the session's cancel mark to the latest accept sequence

1953− // assigned so far. Every prompt currently accepted-but-not-yet-

1954− // active has a sequence at or below that value, so one cancel covers

1955− // all of them; a prompt accepted after this cancel gets a strictly

1956− // higher sequence and is never poisoned. Using max keeps repeated

1957− // cancels idempotent while the same prompts are in flight and lets a

1958− // later cancel extend coverage to prompts accepted since.

1959− a.acceptedMu.Lock()

1960− count, ok := a.acceptedRuns.Get(sessionID)

1961− mark := a.acceptSeqGen

1962− a.acceptedMu.Unlock()

1963− if ok && count > 0 {

1964− slog.Debug("Recording cancel mark for accepted runs", "session_id", sessionID, "count", count, "mark", mark)

1965− existing, _ := a.cancelMark.Get(sessionID)

1966− a.cancelMark.Set(sessionID, max(existing, mark))

1967− }

1968−

1969 if a.QueuedPrompts(sessionID) > 0 {

934 if a.QueuedPrompts(sessionID) > 0 {

1970 slog.Debug("Clearing queued prompts", "session_id", sessionID)

935 slog.Debug("Clearing queued prompts", "session_id", sessionID)

1971− a.clearQueueAndNotify(sessionID)

936+ a.messageQueue.Del(sessionID)

1972 }

937 }

1973}

938}

1974

939

1975func (a *sessionAgent) ClearQueue(sessionID string) {

940func (a *sessionAgent) ClearQueue(sessionID string) {

1976 if a.QueuedPrompts(sessionID) > 0 {

941 if a.QueuedPrompts(sessionID) > 0 {

1977 slog.Debug("Clearing queued prompts", "session_id", sessionID)

942 slog.Debug("Clearing queued prompts", "session_id", sessionID)

1978− a.clearQueueAndNotify(sessionID)

943+ a.messageQueue.Del(sessionID)

1979 }

944 }

1980}

945}

1981

946

@@ -1999,10 +964,14 @@ func (a *sessionAgent) CancelAll() {

1999}

964}

2000

965

2001func (a *sessionAgent) IsBusy() bool {

966func (a *sessionAgent) IsBusy() bool {

2002− // activeRequests only ever holds non-nil cancel funcs (Set with a

967+ var busy bool

2003− // real func, removed via Del), so a length check is exact — and it

968+ for cancelFunc := range a.activeRequests.Seq() {

2004− // avoids cloning the whole map, which Seq() does on every call.

969+ if cancelFunc != nil {

2005− return a.activeRequests.Len() > 0

970+ busy = true

971+ break

972+ }

973+ }

974+ return busy

2006}

975}

2007

976

2008func (a *sessionAgent) IsSessionBusy(sessionID string) bool {

977func (a *sessionAgent) IsSessionBusy(sessionID string) bool {

@@ -2037,7 +1006,6 @@ func (a *sessionAgent) SetModels(large Model, small Model) {

2037

1006

2038func (a *sessionAgent) SetTools(tools []fantasy.AgentTool) {

1007func (a *sessionAgent) SetTools(tools []fantasy.AgentTool) {

2039 a.tools.SetSlice(tools)

1008 a.tools.SetSlice(tools)

2040− a.toolsGen.Add(1)

2041}

1009}

2042

1010

2043func (a *sessionAgent) SetSystemPrompt(systemPrompt string) {

1011func (a *sessionAgent) SetSystemPrompt(systemPrompt string) {

@@ -2068,23 +1036,13 @@ func (a *sessionAgent) convertToToolResult(result fantasy.ToolResultContent) mes

2068 }

1036 }

2069 case fantasy.ToolResultContentTypeMedia:

1037 case fantasy.ToolResultContentTypeMedia:

2070 if r, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentMedia](result.Result); ok {

1038 if r, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentMedia](result.Result); ok {

2071− if !stringext.IsValidBase64(r.Data) {

2072− slog.Warn(

2073− "Tool returned media with invalid base64 data, discarding image",

2074− "tool", result.ToolName,

2075− "tool_call_id", result.ToolCallID,

2076− )

2077− baseResult.Content = "Tool returned image data with invalid encoding"

2078− baseResult.IsError = true

2079− } else {

2080− content := r.Text

1039+ content := r.Text

2081− if content == "" {

1040+ if content == "" {

2082− content = fmt.Sprintf("Loaded %s content", r.MediaType)

1041+ content = fmt.Sprintf("Loaded %s content", r.MediaType)

2083− }

2084− baseResult.Content = content

2085− baseResult.Data = r.Data

2086− baseResult.MIMEType = r.MediaType

2087 }

1042 }

1043+ baseResult.Content = content

1044+ baseResult.Data = r.Data

1045+ baseResult.MIMEType = r.MediaType

2088 }

1046 }

2089 }

1047 }

2090

1048

@@ -2114,15 +1072,12 @@ func (a *sessionAgent) convertToToolResult(result fantasy.ToolResultContent) mes

2114// AFTER: [tool result: "Image loaded - see attached"], [user: image attachment]

1072// AFTER: [tool result: "Image loaded - see attached"], [user: image attachment]

2115func (a *sessionAgent) workaroundProviderMediaLimitations(messages []fantasy.Message, largeModel Model) []fantasy.Message {

1073func (a *sessionAgent) workaroundProviderMediaLimitations(messages []fantasy.Message, largeModel Model) []fantasy.Message {

2116 providerSupportsMedia := largeModel.ModelCfg.Provider == string(catwalk.InferenceProviderAnthropic) ||

1074 providerSupportsMedia := largeModel.ModelCfg.Provider == string(catwalk.InferenceProviderAnthropic) ||

2117− largeModel.ModelCfg.Provider == string(catwalk.InferenceProviderBedrock) ||

1075+ largeModel.ModelCfg.Provider == string(catwalk.InferenceProviderBedrock)

2118− largeModel.ModelCfg.Provider == string(catwalk.InferenceProviderBedrockEurope)

2119

1076

2120 if providerSupportsMedia {

1077 if providerSupportsMedia {

2121 return messages

1078 return messages

2122 }

1079 }

2123

1080

2124− supportsImages := largeModel.CatwalkCfg.SupportsImages

2125−

2126 convertedMessages := make([]fantasy.Message, 0, len(messages))

1081 convertedMessages := make([]fantasy.Message, 0, len(messages))

2127

1082

2128 for _, msg := range messages {

1083 for _, msg := range messages {

@@ -2142,21 +1097,6 @@ func (a *sessionAgent) workaroundProviderMediaLimitations(messages []fantasy.Mes

2142 }

1097 }

2143

1098

2144 if media, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentMedia](toolResult.Output); ok {

1099 if media, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentMedia](toolResult.Output); ok {

2145− if !supportsImages {

2146− // Model cannot process images. Replace with a text

2147− // placeholder and skip creating a synthetic user

2148− // message with FilePart, which would brick the

2149− // session on text-only models.

2150− textParts = append(textParts, fantasy.ToolResultPart{

2151− ToolCallID: toolResult.ToolCallID,

2152− Output: fantasy.ToolResultOutputContentText{

2153− Text: "[Image/media content not supported by this model]",

2154− },

2155− ProviderOptions: toolResult.ProviderOptions,

2156− })

2157− continue

2158− }

2159−

2160 decoded, err := base64.StdEncoding.DecodeString(media.Data)

1100 decoded, err := base64.StdEncoding.DecodeString(media.Data)

2161 if err != nil {

1101 if err != nil {

2162 slog.Warn("Failed to decode media data", "error", err)

1102 slog.Warn("Failed to decode media data", "error", err)

@@ -2212,37 +1152,3 @@ func buildSummaryPrompt(todos []session.Todo) string {

2212 }

1152 }

2213 return sb.String()

1153 return sb.String()

2214}

1154}

2215−

2216−func providerRetryLogFields(err *fantasy.ProviderError, delay time.Duration) []any {

2217− fields := []any{

2218− "retry_delay", delay.String(),

2219− }

2220− if err == nil {

2221− return fields

2222− }

2223− fields = append(fields, "status_code", err.StatusCode)

2224− if err.Title != "" {

2225− fields = append(fields, "title", err.Title)

2226− }

2227− if err.Message != "" {

2228− fields = append(fields, "message", err.Message)

2229− }

2230− return fields

2231−}

2232−

2233−// sanitizeToolInput validates tool call JSON from the provider.

2234−// Malformed input is replaced with an empty object to prevent

2235−// stuck conversations from truncated or malformed model output.

2236−// The second return value indicates whether sanitization occurred.

2237−func sanitizeToolInput(toolName, toolCallID, input string) (string, bool) {

2238− if !json.Valid([]byte(input)) {

2239− slog.Warn(

2240− "Malformed tool call JSON from provider, replacing with empty object",

2241− "tool", toolName,

2242− "id", toolCallID,

2243− "input_len", len(input),

2244− )

2245− return "{}", true

2246− }

2247− return input, false

2248−}

internal/agent/agent_test.go+4−366

@@ -1,21 +1,16 @@

1package agent

3import (

4− "encoding/base64"

5 "fmt"

4 "fmt"

6− "log/slog"

7 "os"

5 "os"

8 "path/filepath"

6 "path/filepath"

9 "runtime"

7 "runtime"

10 "strings"

8 "strings"

11 "testing"

9 "testing"

12− "time"

14− "charm.land/catwalk/pkg/catwalk"

15 "charm.land/fantasy"

11 "charm.land/fantasy"

16 "charm.land/x/vcr"

12 "charm.land/x/vcr"

17 "github.com/charmbracelet/crush/internal/agent/tools"

13 "github.com/charmbracelet/crush/internal/agent/tools"

18− "github.com/charmbracelet/crush/internal/config"

19 "github.com/charmbracelet/crush/internal/message"

14 "github.com/charmbracelet/crush/internal/message"

20 "github.com/charmbracelet/crush/internal/session"

15 "github.com/charmbracelet/crush/internal/session"

21 "github.com/stretchr/testify/assert"

16 "github.com/stretchr/testify/assert"

@@ -24,13 +19,11 @@ import (

24 _ "github.com/joho/godotenv/autoload"

19 _ "github.com/joho/godotenv/autoload"

25)

20)

27−func TestMain(m *testing.M) {

28− slog.SetLogLoggerLevel(slog.LevelError)

29− m.Run()

30−}

31−

32var modelPairs = []modelPair{

22var modelPairs = []modelPair{

33− {"glm-5.1", hyperBuilder("glm-5.1"), hyperBuilder("gpt-oss-120b")},

23+ {"anthropic-sonnet", anthropicBuilder("claude-sonnet-4-5-20250929"), anthropicBuilder("claude-3-5-haiku-20241022")},

24+ {"openai-gpt-5", openaiBuilder("gpt-5"), openaiBuilder("gpt-4o")},

25+ {"openrouter-kimi-k2", openRouterBuilder("moonshotai/kimi-k2-0905"), openRouterBuilder("qwen/qwen3-next-80b-a3b-instruct")},

26+ {"zai-glm4.6", zAIBuilder("glm-4.6"), zAIBuilder("glm-4.5-air")},

34}

27}

36func getModels(t *testing.T, r *vcr.Recorder, pair modelPair) (fantasy.LanguageModel, fantasy.LanguageModel) {

29func getModels(t *testing.T, r *vcr.Recorder, pair modelPair) (fantasy.LanguageModel, fantasy.LanguageModel) {

@@ -658,358 +651,3 @@ func BenchmarkBuildSummaryPrompt(b *testing.B) {

658 })

651 })

659 }

652 }

660}

653}

661−

662−func TestPreparePrompt_FiltersImageAttachments(t *testing.T) {

663− env := testEnv(t)

664− sa := testSessionAgent(env, nil, nil, "test prompt")

665− agent := sa.(*sessionAgent)

666−

667− ctx := t.Context()

668− sess, err := env.sessions.Create(ctx, "test")

669− require.NoError(t, err)

670−

671− // User message with text, a text attachment, and an image attachment.

672− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

673− Role: message.User,

674− Parts: []message.ContentPart{

675− message.TextContent{Text: "hello world"},

676− message.BinaryContent{Path: "notes.txt", MIMEType: "text/plain", Data: []byte("important notes")},

677− message.BinaryContent{Path: "image.png", MIMEType: "image/png", Data: []byte("fake-image-data")},

678− },

679− })

680− require.NoError(t, err)

681−

682− msgs, err := env.messages.List(ctx, sess.ID)

683− require.NoError(t, err)

684−

685− // When supportsImages is false, image attachments should be stripped.

686− history, _ := agent.preparePrompt(msgs, false)

687− // First message is the system reminder, second is the user message.

688− require.Len(t, history, 2)

689− require.Len(t, history[1].Content, 1)

690− text, ok := fantasy.AsMessagePart[fantasy.TextPart](history[1].Content[0])

691− require.True(t, ok)

692− require.Contains(t, text.Text, "hello world")

693− require.Contains(t, text.Text, "important notes")

694−

695− // When supportsImages is true, image attachments should remain.

696− history, _ = agent.preparePrompt(msgs, true)

697− require.Len(t, history, 2)

698− require.Len(t, history[1].Content, 2)

699− text, ok = fantasy.AsMessagePart[fantasy.TextPart](history[1].Content[0])

700− require.True(t, ok)

701− require.Contains(t, text.Text, "hello world")

702− file, ok := fantasy.AsMessagePart[fantasy.FilePart](history[1].Content[1])

703− require.True(t, ok)

704− require.Equal(t, "image.png", file.Filename)

705−}

706−

707−func TestPreparePrompt_OrphanedToolUse(t *testing.T) {

708− env := testEnv(t)

709− sa := testSessionAgent(env, nil, nil, "test prompt")

710− agent := sa.(*sessionAgent)

711−

712− ctx := t.Context()

713− sess, err := env.sessions.Create(ctx, "test")

714− require.NoError(t, err)

715−

716− // Create a user message.

717− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

718− Role: message.User,

719− Parts: []message.ContentPart{

720− message.TextContent{Text: "hello"},

721− },

722− })

723− require.NoError(t, err)

724−

725− // Create an assistant message with a tool call but no tool result —

726− // this simulates a cancelled/interrupted agent tool call.

727− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

728− Role: message.Assistant,

729− Parts: []message.ContentPart{

730− message.TextContent{Text: "let me check"},

731− message.ToolCall{

732− ID: "call_orphaned_1",

733− Name: "agent",

734− Input: `{"prompt":"do something"}`,

735− Finished: true,

736− },

737− },

738− })

739− require.NoError(t, err)

740−

741− // Create the next user message (the one that interrupted the tool call).

742− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

743− Role: message.User,

744− Parts: []message.ContentPart{

745− message.TextContent{Text: "Fix #2"},

746− },

747− })

748− require.NoError(t, err)

749−

750− msgs, err := env.messages.List(ctx, sess.ID)

751− require.NoError(t, err)

752−

753− history, _ := agent.preparePrompt(msgs, true)

754−

755− // The history must contain a synthetic tool result for the orphaned call.

756− found := false

757− for _, msg := range history {

758− if msg.Role != fantasy.MessageRoleTool {

759− continue

760− }

761− for _, part := range msg.Content {

762− if tr, ok := fantasy.AsMessagePart[fantasy.ToolResultPart](part); ok {

763− if tr.ToolCallID == "call_orphaned_1" {

764− found = true

765− _, isError := tr.Output.(fantasy.ToolResultOutputContentError)

766− require.True(t, isError, "orphaned tool result should be an error")

767− }

768− }

769− }

770− }

771− require.True(t, found, "expected synthetic tool result for orphaned tool call")

772−}

773−

774−func TestPreparePrompt_OrphanedToolUseMixed(t *testing.T) {

775− env := testEnv(t)

776− sa := testSessionAgent(env, nil, nil, "test prompt")

777− agent := sa.(*sessionAgent)

778−

779− ctx := t.Context()

780− sess, err := env.sessions.Create(ctx, "test")

781− require.NoError(t, err)

782−

783− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

784− Role: message.User,

785− Parts: []message.ContentPart{

786− message.TextContent{Text: "hello"},

787− },

788− })

789− require.NoError(t, err)

790−

791− // Assistant with 2 tool calls: one has a result, one is orphaned.

792− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

793− Role: message.Assistant,

794− Parts: []message.ContentPart{

795− message.ToolCall{

796− ID: "call_ok",

797− Name: "view",

798− Input: `{"path":"/foo"}`,

799− Finished: true,

800− },

801− message.ToolCall{

802− ID: "call_orphaned",

803− Name: "agent",

804− Input: `{"prompt":"search"}`,

805− Finished: true,

806− },

807− },

808− })

809− require.NoError(t, err)

810−

811− // Only one tool result — for call_ok.

812− _, err = env.messages.Create(ctx, sess.ID, message.CreateMessageParams{

813− Role: message.Tool,

814− Parts: []message.ContentPart{

815− message.ToolResult{

816− ToolCallID: "call_ok",

817− Name: "view",

818− Content: "file contents",

819− },

820− },

821− })

822− require.NoError(t, err)

823−

824− msgs, err := env.messages.List(ctx, sess.ID)

825− require.NoError(t, err)

826−

827− history, _ := agent.preparePrompt(msgs, true)

828−

829− // Should have a synthetic result only for the orphaned call.

830− var syntheticCount int

831− for _, msg := range history {

832− if msg.Role != fantasy.MessageRoleTool {

833− continue

834− }

835− for _, part := range msg.Content {

836− if tr, ok := fantasy.AsMessagePart[fantasy.ToolResultPart](part); ok {

837− if tr.ToolCallID == "call_orphaned" {

838− syntheticCount++

839− }

840− }

841− }

842− }

843− require.Equal(t, 1, syntheticCount, "expected exactly one synthetic result for the orphaned call")

844−}

845−

846−func TestWorkaroundProviderMediaLimitations_TextOnlyModel(t *testing.T) {

847− env := testEnv(t)

848− sa := testSessionAgent(env, nil, nil, "test prompt")

849− agent := sa.(*sessionAgent)

850−

851− pngBase64 := base64.StdEncoding.EncodeToString([]byte("fake-png-data"))

852−

853− messages := []fantasy.Message{

854− {

855− Role: fantasy.MessageRoleTool,

856− Content: []fantasy.MessagePart{

857− fantasy.ToolResultPart{

858− ToolCallID: "call_1",

859− Output: fantasy.ToolResultOutputContentMedia{

860− Data: pngBase64,

861− MediaType: "image/png",

862− },

863− },

864− },

865− },

866− }

867−

868− // Non-Anthropic provider, no image support — should replace media with

869− // a text placeholder and not create a synthetic user message.

870− largeModel := Model{

871− ModelCfg: config.SelectedModel{Provider: "openai"},

872− CatwalkCfg: catwalk.Model{

873− SupportsImages: false,

874− },

875− }

876−

877− result := agent.workaroundProviderMediaLimitations(messages, largeModel)

878−

879− // Should produce exactly one message: the tool message with a text

880− // placeholder. No synthetic user message with FilePart.

881− require.Len(t, result, 1)

882− require.Equal(t, fantasy.MessageRoleTool, result[0].Role)

883−

884− tr, ok := fantasy.AsMessagePart[fantasy.ToolResultPart](result[0].Content[0])

885− require.True(t, ok)

886− _, ok = fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentText](tr.Output)

887− require.True(t, ok)

888−}

889−

890−func TestWorkaroundProviderMediaLimitations_VisionModel(t *testing.T) {

891− env := testEnv(t)

892− sa := testSessionAgent(env, nil, nil, "test prompt")

893− agent := sa.(*sessionAgent)

894−

895− pngBase64 := base64.StdEncoding.EncodeToString([]byte("fake-png-data"))

896−

897− messages := []fantasy.Message{

898− {

899− Role: fantasy.MessageRoleTool,

900− Content: []fantasy.MessagePart{

901− fantasy.ToolResultPart{

902− ToolCallID: "call_1",

903− Output: fantasy.ToolResultOutputContentMedia{

904− Data: pngBase64,

905− MediaType: "image/png",

906− },

907− },

908− },

909− },

910− }

911−

912− // Non-Anthropic provider, image support — should create a synthetic

913− // user message with FilePart.

914− largeModel := Model{

915− ModelCfg: config.SelectedModel{Provider: "openai"},

916− CatwalkCfg: catwalk.Model{

917− SupportsImages: true,

918− },

919− }

920−

921− result := agent.workaroundProviderMediaLimitations(messages, largeModel)

922−

923− // Should produce two messages: tool message with placeholder text,

924− // and synthetic user message with FilePart.

925− require.Len(t, result, 2)

926− require.Equal(t, fantasy.MessageRoleTool, result[0].Role)

927− require.Equal(t, fantasy.MessageRoleUser, result[1].Role)

928−

929− // The tool message should have text placeholder.

930− tr, ok := fantasy.AsMessagePart[fantasy.ToolResultPart](result[0].Content[0])

931− require.True(t, ok)

932− textOutput, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentText](tr.Output)

933− require.True(t, ok)

934− require.Contains(t, textOutput.Text, "see attached file")

935−

936− // The synthetic user message should contain a TextPart and a FilePart.

937− require.Len(t, result[1].Content, 2)

938− file, ok := fantasy.AsMessagePart[fantasy.FilePart](result[1].Content[1])

939− require.True(t, ok)

940− require.Equal(t, "image/png", file.MediaType)

941−}

942−

943−func TestWorkaroundProviderMediaLimitations_AnthropicProvider(t *testing.T) {

944− env := testEnv(t)

945− sa := testSessionAgent(env, nil, nil, "test prompt")

946− agent := sa.(*sessionAgent)

947−

948− pngBase64 := base64.StdEncoding.EncodeToString([]byte("fake-png-data"))

949−

950− messages := []fantasy.Message{

951− {

952− Role: fantasy.MessageRoleTool,

953− Content: []fantasy.MessagePart{

954− fantasy.ToolResultPart{

955− ToolCallID: "call_1",

956− Output: fantasy.ToolResultOutputContentMedia{

957− Data: pngBase64,

958− MediaType: "image/png",

959− },

960− },

961− },

962− },

963− }

964−

965− // Anthropic provider — should return messages unchanged regardless of

966− // SupportsImages, since Anthropic handles media in tool results natively.

967− largeModel := Model{

968− ModelCfg: config.SelectedModel{Provider: string(catwalk.InferenceProviderAnthropic)},

969− CatwalkCfg: catwalk.Model{

970− SupportsImages: true,

971− },

972− }

973−

974− result := agent.workaroundProviderMediaLimitations(messages, largeModel)

975− require.Len(t, result, 1)

976− require.Equal(t, fantasy.MessageRoleTool, result[0].Role)

977−

978− // The media should still be in the tool result, untouched.

979− tr, ok := fantasy.AsMessagePart[fantasy.ToolResultPart](result[0].Content[0])

980− require.True(t, ok)

981− media, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentMedia](tr.Output)

982− require.True(t, ok)

983− require.Equal(t, "image/png", media.MediaType)

984−}

985−

986−func TestProviderRetryLogFields(t *testing.T) {

987− t.Run("nil provider error", func(t *testing.T) {

988− fields := providerRetryLogFields(nil, 2*time.Second)

989− require.Equal(t, []any{"retry_delay", "2s"}, fields)

990− })

991−

992− t.Run("provider error with title and message", func(t *testing.T) {

993− fields := providerRetryLogFields(&fantasy.ProviderError{

994− StatusCode: 429,

995− Title: "rate limit",

996− Message: "too many requests",

997− }, 1500*time.Millisecond)

998− require.Equal(t, []any{

999− "retry_delay", "1.5s",

1000− "status_code", 429,

1001− "title", "rate limit",

1002− "message", "too many requests",

1003− }, fields)

1004− })

1005−

1006− t.Run("provider error without optional strings", func(t *testing.T) {

1007− fields := providerRetryLogFields(&fantasy.ProviderError{

1008− StatusCode: 503,

1009− }, time.Second)

1010− require.Equal(t, []any{

1011− "retry_delay", "1s",

1012− "status_code", 503,

1013− }, fields)

1014− })

1015−}

internal/agent/agent_tool.go+49−12

@@ -4,6 +4,7 @@ import (

4 "context"

5 _ "embed"

6 "errors"

7+ "fmt"

8 "charm.land/fantasy"

9 "charm.land/fantasy"

@@ -13,7 +14,7 @@ import (

13)

14)

15//go:embed templates/agent_tool.md

16//go:embed templates/agent_tool.md

16−var agentToolDescription string

17+var agentToolDescription []byte

18type AgentParams struct {

19type AgentParams struct {

19 Prompt string `json:"prompt" description:"The task for the agent to perform"`

20 Prompt string `json:"prompt" description:"The task for the agent to perform"`

@@ -24,7 +25,7 @@ const (

24)

25)

26func (c *coordinator) agentTool(ctx context.Context) (fantasy.AgentTool, error) {

27func (c *coordinator) agentTool(ctx context.Context) (fantasy.AgentTool, error) {

27− agentCfg, ok := c.cfg.Config().Agents[config.AgentTask]

28+ agentCfg, ok := c.cfg.Agents[config.AgentTask]

28 if !ok {

29 if !ok {

29 return nil, errors.New("task agent not configured")

30 return nil, errors.New("task agent not configured")

30 }

31 }

@@ -39,7 +40,7 @@ func (c *coordinator) agentTool(ctx context.Context) (fantasy.AgentTool, error)

39 }

40 }

40 return fantasy.NewParallelAgentTool(

41 return fantasy.NewParallelAgentTool(

41 AgentToolName,

42 AgentToolName,

42− agentToolDescription,

43+ string(agentToolDescription),

43 func(ctx context.Context, params AgentParams, call fantasy.ToolCall) (fantasy.ToolResponse, error) {

44 func(ctx context.Context, params AgentParams, call fantasy.ToolCall) (fantasy.ToolResponse, error) {

44 if params.Prompt == "" {

45 if params.Prompt == "" {

45 return fantasy.NewTextErrorResponse("prompt is required"), nil

46 return fantasy.NewTextErrorResponse("prompt is required"), nil

@@ -55,14 +56,50 @@ func (c *coordinator) agentTool(ctx context.Context) (fantasy.AgentTool, error)

55 return fantasy.ToolResponse{}, errors.New("agent message id missing from context")

56 return fantasy.ToolResponse{}, errors.New("agent message id missing from context")

56 }

57 }

58− return c.runSubAgent(ctx, subAgentParams{

59+ agentToolSessionID := c.sessions.CreateAgentToolSessionID(agentMessageID, call.ID)

59− Agent: agent,

60+ session, err := c.sessions.CreateTaskSession(ctx, agentToolSessionID, sessionID, "New Agent Session")

60− SessionID: sessionID,

61+ if err != nil {

61− AgentMessageID: agentMessageID,

62+ return fantasy.ToolResponse{}, fmt.Errorf("error creating session: %s", err)

62− ToolCallID: call.ID,

63+ }

64+ model := agent.Model()

65+ maxTokens := model.CatwalkCfg.DefaultMaxTokens

66+ if model.ModelCfg.MaxTokens != 0 {

67+ maxTokens = model.ModelCfg.MaxTokens

68+ }

69+

70+ providerCfg, ok := c.cfg.Providers.Get(model.ModelCfg.Provider)

71+ if !ok {

72+ return fantasy.ToolResponse{}, errors.New("model provider not configured")

73+ }

74+ result, err := agent.Run(ctx, SessionAgentCall{

75+ SessionID: session.ID,

63− Prompt: params.Prompt,

76+ Prompt: params.Prompt,

64− SessionTitle: "New Agent Session",

77+ MaxOutputTokens: maxTokens,

78+ ProviderOptions: getProviderOptions(model, providerCfg),

79+ Temperature: model.ModelCfg.Temperature,

80+ TopP: model.ModelCfg.TopP,

81+ TopK: model.ModelCfg.TopK,

82+ FrequencyPenalty: model.ModelCfg.FrequencyPenalty,

83+ PresencePenalty: model.ModelCfg.PresencePenalty,

65 })

84 })

66− },

85+ if err != nil {

86+ return fantasy.NewTextErrorResponse("error generating response"), nil

87+ }

88+ updatedSession, err := c.sessions.Get(ctx, session.ID)

89+ if err != nil {

90+ return fantasy.ToolResponse{}, fmt.Errorf("error getting session: %s", err)

91+ }

92+ parentSession, err := c.sessions.Get(ctx, sessionID)

93+ if err != nil {

94+ return fantasy.ToolResponse{}, fmt.Errorf("error getting parent session: %s", err)

95+ }

96+

97+ parentSession.Cost += updatedSession.Cost

98+

99+ _, err = c.sessions.Save(ctx, parentSession)

100+ if err != nil {

101+ return fantasy.ToolResponse{}, fmt.Errorf("error saving parent session: %s", err)

102+ }

103+ return fantasy.NewTextResponse(result.Response.Content.Text()), nil

67− ), nil

104+ }), nil

68}

105}

internal/agent/agentic_fetch_tool.go+63−36

@@ -17,7 +17,7 @@ import (

17)

19//go:embed templates/agentic_fetch.md

20−var agenticFetchToolDescription string

20+var agenticFetchToolDescription []byte

22// agenticFetchValidationResult holds the validated parameters from the tool call context.

23type agenticFetchValidationResult struct {

@@ -52,20 +52,19 @@ var agenticFetchPromptTmpl []byte

53func (c *coordinator) agenticFetchTool(_ context.Context, client *http.Client) (fantasy.AgentTool, error) {

54 if client == nil {

55− transport := http.DefaultTransport.(*http.Transport).Clone()

56− transport.MaxIdleConns = 100

57− transport.MaxIdleConnsPerHost = 10

58− transport.IdleConnTimeout = 90 * time.Second

59−

60 client = &http.Client{

55 client = &http.Client{

61− Timeout: 30 * time.Second,

56+ Timeout: 30 * time.Second,

62− Transport: transport,

57+ Transport: &http.Transport{

58+ MaxIdleConns: 100,

59+ MaxIdleConnsPerHost: 10,

60+ IdleConnTimeout: 90 * time.Second,

61+ },

63 }

62 }

64 }

63 }

66 return fantasy.NewParallelAgentTool(

65 return fantasy.NewParallelAgentTool(

67 tools.AgenticFetchToolName,

66 tools.AgenticFetchToolName,

68− agenticFetchToolDescription,

67+ string(agenticFetchToolDescription),

69 func(ctx context.Context, params tools.AgenticFetchParams, call fantasy.ToolCall) (fantasy.ToolResponse, error) {

68 func(ctx context.Context, params tools.AgenticFetchParams, call fantasy.ToolCall) (fantasy.ToolResponse, error) {

70 validationResult, err := validateAgenticFetchParams(ctx, params)

69 validationResult, err := validateAgenticFetchParams(ctx, params)

71 if err != nil {

70 if err != nil {

@@ -80,8 +79,7 @@ func (c *coordinator) agenticFetchTool(_ context.Context, client *http.Client) (

80 description = "Search the web and analyze results"

79 description = "Search the web and analyze results"

81 }

80 }

83− p, err := c.permissions.Request(

82+ p, err := c.permissions.Request(ctx,

84− ctx,

85 permission.CreatePermissionRequest{

83 permission.CreatePermissionRequest{

86 SessionID: validationResult.SessionID,

84 SessionID: validationResult.SessionID,

87 Path: c.cfg.WorkingDir(),

85 Path: c.cfg.WorkingDir(),

@@ -96,10 +94,10 @@ func (c *coordinator) agenticFetchTool(_ context.Context, client *http.Client) (

96 return fantasy.ToolResponse{}, err

94 return fantasy.ToolResponse{}, err

97 }

95 }

98 if !p {

96 if !p {

99− return tools.NewPermissionDeniedResponse(), nil

97+ return fantasy.ToolResponse{}, permission.ErrorPermissionDenied

100 }

98 }

101

102− tmpDir, err := os.MkdirTemp(c.cfg.Config().Options.DataDirectory, "crush-fetch-*")

100+ tmpDir, err := os.MkdirTemp(c.cfg.Options.DataDirectory, "crush-fetch-*")

103 if err != nil {

101 if err != nil {

104 return fantasy.NewTextErrorResponse(fmt.Sprintf("Failed to create temporary directory: %s", err)), nil

102 return fantasy.NewTextErrorResponse(fmt.Sprintf("Failed to create temporary directory: %s", err)), nil

105 }

103 }

@@ -152,12 +150,12 @@ func (c *coordinator) agenticFetchTool(_ context.Context, client *http.Client) (

152 return fantasy.ToolResponse{}, fmt.Errorf("error building models: %s", err)

150 return fantasy.ToolResponse{}, fmt.Errorf("error building models: %s", err)

153 }

151 }

154

152

155− systemPrompt, err := promptTemplate.Build(ctx, small.Model.Provider(), small.Model.Model(), c.cfg)

153+ systemPrompt, err := promptTemplate.Build(ctx, small.Model.Provider(), small.Model.Model(), *c.cfg)

156 if err != nil {

154 if err != nil {

157 return fantasy.ToolResponse{}, fmt.Errorf("error building system prompt: %s", err)

155 return fantasy.ToolResponse{}, fmt.Errorf("error building system prompt: %s", err)

158 }

156 }

159

157

160− smallProviderCfg, ok := c.cfg.Config().Providers.Get(small.ModelCfg.Provider)

158+ smallProviderCfg, ok := c.cfg.Providers.Get(small.ModelCfg.Provider)

161 if !ok {

159 if !ok {

162 return fantasy.ToolResponse{}, errors.New("small model provider not configured")

160 return fantasy.ToolResponse{}, errors.New("small model provider not configured")

163 }

161 }

@@ -167,40 +165,69 @@ func (c *coordinator) agenticFetchTool(_ context.Context, client *http.Client) (

167 fetchTools := []fantasy.AgentTool{

165 fetchTools := []fantasy.AgentTool{

168 webFetchTool,

166 webFetchTool,

169 webSearchTool,

167 webSearchTool,

170− tools.NewGlobTool(tmpDir, c.cfg.Config().Tools.Glob),

168+ tools.NewGlobTool(tmpDir),

171− tools.NewGrepTool(tmpDir, c.cfg.Config().Tools.Grep),

169+ tools.NewGrepTool(tmpDir),

172 tools.NewSourcegraphTool(client),

170 tools.NewSourcegraphTool(client),

173− tools.NewViewTool(c.lspManager, c.permissions, c.filetracker, nil, tmpDir),

171+ tools.NewViewTool(c.lspClients, c.permissions, c.filetracker, tmpDir),

174 }

172 }

175

173

176− // Sub-agent tools run without hook interception. The top-level

177− // `agentic_fetch` call itself is already wrapped from the coder's

178− // side; firing hooks again for every inner tool call would run

179− // the user's hooks N times per delegated turn.

180−

181 agent := NewSessionAgent(SessionAgentOptions{

174 agent := NewSessionAgent(SessionAgentOptions{

182 LargeModel: small, // Use small model for both (fetch doesn't need large)

175 LargeModel: small, // Use small model for both (fetch doesn't need large)

183 SmallModel: small,

176 SmallModel: small,

184 SystemPromptPrefix: smallProviderCfg.SystemPromptPrefix,

177 SystemPromptPrefix: smallProviderCfg.SystemPromptPrefix,

185 SystemPrompt: systemPrompt,

178 SystemPrompt: systemPrompt,

186− DisableAutoSummarize: c.cfg.Config().Options.DisableAutoSummarize,

179+ DisableAutoSummarize: c.cfg.Options.DisableAutoSummarize,

187 IsYolo: c.permissions.SkipRequests(),

180 IsYolo: c.permissions.SkipRequests(),

188 Sessions: c.sessions,

181 Sessions: c.sessions,

189 Messages: c.messages,

182 Messages: c.messages,

190 Tools: fetchTools,

183 Tools: fetchTools,

191 })

184 })

192

185

193− return c.runSubAgent(ctx, subAgentParams{

186+ agentToolSessionID := c.sessions.CreateAgentToolSessionID(validationResult.AgentMessageID, call.ID)

194− Agent: agent,

187+ session, err := c.sessions.CreateTaskSession(ctx, agentToolSessionID, validationResult.SessionID, "Fetch Analysis")

195− SessionID: validationResult.SessionID,

188+ if err != nil {

196− AgentMessageID: validationResult.AgentMessageID,

189+ return fantasy.ToolResponse{}, fmt.Errorf("error creating session: %s", err)

197− ToolCallID: call.ID,

190+ }

198− Prompt: fullPrompt,

191+

199− SessionTitle: "Fetch Analysis",

200− SessionSetup: func(sessionID string) {

201− c.permissions.AutoApproveSession(sessionID)

192+ c.permissions.AutoApproveSession(session.ID)

193+

194+ // Use small model for web content analysis (faster and cheaper)

195+ maxTokens := small.CatwalkCfg.DefaultMaxTokens

196+ if small.ModelCfg.MaxTokens != 0 {

197+ maxTokens = small.ModelCfg.MaxTokens

202− },

198+ }

199+

200+ result, err := agent.Run(ctx, SessionAgentCall{

201+ SessionID: session.ID,

202+ Prompt: fullPrompt,

203+ MaxOutputTokens: maxTokens,

204+ ProviderOptions: getProviderOptions(small, smallProviderCfg),

205+ Temperature: small.ModelCfg.Temperature,

206+ TopP: small.ModelCfg.TopP,

207+ TopK: small.ModelCfg.TopK,

208+ FrequencyPenalty: small.ModelCfg.FrequencyPenalty,

209+ PresencePenalty: small.ModelCfg.PresencePenalty,

203 })

210 })

204− },

211+ if err != nil {

212+ return fantasy.NewTextErrorResponse("error generating response"), nil

213+ }

214+

215+ updatedSession, err := c.sessions.Get(ctx, session.ID)

216+ if err != nil {

217+ return fantasy.ToolResponse{}, fmt.Errorf("error getting session: %s", err)

218+ }

219+ parentSession, err := c.sessions.Get(ctx, validationResult.SessionID)

220+ if err != nil {

221+ return fantasy.ToolResponse{}, fmt.Errorf("error getting parent session: %s", err)

222+ }

223+

224+ parentSession.Cost += updatedSession.Cost

225+

226+ _, err = c.sessions.Save(ctx, parentSession)

227+ if err != nil {

228+ return fantasy.ToolResponse{}, fmt.Errorf("error saving parent session: %s", err)

229+ }

230+

231+ return fantasy.NewTextResponse(result.Response.Content.Text()), nil

205− ), nil

232+ }), nil

206}

233}

internal/agent/agenttest/coordinator.godeleted−80

@@ -1,80 +0,0 @@

1−// Package agenttest provides test-only constructors for wiring a real

2−// production agent.Coordinator without booting a full app.App. It is

3−// imported only from _test.go files (e.g. internal/backend integration

4−// tests) and is never referenced by production code, so it is compiled

5−// only under tests and never ships in the production binary or API.

6−package agenttest

7−

8−import (

9− "context"

10−

11− "charm.land/catwalk/pkg/catwalk"

12− "charm.land/fantasy/providers/openaicompat"

13− "github.com/charmbracelet/crush/internal/agent"

14− "github.com/charmbracelet/crush/internal/config"

15− "github.com/charmbracelet/crush/internal/message"

16− "github.com/charmbracelet/crush/internal/permission"

17− "github.com/charmbracelet/crush/internal/session"

18−)

19−

20−// NewCoordinator builds a real agent.Coordinator through the production

21−// agent.NewCoordinator constructor so the RunAccepted / BeginAccepted /

22−// run path (including UpdateModels) is the actual code under test.

23−//

24−// It installs a minimal config with a single openai-compatible provider

25−// whose model resolves offline. run rebuilds the model on every call, so

26−// the provider must construct without network I/O; the cancel-on-entry

27−// path this helper is built to exercise returns before any model call,

28−// so no request is ever issued. The coder agent's allowed-tools list is

29−// cleared to keep tool construction cheap and free of sub-agent wiring.

30−//

31−// The optional coordinator dependencies (history, filetracker, LSP,

32−// notify, runComplete, skills) are nil: run guards the publisher fields

33−// and the cancel-on-entry path never touches the others.

34−func NewCoordinator(

35− ctx context.Context,

36− workingDir string,

37− sessions session.Service,

38− messages message.Service,

39−) (agent.Coordinator, error) {

40− cfg, err := config.Init(workingDir, "", false)

41− if err != nil {

42− return nil, err

43− }

44−

45− const (

46− providerID = "test-openai-compat"

47− modelID = "test-model"

48− )

49− cfg.Config().Providers.Set(providerID, config.ProviderConfig{

50− ID: providerID,

51− Name: "Test",

52− Type: openaicompat.Name,

53− BaseURL: "http://127.0.0.1:0/v1",

54− APIKey: "test",

55− Models: []catwalk.Model{{ID: modelID, DefaultMaxTokens: 4096}},

56− })

57− selected := config.SelectedModel{Provider: providerID, Model: modelID}

58− cfg.OverridePreferredModel(config.SelectedModelTypeLarge, selected)

59− cfg.OverridePreferredModel(config.SelectedModelTypeSmall, selected)

60− cfg.SetupAgents()

61−

62− // Keep buildTools light: no sub-agent or agentic-fetch construction.

63− coderCfg := cfg.Config().Agents[config.AgentCoder]

64− coderCfg.AllowedTools = nil

65− cfg.Config().Agents[config.AgentCoder] = coderCfg

66−

67− return agent.NewCoordinator(

68− ctx,

69− cfg,

70− sessions,

71− messages,

72− permission.NewPermissionService(workingDir, true, nil),

73− nil,

74− nil,

75− nil,

76− nil,

77− nil,

78− nil,

79− )

80−}

internal/agent/common_test.go+65−30

@@ -8,10 +8,13 @@ import (

8 "testing"

9 "time"

11− "charm.land/catwalk/pkg/catwalk"

12 "charm.land/fantasy"

11 "charm.land/fantasy"

12+ "charm.land/fantasy/providers/anthropic"

13+ "charm.land/fantasy/providers/openai"

13 "charm.land/fantasy/providers/openaicompat"

14 "charm.land/fantasy/providers/openaicompat"

15+ "charm.land/fantasy/providers/openrouter"

14 "charm.land/x/vcr"

16 "charm.land/x/vcr"

17+ "github.com/charmbracelet/catwalk/pkg/catwalk"

15 "github.com/charmbracelet/crush/internal/agent/prompt"

18 "github.com/charmbracelet/crush/internal/agent/prompt"

16 "github.com/charmbracelet/crush/internal/agent/tools"

19 "github.com/charmbracelet/crush/internal/agent/tools"

17 "github.com/charmbracelet/crush/internal/config"

20 "github.com/charmbracelet/crush/internal/config"

@@ -47,11 +50,50 @@ type modelPair struct {

47 smallModel builderFunc

50 smallModel builderFunc

48}

51}

53+func anthropicBuilder(model string) builderFunc {

54+ return func(t *testing.T, r *vcr.Recorder) (fantasy.LanguageModel, error) {

55+ provider, err := anthropic.New(

56+ anthropic.WithAPIKey(os.Getenv("CRUSH_ANTHROPIC_API_KEY")),

57+ anthropic.WithHTTPClient(&http.Client{Transport: r}),

58+ )

59+ if err != nil {

60+ return nil, err

61+ }

62+ return provider.LanguageModel(t.Context(), model)

63+ }

64+}

65+

66+func openaiBuilder(model string) builderFunc {

67+ return func(t *testing.T, r *vcr.Recorder) (fantasy.LanguageModel, error) {

68+ provider, err := openai.New(

69+ openai.WithAPIKey(os.Getenv("CRUSH_OPENAI_API_KEY")),

70+ openai.WithHTTPClient(&http.Client{Transport: r}),

71+ )

72+ if err != nil {

73+ return nil, err

74+ }

75+ return provider.LanguageModel(t.Context(), model)

76+ }

77+}

78+

79+func openRouterBuilder(model string) builderFunc {

80+ return func(t *testing.T, r *vcr.Recorder) (fantasy.LanguageModel, error) {

81+ provider, err := openrouter.New(

82+ openrouter.WithAPIKey(os.Getenv("CRUSH_OPENROUTER_API_KEY")),

83+ openrouter.WithHTTPClient(&http.Client{Transport: r}),

84+ )

85+ if err != nil {

86+ return nil, err

87+ }

88+ return provider.LanguageModel(t.Context(), model)

89+ }

90+}

91+

50−func hyperBuilder(model string) builderFunc {

92+func zAIBuilder(model string) builderFunc {

51 return func(t *testing.T, r *vcr.Recorder) (fantasy.LanguageModel, error) {

93 return func(t *testing.T, r *vcr.Recorder) (fantasy.LanguageModel, error) {

52 provider, err := openaicompat.New(

94 provider, err := openaicompat.New(

53− openaicompat.WithBaseURL("https://hyper.charm.land/v1"),

95+ openaicompat.WithBaseURL("https://api.z.ai/api/coding/paas/v4"),

54− openaicompat.WithAPIKey(os.Getenv("CRUSH_HYPER_API_KEY")),

96+ openaicompat.WithAPIKey(os.Getenv("CRUSH_ZAI_API_KEY")),

55 openaicompat.WithHTTPClient(&http.Client{Transport: r}),

97 openaicompat.WithHTTPClient(&http.Client{Transport: r}),

56 )

98 )

57 if err != nil {

99 if err != nil {

@@ -111,15 +153,7 @@ func testSessionAgent(env fakeEnv, large, small fantasy.LanguageModel, systemPro

111 DefaultMaxTokens: 10000,

153 DefaultMaxTokens: 10000,

112 },

154 },

113 }

155 }

114− agent := NewSessionAgent(SessionAgentOptions{

156+ agent := NewSessionAgent(SessionAgentOptions{largeModel, smallModel, "", systemPrompt, false, false, true, env.sessions, env.messages, tools})

115− LargeModel: largeModel,

116− SmallModel: smallModel,

117− SystemPrompt: systemPrompt,

118− IsYolo: true,

119− Sessions: env.sessions,

120− Messages: env.messages,

121− Tools: tools,

122− })

123 return agent

157 return agent

124}

158}

125

159

@@ -143,41 +177,42 @@ func coderAgent(r *vcr.Recorder, env fakeEnv, large, small fantasy.LanguageModel

143

177

144 // NOTE(@andreynering): Set a fixed config to ensure cassettes match

178 // NOTE(@andreynering): Set a fixed config to ensure cassettes match

145 // independently of user config on `$HOME/.config/crush/crush.json`.

179 // independently of user config on `$HOME/.config/crush/crush.json`.

146− cfg.Config().Options.Attribution = &config.Attribution{

180+ cfg.Options.Attribution = &config.Attribution{

147 TrailerStyle: "co-authored-by",

181 TrailerStyle: "co-authored-by",

148 GeneratedWith: true,

182 GeneratedWith: true,

149 }

183 }

150

184

151− // Clear some fields to avoid issues with VCR cassette matching.

185+ // Clear skills paths to ensure test reproducibility - user's skills

186+ // would be included in prompt and break VCR cassette matching.

152− cfg.Config().Options.SkillsPaths = nil

187+ cfg.Options.SkillsPaths = []string{}

153− cfg.Config().Options.DisabledSkills = []string{"crush-config"}

188+

154− cfg.Config().Options.ContextPaths = nil

189+ // Clear LSP config to ensure test reproducibility - user's LSP config

155− cfg.Config().Options.GlobalContextPaths = nil

190+ // would be included in prompt and break VCR cassette matching.

156− cfg.Config().LSP = nil

191+ cfg.LSP = nil

157

192

158− systemPrompt, err := prompt.Build(context.TODO(), large.Provider(), large.Model(), cfg)

193+ systemPrompt, err := prompt.Build(context.TODO(), large.Provider(), large.Model(), *cfg)

159 if err != nil {

194 if err != nil {

160 return nil, err

195 return nil, err

161 }

196 }

162

197

163 // Get the model name for the bash tool

198 // Get the model name for the bash tool

164 modelName := large.Model() // fallback to ID if Name not available

199 modelName := large.Model() // fallback to ID if Name not available

165− if model := cfg.Config().GetModel(large.Provider(), large.Model()); model != nil {

200+ if model := cfg.GetModel(large.Provider(), large.Model()); model != nil {

166 modelName = model.Name

201 modelName = model.Name

167 }

202 }

168

203

169 allTools := []fantasy.AgentTool{

204 allTools := []fantasy.AgentTool{

170− tools.NewBashTool(env.permissions, env.workingDir, cfg.Config().Options.Attribution, modelName),

205+ tools.NewBashTool(env.permissions, env.workingDir, cfg.Options.Attribution, modelName),

171 tools.NewDownloadTool(env.permissions, env.workingDir, r.GetDefaultClient()),

206 tools.NewDownloadTool(env.permissions, env.workingDir, r.GetDefaultClient()),

172− tools.NewEditTool(nil, env.permissions, env.history, *env.filetracker, env.workingDir),

207+ tools.NewEditTool(env.lspClients, env.permissions, env.history, *env.filetracker, env.workingDir),

173− tools.NewMultiEditTool(nil, env.permissions, env.history, *env.filetracker, env.workingDir),

208+ tools.NewMultiEditTool(env.lspClients, env.permissions, env.history, *env.filetracker, env.workingDir),

174 tools.NewFetchTool(env.permissions, env.workingDir, r.GetDefaultClient()),

209 tools.NewFetchTool(env.permissions, env.workingDir, r.GetDefaultClient()),

175− tools.NewGlobTool(env.workingDir, cfg.Config().Tools.Glob),

210+ tools.NewGlobTool(env.workingDir),

176− tools.NewGrepTool(env.workingDir, cfg.Config().Tools.Grep),

211+ tools.NewGrepTool(env.workingDir),

177− tools.NewLsTool(env.permissions, env.workingDir, cfg.Config().Tools.Ls),

212+ tools.NewLsTool(env.permissions, env.workingDir, cfg.Tools.Ls),

178 tools.NewSourcegraphTool(r.GetDefaultClient()),

213 tools.NewSourcegraphTool(r.GetDefaultClient()),

179− tools.NewViewTool(nil, env.permissions, *env.filetracker, nil, env.workingDir),

214+ tools.NewViewTool(env.lspClients, env.permissions, *env.filetracker, env.workingDir),

180− tools.NewWriteTool(nil, env.permissions, env.history, *env.filetracker, env.workingDir),

215+ tools.NewWriteTool(env.lspClients, env.permissions, env.history, *env.filetracker, env.workingDir),

181 }

216 }

182

217

183 return testSessionAgent(env, large, small, systemPrompt, allTools...), nil

218 return testSessionAgent(env, large, small, systemPrompt, allTools...), nil

internal/agent/convert_test.godeleted−94

@@ -1,94 +0,0 @@

1−package agent

2−

3−import (

4− "encoding/base64"

5− "testing"

6−

7− "charm.land/fantasy"

8− "github.com/stretchr/testify/require"

9−)

10−

11−func TestConvertToToolResult_InvalidBase64(t *testing.T) {

12− t.Parallel()

13−

14− a := &sessionAgent{}

15− result := fantasy.ToolResultContent{

16− ToolCallID: "call_123",

17− ToolName: "test_tool",

18− Result: fantasy.ToolResultOutputContentMedia{

19− Data: "abc\x80def",

20− MediaType: "image/png",

21− },

22− }

23−

24− tr := a.convertToToolResult(result)

25− require.True(t, tr.IsError)

26− require.Empty(t, tr.Data)

27− require.Contains(t, tr.Content, "invalid encoding")

28− require.Equal(t, "call_123", tr.ToolCallID)

29− require.Equal(t, "test_tool", tr.Name)

30−}

31−

32−func TestConvertToToolResult_ValidMedia(t *testing.T) {

33− t.Parallel()

34−

35− a := &sessionAgent{}

36− validData := base64.StdEncoding.EncodeToString([]byte("test image data"))

37−

38− result := fantasy.ToolResultContent{

39− ToolCallID: "call_456",

40− ToolName: "screenshot",

41− Result: fantasy.ToolResultOutputContentMedia{

42− Data: validData,

43− MediaType: "image/png",

44− Text: "Screenshot captured",

45− },

46− }

47−

48− tr := a.convertToToolResult(result)

49− require.False(t, tr.IsError)

50− require.Equal(t, validData, tr.Data)

51− require.Equal(t, "image/png", tr.MIMEType)

52− require.Equal(t, "Screenshot captured", tr.Content)

53−}

54−

55−func TestConvertToToolResult_ValidMediaNoText(t *testing.T) {

56− t.Parallel()

57−

58− a := &sessionAgent{}

59− validData := base64.StdEncoding.EncodeToString([]byte("test image data"))

60−

61− result := fantasy.ToolResultContent{

62− ToolCallID: "call_789",

63− ToolName: "view",

64− Result: fantasy.ToolResultOutputContentMedia{

65− Data: validData,

66− MediaType: "image/jpeg",

67− },

68− }

69−

70− tr := a.convertToToolResult(result)

71− require.False(t, tr.IsError)

72− require.Equal(t, validData, tr.Data)

73− require.Equal(t, "image/jpeg", tr.MIMEType)

74− require.Equal(t, "Loaded image/jpeg content", tr.Content)

75−}

76−

77−func TestConvertToToolResult_ASCIIButInvalidBase64(t *testing.T) {

78− t.Parallel()

79−

80− a := &sessionAgent{}

81− result := fantasy.ToolResultContent{

82− ToolCallID: "call_abc",

83− ToolName: "mcp_tool",

84− Result: fantasy.ToolResultOutputContentMedia{

85− Data: "not-valid-base64!!!",

86− MediaType: "image/png",

87− },

88− }

89−

90− tr := a.convertToToolResult(result)

91− require.True(t, tr.IsError)

92− require.Empty(t, tr.Data)

93− require.Contains(t, tr.Content, "invalid encoding")

94−}

internal/agent/coordinator.go+159−944

@@ -12,34 +12,24 @@ import (

12 "maps"

13 "net/http"

14 "os"

15− "path/filepath"

16 "slices"

15 "slices"

17 "strings"

16 "strings"

18− "sync"

19− "sync/atomic"

21− "charm.land/catwalk/pkg/catwalk"

22 "charm.land/fantasy"

18 "charm.land/fantasy"

19+ "github.com/charmbracelet/catwalk/pkg/catwalk"

23 "github.com/charmbracelet/crush/internal/agent/hyper"

20 "github.com/charmbracelet/crush/internal/agent/hyper"

24− "github.com/charmbracelet/crush/internal/agent/notify"

25 "github.com/charmbracelet/crush/internal/agent/prompt"

21 "github.com/charmbracelet/crush/internal/agent/prompt"

26 "github.com/charmbracelet/crush/internal/agent/tools"

22 "github.com/charmbracelet/crush/internal/agent/tools"

27− "github.com/charmbracelet/crush/internal/agent/tools/mcp"

28 "github.com/charmbracelet/crush/internal/config"

23 "github.com/charmbracelet/crush/internal/config"

29 "github.com/charmbracelet/crush/internal/csync"

24 "github.com/charmbracelet/crush/internal/csync"

30− "github.com/charmbracelet/crush/internal/discover"

31− "github.com/charmbracelet/crush/internal/event"

32 "github.com/charmbracelet/crush/internal/filetracker"

25 "github.com/charmbracelet/crush/internal/filetracker"

33 "github.com/charmbracelet/crush/internal/history"

26 "github.com/charmbracelet/crush/internal/history"

34− "github.com/charmbracelet/crush/internal/hooks"

35 "github.com/charmbracelet/crush/internal/log"

27 "github.com/charmbracelet/crush/internal/log"

36 "github.com/charmbracelet/crush/internal/lsp"

28 "github.com/charmbracelet/crush/internal/lsp"

37 "github.com/charmbracelet/crush/internal/message"

29 "github.com/charmbracelet/crush/internal/message"

38 "github.com/charmbracelet/crush/internal/oauth/copilot"

30 "github.com/charmbracelet/crush/internal/oauth/copilot"

39 "github.com/charmbracelet/crush/internal/permission"

31 "github.com/charmbracelet/crush/internal/permission"

40− "github.com/charmbracelet/crush/internal/pubsub"

41 "github.com/charmbracelet/crush/internal/session"

32 "github.com/charmbracelet/crush/internal/session"

42− "github.com/charmbracelet/crush/internal/skills"

43 "golang.org/x/sync/errgroup"

33 "golang.org/x/sync/errgroup"

45 "charm.land/fantasy/providers/anthropic"

35 "charm.land/fantasy/providers/anthropic"

@@ -49,52 +39,14 @@ import (

49 "charm.land/fantasy/providers/openai"

39 "charm.land/fantasy/providers/openai"

50 "charm.land/fantasy/providers/openaicompat"

40 "charm.land/fantasy/providers/openaicompat"

51 "charm.land/fantasy/providers/openrouter"

41 "charm.land/fantasy/providers/openrouter"

52− "charm.land/fantasy/providers/vercel"

53− openaisdk "github.com/charmbracelet/openai-go/option"

42+ openaisdk "github.com/openai/openai-go/v2/option"

54 "github.com/qjebbs/go-jsons"

43 "github.com/qjebbs/go-jsons"

55)

44)

57−// Coordinator errors.

58−var (

59− errCoderAgentNotConfigured = errors.New("coder agent not configured")

60− errModelProviderNotConfigured = errors.New("model provider not configured")

61− errLargeModelNotSelected = errors.New("large model not selected")

62− errSmallModelNotSelected = errors.New("small model not selected")

63− errLargeModelProviderNotConfigured = errors.New("large model provider not configured")

64− errSmallModelProviderNotConfigured = errors.New("small model provider not configured")

65− errLargeModelNotFound = errors.New("large model not found in provider config")

66− errSmallModelNotFound = errors.New("small model not found in provider config")

67−)

68−

69−// Copilot models that use the Responses API instead of Chat Completions.

70−var copilotResponsesModels = map[string]bool{

71− "gpt-5.2": true,

72− "gpt-5.2-codex": true,

73− "gpt-5.3-codex": true,

74− "gpt-5.4": true,

75− "gpt-5.4-mini": true,

76− "gpt-5.5": true,

77− "gpt-5-mini": true,

78−}

79−

80−// OpenCode models that user Anthropic Messages API instead of Chat Completions.

81−var opencodeMessagesModels = map[string]bool{

82− "qwen3.7-max": true,

83−}

84−

85type Coordinator interface {

46type Coordinator interface {

86 // INFO: (kujtim) this is not used yet we will use this when we have multiple agents

47 // INFO: (kujtim) this is not used yet we will use this when we have multiple agents

87 // SetMainAgent(string)

48 // SetMainAgent(string)

88 Run(ctx context.Context, sessionID, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error)

49 Run(ctx context.Context, sessionID, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error)

89− // RunAccepted runs a call that was already accepted via

90− // BeginAccepted on the fire-and-forget dispatch path. The handle is

91− // the only carrier of accept-state across the backend.runAgent /

92− // Coordinator / sessionAgent.Run layers: it reaches

93− // sessionAgent.Run as SessionAgentCall.Accepted, where it is

94− // consumed under dispatchMu once the accepted -> (cancel-on-entry |

95− // queued | active) transition is chosen.

96− RunAccepted(ctx context.Context, accept *AcceptedRun, sessionID, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error)

97− BeginAccepted(sessionID string) *AcceptedRun

98 Cancel(sessionID string)

50 Cancel(sessionID string)

99 CancelAll()

51 CancelAll()

100 IsSessionBusy(sessionID string) bool

52 IsSessionBusy(sessionID string) bool

@@ -105,104 +57,47 @@ type Coordinator interface {

105 Summarize(context.Context, string) error

57 Summarize(context.Context, string) error

106 Model() Model

58 Model() Model

107 UpdateModels(ctx context.Context) error

59 UpdateModels(ctx context.Context) error

108− GenerateTitle(ctx context.Context, sessionID, prompt string)

109}

60}

110

111type coordinator struct {

62type coordinator struct {

112− cfg *config.ConfigStore

63+ cfg *config.Config

113 sessions session.Service

64 sessions session.Service

114 messages message.Service

65 messages message.Service

115 permissions permission.Service

66 permissions permission.Service

116 history history.Service

67 history history.Service

117 filetracker filetracker.Service

68 filetracker filetracker.Service

118− lspManager *lsp.Manager

69+ lspClients *csync.Map[string, *lsp.Client]

119− notify pubsub.Publisher[notify.Notification]

120− runComplete pubsub.Publisher[notify.RunComplete]

121

122 currentAgent SessionAgent

71 currentAgent SessionAgent

123 agents map[string]SessionAgent

72 agents map[string]SessionAgent

124

125− // Skills discovery results (session-start snapshot).

126− allSkills []*skills.Skill // Pre-filter: all discovered after dedup.

127− activeSkills []*skills.Skill // Post-filter: active skills only.

128− skillTracker *skills.Tracker

129−

130− // providerClients caches built provider clients keyed by their fully

131− // resolved configuration. UpdateModels rebuilds models every turn;

132− // without this, each turn discards the provider's HTTP client and its

133− // keep-alive connection pool, forcing a fresh TLS handshake. The key

134− // captures every input buildProvider depends on, so any config/model/

135− // credential change misses the cache and rebuilds.

136− providerClients *csync.Map[string, fantasy.Provider]

137−

138− // toolsCache avoids rebuilding the tool set on every UpdateModels call.

139− // The cache is invalidated when the config pointer changes (reload or

140− // mutation) or when mcpToolsGen is bumped (MCP server connect/disconnect

141− // or tools/list_changed notification).

142− toolsCacheMu sync.Mutex

143− toolsCache []fantasy.AgentTool

144− toolsCacheCfg *config.Config

145− toolsCacheMCP uint64

146− mcpToolsGen atomic.Uint64

147−

148− // providerOptsCache avoids repeating the 3× JSON marshal + merge +

149− // unmarshal in getProviderOptions on every turn when the model and

150− // config have not changed. Keyed by config pointer + model CatwalkCfg

151− // ID + provider name.

152− providerOptsCacheMu sync.Mutex

153− providerOptsCache fantasy.ProviderOptions

154− providerOptsCacheKey string

155−

156 readyWg errgroup.Group

74 readyWg errgroup.Group

157}

75}

158

159func NewCoordinator(

77func NewCoordinator(

160 ctx context.Context,

78 ctx context.Context,

161− cfg *config.ConfigStore,

79+ cfg *config.Config,

162 sessions session.Service,

80 sessions session.Service,

163 messages message.Service,

81 messages message.Service,

164 permissions permission.Service,

82 permissions permission.Service,

165 history history.Service,

83 history history.Service,

166 filetracker filetracker.Service,

84 filetracker filetracker.Service,

167− lspManager *lsp.Manager,

85+ lspClients *csync.Map[string, *lsp.Client],

168− notify pubsub.Publisher[notify.Notification],

169− runComplete pubsub.Publisher[notify.RunComplete],

170− skillsMgr *skills.Manager,

171) (Coordinator, error) {

86) (Coordinator, error) {

172− // Skills are pre-discovered by the caller (see app.New /

173− // backend.CreateWorkspace) and passed in via the manager. If no

174− // manager was provided (legacy callers), fall back to an in-line

175− // discovery so the coordinator still works.

176− var allSkills, activeSkills []*skills.Skill

177− if skillsMgr != nil {

178− allSkills = skillsMgr.AllSkills()

179− activeSkills = skillsMgr.ActiveSkills()

180− } else {

181− allSkills, activeSkills = discoverSkills(cfg)

182− }

183− skillTracker := skills.NewTracker(activeSkills)

184−

185 c := &coordinator{

87 c := &coordinator{

186− cfg: cfg,

88+ cfg: cfg,

187− sessions: sessions,

89+ sessions: sessions,

188− messages: messages,

90+ messages: messages,

189− permissions: permissions,

91+ permissions: permissions,

190− history: history,

92+ history: history,

191− filetracker: filetracker,

93+ filetracker: filetracker,

192− lspManager: lspManager,

94+ lspClients: lspClients,

193− notify: notify,

194− runComplete: runComplete,

195− agents: make(map[string]SessionAgent),

95+ agents: make(map[string]SessionAgent),

196− allSkills: allSkills,

197− activeSkills: activeSkills,

198− skillTracker: skillTracker,

199−

200− providerClients: csync.NewMap[string, fantasy.Provider](),

201− }

96+ }

202−

97+

203− agentCfg, ok := cfg.Config().Agents[config.AgentCoder]

98+ agentCfg, ok := cfg.Agents[config.AgentCoder]

204 if !ok {

99 if !ok {

205− return nil, errCoderAgentNotConfigured

100+ return nil, errors.New("coder agent not configured")

206 }

101 }

207

102

208 // TODO: make this dynamic when we support multiple agents

103 // TODO: make this dynamic when we support multiple agents

@@ -217,36 +112,11 @@ func NewCoordinator(

217 }

112 }

218 c.currentAgent = agent

113 c.currentAgent = agent

219 c.agents[config.AgentCoder] = agent

114 c.agents[config.AgentCoder] = agent

220−

221− // Subscribe to MCP tool-list changes so the tools cache is invalidated

222− // when MCP servers connect/disconnect or announce new tools.

223− go func() {

224− for ev := range mcp.SubscribeEvents(ctx) {

225− if ev.Payload.Type == mcp.EventToolsListChanged || ev.Payload.Type == mcp.EventStateChanged {

226− c.mcpToolsGen.Add(1)

227− }

228− }

229− }()

230−

231 return c, nil

115 return c, nil

232}

116}

233

117

234// Run implements Coordinator.

118// Run implements Coordinator.

235func (c *coordinator) Run(ctx context.Context, sessionID string, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error) {

119func (c *coordinator) Run(ctx context.Context, sessionID string, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error) {

236− return c.run(ctx, nil, sessionID, prompt, attachments...)

237−}

238−

239−// RunAccepted implements Coordinator.

240−func (c *coordinator) RunAccepted(ctx context.Context, accept *AcceptedRun, sessionID string, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error) {

241− return c.run(ctx, accept, sessionID, prompt, attachments...)

242−}

243−

244−// run is the shared implementation behind Run and RunAccepted. When

245−// accept is non-nil it is threaded onto the SessionAgentCall as

246−// Accepted so sessionAgent.Run can consume the accept reservation under

247−// dispatchMu; when nil (the in-process/local path) no accept tracking

248−// applies.

249−func (c *coordinator) run(ctx context.Context, accept *AcceptedRun, sessionID string, prompt string, attachments ...message.Attachment) (*fantasy.AgentResult, error) {

250 if err := c.readyWg.Wait(); err != nil {

120 if err := c.readyWg.Wait(); err != nil {

251 return nil, err

121 return nil, err

252 }

122 }

@@ -273,47 +143,23 @@ func (c *coordinator) run(ctx context.Context, accept *AcceptedRun, sessionID st

273 attachments = filteredAttachments

143 attachments = filteredAttachments

274 }

144 }

275

145

276− providerCfg, ok := c.cfg.Config().Providers.Get(model.ModelCfg.Provider)

146+ providerCfg, ok := c.cfg.Providers.Get(model.ModelCfg.Provider)

277 if !ok {

147 if !ok {

278− return nil, errModelProviderNotConfigured

148+ return nil, errors.New("model provider not configured")

279 }

149 }

280

150

281− mergedOptions, temp, topP, topK, freqPenalty, presPenalty := c.cachedMergeCallOptions(model, providerCfg)

151+ mergedOptions, temp, topP, topK, freqPenalty, presPenalty := mergeCallOptions(model, providerCfg)

282

152

153+ if providerCfg.OAuthToken != nil && providerCfg.OAuthToken.IsExpired() {

154+ slog.Debug("Token needs to be refreshed", "provider", providerCfg.ID)

283− if err := c.refreshTokenIfExpired(ctx, providerCfg); err != nil {

155+ if err := c.refreshOAuth2Token(ctx, providerCfg); err != nil {

284− // NOTE(@andreynering): We don't return here because the event handling to ask the user to reauthenticate

156+ return nil, err

285− // depends on the flow below. If refresh fails, proceed with the token we have.

157+ }

286− slog.Error("Failed to refresh OAuth2 token. Proceeding with existing token.", "error", err)

287 }

158 }

288

159

289− // Coalesce per-attempt RunComplete payloads so only the final

290− // outcome reaches subscribers. Without this, the first attempt's

291− // failed RunComplete (unauthorized) would race ahead of the

292− // retry's success, and `crush run` would exit on the stale error

293− // before ever seeing the retry result. Each attempt's

294− // SessionAgentCall.OnComplete hook overwrites latest; we publish

295− // exactly once after retries resolve, via PublishMustDeliver, so

296− // a momentarily-full subscriber buffer can't silently drop the

297− // terminal event.

298− var (

299− latest notify.RunComplete

300− hasLatest bool

301− )

302− onComplete := func(rc notify.RunComplete) {

303− latest = rc

304− hasLatest = true

305− }

306− // Propagate the caller-supplied RunID (set via agent.WithRunID

307− // at the HTTP boundary in backend.SendMessage) onto the

308− // SessionAgentCall so the terminal RunComplete event echoes it

309− // back. Both attempts in the retry chain reuse the same RunID;

310− // the coalesce closure publishes the final outcome under that

311− // same correlator.

312− runID := RunIDFromContext(ctx)

313 run := func() (*fantasy.AgentResult, error) {

160 run := func() (*fantasy.AgentResult, error) {

314 return c.currentAgent.Run(ctx, SessionAgentCall{

161 return c.currentAgent.Run(ctx, SessionAgentCall{

315 SessionID: sessionID,

162 SessionID: sessionID,

316− RunID: runID,

317 Prompt: prompt,

163 Prompt: prompt,

318 Attachments: attachments,

164 Attachments: attachments,

319 MaxOutputTokens: maxTokens,

165 MaxOutputTokens: maxTokens,

@@ -323,57 +169,30 @@ func (c *coordinator) run(ctx context.Context, accept *AcceptedRun, sessionID st

323 TopK: topK,

169 TopK: topK,

324 FrequencyPenalty: freqPenalty,

170 FrequencyPenalty: freqPenalty,

325 PresencePenalty: presPenalty,

171 PresencePenalty: presPenalty,

326− OnComplete: onComplete,

327− Accepted: accept,

328− })

329− }

330− beforeLoaded := c.skillTracker.LoadedNames()

331− var result *fantasy.AgentResult

332− originalErr := c.runWithUnauthorizedRetry(ctx, providerCfg, func() error {

333− var err error

334− result, err = run()

335− return err

336− })

337− logTurnSkillUsage(sessionID, prompt, c.activeSkills, c.skillTracker, beforeLoaded)

338−

339− // Notify only if still unauthorized after retry — a successful

340− // retry means the user doesn't need to re-authenticate.

341− if originalErr != nil && c.isUnauthorized(originalErr) && c.notify != nil && model.ModelCfg.Provider == hyper.Name {

342− c.notify.Publish(pubsub.CreatedEvent, notify.Notification{

343− Type: notify.TypeReAuthenticate,

344− ProviderID: model.ModelCfg.Provider,

345 })

172 })

346 }

173 }

174+ result, originalErr := run()

347

175

348− if hasLatest && c.runComplete != nil {

176+ if c.isUnauthorized(originalErr) {

349− c.runComplete.PublishMustDeliver(ctx, pubsub.UpdatedEvent, latest)

177+ switch {

350− // Signal to the dispatcher (backend.runAgent) that the

178+ case providerCfg.OAuthToken != nil:

351− // authoritative terminal RunComplete for this run was already

179+ slog.Debug("Received 401. Refreshing token and retrying", "provider", providerCfg.ID)

352− // emitted, so it does not publish a duplicate fallback for the

180+ if err := c.refreshOAuth2Token(ctx, providerCfg); err != nil {

353− // error it is about to receive.

181+ return nil, originalErr

354− MarkRunCompletePublished(ctx)

182+ }

355− }

183+ slog.Debug("Retrying request with refreshed OAuth token", "provider", providerCfg.ID)

356− return result, originalErr

184+ return run()

357−}

185+ case strings.Contains(providerCfg.APIKeyTemplate, "$"):

358−

186+ slog.Debug("Received 401. Refreshing API Key template and retrying", "provider", providerCfg.ID)

359−// effectiveReasoningEffort returns the reasoning effort to apply for provider calls.

187+ if err := c.refreshApiKeyTemplate(ctx, providerCfg); err != nil {

360−// It prefers the user-selected effort when valid, otherwise the model default when

188+ return nil, originalErr

361−// valid, and finally falls back to the first configured reasoning level.

189+ }

362−func effectiveReasoningEffort(model Model) string {

190+ slog.Debug("Retrying request with refreshed API key", "provider", providerCfg.ID)

363− if !model.CatwalkCfg.CanReason {

191+ return run()

364− return ""

192+ }

365 }

193 }

366

194

367− if effort := model.ModelCfg.ReasoningEffort; effort != "" && slices.Contains(model.CatwalkCfg.ReasoningLevels, effort) {

195+ return result, originalErr

368− return effort

369− }

370− if effort := model.CatwalkCfg.DefaultReasoningEffort; effort != "" && slices.Contains(model.CatwalkCfg.ReasoningLevels, effort) {

371− return effort

372− }

373− if len(model.CatwalkCfg.ReasoningLevels) > 0 {

374− return model.CatwalkCfg.ReasoningLevels[0]

375− }

376− return ""

377}

196}

378

197

379func getProviderOptions(model Model, providerCfg config.ProviderConfig) fantasy.ProviderOptions {

198func getProviderOptions(model Model, providerCfg config.ProviderConfig) fantasy.ProviderOptions {

@@ -424,16 +243,11 @@ func getProviderOptions(model Model, providerCfg config.ProviderConfig) fantasy.

424 return options

243 return options

425 }

244 }

426

245

427− reasoningEffort := effectiveReasoningEffort(model)

428− shouldSetEffort := model.CatwalkCfg.CanReason &&

429− reasoningEffort != "" &&

430− slices.Contains(model.CatwalkCfg.ReasoningLevels, reasoningEffort)

431−

432 switch providerCfg.Type {

246 switch providerCfg.Type {

433 case openai.Name, azure.Name:

247 case openai.Name, azure.Name:

434 _, hasReasoningEffort := mergedOptions["reasoning_effort"]

248 _, hasReasoningEffort := mergedOptions["reasoning_effort"]

435− if !hasReasoningEffort && shouldSetEffort {

249+ if !hasReasoningEffort && model.ModelCfg.ReasoningEffort != "" {

436− mergedOptions["reasoning_effort"] = reasoningEffort

250+ mergedOptions["reasoning_effort"] = model.ModelCfg.ReasoningEffort

437 }

251 }

438 if openai.IsResponsesModel(model.CatwalkCfg.ID) {

252 if openai.IsResponsesModel(model.CatwalkCfg.ID) {

439 if openai.IsResponsesReasoningModel(model.CatwalkCfg.ID) {

253 if openai.IsResponsesReasoningModel(model.CatwalkCfg.ID) {

@@ -450,37 +264,14 @@ func getProviderOptions(model Model, providerCfg config.ProviderConfig) fantasy.

450 options[openai.Name] = parsed

264 options[openai.Name] = parsed

451 }

265 }

452 }

266 }

453−

454− case anthropic.Name, bedrock.Name:

267+ case anthropic.Name:

455− var (

456− _, hasEffort = mergedOptions["effort"]

268+ _, hasThink := mergedOptions["thinking"]

269+ if !hasThink && model.ModelCfg.Think {

457− _, hasThink = mergedOptions["thinking"]

270+ mergedOptions["thinking"] = map[string]any{

458− extraBody = make(map[string]any)

271+ // TODO: kujtim see if we need to make this dynamic

459− )

460−

461− switch providerCfg.ID {

462− case string(catwalk.InferenceProviderAlibabaSingapore), string(catwalk.InferenceProviderAlibabaUS):

463− switch {

464− case !hasEffort && shouldSetEffort:

465− extraBody["reasoning_effort"] = reasoningEffort

466− case !hasThink && model.CatwalkCfg.CanReason:

467− if model.ModelCfg.Think {

468− extraBody["thinking"] = map[string]any{"type": "enabled"}

469− } else {

470− extraBody["thinking"] = map[string]any{"type": "disabled"}

471− }

472− }

473− mergedOptions["extra_body"] = extraBody

474−

475− default:

476− switch {

477− case !hasEffort && shouldSetEffort:

478− mergedOptions["effort"] = reasoningEffort

479− case !hasThink && model.ModelCfg.Think:

480− mergedOptions["thinking"] = map[string]any{"budget_tokens": 2000}

272+ "budget_tokens": 2000,

481 }

273 }

482 }

274 }

483−

484 parsed, err := anthropic.ParseOptions(mergedOptions)

275 parsed, err := anthropic.ParseOptions(mergedOptions)

485 if err == nil {

276 if err == nil {

486 options[anthropic.Name] = parsed

277 options[anthropic.Name] = parsed

@@ -488,123 +279,37 @@ func getProviderOptions(model Model, providerCfg config.ProviderConfig) fantasy.

488

279

489 case openrouter.Name:

280 case openrouter.Name:

490 _, hasReasoning := mergedOptions["reasoning"]

281 _, hasReasoning := mergedOptions["reasoning"]

491− if !hasReasoning && shouldSetEffort {

282+ if !hasReasoning && model.ModelCfg.ReasoningEffort != "" {

492 mergedOptions["reasoning"] = map[string]any{

283 mergedOptions["reasoning"] = map[string]any{

493 "enabled": true,

284 "enabled": true,

494− "effort": reasoningEffort,

285+ "effort": model.ModelCfg.ReasoningEffort,

495 }

286 }

496 }

287 }

497 parsed, err := openrouter.ParseOptions(mergedOptions)

288 parsed, err := openrouter.ParseOptions(mergedOptions)

498 if err == nil {

289 if err == nil {

499 options[openrouter.Name] = parsed

290 options[openrouter.Name] = parsed

500 }

291 }

501−

502− case vercel.Name:

503− _, hasReasoning := mergedOptions["reasoning"]

504− if !hasReasoning && shouldSetEffort {

505− mergedOptions["reasoning"] = map[string]any{

506− "enabled": true,

507− "effort": reasoningEffort,

508− }

509− }

510− parsed, err := vercel.ParseOptions(mergedOptions)

511− if err == nil {

512− options[vercel.Name] = parsed

513− }

514−

515 case google.Name:

292 case google.Name:

516 _, hasReasoning := mergedOptions["thinking_config"]

293 _, hasReasoning := mergedOptions["thinking_config"]

517 if !hasReasoning {

294 if !hasReasoning {

518− if strings.HasPrefix(model.CatwalkCfg.ID, "gemini-2") {

519− mergedOptions["thinking_config"] = map[string]any{

295+ mergedOptions["thinking_config"] = map[string]any{

520− "thinking_budget": 2000,

296+ "thinking_budget": 2000,

521− "include_thoughts": true,

522− }

523− } else {

524− mergedOptions["thinking_config"] = map[string]any{

525− "thinking_level": reasoningEffort,

526− "include_thoughts": true,

297+ "include_thoughts": true,

527− }

528 }

298 }

529 }

299 }

530 parsed, err := google.ParseOptions(mergedOptions)

300 parsed, err := google.ParseOptions(mergedOptions)

531 if err == nil {

301 if err == nil {

532 options[google.Name] = parsed

302 options[google.Name] = parsed

533 }

303 }

534−

535− case openaicompat.Name, hyper.Name:

304+ case openaicompat.Name:

536− extraBody := make(map[string]any)

537−

538 _, hasReasoningEffort := mergedOptions["reasoning_effort"]

305 _, hasReasoningEffort := mergedOptions["reasoning_effort"]

539− if !hasReasoningEffort && shouldSetEffort {

306+ if !hasReasoningEffort && model.ModelCfg.ReasoningEffort != "" {

540− switch providerCfg.ID {

541− case string(catwalk.InferenceProviderIoNet):

542− extraBody["reasoning"] = map[string]string{"effort": reasoningEffort}

543− default:

544− mergedOptions["reasoning_effort"] = reasoningEffort

307+ mergedOptions["reasoning_effort"] = model.ModelCfg.ReasoningEffort

545− }

546− }

547−

548− // "reasoning effort" is a standard OpenAI field, but "thinking" is not.

549− // Setting it in the right way for each provider.

550− // TODO: Abstract this in Fantasy somehow?

551− // TODO: Allow custom providers to specify how to set this?

552− switch providerCfg.ID {

553− case hyper.Name:

554− extraBody["thinking"] = model.ModelCfg.Think

555− case string(catwalk.InferenceProviderIoNet):

556− if _, ok := extraBody["reasoning"]; !ok && model.CatwalkCfg.CanReason {

557− if model.ModelCfg.Think {

558− extraBody["reasoning"] = map[string]string{"effort": "medium"}

559− } else {

560− extraBody["reasoning"] = map[string]string{"effort": "none"}

561− }

562− }

563−

564− case string(catwalk.InferenceProviderZAI), string(catwalk.InferenceProviderDeepSeek):

565− if model.ModelCfg.Think || reasoningEffort != "" {

566− extraBody["thinking"] = map[string]any{"type": "enabled"}

567− } else {

568− extraBody["thinking"] = map[string]any{"type": "disabled"}

569− }

570−

571− case string(catwalk.InferenceProviderFireworks):

572− // NOTE: Fireworks break if we set both `reasoning_effort` and `thinking`.

573− if reasoningEffort == "" {

574− if model.ModelCfg.Think {

575− extraBody["thinking"] = map[string]any{"type": "enabled"}

576− } else {

577− extraBody["thinking"] = map[string]any{"type": "disabled"}

578− }

579− }

580−

581− case string(catwalk.InferenceProviderBaseten):

582− extraBody["chat_template_args"] = map[string]any{

583− "enable_thinking": model.ModelCfg.Think || reasoningEffort != "",

584− }

585−

586− case string(catwalk.InferenceProviderAlibabaSingapore), string(catwalk.InferenceProviderAlibabaUS):

587− if model.CatwalkCfg.CanReason {

588− extraBody["enable_thinking"] = model.ModelCfg.Think || reasoningEffort != ""

589− }

590 }

308 }

591−

592− mergedOptions["extra_body"] = extraBody

593−

594 parsed, err := openaicompat.ParseOptions(mergedOptions)

309 parsed, err := openaicompat.ParseOptions(mergedOptions)

595 if err == nil {

310 if err == nil {

596 options[openaicompat.Name] = parsed

311 options[openaicompat.Name] = parsed

597 }

312 }

598−

599− default:

600− // Known custom providers (litellm, ollama, omlx) are

601− // openai-compat under the hood.

602− if discover.IsKnownCustomProvider(string(providerCfg.Type)) {

603− parsed, err := openaicompat.ParseOptions(mergedOptions)

604− if err == nil {

605− options[openaicompat.Name] = parsed

606− }

607− }

608 }

313 }

609

314

610 return options

315 return options

@@ -620,63 +325,28 @@ func mergeCallOptions(model Model, cfg config.ProviderConfig) (fantasy.ProviderO

620 return modelOptions, temp, topP, topK, freqPenalty, presPenalty

325 return modelOptions, temp, topP, topK, freqPenalty, presPenalty

621}

326}

622

327

623−// cachedMergeCallOptions wraps mergeCallOptions with a memoization cache.

624−// The provider options JSON merge is expensive (3× marshal + merge +

625−// unmarshal) and the inputs rarely change between turns. The cache key

626−// combines the config pointer identity with the model's catwalk ID and

627−// provider name so any config reload or model switch misses the cache.

628−func (c *coordinator) cachedMergeCallOptions(model Model, cfg config.ProviderConfig) (fantasy.ProviderOptions, *float64, *float64, *int64, *float64, *float64) {

629− key := fmt.Sprintf("%p|%s|%s", c.cfg.Config(), model.CatwalkCfg.ID, model.ModelCfg.Provider)

630−

631− c.providerOptsCacheMu.Lock()

632− if c.providerOptsCacheKey == key {

633− opts := c.providerOptsCache

634− c.providerOptsCacheMu.Unlock()

635− // Recompute the scalar overrides (cheap) since they don't need

636− // the JSON merge — they're just cmp.Or calls.

637− temp := cmp.Or(model.ModelCfg.Temperature, model.CatwalkCfg.Options.Temperature)

638− topP := cmp.Or(model.ModelCfg.TopP, model.CatwalkCfg.Options.TopP)

639− topK := cmp.Or(model.ModelCfg.TopK, model.CatwalkCfg.Options.TopK)

640− freqPenalty := cmp.Or(model.ModelCfg.FrequencyPenalty, model.CatwalkCfg.Options.FrequencyPenalty)

641− presPenalty := cmp.Or(model.ModelCfg.PresencePenalty, model.CatwalkCfg.Options.PresencePenalty)

642− return opts, temp, topP, topK, freqPenalty, presPenalty

643− }

644− c.providerOptsCacheMu.Unlock()

645−

646− opts, temp, topP, topK, freqPenalty, presPenalty := mergeCallOptions(model, cfg)

647−

648− c.providerOptsCacheMu.Lock()

649− c.providerOptsCache = opts

650− c.providerOptsCacheKey = key

651− c.providerOptsCacheMu.Unlock()

652−

653− return opts, temp, topP, topK, freqPenalty, presPenalty

654−}

655−

656func (c *coordinator) buildAgent(ctx context.Context, prompt *prompt.Prompt, agent config.Agent, isSubAgent bool) (SessionAgent, error) {

328func (c *coordinator) buildAgent(ctx context.Context, prompt *prompt.Prompt, agent config.Agent, isSubAgent bool) (SessionAgent, error) {

657 large, small, err := c.buildAgentModels(ctx, isSubAgent)

329 large, small, err := c.buildAgentModels(ctx, isSubAgent)

658 if err != nil {

330 if err != nil {

659 return nil, err

331 return nil, err

660 }

332 }

661

333

662− largeProviderCfg, _ := c.cfg.Config().Providers.Get(large.ModelCfg.Provider)

334+ largeProviderCfg, _ := c.cfg.Providers.Get(large.ModelCfg.Provider)

663 result := NewSessionAgent(SessionAgentOptions{

335 result := NewSessionAgent(SessionAgentOptions{

664− LargeModel: large,

336+ large,

665− SmallModel: small,

337+ small,

666− SystemPromptPrefix: largeProviderCfg.SystemPromptPrefix,

338+ largeProviderCfg.SystemPromptPrefix,

667− SystemPrompt: "",

339+ "",

668− IsSubAgent: isSubAgent,

340+ isSubAgent,

669− DisableAutoSummarize: c.cfg.Config().Options.DisableAutoSummarize,

341+ c.cfg.Options.DisableAutoSummarize,

670− IsYolo: c.permissions.SkipRequests(),

342+ c.permissions.SkipRequests(),

671− Sessions: c.sessions,

343+ c.sessions,

672− Messages: c.messages,

344+ c.messages,

673− Tools: nil,

345+ nil,

674− Notify: c.notify,

675− RunComplete: c.runComplete,

676 })

346 })

677

347

678 c.readyWg.Go(func() error {

348 c.readyWg.Go(func() error {

679− systemPrompt, err := prompt.Build(ctx, large.Model.Provider(), large.Model.Model(), c.cfg)

349+ systemPrompt, err := prompt.Build(ctx, large.Model.Provider(), large.Model.Model(), *c.cfg)

680 if err != nil {

350 if err != nil {

681 return err

351 return err

682 }

352 }

@@ -685,7 +355,7 @@ func (c *coordinator) buildAgent(ctx context.Context, prompt *prompt.Prompt, age

685 })

355 })

686

356

687 c.readyWg.Go(func() error {

357 c.readyWg.Go(func() error {

688− tools, err := c.buildTools(ctx, agent, isSubAgent)

358+ tools, err := c.buildTools(ctx, agent)

689 if err != nil {

359 if err != nil {

690 return err

360 return err

691 }

361 }

@@ -696,7 +366,7 @@ func (c *coordinator) buildAgent(ctx context.Context, prompt *prompt.Prompt, age

696 return result, nil

366 return result, nil

697}

367}

698

368

699−func (c *coordinator) buildTools(ctx context.Context, agent config.Agent, isSubAgent bool) ([]fantasy.AgentTool, error) {

369+func (c *coordinator) buildTools(ctx context.Context, agent config.Agent) ([]fantasy.AgentTool, error) {

700 var allTools []fantasy.AgentTool

370 var allTools []fantasy.AgentTool

701 if slices.Contains(agent.AllowedTools, AgentToolName) {

371 if slices.Contains(agent.AllowedTools, AgentToolName) {

702 agentTool, err := c.agentTool(ctx)

372 agentTool, err := c.agentTool(ctx)

@@ -715,52 +385,32 @@ func (c *coordinator) buildTools(ctx context.Context, agent config.Agent, isSubA

715 }

385 }

716

386

717 // Get the model name for the agent

387 // Get the model name for the agent

718− modelID := ""

388+ modelName := ""

719− if modelCfg, ok := c.cfg.Config().Models[agent.Model]; ok {

389+ if modelCfg, ok := c.cfg.Models[agent.Model]; ok {

720− if model := c.cfg.Config().GetModel(modelCfg.Provider, modelCfg.Model); model != nil {

390+ if model := c.cfg.GetModel(modelCfg.Provider, modelCfg.Model); model != nil {

721− modelID = model.ID

391+ modelName = model.Name

722 }

392 }

723 }

393 }

724

394

725− logFile := filepath.Join(c.cfg.Config().Options.DataDirectory, "logs", "crush.log")

726−

727− // Build hook runner if PreToolUse hooks are configured.

728− var hookRunner *hooks.Runner

729− if preToolHooks := c.cfg.Config().Hooks[hooks.EventPreToolUse]; len(preToolHooks) > 0 {

730− hookRunner = hooks.NewRunner(preToolHooks, c.cfg.WorkingDir(), c.cfg.WorkingDir())

731− }

732−

733− allTools = append(

395+ allTools = append(allTools,

734− allTools,

735− tools.NewBashTool(c.permissions, c.cfg.WorkingDir(), c.cfg.Config().Options.Attribution, modelID),

396+ tools.NewBashTool(c.permissions, c.cfg.WorkingDir(), c.cfg.Options.Attribution, modelName),

736− tools.NewCrushInfoTool(c.cfg, c.lspManager, c.allSkills, c.activeSkills, c.skillTracker),

737− tools.NewCrushLogsTool(logFile),

738 tools.NewJobOutputTool(),

397 tools.NewJobOutputTool(),

739 tools.NewJobKillTool(),

398 tools.NewJobKillTool(),

740 tools.NewDownloadTool(c.permissions, c.cfg.WorkingDir(), nil),

399 tools.NewDownloadTool(c.permissions, c.cfg.WorkingDir(), nil),

741− tools.NewEditTool(c.lspManager, c.permissions, c.history, c.filetracker, c.cfg.WorkingDir()),

400+ tools.NewEditTool(c.lspClients, c.permissions, c.history, c.filetracker, c.cfg.WorkingDir()),

742− tools.NewMultiEditTool(c.lspManager, c.permissions, c.history, c.filetracker, c.cfg.WorkingDir()),

401+ tools.NewMultiEditTool(c.lspClients, c.permissions, c.history, c.filetracker, c.cfg.WorkingDir()),

743 tools.NewFetchTool(c.permissions, c.cfg.WorkingDir(), nil),

402 tools.NewFetchTool(c.permissions, c.cfg.WorkingDir(), nil),

744− tools.NewGlobTool(c.cfg.WorkingDir(), c.cfg.Config().Tools.Glob),

403+ tools.NewGlobTool(c.cfg.WorkingDir()),

745− tools.NewGrepTool(c.cfg.WorkingDir(), c.cfg.Config().Tools.Grep),

404+ tools.NewGrepTool(c.cfg.WorkingDir()),

746− tools.NewLsTool(c.permissions, c.cfg.WorkingDir(), c.cfg.Config().Tools.Ls),

405+ tools.NewLsTool(c.permissions, c.cfg.WorkingDir(), c.cfg.Tools.Ls),

747 tools.NewSourcegraphTool(nil),

406 tools.NewSourcegraphTool(nil),

748 tools.NewTodosTool(c.sessions),

407 tools.NewTodosTool(c.sessions),

749− tools.NewViewTool(c.lspManager, c.permissions, c.filetracker, c.skillTracker, c.cfg.WorkingDir(), c.cfg.Config().Options.SkillsPaths...),

408+ tools.NewViewTool(c.lspClients, c.permissions, c.filetracker, c.cfg.WorkingDir(), c.cfg.Options.SkillsPaths...),

750− tools.NewWriteTool(c.lspManager, c.permissions, c.history, c.filetracker, c.cfg.WorkingDir()),

409+ tools.NewWriteTool(c.lspClients, c.permissions, c.history, c.filetracker, c.cfg.WorkingDir()),

751 )

410 )

752

411

753− // Add LSP tools if user has configured LSPs or auto_lsp is enabled (nil or true).

412+ if len(c.cfg.LSP) > 0 {

754− if len(c.cfg.Config().LSP) > 0 || c.cfg.Config().Options.AutoLSP == nil || *c.cfg.Config().Options.AutoLSP {

755− allTools = append(allTools, tools.NewDiagnosticsTool(c.lspManager), tools.NewReferencesTool(c.lspManager), tools.NewLSPRestartTool(c.lspManager))

413+ allTools = append(allTools, tools.NewDiagnosticsTool(c.lspClients), tools.NewReferencesTool(c.lspClients), tools.NewLSPRestartTool(c.lspClients))

756− }

757−

758− if len(c.cfg.Config().MCP) > 0 {

759− allTools = append(

760− allTools,

761− tools.NewListMCPResourcesTool(c.cfg, c.permissions),

762− tools.NewReadMCPResourceTool(c.cfg, c.permissions),

763− )

764 }

414 }

765

415

766 var filteredTools []fantasy.AgentTool

416 var filteredTools []fantasy.AgentTool

@@ -770,7 +420,7 @@ func (c *coordinator) buildTools(ctx context.Context, agent config.Agent, isSubA

770 }

420 }

771 }

421 }

772

422

773− for _, tool := range tools.GetMCPTools(c.permissions, c.cfg, c.cfg.WorkingDir()) {

423+ for _, tool := range tools.GetMCPTools(c.permissions, c.cfg.WorkingDir()) {

774 if agent.AllowedMCP == nil {

424 if agent.AllowedMCP == nil {

775 // No MCP restrictions

425 // No MCP restrictions

776 filteredTools = append(filteredTools, tool)

426 filteredTools = append(filteredTools, tool)

@@ -788,39 +438,30 @@ func (c *coordinator) buildTools(ctx context.Context, agent config.Agent, isSubA

788 }

438 }

789 if len(tools) == 0 || slices.Contains(tools, tool.MCPToolName()) {

439 if len(tools) == 0 || slices.Contains(tools, tool.MCPToolName()) {

790 filteredTools = append(filteredTools, tool)

440 filteredTools = append(filteredTools, tool)

791− break

792 }

441 }

793− slog.Debug("MCP not allowed", "tool", tool.Name(), "agent", agent.Name)

794 }

442 }

443+ slog.Debug("MCP not allowed", "tool", tool.Name(), "agent", agent.Name)

795 }

444 }

796 slices.SortFunc(filteredTools, func(a, b fantasy.AgentTool) int {

445 slices.SortFunc(filteredTools, func(a, b fantasy.AgentTool) int {

797 return strings.Compare(a.Info().Name, b.Info().Name)

446 return strings.Compare(a.Info().Name, b.Info().Name)

798 })

447 })

799−

800− // Wrap tools with hook interception for the top-level agent only.

801− // Sub-agents (the `agent` task tool, `agentic_fetch`, etc.) run

802− // without hook interception to avoid firing the user's hook N times

803− // per delegated turn. The top-level invocation of the sub-agent tool

804− // itself is still wrapped from the coder's side.

805− filteredTools = wrapToolsWithHooks(filteredTools, hookRunner, isSubAgent)

806−

807 return filteredTools, nil

448 return filteredTools, nil

808}

449}

809

450

810// TODO: when we support multiple agents we need to change this so that we pass in the agent specific model config

451// TODO: when we support multiple agents we need to change this so that we pass in the agent specific model config

811func (c *coordinator) buildAgentModels(ctx context.Context, isSubAgent bool) (Model, Model, error) {

452func (c *coordinator) buildAgentModels(ctx context.Context, isSubAgent bool) (Model, Model, error) {

812− largeModelCfg, ok := c.cfg.Config().Models[config.SelectedModelTypeLarge]

453+ largeModelCfg, ok := c.cfg.Models[config.SelectedModelTypeLarge]

813 if !ok {

454 if !ok {

814− return Model{}, Model{}, errLargeModelNotSelected

455+ return Model{}, Model{}, errors.New("large model not selected")

815 }

456 }

816− smallModelCfg, ok := c.cfg.Config().Models[config.SelectedModelTypeSmall]

457+ smallModelCfg, ok := c.cfg.Models[config.SelectedModelTypeSmall]

817 if !ok {

458 if !ok {

818− return Model{}, Model{}, errSmallModelNotSelected

459+ return Model{}, Model{}, errors.New("small model not selected")

819 }

460 }

820

461

821− largeProviderCfg, ok := c.cfg.Config().Providers.Get(largeModelCfg.Provider)

462+ largeProviderCfg, ok := c.cfg.Providers.Get(largeModelCfg.Provider)

822 if !ok {

463 if !ok {

823− return Model{}, Model{}, errLargeModelProviderNotConfigured

464+ return Model{}, Model{}, errors.New("large model provider not configured")

824 }

465 }

825

466

826 largeProvider, err := c.buildProvider(largeProviderCfg, largeModelCfg, isSubAgent)

467 largeProvider, err := c.buildProvider(largeProviderCfg, largeModelCfg, isSubAgent)

@@ -828,12 +469,12 @@ func (c *coordinator) buildAgentModels(ctx context.Context, isSubAgent bool) (Mo

828 return Model{}, Model{}, err

469 return Model{}, Model{}, err

829 }

470 }

830

471

831− smallProviderCfg, ok := c.cfg.Config().Providers.Get(smallModelCfg.Provider)

472+ smallProviderCfg, ok := c.cfg.Providers.Get(smallModelCfg.Provider)

832 if !ok {

473 if !ok {

833− return Model{}, Model{}, errSmallModelProviderNotConfigured

474+ return Model{}, Model{}, errors.New("large model provider not configured")

834 }

475 }

835

476

836− smallProvider, err := c.buildProvider(smallProviderCfg, smallModelCfg, true)

477+ smallProvider, err := c.buildProvider(smallProviderCfg, largeModelCfg, true)

837 if err != nil {

478 if err != nil {

838 return Model{}, Model{}, err

479 return Model{}, Model{}, err

839 }

480 }

@@ -853,11 +494,11 @@ func (c *coordinator) buildAgentModels(ctx context.Context, isSubAgent bool) (Mo

853 }

494 }

854

495

855 if largeCatwalkModel == nil {

496 if largeCatwalkModel == nil {

856− return Model{}, Model{}, errLargeModelNotFound

497+ return Model{}, Model{}, errors.New("large model not found in provider config")

857 }

498 }

858

499

859 if smallCatwalkModel == nil {

500 if smallCatwalkModel == nil {

860− return Model{}, Model{}, errSmallModelNotFound

501+ return Model{}, Model{}, errors.New("small model not found in provider config")

861 }

502 }

862

503

863 largeModelID := largeModelCfg.Model

504 largeModelID := largeModelCfg.Model

@@ -884,28 +525,21 @@ func (c *coordinator) buildAgentModels(ctx context.Context, isSubAgent bool) (Mo

884 Model: largeModel,

525 Model: largeModel,

885 CatwalkCfg: *largeCatwalkModel,

526 CatwalkCfg: *largeCatwalkModel,

886 ModelCfg: largeModelCfg,

527 ModelCfg: largeModelCfg,

887− FlatRate: largeProviderCfg.FlatRate,

888 }, Model{

528 }, Model{

889 Model: smallModel,

529 Model: smallModel,

890 CatwalkCfg: *smallCatwalkModel,

530 CatwalkCfg: *smallCatwalkModel,

891 ModelCfg: smallModelCfg,

531 ModelCfg: smallModelCfg,

892− FlatRate: smallProviderCfg.FlatRate,

893 }, nil

532 }, nil

894}

533}

895

534

896−func (c *coordinator) buildAnthropicProvider(baseURL, apiKey string, headers map[string]string, providerID string) (fantasy.Provider, error) {

535+func (c *coordinator) buildAnthropicProvider(baseURL, apiKey string, headers map[string]string) (fantasy.Provider, error) {

897 var opts []anthropic.Option

536 var opts []anthropic.Option

898

537

899− switch {

900− case strings.HasPrefix(apiKey, "Bearer "):

538+ if strings.HasPrefix(apiKey, "Bearer ") {

901 // NOTE: Prevent the SDK from picking up the API key from env.

539 // NOTE: Prevent the SDK from picking up the API key from env.

902 os.Setenv("ANTHROPIC_API_KEY", "")

540 os.Setenv("ANTHROPIC_API_KEY", "")

903 headers["Authorization"] = apiKey

541 headers["Authorization"] = apiKey

904− case providerID == string(catwalk.InferenceProviderMiniMax) || providerID == string(catwalk.InferenceProviderMiniMaxChina):

905− // NOTE: Prevent the SDK from picking up the API key from env.

906− os.Setenv("ANTHROPIC_API_KEY", "")

907− headers["Authorization"] = "Bearer " + apiKey

908− case apiKey != "":

542+ } else if apiKey != "" {

909 // X-Api-Key header

543 // X-Api-Key header

910 opts = append(opts, anthropic.WithAPIKey(apiKey))

544 opts = append(opts, anthropic.WithAPIKey(apiKey))

911 }

545 }

@@ -918,7 +552,7 @@ func (c *coordinator) buildAnthropicProvider(baseURL, apiKey string, headers map

918 opts = append(opts, anthropic.WithBaseURL(baseURL))

552 opts = append(opts, anthropic.WithBaseURL(baseURL))

919 }

553 }

920

554

921− if c.cfg.Config().Options.Debug {

555+ if c.cfg.Options.Debug {

922 httpClient := log.NewHTTPClient()

556 httpClient := log.NewHTTPClient()

923 opts = append(opts, anthropic.WithHTTPClient(httpClient))

557 opts = append(opts, anthropic.WithHTTPClient(httpClient))

924 }

558 }

@@ -930,7 +564,7 @@ func (c *coordinator) buildOpenaiProvider(baseURL, apiKey string, headers map[st

930 openai.WithAPIKey(apiKey),

564 openai.WithAPIKey(apiKey),

931 openai.WithUseResponsesAPI(),

565 openai.WithUseResponsesAPI(),

932 }

566 }

933− if c.cfg.Config().Options.Debug {

567+ if c.cfg.Options.Debug {

934 httpClient := log.NewHTTPClient()

568 httpClient := log.NewHTTPClient()

935 opts = append(opts, openai.WithHTTPClient(httpClient))

569 opts = append(opts, openai.WithHTTPClient(httpClient))

936 }

570 }

@@ -947,7 +581,7 @@ func (c *coordinator) buildOpenrouterProvider(_, apiKey string, headers map[stri

947 opts := []openrouter.Option{

581 opts := []openrouter.Option{

948 openrouter.WithAPIKey(apiKey),

582 openrouter.WithAPIKey(apiKey),

949 }

583 }

950− if c.cfg.Config().Options.Debug {

584+ if c.cfg.Options.Debug {

951 httpClient := log.NewHTTPClient()

585 httpClient := log.NewHTTPClient()

952 opts = append(opts, openrouter.WithHTTPClient(httpClient))

586 opts = append(opts, openrouter.WithHTTPClient(httpClient))

953 }

587 }

@@ -957,20 +591,6 @@ func (c *coordinator) buildOpenrouterProvider(_, apiKey string, headers map[stri

957 return openrouter.New(opts...)

591 return openrouter.New(opts...)

958}

592}

959

593

960−func (c *coordinator) buildVercelProvider(_, apiKey string, headers map[string]string) (fantasy.Provider, error) {

961− opts := []vercel.Option{

962− vercel.WithAPIKey(apiKey),

963− }

964− if c.cfg.Config().Options.Debug {

965− httpClient := log.NewHTTPClient()

966− opts = append(opts, vercel.WithHTTPClient(httpClient))

967− }

968− if len(headers) > 0 {

969− opts = append(opts, vercel.WithHeaders(headers))

970− }

971− return vercel.New(opts...)

972−}

973−

974func (c *coordinator) buildOpenaiCompatProvider(baseURL, apiKey string, headers map[string]string, extraBody map[string]any, providerID string, isSubAgent bool) (fantasy.Provider, error) {

594func (c *coordinator) buildOpenaiCompatProvider(baseURL, apiKey string, headers map[string]string, extraBody map[string]any, providerID string, isSubAgent bool) (fantasy.Provider, error) {

975 opts := []openaicompat.Option{

595 opts := []openaicompat.Option{

976 openaicompat.WithBaseURL(baseURL),

596 openaicompat.WithBaseURL(baseURL),

@@ -979,18 +599,10 @@ func (c *coordinator) buildOpenaiCompatProvider(baseURL, apiKey string, headers

979

599

980 // Set HTTP client based on provider and debug mode.

600 // Set HTTP client based on provider and debug mode.

981 var httpClient *http.Client

601 var httpClient *http.Client

982− switch providerID {

983− case string(catwalk.InferenceProviderCopilot):

602+ if providerID == string(catwalk.InferenceProviderCopilot) {

984− opts = append(

985− opts,

986− openaicompat.WithUseResponsesAPI(),

603+ opts = append(opts, openaicompat.WithUseResponsesAPI())

987− openaicompat.WithResponsesAPIFunc(func(modelID string) bool {

988− return copilotResponsesModels[modelID]

989− }),

990− )

991− httpClient = copilot.NewClient(isSubAgent, c.cfg.Config().Options.Debug)

604+ httpClient = copilot.NewClient(isSubAgent, c.cfg.Options.Debug)

992− }

993− if httpClient == nil && c.cfg.Config().Options.Debug {

605+ } else if c.cfg.Options.Debug {

994 httpClient = log.NewHTTPClient()

606 httpClient = log.NewHTTPClient()

995 }

607 }

996 if httpClient != nil {

608 if httpClient != nil {

@@ -1014,7 +626,7 @@ func (c *coordinator) buildAzureProvider(baseURL, apiKey string, headers map[str

1014 azure.WithAPIKey(apiKey),

626 azure.WithAPIKey(apiKey),

1015 azure.WithUseResponsesAPI(),

627 azure.WithUseResponsesAPI(),

1016 }

628 }

1017− if c.cfg.Config().Options.Debug {

629+ if c.cfg.Options.Debug {

1018 httpClient := log.NewHTTPClient()

630 httpClient := log.NewHTTPClient()

1019 opts = append(opts, azure.WithHTTPClient(httpClient))

631 opts = append(opts, azure.WithHTTPClient(httpClient))

1020 }

632 }

@@ -1031,32 +643,19 @@ func (c *coordinator) buildAzureProvider(baseURL, apiKey string, headers map[str

1031 return azure.New(opts...)

643 return azure.New(opts...)

1032}

644}

1033

645

1034−func (c *coordinator) buildBedrockProvider(apiKey string, headers map[string]string, providerID string) (fantasy.Provider, error) {

646+func (c *coordinator) buildBedrockProvider(headers map[string]string) (fantasy.Provider, error) {

1035 var opts []bedrock.Option

647 var opts []bedrock.Option

1036− if c.cfg.Config().Options.Debug {

648+ if c.cfg.Options.Debug {

1037 httpClient := log.NewHTTPClient()

649 httpClient := log.NewHTTPClient()

1038 opts = append(opts, bedrock.WithHTTPClient(httpClient))

650 opts = append(opts, bedrock.WithHTTPClient(httpClient))

1039 }

651 }

1040 if len(headers) > 0 {

652 if len(headers) > 0 {

1041 opts = append(opts, bedrock.WithHeaders(headers))

653 opts = append(opts, bedrock.WithHeaders(headers))

1042 }

654 }

1043−

1044− switch {

1045− case apiKey != "":

1046− opts = append(opts, bedrock.WithAPIKey(apiKey))

1047− case os.Getenv("AWS_BEARER_TOKEN_BEDROCK") != "":

655+ bearerToken := os.Getenv("AWS_BEARER_TOKEN_BEDROCK")

1048− opts = append(opts, bedrock.WithAPIKey(os.Getenv("AWS_BEARER_TOKEN_BEDROCK")))

656+ if bearerToken != "" {

1049− default:

1050− // Skip, let the SDK do authentication.

1051− }

1052−

1053− switch providerID {

1054− case string(catwalk.InferenceProviderBedrockEurope):

1055− opts = append(opts, bedrock.WithRegion("eu-west-1"))

657+ opts = append(opts, bedrock.WithAPIKey(bearerToken))

1056− default:

1057− opts = append(opts, bedrock.WithRegion("us-east-1"))

1058 }

658 }

1059−

1060 return bedrock.New(opts...)

659 return bedrock.New(opts...)

1061}

660}

1062

661

@@ -1065,7 +664,7 @@ func (c *coordinator) buildGoogleProvider(baseURL, apiKey string, headers map[st

1065 google.WithBaseURL(baseURL),

664 google.WithBaseURL(baseURL),

1066 google.WithGeminiAPIKey(apiKey),

665 google.WithGeminiAPIKey(apiKey),

1067 }

666 }

1068− if c.cfg.Config().Options.Debug {

667+ if c.cfg.Options.Debug {

1069 httpClient := log.NewHTTPClient()

668 httpClient := log.NewHTTPClient()

1070 opts = append(opts, google.WithHTTPClient(httpClient))

669 opts = append(opts, google.WithHTTPClient(httpClient))

1071 }

670 }

@@ -1077,7 +676,7 @@ func (c *coordinator) buildGoogleProvider(baseURL, apiKey string, headers map[st

1077

676

1078func (c *coordinator) buildGoogleVertexProvider(headers map[string]string, options map[string]string) (fantasy.Provider, error) {

677func (c *coordinator) buildGoogleVertexProvider(headers map[string]string, options map[string]string) (fantasy.Provider, error) {

1079 opts := []google.Option{}

678 opts := []google.Option{}

1080− if c.cfg.Config().Options.Debug {

679+ if c.cfg.Options.Debug {

1081 httpClient := log.NewHTTPClient()

680 httpClient := log.NewHTTPClient()

1082 opts = append(opts, google.WithHTTPClient(httpClient))

681 opts = append(opts, google.WithHTTPClient(httpClient))

1083 }

682 }

@@ -1093,12 +692,35 @@ func (c *coordinator) buildGoogleVertexProvider(headers map[string]string, optio

1093 return google.New(opts...)

692 return google.New(opts...)

1094}

693}

1095

694

695+func (c *coordinator) buildHyperProvider(baseURL, apiKey string) (fantasy.Provider, error) {

696+ opts := []hyper.Option{

697+ hyper.WithBaseURL(baseURL),

698+ hyper.WithAPIKey(apiKey),

699+ }

700+ if c.cfg.Options.Debug {

701+ httpClient := log.NewHTTPClient()

702+ opts = append(opts, hyper.WithHTTPClient(httpClient))

703+ }

704+ return hyper.New(opts...)

705+}

706+

1096func (c *coordinator) isAnthropicThinking(model config.SelectedModel) bool {

707func (c *coordinator) isAnthropicThinking(model config.SelectedModel) bool {

1097 if model.Think {

708 if model.Think {

1098 return true

709 return true

1099 }

710 }

711+

712+ if model.ProviderOptions == nil {

713+ return false

714+ }

715+

1100 opts, err := anthropic.ParseOptions(model.ProviderOptions)

716 opts, err := anthropic.ParseOptions(model.ProviderOptions)

717+ if err != nil {

718+ return false

719+ }

1101− return err == nil && opts.Thinking != nil

720+ if opts.Thinking != nil {

721+ return true

722+ }

723+ return false

1102}

724}

1103

725

1104func (c *coordinator) buildProvider(providerCfg config.ProviderConfig, model config.SelectedModel, isSubAgent bool) (fantasy.Provider, error) {

726func (c *coordinator) buildProvider(providerCfg config.ProviderConfig, model config.SelectedModel, isSubAgent bool) (fantasy.Provider, error) {

@@ -1119,117 +741,32 @@ func (c *coordinator) buildProvider(providerCfg config.ProviderConfig, model con

1119 apiKey, _ := c.cfg.Resolve(providerCfg.APIKey)

741 apiKey, _ := c.cfg.Resolve(providerCfg.APIKey)

1120 baseURL, _ := c.cfg.Resolve(providerCfg.BaseURL)

742 baseURL, _ := c.cfg.Resolve(providerCfg.BaseURL)

1121

743

1122− // Reuse a previously built client when nothing that affects it has

1123− // changed, so the HTTP keep-alive pool survives across turns. The key

1124− // is derived from the resolved inputs above (plus the debug flag); the

1125− // per-provider adjustments inside the switch below are deterministic

1126− // functions of these, so identical inputs always yield an identical

1127− // client.

1128− cacheKey := c.providerCacheKey(providerCfg, model, apiKey, baseURL, headers, isSubAgent)

1129− if p, ok := c.providerClients.Get(cacheKey); ok {

1130− return p, nil

1131− }

1132− provider, err := c.buildProviderUncached(providerCfg, model, apiKey, baseURL, headers, isSubAgent)

1133− if err != nil {

1134− return nil, err

1135− }

1136− c.providerClients.Set(cacheKey, provider)

1137− return provider, nil

1138−}

1139−

1140−// providerCacheKey builds a stable key from every input buildProvider's

1141−// switch consumes, so any credential, endpoint, header, or option change

1142−// produces a distinct key (and thus a rebuilt client).

1143−func (c *coordinator) providerCacheKey(providerCfg config.ProviderConfig, model config.SelectedModel, apiKey, baseURL string, headers map[string]string, isSubAgent bool) string {

1144− var b strings.Builder

1145− b.WriteString(providerCfg.ID)

1146− b.WriteByte(0)

1147− b.WriteString(string(providerCfg.Type))

1148− b.WriteByte(0)

1149− // The opencode providers pick their client type from the model, so the

1150− // model must participate in the key.

1151− b.WriteString(model.Model)

1152− b.WriteByte(0)

1153− b.WriteString(apiKey)

1154− b.WriteByte(0)

1155− // Bedrock falls back to AWS_BEARER_TOKEN_BEDROCK when no apiKey is

1156− // configured, and that token is baked into the built client. Fold it

1157− // into the key so a rotated token rebuilds the client instead of

1158− // reusing a stale one.

1159− if apiKey == "" && providerCfg.Type == bedrock.Name {

1160− b.WriteString(os.Getenv("AWS_BEARER_TOKEN_BEDROCK"))

1161− }

1162− b.WriteByte(0)

1163− b.WriteString(baseURL)

1164− b.WriteByte(0)

1165− for _, k := range slices.Sorted(maps.Keys(headers)) {

1166− b.WriteString(k)

1167− b.WriteByte('=')

1168− b.WriteString(headers[k])

1169− b.WriteByte(1)

1170− }

1171− b.WriteByte(0)

1172− if extra, err := json.Marshal(providerCfg.ExtraBody); err == nil {

1173− b.Write(extra)

1174− }

1175− b.WriteByte(0)

1176− if extra, err := json.Marshal(providerCfg.ExtraParams); err == nil {

1177− b.Write(extra)

1178− }

1179− b.WriteByte(0)

1180− if isSubAgent {

1181− b.WriteByte('s')

1182− }

1183− if c.cfg.Config().Options.Debug {

1184− b.WriteByte('d')

1185− }

1186− return b.String()

1187−}

1188−

1189−func (c *coordinator) buildProviderUncached(providerCfg config.ProviderConfig, model config.SelectedModel, apiKey, baseURL string, headers map[string]string, isSubAgent bool) (fantasy.Provider, error) {

1190− switch providerCfg.ID {

1191− case string(catwalk.InferenceProviderOpenCodeGo), string(catwalk.InferenceProviderOpenCodeZen):

1192− if opencodeMessagesModels[model.Model] {

1193− baseURL = strings.TrimSuffix(baseURL, "/v1")

1194− return c.buildAnthropicProvider(baseURL, apiKey, headers, providerCfg.ID)

1195− }

1196− }

1197−

1198 switch providerCfg.Type {

744 switch providerCfg.Type {

1199 case openai.Name:

745 case openai.Name:

1200 return c.buildOpenaiProvider(baseURL, apiKey, headers)

746 return c.buildOpenaiProvider(baseURL, apiKey, headers)

1201 case anthropic.Name:

747 case anthropic.Name:

1202− return c.buildAnthropicProvider(baseURL, apiKey, headers, providerCfg.ID)

748+ return c.buildAnthropicProvider(baseURL, apiKey, headers)

1203 case openrouter.Name:

749 case openrouter.Name:

1204 return c.buildOpenrouterProvider(baseURL, apiKey, headers)

750 return c.buildOpenrouterProvider(baseURL, apiKey, headers)

1205− case vercel.Name:

1206− return c.buildVercelProvider(baseURL, apiKey, headers)

1207 case azure.Name:

751 case azure.Name:

1208 return c.buildAzureProvider(baseURL, apiKey, headers, providerCfg.ExtraParams)

752 return c.buildAzureProvider(baseURL, apiKey, headers, providerCfg.ExtraParams)

1209 case bedrock.Name:

753 case bedrock.Name:

1210− return c.buildBedrockProvider(apiKey, headers, providerCfg.ID)

754+ return c.buildBedrockProvider(headers)

1211 case google.Name:

755 case google.Name:

1212 return c.buildGoogleProvider(baseURL, apiKey, headers)

756 return c.buildGoogleProvider(baseURL, apiKey, headers)

1213 case "google-vertex":

757 case "google-vertex":

1214 return c.buildGoogleVertexProvider(headers, providerCfg.ExtraParams)

758 return c.buildGoogleVertexProvider(headers, providerCfg.ExtraParams)

1215− case openaicompat.Name, hyper.Name:

759+ case openaicompat.Name:

1216− switch providerCfg.ID {

1217− case hyper.Name:

1218− baseURL = hyper.BaseURL() + "/v1"

1219− headers["x-crush-id"] = event.GetID()

1220− case string(catwalk.InferenceProviderZAI):

760+ if providerCfg.ID == string(catwalk.InferenceProviderZAI) {

1221 if providerCfg.ExtraBody == nil {

761 if providerCfg.ExtraBody == nil {

1222 providerCfg.ExtraBody = map[string]any{}

762 providerCfg.ExtraBody = map[string]any{}

1223 }

763 }

1224 providerCfg.ExtraBody["tool_stream"] = true

764 providerCfg.ExtraBody["tool_stream"] = true

1225 }

765 }

1226 return c.buildOpenaiCompatProvider(baseURL, apiKey, headers, providerCfg.ExtraBody, providerCfg.ID, isSubAgent)

766 return c.buildOpenaiCompatProvider(baseURL, apiKey, headers, providerCfg.ExtraBody, providerCfg.ID, isSubAgent)

767+ case hyper.Name:

768+ return c.buildHyperProvider(baseURL, apiKey)

1227 default:

769 default:

1228− // Known custom providers (litellm, ollama, omlx) are

1229− // openai-compat under the hood.

1230− if discover.IsKnownCustomProvider(string(providerCfg.Type)) {

1231− return c.buildOpenaiCompatProvider(baseURL, apiKey, headers, providerCfg.ExtraBody, providerCfg.ID, isSubAgent)

1232− }

1233 return nil, fmt.Errorf("provider type not supported: %q", providerCfg.Type)

770 return nil, fmt.Errorf("provider type not supported: %q", providerCfg.Type)

1234 }

771 }

1235}

772}

@@ -1245,15 +782,6 @@ func isExactoSupported(modelID string) bool {

1245 return slices.Contains(supportedModels, modelID)

782 return slices.Contains(supportedModels, modelID)

1246}

783}

1247

784

1248−// BeginAccepted reserves an accept slot for sessionID on the active

1249−// agent and returns the ownership handle. It is the fire-and-forget

1250−// dispatch path's only way to mark a run as accepted-but-not-yet-active

1251−// so a cancel arriving before the run registers in activeRequests is not

1252−// lost.

1253−func (c *coordinator) BeginAccepted(sessionID string) *AcceptedRun {

1254− return c.currentAgent.BeginAccepted(sessionID)

1255−}

1256−

1257func (c *coordinator) Cancel(sessionID string) {

785func (c *coordinator) Cancel(sessionID string) {

1258 c.currentAgent.Cancel(sessionID)

786 c.currentAgent.Cancel(sessionID)

1259}

787}

@@ -1286,34 +814,15 @@ func (c *coordinator) UpdateModels(ctx context.Context) error {

1286 }

814 }

1287 c.currentAgent.SetModels(large, small)

815 c.currentAgent.SetModels(large, small)

1288

816

1289− cfg := c.cfg.Config()

1290− agentCfg, ok := cfg.Agents[config.AgentCoder]

817+ agentCfg, ok := c.cfg.Agents[config.AgentCoder]

1291 if !ok {

818 if !ok {

1292− return errCoderAgentNotConfigured

819+ return errors.New("coder agent not configured")

1293 }

820 }

1294

821

1295− // Fast path: reuse cached tools when neither the config nor the MCP

1296− // tool set has changed since the last build.

1297− mcpGen := c.mcpToolsGen.Load()

1298− c.toolsCacheMu.Lock()

1299− if c.toolsCache != nil && c.toolsCacheCfg == cfg && c.toolsCacheMCP == mcpGen {

1300− c.currentAgent.SetTools(c.toolsCache)

1301− c.toolsCacheMu.Unlock()

1302− return nil

1303− }

1304− c.toolsCacheMu.Unlock()

1305−

1306− tools, err := c.buildTools(ctx, agentCfg, false)

822+ tools, err := c.buildTools(ctx, agentCfg)

1307 if err != nil {

823 if err != nil {

1308 return err

824 return err

1309 }

825 }

1310−

1311− c.toolsCacheMu.Lock()

1312− c.toolsCache = tools

1313− c.toolsCacheCfg = cfg

1314− c.toolsCacheMCP = mcpGen

1315− c.toolsCacheMu.Unlock()

1316−

1317 c.currentAgent.SetTools(tools)

826 c.currentAgent.SetTools(tools)

1318 return nil

827 return nil

1319}

828}

@@ -1327,67 +836,11 @@ func (c *coordinator) QueuedPromptsList(sessionID string) []string {

1327}

836}

1328

837

1329func (c *coordinator) Summarize(ctx context.Context, sessionID string) error {

838func (c *coordinator) Summarize(ctx context.Context, sessionID string) error {

1330− providerCfg, ok := c.cfg.Config().Providers.Get(c.currentAgent.Model().ModelCfg.Provider)

839+ providerCfg, ok := c.cfg.Providers.Get(c.currentAgent.Model().ModelCfg.Provider)

1331 if !ok {

840 if !ok {

1332− return errModelProviderNotConfigured

841+ return errors.New("model provider not configured")

1333− }

1334−

1335− if err := c.refreshTokenIfExpired(ctx, providerCfg); err != nil {

1336− slog.Error("Failed to refresh OAuth2 token before summarize. Proceeding with existing token.", "error", err)

1337− }

1338−

1339− summarize := func() error {

1340− return c.currentAgent.Summarize(ctx, sessionID, getProviderOptions(c.currentAgent.Model(), providerCfg))

1341− }

1342−

1343− return c.runWithUnauthorizedRetry(ctx, providerCfg, summarize)

1344−}

1345−

1346−// GenerateTitle generates a session title using the current agent.

1347−func (c *coordinator) GenerateTitle(ctx context.Context, sessionID, prompt string) {

1348− if c.currentAgent == nil {

1349− return

1350− }

1351− c.currentAgent.GenerateTitle(ctx, sessionID, prompt)

1352−}

1353−

1354−// refreshTokenIfExpired proactively refreshes the OAuth token if it has expired.

1355−func (c *coordinator) refreshTokenIfExpired(ctx context.Context, providerCfg config.ProviderConfig) error {

1356− if providerCfg.OAuthToken == nil || !providerCfg.OAuthToken.IsExpired() {

1357− return nil

1358− }

1359− slog.Debug("Token needs to be refreshed", "provider", providerCfg.ID)

1360− return c.refreshOAuth2Token(ctx, providerCfg)

1361−}

1362−

1363−// runWithUnauthorizedRetry executes fn. If fn returns a 401 error, it

1364−// attempts to refresh credentials and re-runs fn once. Returns the

1365−// final error: from the retry if a retry was attempted, otherwise from

1366−// the original run. Callers that need to notify the user on persistent

1367−// failure should check isUnauthorized on the returned error.

1368−func (c *coordinator) runWithUnauthorizedRetry(ctx context.Context, providerCfg config.ProviderConfig, fn func() error) error {

1369− err := fn()

1370− if err != nil && c.isUnauthorized(err) {

1371− if retryErr := c.retryAfterUnauthorized(ctx, providerCfg); retryErr == nil {

1372− return fn()

1373− }

1374− }

1375− return err

1376−}

1377−

1378−// retryAfterUnauthorized attempts to refresh credentials after receiving a 401

1379−// and returns nil if retry should be attempted.

1380−func (c *coordinator) retryAfterUnauthorized(ctx context.Context, providerCfg config.ProviderConfig) error {

1381− switch {

1382− case providerCfg.OAuthToken != nil:

1383− slog.Debug("Received 401. Refreshing token and retrying", "provider", providerCfg.ID)

1384− return c.refreshOAuth2Token(ctx, providerCfg)

1385− case strings.Contains(providerCfg.APIKeyTemplate, "$"):

1386− slog.Debug("Received 401. Refreshing API Key template and retrying", "provider", providerCfg.ID)

1387− return c.refreshApiKeyTemplate(ctx, providerCfg)

1388− default:

1389− return nil

1390 }

842 }

843+ return c.currentAgent.Summarize(ctx, sessionID, getProviderOptions(c.currentAgent.Model(), providerCfg))

1391}

844}

1392

845

1393func (c *coordinator) isUnauthorized(err error) bool {

846func (c *coordinator) isUnauthorized(err error) bool {

@@ -1396,7 +849,7 @@ func (c *coordinator) isUnauthorized(err error) bool {

1396}

849}

1397

850

1398func (c *coordinator) refreshOAuth2Token(ctx context.Context, providerCfg config.ProviderConfig) error {

851func (c *coordinator) refreshOAuth2Token(ctx context.Context, providerCfg config.ProviderConfig) error {

1399− if err := c.cfg.RefreshOAuthToken(ctx, config.ScopeGlobal, providerCfg.ID); err != nil {

852+ if err := c.cfg.RefreshOAuthToken(ctx, providerCfg.ID); err != nil {

1400 slog.Error("Failed to refresh OAuth token after 401 error", "provider", providerCfg.ID, "error", err)

853 slog.Error("Failed to refresh OAuth token after 401 error", "provider", providerCfg.ID, "error", err)

1401 return err

854 return err

1402 }

855 }

@@ -1414,248 +867,10 @@ func (c *coordinator) refreshApiKeyTemplate(ctx context.Context, providerCfg con

1414 }

867 }

1415

868

1416 providerCfg.APIKey = newAPIKey

869 providerCfg.APIKey = newAPIKey

1417− c.cfg.Config().Providers.Set(providerCfg.ID, providerCfg)

870+ c.cfg.Providers.Set(providerCfg.ID, providerCfg)

1418

871

1419 if err := c.UpdateModels(ctx); err != nil {

872 if err := c.UpdateModels(ctx); err != nil {

1420 return err

873 return err

1421 }

874 }

1422 return nil

875 return nil

1423}

876}

1424−

1425−// subAgentParams holds the parameters for running a sub-agent.

1426−type subAgentParams struct {

1427− Agent SessionAgent

1428− SessionID string

1429− AgentMessageID string

1430− ToolCallID string

1431− Prompt string

1432− SessionTitle string

1433− // SessionSetup is an optional callback invoked after session creation

1434− // but before agent execution, for custom session configuration.

1435− SessionSetup func(sessionID string)

1436−}

1437−

1438−// runSubAgent runs a sub-agent and handles session management and cost accumulation.

1439−// It creates a sub-session, runs the agent with the given prompt, and propagates

1440−// the cost to the parent session.

1441−func (c *coordinator) runSubAgent(ctx context.Context, params subAgentParams) (fantasy.ToolResponse, error) {

1442− // Create sub-session

1443− agentToolSessionID := c.sessions.CreateAgentToolSessionID(params.AgentMessageID, params.ToolCallID)

1444− session, err := c.sessions.CreateTaskSession(ctx, agentToolSessionID, params.SessionID, params.SessionTitle)

1445− if err != nil {

1446− return fantasy.ToolResponse{}, fmt.Errorf("create session: %w", err)

1447− }

1448−

1449− // Call session setup function if provided

1450− if params.SessionSetup != nil {

1451− params.SessionSetup(session.ID)

1452− }

1453−

1454− // Get model configuration

1455− model := params.Agent.Model()

1456− maxTokens := model.CatwalkCfg.DefaultMaxTokens

1457− if model.ModelCfg.MaxTokens != 0 {

1458− maxTokens = model.ModelCfg.MaxTokens

1459− }

1460−

1461− providerCfg, ok := c.cfg.Config().Providers.Get(model.ModelCfg.Provider)

1462− if !ok {

1463− return fantasy.ToolResponse{}, errModelProviderNotConfigured

1464− }

1465−

1466− // Run the agent

1467− run := func() (*fantasy.AgentResult, error) {

1468− return params.Agent.Run(ctx, SessionAgentCall{

1469− SessionID: session.ID,

1470− Prompt: params.Prompt,

1471− MaxOutputTokens: maxTokens,

1472− ProviderOptions: getProviderOptions(model, providerCfg),

1473− Temperature: model.ModelCfg.Temperature,

1474− TopP: model.ModelCfg.TopP,

1475− TopK: model.ModelCfg.TopK,

1476− FrequencyPenalty: model.ModelCfg.FrequencyPenalty,

1477− PresencePenalty: model.ModelCfg.PresencePenalty,

1478− NonInteractive: true,

1479− })

1480− }

1481− var result *fantasy.AgentResult

1482− err = c.runWithUnauthorizedRetry(ctx, providerCfg, func() error {

1483− var runErr error

1484− result, runErr = run()

1485− return runErr

1486− })

1487− // Notify only if still unauthorized after retry.

1488− if err != nil && c.isUnauthorized(err) && c.notify != nil && model.ModelCfg.Provider == hyper.Name {

1489− c.notify.Publish(pubsub.CreatedEvent, notify.Notification{

1490− Type: notify.TypeReAuthenticate,

1491− ProviderID: model.ModelCfg.Provider,

1492− })

1493− }

1494− if err != nil {

1495− return fantasy.NewTextErrorResponse(fmt.Sprintf("Failed to generate response: %s", err)), nil

1496− }

1497−

1498− // Update parent session cost on a best-effort basis. A failure here must

1499− // not discard the sub-agent output that was already produced.

1500− if err := c.updateParentSessionCost(ctx, session.ID, params.SessionID); err != nil {

1501− slog.Warn(

1502− "Failed to update parent session cost",

1503− "child_session", session.ID,

1504− "parent_session", params.SessionID,

1505− "error", err,

1506− )

1507− }

1508−

1509− output := subAgentOutput(result)

1510− if output == "" {

1511− return fantasy.NewTextErrorResponse("Sub-agent completed but produced no text output."), nil

1512− }

1513− return fantasy.NewTextResponse(output), nil

1514−}

1515−

1516−func subAgentOutput(result *fantasy.AgentResult) string {

1517− if result == nil {

1518− return ""

1519− }

1520− return result.Response.Content.Text()

1521−}

1522−

1523−// updateParentSessionCost accumulates the cost from a child session to its parent session.

1524−func (c *coordinator) updateParentSessionCost(ctx context.Context, childSessionID, parentSessionID string) error {

1525− childSession, err := c.sessions.Get(ctx, childSessionID)

1526− if err != nil {

1527− return fmt.Errorf("get child session: %w", err)

1528− }

1529−

1530− parentSession, err := c.sessions.Get(ctx, parentSessionID)

1531− if err != nil {

1532− return fmt.Errorf("get parent session: %w", err)

1533− }

1534−

1535− parentSession.Cost += childSession.Cost

1536−

1537− if _, err := c.sessions.Save(ctx, parentSession); err != nil {

1538− return fmt.Errorf("save parent session: %w", err)

1539− }

1540−

1541− return nil

1542−}

1543−

1544−// discoverSkills is a thin fallback wrapper used only when no

1545−// skills.Manager has been threaded through to the coordinator. All

1546−// production call sites (backend.CreateWorkspace, setupLocalWorkspace)

1547−// run discovery in advance and pass the results via the manager;

1548−// reaching this path means a caller bypassed both. It deliberately does

1549−// NOT publish to the package-level broker — there are no subscribers in

1550−// that case, so doing so would be misleading without delivering the

1551−// snapshot anywhere useful.

1552−func discoverSkills(cfg *config.ConfigStore) (allSkills, activeSkills []*skills.Skill) {

1553− opts := cfg.Config().Options

1554− var paths, disabled []string

1555− if opts != nil {

1556− paths = opts.SkillsPaths

1557− disabled = opts.DisabledSkills

1558− }

1559− var resolver func(string) (string, error)

1560− if r := cfg.Resolver(); r != nil {

1561− resolver = r.ResolveValue

1562− }

1563− allSkills, activeSkills, states := skills.DiscoverFromConfig(skills.DiscoveryConfig{

1564− SkillsPaths: paths,

1565− DisabledSkills: disabled,

1566− Resolver: resolver,

1567− })

1568− logDiscoveryStats(states, paths, allSkills, activeSkills, disabled)

1569− return allSkills, activeSkills

1570−}

1571−

1572−// logTurnSkillUsage emits a per-turn diagnostic line showing which skills

1573−// (if any) were loaded during this turn and which looked relevant based on

1574−// a cheap keyword match against the user prompt. The goal is to surface

1575−// "should-have-loaded but didn't" situations for later analysis.

1576−//

1577−// Logged at Info level under component=skills; heavy fields are elided when

1578−// there is nothing interesting to report.

1579−func logTurnSkillUsage(

1580− sessionID string,

1581− prompt string,

1582− activeSkills []*skills.Skill,

1583− tracker *skills.Tracker,

1584− before []string,

1585−) {

1586− if tracker == nil || len(activeSkills) == 0 {

1587− return

1588− }

1589−

1590− after := tracker.LoadedNames()

1591−

1592− beforeSet := make(map[string]bool, len(before))

1593− for _, n := range before {

1594− beforeSet[n] = true

1595− }

1596− var loadedThisTurn []string

1597− for _, n := range after {

1598− if !beforeSet[n] {

1599− loadedThisTurn = append(loadedThisTurn, n)

1600− }

1601− }

1602−

1603− slog.Info(

1604− "Skill turn summary",

1605− "component", "skills",

1606− "session_id", sessionID,

1607− "prompt_len", len(prompt),

1608− "active_total", len(activeSkills),

1609− "loaded_total", len(after),

1610− "loaded_this_turn", loadedThisTurn,

1611− )

1612−}

1613−

1614−// logDiscoveryStats emits a single structured log line summarising skill

1615−// discovery for the current session. It is intentionally low-volume: one

1616−// line per session start. Builtin vs user counts are derived from the

1617−// SkillState.Path — builtin states use the "builtin/" embed prefix.

1618−func logDiscoveryStats(

1619− states []*skills.SkillState,

1620− userPaths []string,

1621− allSkills, activeSkills []*skills.Skill,

1622− disabled []string,

1623−) {

1624− var builtinOK, builtinErr, userOK, userErr int

1625− for _, s := range states {

1626− isBuiltin := strings.HasPrefix(s.Path, "builtin/")

1627− switch {

1628− case isBuiltin && s.State == skills.StateNormal:

1629− builtinOK++

1630− case isBuiltin && s.State == skills.StateError:

1631− builtinErr++

1632− case !isBuiltin && s.State == skills.StateNormal:

1633− userOK++

1634− case !isBuiltin && s.State == skills.StateError:

1635− userErr++

1636− }

1637− }

1638−

1639− activeNames := make([]string, 0, len(activeSkills))

1640− for _, s := range activeSkills {

1641− activeNames = append(activeNames, s.Name)

1642− }

1643−

1644− xml := skills.ToPromptXML(activeSkills)

1645−

1646− slog.Info(

1647− "Skill discovery complete",

1648− "component", "skills",

1649− "builtin_ok", builtinOK,

1650− "builtin_errors", builtinErr,

1651− "user_ok", userOK,

1652− "user_errors", userErr,

1653− "user_paths", len(userPaths),

1654− "deduped_total", len(allSkills),

1655− "active", len(activeSkills),

1656− "disabled", len(disabled),

1657− "prompt_bytes", len(xml),

1658− "prompt_tok_est", skills.ApproxTokenCount(xml),

1659− "active_names", activeNames,

1660− )

1661−}

internal/agent/coordinator_test.godeleted−554

@@ -1,554 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "errors"

6− "testing"

7−

8− "charm.land/catwalk/pkg/catwalk"

9− "charm.land/fantasy"

10− "charm.land/fantasy/providers/anthropic"

11− "charm.land/fantasy/providers/bedrock"

12− "charm.land/fantasy/providers/openaicompat"

13− "github.com/charmbracelet/crush/internal/config"

14− "github.com/stretchr/testify/assert"

15− "github.com/stretchr/testify/require"

16−)

17−

18−// mockSessionAgent is a minimal mock for the SessionAgent interface.

19−type mockSessionAgent struct {

20− model Model

21− runFunc func(ctx context.Context, call SessionAgentCall) (*fantasy.AgentResult, error)

22− cancelled []string

23−}

24−

25−func (m *mockSessionAgent) Run(ctx context.Context, call SessionAgentCall) (*fantasy.AgentResult, error) {

26− return m.runFunc(ctx, call)

27−}

28−

29−func (m *mockSessionAgent) BeginAccepted(sessionID string) *AcceptedRun {

30− return &AcceptedRun{sessionID: sessionID}

31−}

32−

33−func (m *mockSessionAgent) Model() Model { return m.model }

34−func (m *mockSessionAgent) SetModels(large, small Model) {}

35−func (m *mockSessionAgent) SetTools(tools []fantasy.AgentTool) {}

36−func (m *mockSessionAgent) SetSystemPrompt(systemPrompt string) {}

37−func (m *mockSessionAgent) Cancel(sessionID string) {

38− m.cancelled = append(m.cancelled, sessionID)

39−}

40−func (m *mockSessionAgent) CancelAll() {}

41−func (m *mockSessionAgent) IsSessionBusy(sessionID string) bool { return false }

42−func (m *mockSessionAgent) IsBusy() bool { return false }

43−func (m *mockSessionAgent) QueuedPrompts(sessionID string) int { return 0 }

44−func (m *mockSessionAgent) QueuedPromptsList(sessionID string) []string { return nil }

45−func (m *mockSessionAgent) ClearQueue(sessionID string) {}

46−func (m *mockSessionAgent) Summarize(context.Context, string, fantasy.ProviderOptions) error {

47− return nil

48−}

49−func (m *mockSessionAgent) GenerateTitle(context.Context, string, string) {}

50−

51−// newTestCoordinator creates a minimal coordinator for unit testing runSubAgent.

52−func newTestCoordinator(t *testing.T, env fakeEnv, providerID string, providerCfg config.ProviderConfig) *coordinator {

53− cfg, err := config.Init(env.workingDir, "", false)

54− require.NoError(t, err)

55− cfg.Config().Providers.Set(providerID, providerCfg)

56− return &coordinator{

57− cfg: cfg,

58− sessions: env.sessions,

59− messages: env.messages,

60− }

61−}

62−

63−// newMockAgent creates a mockSessionAgent with the given provider and run function.

64−func newMockAgent(providerID string, maxTokens int64, runFunc func(context.Context, SessionAgentCall) (*fantasy.AgentResult, error)) *mockSessionAgent {

65− return &mockSessionAgent{

66− model: Model{

67− CatwalkCfg: catwalk.Model{

68− DefaultMaxTokens: maxTokens,

69− },

70− ModelCfg: config.SelectedModel{

71− Provider: providerID,

72− },

73− },

74− runFunc: runFunc,

75− }

76−}

77−

78−// agentResultWithText creates a minimal AgentResult with the given text response.

79−func agentResultWithText(text string) *fantasy.AgentResult {

80− return &fantasy.AgentResult{

81− Response: fantasy.Response{

82− Content: fantasy.ResponseContent{

83− fantasy.TextContent{Text: text},

84− },

85− },

86− }

87−}

88−

89−func TestRunSubAgent(t *testing.T) {

90− const providerID = "test-provider"

91− providerCfg := config.ProviderConfig{ID: providerID}

92−

93− t.Run("happy path", func(t *testing.T) {

94− env := testEnv(t)

95− coord := newTestCoordinator(t, env, providerID, providerCfg)

96−

97− parentSession, err := env.sessions.Create(t.Context(), "Parent")

98− require.NoError(t, err)

99−

100− agent := newMockAgent(providerID, 4096, func(_ context.Context, call SessionAgentCall) (*fantasy.AgentResult, error) {

101− assert.Equal(t, "do something", call.Prompt)

102− assert.Equal(t, int64(4096), call.MaxOutputTokens)

103− return agentResultWithText("done"), nil

104− })

105−

106− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

107− Agent: agent,

108− SessionID: parentSession.ID,

109− AgentMessageID: "msg-1",

110− ToolCallID: "call-1",

111− Prompt: "do something",

112− SessionTitle: "Test Session",

113− })

114− require.NoError(t, err)

115− assert.Equal(t, "done", resp.Content)

116− assert.False(t, resp.IsError)

117− })

118−

119− t.Run("cost update failure preserves output", func(t *testing.T) {

120− env := testEnv(t)

121− coord := newTestCoordinator(t, env, providerID, providerCfg)

122−

123− agent := newMockAgent(providerID, 4096, func(_ context.Context, _ SessionAgentCall) (*fantasy.AgentResult, error) {

124− return agentResultWithText("output before cost failure"), nil

125− })

126−

127− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

128− Agent: agent,

129− SessionID: "missing-parent-session",

130− AgentMessageID: "msg-1",

131− ToolCallID: "call-1",

132− Prompt: "test",

133− SessionTitle: "Test",

134− })

135− require.NoError(t, err)

136− assert.False(t, resp.IsError)

137− assert.Equal(t, "output before cost failure", resp.Content)

138− })

139−

140− t.Run("response with text returns it", func(t *testing.T) {

141− env := testEnv(t)

142− coord := newTestCoordinator(t, env, providerID, providerCfg)

143−

144− parentSession, err := env.sessions.Create(t.Context(), "Parent")

145− require.NoError(t, err)

146−

147− agent := newMockAgent(providerID, 4096, func(_ context.Context, _ SessionAgentCall) (*fantasy.AgentResult, error) {

148− return agentResultWithText("the answer"), nil

149− })

150−

151− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

152− Agent: agent,

153− SessionID: parentSession.ID,

154− AgentMessageID: "msg-1",

155− ToolCallID: "call-1",

156− Prompt: "test",

157− SessionTitle: "Test",

158− })

159− require.NoError(t, err)

160− assert.False(t, resp.IsError)

161− assert.Equal(t, "the answer", resp.Content)

162− })

163−

164− t.Run("nil result returns error response", func(t *testing.T) {

165− env := testEnv(t)

166− coord := newTestCoordinator(t, env, providerID, providerCfg)

167−

168− parentSession, err := env.sessions.Create(t.Context(), "Parent")

169− require.NoError(t, err)

170−

171− agent := newMockAgent(providerID, 4096, func(_ context.Context, _ SessionAgentCall) (*fantasy.AgentResult, error) {

172− return nil, nil

173− })

174−

175− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

176− Agent: agent,

177− SessionID: parentSession.ID,

178− AgentMessageID: "msg-1",

179− ToolCallID: "call-1",

180− Prompt: "test",

181− SessionTitle: "Test",

182− })

183− require.NoError(t, err)

184− assert.True(t, resp.IsError)

185− assert.Equal(t, "Sub-agent completed but produced no text output.", resp.Content)

186− })

187−

188− t.Run("empty result returns error response", func(t *testing.T) {

189− env := testEnv(t)

190− coord := newTestCoordinator(t, env, providerID, providerCfg)

191−

192− parentSession, err := env.sessions.Create(t.Context(), "Parent")

193− require.NoError(t, err)

194−

195− agent := newMockAgent(providerID, 4096, func(_ context.Context, _ SessionAgentCall) (*fantasy.AgentResult, error) {

196− return &fantasy.AgentResult{}, nil

197− })

198−

199− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

200− Agent: agent,

201− SessionID: parentSession.ID,

202− AgentMessageID: "msg-1",

203− ToolCallID: "call-1",

204− Prompt: "test",

205− SessionTitle: "Test",

206− })

207− require.NoError(t, err)

208− assert.True(t, resp.IsError)

209− assert.Equal(t, "Sub-agent completed but produced no text output.", resp.Content)

210− })

211−

212− t.Run("ModelCfg.MaxTokens overrides default", func(t *testing.T) {

213− env := testEnv(t)

214− coord := newTestCoordinator(t, env, providerID, providerCfg)

215−

216− parentSession, err := env.sessions.Create(t.Context(), "Parent")

217− require.NoError(t, err)

218−

219− agent := &mockSessionAgent{

220− model: Model{

221− CatwalkCfg: catwalk.Model{

222− DefaultMaxTokens: 4096,

223− },

224− ModelCfg: config.SelectedModel{

225− Provider: providerID,

226− MaxTokens: 8192,

227− },

228− },

229− runFunc: func(_ context.Context, call SessionAgentCall) (*fantasy.AgentResult, error) {

230− assert.Equal(t, int64(8192), call.MaxOutputTokens)

231− return agentResultWithText("ok"), nil

232− },

233− }

234−

235− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

236− Agent: agent,

237− SessionID: parentSession.ID,

238− AgentMessageID: "msg-1",

239− ToolCallID: "call-1",

240− Prompt: "test",

241− SessionTitle: "Test",

242− })

243− require.NoError(t, err)

244− assert.Equal(t, "ok", resp.Content)

245− })

246−

247− t.Run("session creation failure with canceled context", func(t *testing.T) {

248− env := testEnv(t)

249− coord := newTestCoordinator(t, env, providerID, providerCfg)

250−

251− parentSession, err := env.sessions.Create(t.Context(), "Parent")

252− require.NoError(t, err)

253−

254− agent := newMockAgent(providerID, 4096, nil)

255−

256− // Use a canceled context to trigger CreateTaskSession failure.

257− ctx, cancel := context.WithCancel(t.Context())

258− cancel()

259−

260− _, err = coord.runSubAgent(ctx, subAgentParams{

261− Agent: agent,

262− SessionID: parentSession.ID,

263− AgentMessageID: "msg-1",

264− ToolCallID: "call-1",

265− Prompt: "test",

266− SessionTitle: "Test",

267− })

268− require.Error(t, err)

269− })

270−

271− t.Run("provider not configured", func(t *testing.T) {

272− env := testEnv(t)

273− coord := newTestCoordinator(t, env, providerID, providerCfg)

274−

275− parentSession, err := env.sessions.Create(t.Context(), "Parent")

276− require.NoError(t, err)

277−

278− // Agent references a provider that doesn't exist in config.

279− agent := newMockAgent("unknown-provider", 4096, nil)

280−

281− _, err = coord.runSubAgent(t.Context(), subAgentParams{

282− Agent: agent,

283− SessionID: parentSession.ID,

284− AgentMessageID: "msg-1",

285− ToolCallID: "call-1",

286− Prompt: "test",

287− SessionTitle: "Test",

288− })

289− require.Error(t, err)

290− assert.Contains(t, err.Error(), "model provider not configured")

291− })

292−

293− t.Run("agent run error returns error response", func(t *testing.T) {

294− env := testEnv(t)

295− coord := newTestCoordinator(t, env, providerID, providerCfg)

296−

297− parentSession, err := env.sessions.Create(t.Context(), "Parent")

298− require.NoError(t, err)

299−

300− agent := newMockAgent(providerID, 4096, func(_ context.Context, _ SessionAgentCall) (*fantasy.AgentResult, error) {

301− return nil, errors.New("provider request failed")

302− })

303−

304− resp, err := coord.runSubAgent(t.Context(), subAgentParams{

305− Agent: agent,

306− SessionID: parentSession.ID,

307− AgentMessageID: "msg-1",

308− ToolCallID: "call-1",

309− Prompt: "test",

310− SessionTitle: "Test",

311− })

312− // runSubAgent returns (errorResponse, nil) when agent.Run fails — not a Go error.

313− require.NoError(t, err)

314− assert.True(t, resp.IsError)

315− assert.Equal(t, "Failed to generate response: provider request failed", resp.Content)

316− })

317−

318− t.Run("session setup callback is invoked", func(t *testing.T) {

319− env := testEnv(t)

320− coord := newTestCoordinator(t, env, providerID, providerCfg)

321−

322− parentSession, err := env.sessions.Create(t.Context(), "Parent")

323− require.NoError(t, err)

324−

325− var setupCalledWith string

326− agent := newMockAgent(providerID, 4096, func(_ context.Context, _ SessionAgentCall) (*fantasy.AgentResult, error) {

327− return agentResultWithText("ok"), nil

328− })

329−

330− _, err = coord.runSubAgent(t.Context(), subAgentParams{

331− Agent: agent,

332− SessionID: parentSession.ID,

333− AgentMessageID: "msg-1",

334− ToolCallID: "call-1",

335− Prompt: "test",

336− SessionTitle: "Test",

337− SessionSetup: func(sessionID string) {

338− setupCalledWith = sessionID

339− },

340− })

341− require.NoError(t, err)

342− assert.NotEmpty(t, setupCalledWith, "SessionSetup should have been called")

343− })

344−

345− t.Run("cost propagation to parent session", func(t *testing.T) {

346− env := testEnv(t)

347− coord := newTestCoordinator(t, env, providerID, providerCfg)

348−

349− parentSession, err := env.sessions.Create(t.Context(), "Parent")

350− require.NoError(t, err)

351−

352− agent := newMockAgent(providerID, 4096, func(ctx context.Context, call SessionAgentCall) (*fantasy.AgentResult, error) {

353− // Simulate the agent incurring cost by updating the child session.

354− childSession, err := env.sessions.Get(ctx, call.SessionID)

355− if err != nil {

356− return nil, err

357− }

358− childSession.Cost = 0.05

359− _, err = env.sessions.Save(ctx, childSession)

360− if err != nil {

361− return nil, err

362− }

363− return agentResultWithText("ok"), nil

364− })

365−

366− _, err = coord.runSubAgent(t.Context(), subAgentParams{

367− Agent: agent,

368− SessionID: parentSession.ID,

369− AgentMessageID: "msg-1",

370− ToolCallID: "call-1",

371− Prompt: "test",

372− SessionTitle: "Test",

373− })

374− require.NoError(t, err)

375−

376− updated, err := env.sessions.Get(t.Context(), parentSession.ID)

377− require.NoError(t, err)

378− assert.InDelta(t, 0.05, updated.Cost, 1e-9)

379− })

380−}

381−

382−func TestUpdateParentSessionCost(t *testing.T) {

383− t.Run("accumulates cost correctly", func(t *testing.T) {

384− env := testEnv(t)

385− cfg, err := config.Init(env.workingDir, "", false)

386− require.NoError(t, err)

387− coord := &coordinator{cfg: cfg, sessions: env.sessions}

388−

389− parent, err := env.sessions.Create(t.Context(), "Parent")

390− require.NoError(t, err)

391−

392− child, err := env.sessions.CreateTaskSession(t.Context(), "tool-1", parent.ID, "Child")

393− require.NoError(t, err)

394−

395− // Set child cost.

396− child.Cost = 0.10

397− _, err = env.sessions.Save(t.Context(), child)

398− require.NoError(t, err)

399−

400− err = coord.updateParentSessionCost(t.Context(), child.ID, parent.ID)

401− require.NoError(t, err)

402−

403− updated, err := env.sessions.Get(t.Context(), parent.ID)

404− require.NoError(t, err)

405− assert.InDelta(t, 0.10, updated.Cost, 1e-9)

406− })

407−

408− t.Run("accumulates multiple child costs", func(t *testing.T) {

409− env := testEnv(t)

410− cfg, err := config.Init(env.workingDir, "", false)

411− require.NoError(t, err)

412− coord := &coordinator{cfg: cfg, sessions: env.sessions}

413−

414− parent, err := env.sessions.Create(t.Context(), "Parent")

415− require.NoError(t, err)

416−

417− child1, err := env.sessions.CreateTaskSession(t.Context(), "tool-1", parent.ID, "Child1")

418− require.NoError(t, err)

419− child1.Cost = 0.05

420− _, err = env.sessions.Save(t.Context(), child1)

421− require.NoError(t, err)

422−

423− child2, err := env.sessions.CreateTaskSession(t.Context(), "tool-2", parent.ID, "Child2")

424− require.NoError(t, err)

425− child2.Cost = 0.03

426− _, err = env.sessions.Save(t.Context(), child2)

427− require.NoError(t, err)

428−

429− err = coord.updateParentSessionCost(t.Context(), child1.ID, parent.ID)

430− require.NoError(t, err)

431− err = coord.updateParentSessionCost(t.Context(), child2.ID, parent.ID)

432− require.NoError(t, err)

433−

434− updated, err := env.sessions.Get(t.Context(), parent.ID)

435− require.NoError(t, err)

436− assert.InDelta(t, 0.08, updated.Cost, 1e-9)

437− })

438−

439− t.Run("child session not found", func(t *testing.T) {

440− env := testEnv(t)

441− cfg, err := config.Init(env.workingDir, "", false)

442− require.NoError(t, err)

443− coord := &coordinator{cfg: cfg, sessions: env.sessions}

444−

445− parent, err := env.sessions.Create(t.Context(), "Parent")

446− require.NoError(t, err)

447−

448− err = coord.updateParentSessionCost(t.Context(), "non-existent", parent.ID)

449− require.Error(t, err)

450− assert.Contains(t, err.Error(), "get child session")

451− })

452−

453− t.Run("parent session not found", func(t *testing.T) {

454− env := testEnv(t)

455− cfg, err := config.Init(env.workingDir, "", false)

456− require.NoError(t, err)

457− coord := &coordinator{cfg: cfg, sessions: env.sessions}

458−

459− parent, err := env.sessions.Create(t.Context(), "Parent")

460− require.NoError(t, err)

461− child, err := env.sessions.CreateTaskSession(t.Context(), "tool-1", parent.ID, "Child")

462− require.NoError(t, err)

463−

464− err = coord.updateParentSessionCost(t.Context(), child.ID, "non-existent")

465− require.Error(t, err)

466− assert.Contains(t, err.Error(), "get parent session")

467− })

468−

469− t.Run("zero cost handled correctly", func(t *testing.T) {

470− env := testEnv(t)

471− cfg, err := config.Init(env.workingDir, "", false)

472− require.NoError(t, err)

473− coord := &coordinator{cfg: cfg, sessions: env.sessions}

474−

475− parent, err := env.sessions.Create(t.Context(), "Parent")

476− require.NoError(t, err)

477− child, err := env.sessions.CreateTaskSession(t.Context(), "tool-1", parent.ID, "Child")

478− require.NoError(t, err)

479−

480− err = coord.updateParentSessionCost(t.Context(), child.ID, parent.ID)

481− require.NoError(t, err)

482−

483− updated, err := env.sessions.Get(t.Context(), parent.ID)

484− require.NoError(t, err)

485− assert.InDelta(t, 0.0, updated.Cost, 1e-9)

486− })

487−}

488−

489−func TestGetProviderOptionsReasoningEffort(t *testing.T) {

490− // Bedrock is Fantasy's Anthropic under a different provider name; options

491− // must land under anthropic.Name so the Anthropic language model picks them up.

492− tests := []struct {

493− name string

494− providerType catwalk.Type

495− }{

496− {"anthropic honors reasoning_effort", catwalk.Type(anthropic.Name)},

497− {"bedrock honors reasoning_effort", catwalk.Type(bedrock.Name)},

498− }

499− for _, tc := range tests {

500− t.Run(tc.name, func(t *testing.T) {

501− model := Model{

502− CatwalkCfg: catwalk.Model{

503− ID: "claude-opus-4-7",

504− CanReason: true,

505− ReasoningLevels: []string{"max"},

506− },

507− ModelCfg: config.SelectedModel{

508− Provider: "test",

509− ReasoningEffort: "max",

510− },

511− }

512− providerCfg := config.ProviderConfig{ID: "test", Type: tc.providerType}

513−

514− opts := getProviderOptions(model, providerCfg)

515−

516− raw, ok := opts[anthropic.Name]

517− require.True(t, ok, "options should be keyed under anthropic.Name for type %q", tc.providerType)

518− parsed, ok := raw.(*anthropic.ProviderOptions)

519− require.True(t, ok)

520− require.NotNil(t, parsed.Effort)

521− assert.Equal(t, anthropic.Effort("max"), *parsed.Effort)

522− })

523− }

524−}

525−

526−func TestGetProviderOptionsReasoningEffortFallback(t *testing.T) {

527− model := Model{

528− CatwalkCfg: catwalk.Model{

529− ID: "glm-5.2",

530− CanReason: true,

531− ReasoningLevels: []string{"high", "max"},

532− },

533− ModelCfg: config.SelectedModel{

534− Provider: "zai",

535− },

536− }

537− providerCfg := config.ProviderConfig{

538− ID: string(catwalk.InferenceProviderZAI),

539− Type: openaicompat.Name,

540− }

541−

542− opts := getProviderOptions(model, providerCfg)

543−

544− raw, ok := opts[openaicompat.Name]

545− require.True(t, ok)

546− parsed, ok := raw.(*openaicompat.ProviderOptions)

547− require.True(t, ok)

548− require.NotNil(t, parsed.ReasoningEffort)

549− assert.Equal(t, "high", string(*parsed.ReasoningEffort))

550−

551− thinking, ok := parsed.ExtraBody["thinking"].(map[string]any)

552− require.True(t, ok)

553− assert.Equal(t, "enabled", thinking["type"])

554−}

internal/agent/dispatch_cancel_test.godeleted−447

@@ -1,447 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "errors"

6− "sync/atomic"

7− "testing"

8− "time"

9−

10− "charm.land/fantasy"

11− "github.com/charmbracelet/crush/internal/agent/notify"

12− "github.com/charmbracelet/crush/internal/message"

13− "github.com/charmbracelet/crush/internal/pubsub"

14− "github.com/stretchr/testify/assert"

15− "github.com/stretchr/testify/require"

16−)

17−

18−// finishStreamModel is a minimal fantasy.LanguageModel that streams a

19−// single text part followed by a normal (FinishReasonStop) finish. It

20−// is enough to drive sessionAgent.Run through PrepareStep and a clean

21−// completion without a recorded provider cassette.

22−type finishStreamModel struct {

23− text string

24−}

25−

26−func (m *finishStreamModel) Provider() string { return "fake" }

27−func (m *finishStreamModel) Model() string { return "fake-model" }

28−

29−func (m *finishStreamModel) Generate(ctx context.Context, call fantasy.Call) (*fantasy.Response, error) {

30− return &fantasy.Response{

31− Content: fantasy.ResponseContent{fantasy.TextContent{Text: m.text}},

32− FinishReason: fantasy.FinishReasonStop,

33− }, nil

34−}

35−

36−func (m *finishStreamModel) Stream(ctx context.Context, call fantasy.Call) (fantasy.StreamResponse, error) {

37− text := m.text

38− return func(yield func(fantasy.StreamPart) bool) {

39− if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextStart, ID: "1"}) {

40− return

41− }

42− if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextDelta, ID: "1", Delta: text}) {

43− return

44− }

45− if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextEnd, ID: "1"}) {

46− return

47− }

48− yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonStop})

49− }, nil

50−}

51−

52−func (m *finishStreamModel) GenerateObject(ctx context.Context, call fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {

53− return nil, errors.New("not implemented")

54−}

55−

56−func (m *finishStreamModel) StreamObject(ctx context.Context, call fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {

57− return nil, errors.New("not implemented")

58−}

59−

60−func newStreamTestAgent(t *testing.T) (*sessionAgent, fakeEnv) {

61− t.Helper()

62− env := testEnv(t)

63− model := &finishStreamModel{text: "done"}

64− sa := testSessionAgent(env, model, model, "system").(*sessionAgent)

65− return sa, env

66−}

67−

68−// TestCancel_ActiveAndAcceptedFiresBothBranches covers the case where a

69−// session is actively running (activeRequests set) AND a follow-up has

70−// been accepted (acceptedRuns > 0). A single Cancel must fire both: it

71−// invokes the active cancel func and records a pending cancel for the

72−// accepted follow-up.

73−func TestCancel_ActiveAndAcceptedFiresBothBranches(t *testing.T) {

74− t.Parallel()

75− sa, _ := newCancelTestAgent(t)

76−

77− const sid = "sid"

78− var activeCanceled atomic.Bool

79− sa.activeRequests.Set(sid, func() { activeCanceled.Store(true) })

80−

81− accept := sa.BeginAccepted(sid)

82− defer accept.Close()

83−

84− sa.Cancel(sid)

85−

86− require.True(t, activeCanceled.Load(), "active cancel func must fire")

87− require.True(t, sa.hasPendingCancel(sid), "accepted follow-up must record a pending cancel")

88−}

89−

90−// TestRun_BusyWithPendingCancelTakesCancelOnEntry covers the busy-queue

91−// branch consulting pendingCancels: when the session is busy AND a

92−// cancel has been recorded for an accepted follow-up, Run must take the

93−// cancel-on-entry path (persist a canceled turn) instead of enqueueing

94−// the call behind the active run.

95−func TestRun_BusyWithPendingCancelTakesCancelOnEntry(t *testing.T) {

96− t.Parallel()

97− sa, env := newCancelTestAgent(t)

98−

99− sess, err := env.sessions.Create(t.Context(), "session")

100− require.NoError(t, err)

101−

102− // Make the session look busy: an earlier prompt is active.

103− sa.activeRequests.Set(sess.ID, func() {})

104−

105− accept := sa.BeginAccepted(sess.ID)

106− // A cancel arrives while this follow-up is accepted-but-not-active.

107− sa.Cancel(sess.ID)

108− require.True(t, sa.hasPendingCancel(sess.ID))

109−

110− result, err := sa.Run(t.Context(), SessionAgentCall{

111− SessionID: sess.ID,

112− Prompt: "follow-up",

113− Accepted: accept,

114− })

115− require.NoError(t, err)

116− require.Nil(t, result)

117−

118− // The follow-up was canceled on entry, not enqueued.

119− require.Equal(t, 0, sa.QueuedPrompts(sess.ID),

120− "cancel-on-entry must not enqueue the follow-up behind the active run")

121− require.False(t, sa.hasPendingCancel(sess.ID), "pending cancel must be consumed")

122− require.Equal(t, 0, sa.acceptedCount(sess.ID), "accept reservation must be released")

123−

124− msgs, err := env.messages.List(t.Context(), sess.ID)

125− require.NoError(t, err)

126− require.Len(t, msgs, 2)

127− assert.Equal(t, message.User, msgs[0].Role)

128− assert.Equal(t, message.Assistant, msgs[1].Role)

129− assert.Equal(t, message.FinishReasonCanceled, msgs[1].FinishReason())

130−}

131−

132−// TestRun_PrepareStepDrainSkipsQueuedOnPendingCancel verifies that the

133−// queue drain inside PrepareStep skips queued follow-up prompts when a

134−// cancel has been recorded for the session: the queued prompt must not

135−// be folded into the active turn as an extra user message.

136−func TestRun_PrepareStepDrainSkipsQueuedOnPendingCancel(t *testing.T) {

137− t.Parallel()

138− sa, env := newStreamTestAgent(t)

139−

140− sess, err := env.sessions.Create(t.Context(), "session")

141− require.NoError(t, err)

142−

143− // A follow-up prompt sits queued for the session.

144− sa.enqueueCall(SessionAgentCall{SessionID: sess.ID, Prompt: "queued-followup"})

145− // A cancel was recorded for the session while it sat in the queue.

146− sa.cancelMark.Set(sess.ID, 1)

147−

148− result, err := sa.Run(t.Context(), SessionAgentCall{

149− SessionID: sess.ID,

150− Prompt: "main",

151− })

152− require.NoError(t, err)

153− require.NotNil(t, result)

154−

155− // Only the main prompt produced a user message; the queued

156− // follow-up was skipped, not folded into the turn.

157− msgs, err := env.messages.List(t.Context(), sess.ID)

158− require.NoError(t, err)

159− var userMsgs []message.Message

160− for _, m := range msgs {

161− if m.Role == message.User {

162− userMsgs = append(userMsgs, m)

163− }

164− }

165− require.Len(t, userMsgs, 1, "queued follow-up must not create a user message")

166− assert.Equal(t, "main", userMsgs[0].Content().String())

167−

168− // The queue was drained and the pending cancel consumed.

169− require.Equal(t, 0, sa.QueuedPrompts(sess.ID))

170− require.False(t, sa.hasPendingCancel(sess.ID))

171−}

172−

173−// TestRun_NormalCompletionClearsStalePendingCancel verifies that a Run

174−// which completes normally clears any stale pending-cancel entry for the

175−// session, so it cannot catch a future run.

176−func TestRun_NormalCompletionClearsStalePendingCancel(t *testing.T) {

177− t.Parallel()

178− sa, env := newStreamTestAgent(t)

179−

180− sess, err := env.sessions.Create(t.Context(), "session")

181− require.NoError(t, err)

182−

183− // A stale cancel mark lingers (no queued work, no accepted run).

184− sa.cancelMark.Set(sess.ID, 1)

185−

186− result, err := sa.Run(t.Context(), SessionAgentCall{

187− SessionID: sess.ID,

188− Prompt: "main",

189− })

190− require.NoError(t, err)

191− require.NotNil(t, result)

192−

193− require.False(t, sa.hasPendingCancel(sess.ID),

194− "normal completion must clear the stale pending cancel")

195−

196− msgs, err := env.messages.List(t.Context(), sess.ID)

197− require.NoError(t, err)

198− require.Len(t, msgs, 2)

199− assert.Equal(t, message.Assistant, msgs[1].Role)

200− assert.Equal(t, message.FinishReasonEndTurn, msgs[1].FinishReason())

201−}

202−

203−// newCancelTestAgentWithRunComplete builds a DB-backed sessionAgent wired

204−// to a RunComplete broker so tests can observe the terminal event a

205−// RunID-bearing caller (e.g. `crush run`) blocks on.

206−func newCancelTestAgentWithRunComplete(t *testing.T) (*sessionAgent, fakeEnv, *pubsub.Broker[notify.RunComplete]) {

207− t.Helper()

208− env := testEnv(t)

209− broker := pubsub.NewBroker[notify.RunComplete]()

210− t.Cleanup(broker.Shutdown)

211− sa := NewSessionAgent(SessionAgentOptions{

212− Sessions: env.sessions,

213− Messages: env.messages,

214− RunComplete: broker,

215− }).(*sessionAgent)

216− return sa, env, broker

217−}

218−

219−// TestRun_CancelOnEntryPublishesRunComplete covers the first review

220−// finding: the cancel-on-entry path returned before the streaming defer

221−// that publishes RunComplete was installed. A caller that dispatches a

222−// run with a RunID and blocks on RunComplete (ignoring message events,

223−// like `crush run`) would hang on an immediately-canceled accepted run.

224−// The cancel-on-entry path must publish a terminal RunComplete carrying

225−// the originating RunID.

226−func TestRun_CancelOnEntryPublishesRunComplete(t *testing.T) {

227− t.Parallel()

228− sa, env, broker := newCancelTestAgentWithRunComplete(t)

229−

230− sess, err := env.sessions.Create(t.Context(), "session")

231− require.NoError(t, err)

232−

233− ctx, cancel := context.WithCancel(t.Context())

234− defer cancel()

235− ch := broker.Subscribe(ctx)

236−

237− accept := sa.BeginAccepted(sess.ID)

238− // A cancel arrives in the accepted-but-not-yet-active window.

239− sa.Cancel(sess.ID)

240− require.True(t, sa.hasPendingCancel(sess.ID))

241−

242− result, err := sa.Run(t.Context(), SessionAgentCall{

243− SessionID: sess.ID,

244− RunID: "run-cancel-on-entry",

245− Prompt: "hello",

246− Accepted: accept,

247− })

248− require.NoError(t, err)

249− require.Nil(t, result)

250−

251− select {

252− case got := <-ch:

253− assert.Equal(t, "run-cancel-on-entry", got.Payload.RunID,

254− "RunComplete must echo the originating RunID")

255− assert.Equal(t, sess.ID, got.Payload.SessionID)

256− assert.True(t, got.Payload.Cancelled,

257− "cancel-on-entry RunComplete must be marked cancelled")

258− case <-time.After(2 * time.Second):

259− t.Fatal("cancel-on-entry must publish RunComplete; a RunID caller would hang otherwise")

260− }

261−}

262−

263−// TestCancel_TwoAcceptedBothObserveCancellation covers the second review

264−// finding: a single cancel with two accepted-not-yet-active prompts must

265−// cancel both. The cancel raises the session's high-water mark to the

266−// latest accept sequence, so every prompt accepted-but-not-yet-active at

267−// cancel time is covered and both take the cancel-on-entry path.

268−func TestCancel_TwoAcceptedBothObserveCancellation(t *testing.T) {

269− t.Parallel()

270− sa, env := newCancelTestAgent(t)

271−

272− sess, err := env.sessions.Create(t.Context(), "session")

273− require.NoError(t, err)

274−

275− // Two prompts are accepted-but-not-yet-active for the same session.

276− accept1 := sa.BeginAccepted(sess.ID)

277− accept2 := sa.BeginAccepted(sess.ID)

278− require.Equal(t, 2, sa.acceptedCount(sess.ID))

279−

280− // A single cancel arrives before either becomes active.

281− sa.Cancel(sess.ID)

282− require.Equal(t, accept2.seq, sa.pendingCancelMark(sess.ID),

283− "one cancel must mark every currently-accepted prompt as canceled")

284− require.GreaterOrEqual(t, sa.pendingCancelMark(sess.ID), accept1.seq,

285− "the mark must cover the earlier accepted prompt too")

286−

287− // Both prompts enter Run; each must take cancel-on-entry, not run.

288− r1, err := sa.Run(t.Context(), SessionAgentCall{

289− SessionID: sess.ID,

290− Prompt: "first",

291− Accepted: accept1,

292− })

293− require.NoError(t, err)

294− require.Nil(t, r1)

295−

296− r2, err := sa.Run(t.Context(), SessionAgentCall{

297− SessionID: sess.ID,

298− Prompt: "second",

299− Accepted: accept2,

300− })

301− require.NoError(t, err)

302− require.Nil(t, r2)

303−

304− require.False(t, sa.hasPendingCancel(sess.ID),

305− "both reserved units must be consumed")

306− require.Equal(t, 0, sa.acceptedCount(sess.ID),

307− "both accept reservations must be released")

308−

309− // Each canceled-on-entry turn writes a user + canceled assistant

310− // message, and neither prompt was enqueued to run normally.

311− require.Equal(t, 0, sa.QueuedPrompts(sess.ID),

312− "neither accepted prompt may be enqueued to run normally")

313− msgs, err := env.messages.List(t.Context(), sess.ID)

314− require.NoError(t, err)

315− require.Len(t, msgs, 4, "two canceled turns produce two user + two assistant messages")

316− var canceled int

317− for _, m := range msgs {

318− if m.Role == message.Assistant {

319− assert.Equal(t, message.FinishReasonCanceled, m.FinishReason())

320− canceled++

321− }

322− }

323− require.Equal(t, 2, canceled, "both turns must finish canceled")

324−}

325−

326−// TestRun_IdleCancelDoesNotPoisonNextPrompt covers the idle-cancel

327−// no-op guarantee end-to-end: an Escape on an idle session must not

328−// record a pending cancel that leaks into the next accepted prompt, which

329−// must run normally to completion.

330−func TestRun_IdleCancelDoesNotPoisonNextPrompt(t *testing.T) {

331− t.Parallel()

332− sa, env := newStreamTestAgent(t)

333−

334− sess, err := env.sessions.Create(t.Context(), "session")

335− require.NoError(t, err)

336−

337− // Idle Escape: no accepted run, no active request.

338− sa.Cancel(sess.ID)

339− require.False(t, sa.hasPendingCancel(sess.ID),

340− "idle cancel must not record a pending cancel")

341−

342− // The next accepted prompt must run normally, not cancel on entry.

343− accept := sa.BeginAccepted(sess.ID)

344− result, err := sa.Run(t.Context(), SessionAgentCall{

345− SessionID: sess.ID,

346− Prompt: "next",

347− Accepted: accept,

348− })

349− require.NoError(t, err)

350− require.NotNil(t, result, "next prompt must run normally after an idle cancel")

351−

352− msgs, err := env.messages.List(t.Context(), sess.ID)

353− require.NoError(t, err)

354− require.Len(t, msgs, 2)

355− assert.Equal(t, message.User, msgs[0].Role)

356− assert.Equal(t, message.Assistant, msgs[1].Role)

357− assert.Equal(t, message.FinishReasonEndTurn, msgs[1].FinishReason(),

358− "the prompt must finish normally, not canceled")

359−}

360−

361−// TestCancel_AcceptedAfterCancelIsNotPoisoned is the regression test for

362−// the reviewer's finding: a counted session-level pending cancel let a

363−// prompt accepted after the cancel enter Run first and consume a unit

364−// reserved for the earlier prompts. With a sequence high-water mark, a

365−// single cancel covers exactly the prompts accepted-but-not-yet-active at

366−// cancel time (A and B); a prompt accepted afterwards (C) gets a higher

367−// sequence and must run normally without consuming A or B's cancellation.

368−// C is run first to prove it neither cancels nor drains the mark, then A

369−// and B are run and must both cancel on entry.

370−func TestCancel_AcceptedAfterCancelIsNotPoisoned(t *testing.T) {

371− t.Parallel()

372− sa, env := newStreamTestAgent(t)

373−

374− sess, err := env.sessions.Create(t.Context(), "session")

375− require.NoError(t, err)

376−

377− // A and B are accepted-but-not-yet-active.

378− acceptA := sa.BeginAccepted(sess.ID)

379− acceptB := sa.BeginAccepted(sess.ID)

380−

381− // One cancel arrives covering both A and B.

382− sa.Cancel(sess.ID)

383− require.True(t, sa.hasPendingCancel(sess.ID))

384− require.Equal(t, acceptB.seq, sa.pendingCancelMark(sess.ID),

385− "the mark must cover every prompt accepted before the cancel")

386−

387− // C is accepted AFTER the cancel; its sequence is above the mark.

388− acceptC := sa.BeginAccepted(sess.ID)

389− require.Greater(t, acceptC.seq, sa.pendingCancelMark(sess.ID),

390− "a prompt accepted after the cancel must not be covered by the mark")

391−

392− // Run C first. It must run normally to completion and must not

393− // consume or clear the cancellation reserved for A and B.

394− rc, err := sa.Run(t.Context(), SessionAgentCall{

395− SessionID: sess.ID,

396− Prompt: "C",

397− Accepted: acceptC,

398− })

399− require.NoError(t, err)

400− require.NotNil(t, rc, "C was accepted after the cancel and must run normally")

401− require.True(t, sa.hasPendingCancel(sess.ID),

402− "running C must not drain the cancellation reserved for A and B")

403−

404− // Now A and B run. Both were accepted before the cancel and must

405− // take the cancel-on-entry path.

406− ra, err := sa.Run(t.Context(), SessionAgentCall{

407− SessionID: sess.ID,

408− Prompt: "A",

409− Accepted: acceptA,

410− })

411− require.NoError(t, err)

412− require.Nil(t, ra, "A must cancel on entry, not run")

413−

414− rb, err := sa.Run(t.Context(), SessionAgentCall{

415− SessionID: sess.ID,

416− Prompt: "B",

417− Accepted: acceptB,

418− })

419− require.NoError(t, err)

420− require.Nil(t, rb, "B must cancel on entry, not run")

421−

422− require.False(t, sa.hasPendingCancel(sess.ID),

423− "the mark clears once all covered handles are resolved")

424− require.Equal(t, 0, sa.acceptedCount(sess.ID))

425− require.Equal(t, 0, sa.QueuedPrompts(sess.ID),

426− "neither A nor B may be enqueued to run normally")

427−

428− // C produced a normal turn; A and B each produced a canceled turn.

429− msgs, err := env.messages.List(t.Context(), sess.ID)

430− require.NoError(t, err)

431− require.Len(t, msgs, 6, "C normal + A canceled + B canceled = 3 user + 3 assistant")

432−

433− var normal, canceled int

434− for _, m := range msgs {

435− if m.Role != message.Assistant {

436− continue

437− }

438− switch m.FinishReason() {

439− case message.FinishReasonEndTurn:

440− normal++

441− case message.FinishReasonCanceled:

442− canceled++

443− }

444− }

445− require.Equal(t, 1, normal, "only C finished normally")

446− require.Equal(t, 2, canceled, "both A and B finished canceled")

447−}

internal/agent/hooked_tool.godeleted−142

@@ -1,142 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "encoding/json"

6− "fmt"

7− "log/slog"

8−

9− "charm.land/fantasy"

10− "github.com/charmbracelet/crush/internal/agent/tools"

11− "github.com/charmbracelet/crush/internal/hooks"

12− "github.com/charmbracelet/crush/internal/permission"

13− "github.com/tidwall/sjson"

14−)

15−

16−// hookedTool wraps a fantasy.AgentTool to run PreToolUse hooks before

17−// delegating to the inner tool.

18−type hookedTool struct {

19− inner fantasy.AgentTool

20− runner *hooks.Runner

21−}

22−

23−func newHookedTool(inner fantasy.AgentTool, runner *hooks.Runner) *hookedTool {

24− return &hookedTool{inner: inner, runner: runner}

25−}

26−

27−// wrapToolsWithHooks returns a tool slice with each entry wrapped in a

28−// hookedTool. Returns the original slice unchanged when runner is nil or

29−// when isSubAgent is true — sub-agents never fire hooks, the top-level

30−// invocation of the sub-agent tool itself is wrapped on the caller's side.

31−func wrapToolsWithHooks(tools []fantasy.AgentTool, runner *hooks.Runner, isSubAgent bool) []fantasy.AgentTool {

32− if runner == nil || isSubAgent {

33− return tools

34− }

35− out := make([]fantasy.AgentTool, len(tools))

36− for i, tool := range tools {

37− out[i] = newHookedTool(tool, runner)

38− }

39− return out

40−}

41−

42−func (h *hookedTool) Info() fantasy.ToolInfo {

43− return h.inner.Info()

44−}

45−

46−func (h *hookedTool) ProviderOptions() fantasy.ProviderOptions {

47− return h.inner.ProviderOptions()

48−}

49−

50−func (h *hookedTool) SetProviderOptions(opts fantasy.ProviderOptions) {

51− h.inner.SetProviderOptions(opts)

52−}

53−

54−func (h *hookedTool) Run(ctx context.Context, call fantasy.ToolCall) (fantasy.ToolResponse, error) {

55− sessionID := tools.GetSessionFromContext(ctx)

56− result, err := h.runner.Run(ctx, hooks.EventPreToolUse, sessionID, call.Name, call.Input)

57− if err != nil {

58− slog.Warn("Hook execution error, proceeding with tool call",

59− "tool", call.Name, "error", err)

60− }

61−

62− if result.Decision == hooks.DecisionDeny || result.Halt {

63− reason := fmt.Sprintf("Tool call blocked by hook. Reason: %s", result.Reason)

64− if result.Halt {

65− reason = fmt.Sprintf("Turn halted by hook. Reason: %s", result.Reason)

66− }

67− resp := fantasy.NewTextErrorResponse(reason)

68− // Halt ends the whole turn; a plain deny only blocks this tool

69− // call so the model can see the error and try something else.

70− resp.StopTurn = result.Halt

71− resp.Metadata = hookMetadataJSON(result)

72− return resp, nil

73− }

74−

75− if result.UpdatedInput != "" {

76− call.Input = result.UpdatedInput

77− }

78−

79− // An explicit allow from a hook pre-approves the permission prompt for

80− // this tool call. Deny is already handled above; silence falls through

81− // to the normal permission flow.

82− if result.Decision == hooks.DecisionAllow {

83− ctx = permission.WithHookApproval(ctx, call.ID)

84− }

85−

86− resp, err := h.inner.Run(ctx, call)

87− if err != nil {

88− return resp, err

89− }

90−

91− if result.Context != "" {

92− if resp.Content != "" {

93− resp.Content += "\n"

94− }

95− resp.Content += result.Context

96− }

97−

98− resp.Metadata = mergeHookMetadata(resp.Metadata, result)

99− return resp, nil

100−}

101−

102−// buildHookMetadata creates a HookMetadata from an AggregateResult.

103−func buildHookMetadata(result hooks.AggregateResult) hooks.HookMetadata {

104− return hooks.HookMetadata{

105− HookCount: result.HookCount,

106− Decision: result.Decision.String(),

107− Halt: result.Halt,

108− Reason: result.Reason,

109− InputRewrite: result.UpdatedInput != "",

110− Hooks: result.Hooks,

111− }

112−}

113−

114−// hookMetadataJSON builds a JSON string containing only the hook metadata.

115−func hookMetadataJSON(result hooks.AggregateResult) string {

116− meta := buildHookMetadata(result)

117− data, err := json.Marshal(meta)

118− if err != nil {

119− return ""

120− }

121− return `{"hook":` + string(data) + `}`

122−}

123−

124−// mergeHookMetadata injects hook metadata into existing tool metadata.

125−func mergeHookMetadata(existing string, result hooks.AggregateResult) string {

126− if result.HookCount == 0 {

127− return existing

128− }

129− meta := buildHookMetadata(result)

130− data, err := json.Marshal(meta)

131− if err != nil {

132− return existing

133− }

134− if existing == "" {

135− existing = "{}"

136− }

137− merged, err := sjson.SetRaw(existing, "hook", string(data))

138− if err != nil {

139− return existing

140− }

141− return merged

142−}

internal/agent/hooked_tool_test.godeleted−147

@@ -1,147 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "testing"

6−

7− "charm.land/fantasy"

8− "github.com/charmbracelet/crush/internal/config"

9− "github.com/charmbracelet/crush/internal/hooks"

10− "github.com/charmbracelet/crush/internal/permission"

11− "github.com/stretchr/testify/require"

12−)

13−

14−// fakeTool records the context it was invoked with so tests can assert on

15−// values stamped onto it by the hookedTool decorator.

16−type fakeTool struct {

17− name string

18− called bool

19− gotCtx context.Context

20− resp fantasy.ToolResponse

21−}

22−

23−func (f *fakeTool) Info() fantasy.ToolInfo {

24− return fantasy.ToolInfo{Name: f.name}

25−}

26−

27−func (f *fakeTool) Run(ctx context.Context, _ fantasy.ToolCall) (fantasy.ToolResponse, error) {

28− f.called = true

29− f.gotCtx = ctx

30− return f.resp, nil

31−}

32−

33−func (f *fakeTool) ProviderOptions() fantasy.ProviderOptions { return nil }

34−func (f *fakeTool) SetProviderOptions(_ fantasy.ProviderOptions) {}

35−

36−// newRunner builds a hooks.Runner from a single HookConfig, running the

37−// config-loader path that compiles the matcher regex.

38−func newRunner(t *testing.T, cmd string) *hooks.Runner {

39− t.Helper()

40− cfg := &config.Config{

41− Hooks: map[string][]config.HookConfig{

42− hooks.EventPreToolUse: {{Command: cmd}},

43− },

44− }

45− require.NoError(t, cfg.ValidateHooks())

46− return hooks.NewRunner(cfg.Hooks[hooks.EventPreToolUse], t.TempDir(), t.TempDir())

47−}

48−

49−func TestHookedTool_AllowStampsHookApproval(t *testing.T) {

50− t.Parallel()

51−

52− inner := &fakeTool{name: "view", resp: fantasy.NewTextResponse("ok")}

53− runner := newRunner(t, `echo '{"decision":"allow"}'`)

54− tool := newHookedTool(inner, runner)

55−

56− _, err := tool.Run(t.Context(), fantasy.ToolCall{ID: "call-1", Name: "view"})

57− require.NoError(t, err)

58− require.True(t, inner.called, "inner tool should have run")

59−

60− // The inner tool's permission service can now treat call-1 as pre-approved.

61− svc := permission.NewPermissionService(t.TempDir(), false, nil)

62− granted, err := svc.Request(inner.gotCtx, permission.CreatePermissionRequest{

63− SessionID: "s1",

64− ToolCallID: "call-1",

65− ToolName: "view",

66− Action: "read",

67− Path: t.TempDir(),

68− })

69− require.NoError(t, err)

70− require.True(t, granted, "hook allow should bypass the permission prompt")

71−}

72−

73−func TestHookedTool_SilentDoesNotStampApproval(t *testing.T) {

74− t.Parallel()

75−

76− inner := &fakeTool{name: "view", resp: fantasy.NewTextResponse("ok")}

77− runner := newRunner(t, `exit 0`) // no stdout, no decision

78− tool := newHookedTool(inner, runner)

79−

80− _, err := tool.Run(t.Context(), fantasy.ToolCall{ID: "call-2", Name: "view"})

81− require.NoError(t, err)

82− require.True(t, inner.called)

83−

84− // With no hook opinion, a fresh permission request has nothing stamped

85− // and must fall through to the normal flow. We verify by checking that

86− // the context does not look pre-approved for this call ID: sending a

87− // request that no subscriber resolves will block until cancelled.

88− svc := permission.NewPermissionService(t.TempDir(), false, nil)

89− ctx, cancel := context.WithCancel(inner.gotCtx)

90− cancel()

91− granted, err := svc.Request(ctx, permission.CreatePermissionRequest{

92− SessionID: "s1",

93− ToolCallID: "call-2",

94− ToolName: "view",

95− Action: "read",

96− Path: t.TempDir(),

97− })

98− require.Error(t, err, "no approval stamped => request should reach the prompt path")

99− require.False(t, granted)

100−}

101−

102−func TestHookedTool_DenySkipsInnerTool(t *testing.T) {

103− t.Parallel()

104−

105− inner := &fakeTool{name: "bash"}

106− runner := newRunner(t, `echo "blocked" >&2; exit 2`)

107− tool := newHookedTool(inner, runner)

108−

109− resp, err := tool.Run(t.Context(), fantasy.ToolCall{ID: "call-3", Name: "bash"})

110− require.NoError(t, err)

111− require.False(t, inner.called, "denied call must not reach the inner tool")

112− require.True(t, resp.IsError)

113− require.Contains(t, resp.Content, "blocked")

114−}

115−

116−func TestWrapToolsWithHooks(t *testing.T) {

117− t.Parallel()

118−

119− runner := newRunner(t, `exit 0`)

120− inputs := []fantasy.AgentTool{&fakeTool{name: "a"}, &fakeTool{name: "b"}}

121−

122− t.Run("top-level agent wraps every tool", func(t *testing.T) {

123− t.Parallel()

124− out := wrapToolsWithHooks(inputs, runner, false)

125− require.Len(t, out, len(inputs))

126− for i, tool := range out {

127− _, ok := tool.(*hookedTool)

128− require.Truef(t, ok, "tool %d should be a *hookedTool", i)

129− }

130− })

131−

132− t.Run("sub-agent skips the wrap", func(t *testing.T) {

133− t.Parallel()

134− out := wrapToolsWithHooks(inputs, runner, true)

135− require.Equal(t, inputs, out, "sub-agent tools should be returned unwrapped")

136− for _, tool := range out {

137− _, isHooked := tool.(*hookedTool)

138− require.False(t, isHooked, "sub-agent tool should not be wrapped")

139− }

140− })

141−

142− t.Run("nil runner skips the wrap for both agent kinds", func(t *testing.T) {

143− t.Parallel()

144− require.Equal(t, inputs, wrapToolsWithHooks(inputs, nil, false))

145− require.Equal(t, inputs, wrapToolsWithHooks(inputs, nil, true))

146− })

147−}

internal/agent/hyper/provider.go+273−45

@@ -2,26 +2,49 @@

2package hyper

4import (

5+ "bufio"

6+ "bytes"

5 "cmp"

7 "cmp"

6 "context"

8 "context"

7 _ "embed"

9 _ "embed"

8 "encoding/json"

10 "encoding/json"

11+ "errors"

9 "fmt"

12 "fmt"

13+ "io"

10 "log/slog"

14 "log/slog"

15+ "maps"

11 "net/http"

16 "net/http"

17+ "net/url"

12 "os"

18 "os"

19+ "strconv"

20+ "strings"

13 "sync"

21 "sync"

14− "sync/atomic"

15 "time"

22 "time"

24+ "charm.land/fantasy"

25+ "charm.land/fantasy/object"

17− "charm.land/catwalk/pkg/catwalk"

26+ "github.com/charmbracelet/catwalk/pkg/catwalk"

27+ "github.com/charmbracelet/crush/internal/event"

18)

28)

20−//go:generate wget -O provider.json https://hyper.charm.land/v1/provider

30+//go:generate wget -O provider.json https://console.charm.land/api/v1/provider

22//go:embed provider.json

32//go:embed provider.json

23var embedded []byte

33var embedded []byte

35+// Enabled returns true if hyper is enabled.

36+var Enabled = sync.OnceValue(func() bool {

37+ b, _ := strconv.ParseBool(

38+ cmp.Or(

39+ os.Getenv("HYPER"),

40+ os.Getenv("HYPERCRUSH"),

41+ os.Getenv("HYPER_ENABLE"),

42+ os.Getenv("HYPER_ENABLED"),

43+ ),

44+ )

45+ return b

46+})

47+

25// Embedded returns the embedded Hyper provider.

48// Embedded returns the embedded Hyper provider.

26var Embedded = sync.OnceValue(func() catwalk.Provider {

49var Embedded = sync.OnceValue(func() catwalk.Provider {

27 var provider catwalk.Provider

50 var provider catwalk.Provider

@@ -37,10 +60,9 @@ var Embedded = sync.OnceValue(func() catwalk.Provider {

37const (

60const (

38 // Name is the default name of this meta provider.

61 // Name is the default name of this meta provider.

39 Name = "hyper"

62 Name = "hyper"

40− // DisplayName is the display name of Hyper.

41− DisplayName = "Charm Hyper"

42 // defaultBaseURL is the default proxy URL.

63 // defaultBaseURL is the default proxy URL.

64+ // TODO: change this to production URL when ready.

43− defaultBaseURL = "https://hyper.charm.land"

65+ defaultBaseURL = "https://console.charm.land"

44)

66)

46// BaseURL returns the base URL, which is either $HYPER_URL or the default.

68// BaseURL returns the base URL, which is either $HYPER_URL or the default.

@@ -48,58 +70,264 @@ var BaseURL = sync.OnceValue(func() string {

48 return cmp.Or(os.Getenv("HYPER_URL"), defaultBaseURL)

70 return cmp.Or(os.Getenv("HYPER_URL"), defaultBaseURL)

49})

71})

51−// lastKnownBalance stores the most recently extracted hypercredit balance

73+var ErrNoCredits = errors.New("you're out of credits")

52−// from API response metadata. FetchCredits checks this before making a

74+

53−// separate HTTP call.

75+type options struct {

54−var lastKnownBalance atomic.Int64

76+ baseURL string

77+ apiKey string

78+ name string

79+ headers map[string]string

80+ client *http.Client

81+}

82+

83+// Option configures the proxy provider.

84+type Option = func(*options)

85+

86+// New creates a new proxy provider.

87+func New(opts ...Option) (fantasy.Provider, error) {

88+ o := options{

89+ baseURL: BaseURL() + "/api/v1/fantasy",

90+ name: Name,

91+ headers: map[string]string{

92+ "x-crush-id": event.GetID(),

93+ },

94+ client: &http.Client{Timeout: 0}, // stream-safe

95+ }

96+ for _, opt := range opts {

97+ opt(&o)

98+ }

99+ return &provider{options: o}, nil

100+}

101+

102+// WithBaseURL sets the proxy base URL (e.g. http://localhost:8080).

103+func WithBaseURL(url string) Option { return func(o *options) { o.baseURL = url } }

104

56−// hasBalance tracks whether lastKnownBalance has been set.

105+// WithName sets the provider name.

57−var hasBalance atomic.Bool

106+func WithName(name string) Option { return func(o *options) { o.name = name } }

107

59−// SetBalance stores a credit balance extracted from API response metadata.

108+// WithHeaders sets extra headers sent to the proxy.

60−func SetBalance(balance int) {

109+func WithHeaders(headers map[string]string) Option {

61− lastKnownBalance.Store(int64(balance))

110+ return func(o *options) {

62− hasBalance.Store(true)

111+ maps.Copy(o.headers, headers)

112+ }

63}

113}

114

65−// FetchCredits returns the remaining hypercredit balance. It first checks

115+// WithHTTPClient sets custom HTTP client.

66−// for a balance extracted from the most recent API response's usage

116+func WithHTTPClient(c *http.Client) Option { return func(o *options) { o.client = c } }

67−// metadata. If none is available, it falls back to calling the /v1/credits

117+

68−// endpoint directly.

118+// WithAPIKey sets the API key.

69−func FetchCredits(ctx context.Context, apiKey string) (int, error) {

119+func WithAPIKey(key string) Option {

70− if hasBalance.Load() {

120+ return func(o *options) {

71− hasBalance.Store(false)

121+ o.apiKey = key

72− return int(lastKnownBalance.Load()), nil

122+ }

123+}

124+

125+type provider struct{ options options }

126+

127+func (p *provider) Name() string { return p.options.name }

128+

129+// LanguageModel implements fantasy.Provider.

130+func (p *provider) LanguageModel(_ context.Context, modelID string) (fantasy.LanguageModel, error) {

131+ if modelID == "" {

132+ return nil, errors.New("missing model id")

133+ }

134+ return &languageModel{modelID: modelID, provider: p.options.name, opts: p.options}, nil

135+}

136+

137+type languageModel struct {

138+ provider string

139+ modelID string

140+ opts options

141+}

142+

143+// GenerateObject implements fantasy.LanguageModel.

144+func (m *languageModel) GenerateObject(ctx context.Context, call fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {

145+ return object.GenerateWithTool(ctx, m, call)

146+}

147+

148+// StreamObject implements fantasy.LanguageModel.

149+func (m *languageModel) StreamObject(ctx context.Context, call fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {

150+ return object.StreamWithTool(ctx, m, call)

151+}

152+

153+func (m *languageModel) Provider() string { return m.provider }

154+func (m *languageModel) Model() string { return m.modelID }

155+

156+// Generate implements fantasy.LanguageModel by calling the proxy JSON endpoint.

157+func (m *languageModel) Generate(ctx context.Context, call fantasy.Call) (*fantasy.Response, error) {

158+ resp, err := m.doRequest(ctx, false, call)

159+ if err != nil {

160+ return nil, err

161+ }

162+ defer func() { _ = resp.Body.Close() }()

163+ if resp.StatusCode < 200 || resp.StatusCode >= 300 {

164+ b, _ := ioReadAllLimit(resp.Body, 64*1024)

165+ return nil, fmt.Errorf("proxy generate error: %s", strings.TrimSpace(string(b)))

73− }

166+ }

74−

167+ var out fantasy.Response

75− req, err := http.NewRequestWithContext(

168+ if err := json.NewDecoder(resp.Body).Decode(&out); err != nil {

76− ctx,

169+ return nil, err

77− http.MethodGet,

170+ }

78− BaseURL()+"/v1/credits",

171+ return &out, nil

79− nil,

172+}

80− )

173+

174+// Stream implements fantasy.LanguageModel using SSE from the proxy.

175+func (m *languageModel) Stream(ctx context.Context, call fantasy.Call) (fantasy.StreamResponse, error) {

176+ // Prefer explicit /stream endpoint

177+ resp, err := m.doRequest(ctx, true, call)

178+ if err != nil {

179+ return nil, err

180+ }

181+ switch resp.StatusCode {

182+ case http.StatusTooManyRequests:

183+ _ = resp.Body.Close()

184+ return nil, toProviderError(resp, retryAfter(resp))

185+ case http.StatusUnauthorized:

186+ _ = resp.Body.Close()

187+ return nil, toProviderError(resp, "")

188+ case http.StatusPaymentRequired:

189+ _ = resp.Body.Close()

190+ return nil, ErrNoCredits

191+ }

192+

193+ if resp.StatusCode < 200 || resp.StatusCode >= 300 {

194+ defer func() { _ = resp.Body.Close() }()

195+ b, _ := ioReadAllLimit(resp.Body, 64*1024)

196+ return nil, &fantasy.ProviderError{

197+ Title: "Stream Error",

198+ Message: strings.TrimSpace(string(b)),

199+ StatusCode: resp.StatusCode,

200+ }

201+ }

202+

203+ return func(yield func(fantasy.StreamPart) bool) {

204+ defer func() { _ = resp.Body.Close() }()

205+ scanner := bufio.NewScanner(resp.Body)

206+ buf := make([]byte, 0, 64*1024)

207+ scanner.Buffer(buf, 4*1024*1024)

208+

209+ var (

210+ event string

211+ dataBuf bytes.Buffer

212+ sawFinish bool

213+ dispatch = func() bool {

214+ if dataBuf.Len() == 0 || event == "" {

215+ dataBuf.Reset()

216+ event = ""

217+ return true

218+ }

219+ var part fantasy.StreamPart

220+ if err := json.Unmarshal(dataBuf.Bytes(), &part); err != nil {

221+ return yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeError, Error: err})

222+ }

223+ if part.Type == fantasy.StreamPartTypeFinish {

224+ sawFinish = true

225+ }

226+ ok := yield(part)

227+ dataBuf.Reset()

228+ event = ""

229+ return ok

230+ }

231+ )

232+

233+ for scanner.Scan() {

234+ line := scanner.Text()

235+ if line == "" { // event boundary

236+ if !dispatch() {

237+ return

238+ }

239+ continue

240+ }

241+ if strings.HasPrefix(line, ":") { // comment / ping

242+ continue

243+ }

244+ if strings.HasPrefix(line, "event: ") {

245+ event = strings.TrimSpace(line[len("event: "):])

246+ continue

247+ }

248+ if strings.HasPrefix(line, "data: ") {

249+ if dataBuf.Len() > 0 {

250+ dataBuf.WriteByte('\n')

251+ }

252+ dataBuf.WriteString(line[len("data: "):])

253+ continue

254+ }

255+ }

256+ if err := scanner.Err(); err != nil &&

257+ !errors.Is(err, context.Canceled) &&

258+ !errors.Is(err, context.DeadlineExceeded) {

259+ yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeError, Error: err})

260+ return

261+ }

262+ // flush any pending data

263+ _ = dispatch()

264+ if !sawFinish {

265+ _ = yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish})

266+ }

267+ }, nil

268+}

269+

270+func (m *languageModel) doRequest(ctx context.Context, stream bool, call fantasy.Call) (*http.Response, error) {

271+ addr, err := url.Parse(m.opts.baseURL)

272+ if err != nil {

273+ return nil, err

274+ }

275+ addr = addr.JoinPath(m.modelID)

276+ if stream {

277+ addr = addr.JoinPath("stream")

278+ } else {

279+ addr = addr.JoinPath("generate")

280+ }

281+

282+ body, err := json.Marshal(call)

81 if err != nil {

283 if err != nil {

82− return 0, fmt.Errorf("could not create request: %w", err)

284+ return nil, err

83 }

285 }

84− req.Header.Set("Authorization", "Bearer "+apiKey)

286

86− client := &http.Client{Timeout: 10 * time.Second}

287+ req, err := http.NewRequestWithContext(ctx, http.MethodPost, addr.String(), bytes.NewReader(body))

87− resp, err := client.Do(req)

88 if err != nil {

288 if err != nil {

89− return 0, fmt.Errorf("failed to make request: %w", err)

289+ return nil, err

290+ }

291+ req.Header.Set("Content-Type", "application/json")

292+ if stream {

293+ req.Header.Set("Accept", "text/event-stream")

294+ } else {

295+ req.Header.Set("Accept", "application/json")

296+ }

297+ for k, v := range m.opts.headers {

298+ req.Header.Set(k, v)

90 }

299 }

91− defer resp.Body.Close()

300

93− if resp.StatusCode != http.StatusOK {

301+ if m.opts.apiKey != "" {

94− return 0, fmt.Errorf("unexpected status code: %d", resp.StatusCode)

302+ req.Header.Set("Authorization", m.opts.apiKey)

95 }

303 }

304+ return m.opts.client.Do(req)

305+}

306

97− var result struct {

307+// ioReadAllLimit reads up to n bytes.

98− Balance int `json:"balance"`

308+func ioReadAllLimit(r io.Reader, n int64) ([]byte, error) {

309+ var b bytes.Buffer

310+ if n <= 0 {

311+ n = 1 << 20

99 }

312 }

100− if err := json.NewDecoder(resp.Body).Decode(&result); err != nil {

313+ lr := &io.LimitedReader{R: r, N: n}

101− return 0, fmt.Errorf("failed to decode response: %w", err)

314+ _, err := b.ReadFrom(lr)

315+ return b.Bytes(), err

316+}

317+

318+func toProviderError(resp *http.Response, message string) error {

319+ return &fantasy.ProviderError{

320+ Title: fantasy.ErrorTitleForStatusCode(resp.StatusCode),

321+ Message: message,

322+ StatusCode: resp.StatusCode,

102 }

323 }

324+}

103

325

104− return result.Balance, nil

326+func retryAfter(resp *http.Response) string {

327+ after, err := strconv.Atoi(resp.Header.Get("Retry-After"))

328+ if err == nil && after > 0 {

329+ d := time.Duration(after) * time.Second

330+ return "Try again in " + d.String()

331+ }

332+ return "Try again later"

105}

333}

internal/agent/hyper/provider.json+1−295

@@ -1,295 +1 @@

1−{

1+{"name":"Charm Hyper","id":"hyper","api_endpoint":"https://console.charm.land/api/v1/fantasy","type":"hyper","default_large_model_id":"claude-opus-4-5","default_small_model_id":"claude-haiku-4-5","models":[{"id":"claude-haiku-4-5","name":"Claude Haiku 4.5","cost_per_1m_in":1,"cost_per_1m_out":5,"cost_per_1m_in_cached":1.25,"cost_per_1m_out_cached":0.09999999999999999,"context_window":200000,"default_max_tokens":32000,"can_reason":true,"supports_attachments":true,"options":{}},{"id":"claude-opus-4-5","name":"Claude Opus 4.5","cost_per_1m_in":5,"cost_per_1m_out":25,"cost_per_1m_in_cached":6.25,"cost_per_1m_out_cached":0.5,"context_window":200000,"default_max_tokens":50000,"can_reason":true,"supports_attachments":true,"options":{}},{"id":"claude-sonnet-4-5","name":"Claude Sonnet 4.5","cost_per_1m_in":3,"cost_per_1m_out":15,"cost_per_1m_in_cached":3.75,"cost_per_1m_out_cached":0.3,"context_window":200000,"default_max_tokens":50000,"can_reason":true,"supports_attachments":true,"options":{}},{"id":"gemini-3-flash","name":"Gemini 3 Flash","cost_per_1m_in":0.5,"cost_per_1m_out":3,"cost_per_1m_in_cached":0.049999999999999996,"cost_per_1m_out_cached":0,"context_window":1000000,"default_max_tokens":8000,"can_reason":true,"reasoning_levels":["low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"gemini-3-pro-preview","name":"Gemini 3 Pro","cost_per_1m_in":2,"cost_per_1m_out":12,"cost_per_1m_in_cached":0.19999999999999998,"cost_per_1m_out_cached":0,"context_window":1000000,"default_max_tokens":8000,"can_reason":true,"reasoning_levels":["low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"glm-4.6","name":"GLM 4.6","cost_per_1m_in":0.44999999999999996,"cost_per_1m_out":1.7999999999999998,"cost_per_1m_in_cached":0.11,"cost_per_1m_out_cached":0,"context_window":200000,"default_max_tokens":8000,"can_reason":true,"reasoning_levels":["low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":false,"options":{}},{"id":"glm-4.7","name":"GLM 4.7","cost_per_1m_in":0.43,"cost_per_1m_out":1.75,"cost_per_1m_in_cached":0.08,"cost_per_1m_out_cached":0,"context_window":202752,"default_max_tokens":8000,"can_reason":true,"reasoning_levels":["low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":false,"options":{}},{"id":"gpt-5.1-codex","name":"GPT 5.1 Codex","cost_per_1m_in":1.25,"cost_per_1m_out":10,"cost_per_1m_in_cached":0.125,"cost_per_1m_out_cached":0.125,"context_window":400000,"default_max_tokens":128000,"can_reason":true,"reasoning_levels":["minimal","low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"gpt-5.1-codex-max","name":"GPT 5.1 Codex Max","cost_per_1m_in":1.25,"cost_per_1m_out":10,"cost_per_1m_in_cached":0.125,"cost_per_1m_out_cached":0.125,"context_window":400000,"default_max_tokens":128000,"can_reason":true,"reasoning_levels":["minimal","low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"gpt-5.1-codex-mini","name":"GPT 5.1 Codex Mini","cost_per_1m_in":0.25,"cost_per_1m_out":2,"cost_per_1m_in_cached":0.025,"cost_per_1m_out_cached":0.025,"context_window":400000,"default_max_tokens":128000,"can_reason":true,"reasoning_levels":["low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"gpt-5.2","name":"GPT 5.2","cost_per_1m_in":1.75,"cost_per_1m_out":14,"cost_per_1m_in_cached":0.175,"cost_per_1m_out_cached":0.175,"context_window":400000,"default_max_tokens":128000,"can_reason":true,"reasoning_levels":["minimal","low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"gpt-5.2-codex","name":"GPT 5.2 Codex","cost_per_1m_in":1.75,"cost_per_1m_out":14,"cost_per_1m_in_cached":0.175,"cost_per_1m_out_cached":0.175,"context_window":400000,"default_max_tokens":128000,"can_reason":true,"reasoning_levels":["minimal","low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}},{"id":"grok-4.1-fast-non-reasoning","name":"Grok 4.1 Fast Non Reasoning","cost_per_1m_in":0.2,"cost_per_1m_out":0.5,"cost_per_1m_in_cached":0,"cost_per_1m_out_cached":0.05,"context_window":2000000,"default_max_tokens":200000,"can_reason":false,"supports_attachments":true,"options":{}},{"id":"grok-4.1-fast-reasoning","name":"Grok 4.1 Fast Reasoning","cost_per_1m_in":0.2,"cost_per_1m_out":0.5,"cost_per_1m_in_cached":0,"cost_per_1m_out_cached":0.05,"context_window":2000000,"default_max_tokens":200000,"can_reason":true,"supports_attachments":true,"options":{}},{"id":"grok-code-fast-1","name":"Grok Code Fast","cost_per_1m_in":0.2,"cost_per_1m_out":1.5,"cost_per_1m_in_cached":0,"cost_per_1m_out_cached":0.02,"context_window":256000,"default_max_tokens":20000,"can_reason":true,"supports_attachments":false,"options":{}},{"id":"kimi-k2-0905","name":"Kimi K2","cost_per_1m_in":0.55,"cost_per_1m_out":2.19,"cost_per_1m_in_cached":0,"cost_per_1m_out_cached":0,"context_window":256000,"default_max_tokens":10000,"can_reason":true,"default_reasoning_effort":"medium","supports_attachments":false,"options":{}},{"id":"kimi-k2.5","name":"Kimi K2.5","cost_per_1m_in":0.6,"cost_per_1m_out":3,"cost_per_1m_in_cached":0.09999999999999999,"cost_per_1m_out_cached":0,"context_window":262114,"default_max_tokens":8000,"can_reason":true,"reasoning_levels":["low","medium","high"],"default_reasoning_effort":"medium","supports_attachments":true,"options":{}}]}

2− "name": "Charm Hyper",

3− "id": "hyper",

4− "api_endpoint": "https://hyper.charm.land/api/v1/fantasy",

5− "type": "hyper",

6− "default_large_model_id": "qwen3.7-plus",

7− "default_small_model_id": "deepseek-v4-flash",

8− "models": [

9− {

10− "id": "deepseek-v4-flash",

11− "name": "DeepSeek V4 Flash",

12− "cost_per_1m_in": 0.14,

13− "cost_per_1m_out": 0.28,

14− "cost_per_1m_in_cached": 0.0028,

15− "cost_per_1m_out_cached": 0,

16− "context_window": 1000000,

17− "default_max_tokens": 384000,

18− "can_reason": true,

19− "reasoning_levels": [

20− "high",

21− "xhigh"

22− ],

23− "default_reasoning_effort": "high",

24− "supports_attachments": false

25− },

26− {

27− "id": "deepseek-v4-pro",

28− "name": "DeepSeek V4 Pro",

29− "cost_per_1m_in": 0.435,

30− "cost_per_1m_out": 0.87,

31− "cost_per_1m_in_cached": 0.0036,

32− "cost_per_1m_out_cached": 0,

33− "context_window": 1000000,

34− "default_max_tokens": 384000,

35− "can_reason": true,

36− "reasoning_levels": [

37− "high",

38− "xhigh"

39− ],

40− "default_reasoning_effort": "high",

41− "supports_attachments": false

42− },

43− {

44− "id": "gemma-4-26b-a4b-it",

45− "name": "Gemma 4 26B A4B",

46− "cost_per_1m_in": 0.126,

47− "cost_per_1m_out": 0.42,

48− "cost_per_1m_in_cached": 0.063,

49− "cost_per_1m_out_cached": 0,

50− "context_window": 256000,

51− "default_max_tokens": 25600,

52− "can_reason": true,

53− "supports_attachments": false

54− },

55− {

56− "id": "glm-5",

57− "name": "GLM-5",

58− "cost_per_1m_in": 0.86,

59− "cost_per_1m_out": 2.784,

60− "cost_per_1m_in_cached": 0.43,

61− "cost_per_1m_out_cached": 0,

62− "context_window": 202752,

63− "default_max_tokens": 20275,

64− "can_reason": true,

65− "supports_attachments": false

66− },

67− {

68− "id": "glm-5.1",

69− "name": "GLM-5.1",

70− "cost_per_1m_in": 1.3,

71− "cost_per_1m_out": 4.3,

72− "cost_per_1m_in_cached": 0.26,

73− "cost_per_1m_out_cached": 0,

74− "context_window": 202000,

75− "default_max_tokens": 202000,

76− "can_reason": true,

77− "reasoning_levels": [

78− "low",

79− "medium",

80− "high"

81− ],

82− "default_reasoning_effort": "medium",

83− "supports_attachments": false

84− },

85− {

86− "id": "glm-5.2",

87− "name": "GLM-5.2",

88− "cost_per_1m_in": 1.5,

89− "cost_per_1m_out": 4.5,

90− "cost_per_1m_in_cached": 0.3,

91− "cost_per_1m_out_cached": 0,

92− "context_window": 1000000,

93− "default_max_tokens": 128000,

94− "can_reason": true,

95− "reasoning_levels": [

96− "high",

97− "xhigh"

98− ],

99− "default_reasoning_effort": "high",

100− "supports_attachments": false

101− },

102− {

103− "id": "gpt-oss-120b",

104− "name": "gpt-oss-120b",

105− "cost_per_1m_in": 0.168,

106− "cost_per_1m_out": 0.562,

107− "cost_per_1m_in_cached": 0.084,

108− "cost_per_1m_out_cached": 0,

109− "context_window": 131072,

110− "default_max_tokens": 13107,

111− "can_reason": true,

112− "reasoning_levels": [

113− "low",

114− "medium",

115− "high"

116− ],

117− "default_reasoning_effort": "medium",

118− "supports_attachments": false

119− },

120− {

121− "id": "kimi-k2.5",

122− "name": "Kimi K2.5",

123− "cost_per_1m_in": 0.51,

124− "cost_per_1m_out": 2.67,

125− "cost_per_1m_in_cached": 0.255,

126− "cost_per_1m_out_cached": 0,

127− "context_window": 262144,

128− "default_max_tokens": 26214,

129− "can_reason": true,

130− "supports_attachments": false

131− },

132− {

133− "id": "kimi-k2.6",

134− "name": "Kimi K2.6",

135− "cost_per_1m_in": 0.95,

136− "cost_per_1m_out": 4,

137− "cost_per_1m_in_cached": 0.16,

138− "cost_per_1m_out_cached": 0,

139− "context_window": 262000,

140− "default_max_tokens": 262000,

141− "can_reason": true,

142− "reasoning_levels": [

143− "low",

144− "medium",

145− "high"

146− ],

147− "default_reasoning_effort": "medium",

148− "supports_attachments": true

149− },

150− {

151− "id": "kimi-k2.7-code",

152− "name": "Kimi K2.7 Code",

153− "cost_per_1m_in": 0.95,

154− "cost_per_1m_out": 4,

155− "cost_per_1m_in_cached": 0.19,

156− "cost_per_1m_out_cached": 0,

157− "context_window": 256000,

158− "default_max_tokens": 32768,

159− "can_reason": true,

160− "reasoning_levels": [

161− "low",

162− "medium",

163− "high"

164− ],

165− "default_reasoning_effort": "medium",

166− "supports_attachments": true

167− },

168− {

169− "id": "llama-3.3-70b-instruct",

170− "name": "Llama 3.3 70B Instruct",

171− "cost_per_1m_in": 0.6066,

172− "cost_per_1m_out": 1.0386,

173− "cost_per_1m_in_cached": 0.3033,

174− "cost_per_1m_out_cached": 0,

175− "context_window": 128000,

176− "default_max_tokens": 12800,

177− "can_reason": true,

178− "supports_attachments": false

179− },

180− {

181− "id": "llama-4-maverick-17b-128e-instruct-fp8",

182− "name": "Llama 4 Maverick 17B 128E Instruct FP8",

183− "cost_per_1m_in": 0.274,

184− "cost_per_1m_out": 0.894,

185− "cost_per_1m_in_cached": 0.137,

186− "cost_per_1m_out_cached": 0,

187− "context_window": 430000,

188− "default_max_tokens": 43000,

189− "can_reason": true,

190− "supports_attachments": false

191− },

192− {

193− "id": "minimax-m2.7",

194− "name": "MiniMax M2.7",

195− "cost_per_1m_in": 0.3,

196− "cost_per_1m_out": 1.2,

197− "cost_per_1m_in_cached": 0.06,

198− "cost_per_1m_out_cached": 0.375,

199− "context_window": 204800,

200− "default_max_tokens": 131000,

201− "can_reason": true,

202− "reasoning_levels": [

203− "low",

204− "medium",

205− "high"

206− ],

207− "default_reasoning_effort": "medium",

208− "supports_attachments": false

209− },

210− {

211− "id": "qwen3.6-flash",

212− "name": "Qwen3.6-Flash",

213− "cost_per_1m_in": 1,

214− "cost_per_1m_out": 4,

215− "cost_per_1m_in_cached": 1.25,

216− "cost_per_1m_out_cached": 0.1,

217− "context_window": 1000000,

218− "default_max_tokens": 64000,

219− "can_reason": true,

220− "supports_attachments": true

221− },

222− {

223− "id": "qwen3.6-max",

224− "name": "Qwen3.6-Max",

225− "cost_per_1m_in": 2,

226− "cost_per_1m_out": 12,

227− "cost_per_1m_in_cached": 2.5,

228− "cost_per_1m_out_cached": 0.2,

229− "context_window": 256000,

230− "default_max_tokens": 64000,

231− "can_reason": true,

232− "supports_attachments": false

233− },

234− {

235− "id": "qwen3.6-plus",

236− "name": "Qwen3.6-Plus",

237− "cost_per_1m_in": 2,

238− "cost_per_1m_out": 6,

239− "cost_per_1m_in_cached": 2.5,

240− "cost_per_1m_out_cached": 0.2,

241− "context_window": 1000000,

242− "default_max_tokens": 64000,

243− "can_reason": true,

244− "supports_attachments": true

245− },

246− {

247− "id": "qwen3.7-max",

248− "name": "Qwen3.7-Max",

249− "cost_per_1m_in": 2.5,

250− "cost_per_1m_out": 7.5,

251− "cost_per_1m_in_cached": 3.125,

252− "cost_per_1m_out_cached": 0.25,

253− "context_window": 1000000,

254− "default_max_tokens": 64000,

255− "can_reason": true,

256− "supports_attachments": false

257− },

258− {

259− "id": "qwen3.7-plus",

260− "name": "Qwen3.7-Plus",

261− "cost_per_1m_in": 1.2,

262− "cost_per_1m_out": 4.8,

263− "cost_per_1m_in_cached": 1.5,

264− "cost_per_1m_out_cached": 0.12,

265− "context_window": 1000000,

266− "default_max_tokens": 64000,

267− "can_reason": true,

268− "supports_attachments": true

269− },

270− {

271− "id": "qwen3-coder-480b-a35b-instruct-int4-mixed-ar",

272− "name": "Qwen3 Coder 480B A35B Instruct INT4 Mixed AR",

273− "cost_per_1m_in": 0.557,

274− "cost_per_1m_out": 2.545,

275− "cost_per_1m_in_cached": 0.2785,

276− "cost_per_1m_out_cached": 0,

277− "context_window": 106000,

278− "default_max_tokens": 10600,

279− "can_reason": false,

280− "supports_attachments": false

281− },

282− {

283− "id": "qwen3-next-80b-a3b-instruct",

284− "name": "Qwen3 Next 80B A3B Instruct",

285− "cost_per_1m_in": 0.128,

286− "cost_per_1m_out": 1.28,

287− "cost_per_1m_in_cached": 0.064,

288− "cost_per_1m_out_cached": 0,

289− "context_window": 262144,

290− "default_max_tokens": 26214,

291− "can_reason": false,

292− "supports_attachments": false

293− }

294− ]

295−}

internal/agent/loop_detection.godeleted−92

@@ -1,92 +0,0 @@

1−package agent

2−

3−import (

4− "hash/fnv"

5− "io"

6−

7− "charm.land/fantasy"

8−)

9−

10−const (

11− loopDetectionWindowSize = 10

12− loopDetectionMaxRepeats = 5

13−)

14−

15−// hasRepeatedToolCalls checks whether the agent is stuck in a loop by looking

16−// at recent steps. It examines the last windowSize steps and returns true if

17−// any tool-call signature appears more than maxRepeats times.

18−func hasRepeatedToolCalls(steps []fantasy.StepResult, windowSize, maxRepeats int) bool {

19− if len(steps) < windowSize {

20− return false

21− }

22−

23− window := steps[len(steps)-windowSize:]

24− counts := make(map[uint64]int, windowSize)

25−

26− for _, step := range window {

27− sig := getToolInteractionSignature(step.Content)

28− if sig == 0 {

29− continue

30− }

31− counts[sig]++

32− if counts[sig] > maxRepeats {

33− return true

34− }

35− }

36−

37− return false

38−}

39−

40−// getToolInteractionSignature computes a hash signature for the tool

41−// interactions in a single step's content. It pairs tool calls with their

42−// results (matched by ToolCallID) and returns an FNV-1a hash. FNV is ~10x

43−// faster than SHA-256 and collision risk is negligible for loop detection

44−// over a 10-step window. If the step contains no tool calls, it returns 0.

45−func getToolInteractionSignature(content fantasy.ResponseContent) uint64 {

46− toolCalls := content.ToolCalls()

47− if len(toolCalls) == 0 {

48− return 0

49− }

50−

51− // Index tool results by their ToolCallID for fast lookup.

52− resultsByID := make(map[string]fantasy.ToolResultContent, len(toolCalls))

53− for _, tr := range content.ToolResults() {

54− resultsByID[tr.ToolCallID] = tr

55− }

56−

57− h := fnv.New64a()

58− for _, tc := range toolCalls {

59− output := ""

60− if tr, ok := resultsByID[tc.ToolCallID]; ok {

61− output = toolResultOutputString(tr.Result)

62− }

63− _, _ = io.WriteString(h, tc.ToolName)

64− h.Write([]byte{0})

65− _, _ = io.WriteString(h, tc.Input)

66− h.Write([]byte{0})

67− _, _ = io.WriteString(h, output)

68− h.Write([]byte{0})

69− }

70− return h.Sum64()

71−}

72−

73−// toolResultOutputString converts a ToolResultOutputContent to a stable string

74−// representation for signature comparison.

75−func toolResultOutputString(result fantasy.ToolResultOutputContent) string {

76− if result == nil {

77− return ""

78− }

79− if text, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentText](result); ok {

80− return text.Text

81− }

82− if errResult, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentError](result); ok {

83− if errResult.Error != nil {

84− return errResult.Error.Error()

85− }

86− return ""

87− }

88− if media, ok := fantasy.AsToolResultOutputType[fantasy.ToolResultOutputContentMedia](result); ok {

89− return media.Data

90− }

91− return ""

92−}

internal/agent/loop_detection_test.godeleted−205

@@ -1,205 +0,0 @@

1−package agent

2−

3−import (

4− "fmt"

5− "testing"

6−

7− "charm.land/fantasy"

8−)

9−

10−// makeStep creates a StepResult with the given tool calls and results in its Content.

11−func makeStep(calls []fantasy.ToolCallContent, results []fantasy.ToolResultContent) fantasy.StepResult {

12− var content fantasy.ResponseContent

13− for _, c := range calls {

14− content = append(content, c)

15− }

16− for _, r := range results {

17− content = append(content, r)

18− }

19− return fantasy.StepResult{

20− Response: fantasy.Response{

21− Content: content,

22− },

23− }

24−}

25−

26−// makeToolStep creates a step with a single tool call and matching text result.

27−func makeToolStep(name, input, output string) fantasy.StepResult {

28− callID := fmt.Sprintf("call_%s_%s", name, input)

29− return makeStep(

30− []fantasy.ToolCallContent{

31− {ToolCallID: callID, ToolName: name, Input: input},

32− },

33− []fantasy.ToolResultContent{

34− {ToolCallID: callID, ToolName: name, Result: fantasy.ToolResultOutputContentText{Text: output}},

35− },

36− )

37−}

38−

39−// makeEmptyStep creates a step with no tool calls (e.g. a text-only response).

40−func makeEmptyStep() fantasy.StepResult {

41− return fantasy.StepResult{

42− Response: fantasy.Response{

43− Content: fantasy.ResponseContent{

44− fantasy.TextContent{Text: "thinking..."},

45− },

46− },

47− }

48−}

49−

50−func TestHasRepeatedToolCalls(t *testing.T) {

51− t.Run("no steps", func(t *testing.T) {

52− result := hasRepeatedToolCalls(nil, 10, 5)

53− if result {

54− t.Error("expected false for empty steps")

55− }

56− })

57−

58− t.Run("fewer steps than window", func(t *testing.T) {

59− steps := make([]fantasy.StepResult, 5)

60− for i := range steps {

61− steps[i] = makeToolStep("read", `{"file":"a.go"}`, "content")

62− }

63− result := hasRepeatedToolCalls(steps, 10, 5)

64− if result {

65− t.Error("expected false when fewer steps than window size")

66− }

67− })

68−

69− t.Run("all different signatures", func(t *testing.T) {

70− steps := make([]fantasy.StepResult, 10)

71− for i := range steps {

72− steps[i] = makeToolStep("tool", fmt.Sprintf(`{"i":%d}`, i), fmt.Sprintf("result-%d", i))

73− }

74− result := hasRepeatedToolCalls(steps, 10, 5)

75− if result {

76− t.Error("expected false when all signatures are different")

77− }

78− })

79−

80− t.Run("exact repeat at threshold not detected", func(t *testing.T) {

81− // maxRepeats=5 means > 5 is needed, so exactly 5 should return false

82− steps := make([]fantasy.StepResult, 10)

83− for i := range 5 {

84− steps[i] = makeToolStep("read", `{"file":"a.go"}`, "content")

85− }

86− for i := 5; i < 10; i++ {

87− steps[i] = makeToolStep("tool", fmt.Sprintf(`{"i":%d}`, i), fmt.Sprintf("result-%d", i))

88− }

89− result := hasRepeatedToolCalls(steps, 10, 5)

90− if result {

91− t.Error("expected false when count equals maxRepeats (threshold is >)")

92− }

93− })

94−

95− t.Run("loop detected", func(t *testing.T) {

96− // 6 identical steps in a window of 10 with maxRepeats=5 → detected

97− steps := make([]fantasy.StepResult, 10)

98− for i := range 6 {

99− steps[i] = makeToolStep("read", `{"file":"a.go"}`, "content")

100− }

101− for i := 6; i < 10; i++ {

102− steps[i] = makeToolStep("tool", fmt.Sprintf(`{"i":%d}`, i), fmt.Sprintf("result-%d", i))

103− }

104− result := hasRepeatedToolCalls(steps, 10, 5)

105− if !result {

106− t.Error("expected true when same signature appears more than maxRepeats times")

107− }

108− })

109−

110− t.Run("steps without tool calls are skipped", func(t *testing.T) {

111− // Mix of tool steps and empty steps — empty ones should not affect counts

112− steps := make([]fantasy.StepResult, 10)

113− for i := range 4 {

114− steps[i] = makeToolStep("read", `{"file":"a.go"}`, "content")

115− }

116− for i := 4; i < 8; i++ {

117− steps[i] = makeEmptyStep()

118− }

119− for i := 8; i < 10; i++ {

120− steps[i] = makeToolStep("write", `{"file":"b.go"}`, "ok")

121− }

122− result := hasRepeatedToolCalls(steps, 10, 5)

123− if result {

124− t.Error("expected false: only 4 repeated tool calls, empty steps should be skipped")

125− }

126− })

127−

128− t.Run("multiple different patterns alternating", func(t *testing.T) {

129− // Two patterns alternating: each appears 5 times — not above threshold

130− steps := make([]fantasy.StepResult, 10)

131− for i := range steps {

132− if i%2 == 0 {

133− steps[i] = makeToolStep("read", `{"file":"a.go"}`, "content-a")

134− } else {

135− steps[i] = makeToolStep("write", `{"file":"b.go"}`, "content-b")

136− }

137− }

138− result := hasRepeatedToolCalls(steps, 10, 5)

139− if result {

140− t.Error("expected false: two patterns each appearing 5 times (not > 5)")

141− }

142− })

143−}

144−

145−func TestGetToolInteractionSignature(t *testing.T) {

146− t.Run("empty content returns empty string", func(t *testing.T) {

147− sig := getToolInteractionSignature(fantasy.ResponseContent{})

148− if sig != 0 {

149− t.Errorf("expected zero signature, got %d", sig)

150− }

151− })

152−

153− t.Run("text only content returns zero signature", func(t *testing.T) {

154− content := fantasy.ResponseContent{

155− fantasy.TextContent{Text: "hello"},

156− }

157− sig := getToolInteractionSignature(content)

158− if sig != 0 {

159− t.Errorf("expected zero signature, got %d", sig)

160− }

161− })

162−

163− t.Run("tool call with result produces signature", func(t *testing.T) {

164− content := fantasy.ResponseContent{

165− fantasy.ToolCallContent{ToolCallID: "1", ToolName: "read", Input: `{"file":"a.go"}`},

166− fantasy.ToolResultContent{ToolCallID: "1", ToolName: "read", Result: fantasy.ToolResultOutputContentText{Text: "content"}},

167− }

168− sig := getToolInteractionSignature(content)

169− if sig == 0 {

170− t.Error("expected non-zero signature")

171− }

172− })

173−

174− t.Run("same interactions produce same signature", func(t *testing.T) {

175− content1 := fantasy.ResponseContent{

176− fantasy.ToolCallContent{ToolCallID: "1", ToolName: "read", Input: `{"file":"a.go"}`},

177− fantasy.ToolResultContent{ToolCallID: "1", ToolName: "read", Result: fantasy.ToolResultOutputContentText{Text: "content"}},

178− }

179− content2 := fantasy.ResponseContent{

180− fantasy.ToolCallContent{ToolCallID: "2", ToolName: "read", Input: `{"file":"a.go"}`},

181− fantasy.ToolResultContent{ToolCallID: "2", ToolName: "read", Result: fantasy.ToolResultOutputContentText{Text: "content"}},

182− }

183− sig1 := getToolInteractionSignature(content1)

184− sig2 := getToolInteractionSignature(content2)

185− if sig1 != sig2 {

186− t.Errorf("expected same signature for same interactions, got %d and %d", sig1, sig2)

187− }

188− })

189−

190− t.Run("different inputs produce different signatures", func(t *testing.T) {

191− content1 := fantasy.ResponseContent{

192− fantasy.ToolCallContent{ToolCallID: "1", ToolName: "read", Input: `{"file":"a.go"}`},

193− fantasy.ToolResultContent{ToolCallID: "1", ToolName: "read", Result: fantasy.ToolResultOutputContentText{Text: "content"}},

194− }

195− content2 := fantasy.ResponseContent{

196− fantasy.ToolCallContent{ToolCallID: "1", ToolName: "read", Input: `{"file":"b.go"}`},

197− fantasy.ToolResultContent{ToolCallID: "1", ToolName: "read", Result: fantasy.ToolResultOutputContentText{Text: "content"}},

198− }

199− sig1 := getToolInteractionSignature(content1)

200− sig2 := getToolInteractionSignature(content2)

201− if sig1 == sig2 {

202− t.Error("expected different signatures for different inputs")

203− }

204− })

205−}

internal/agent/notify/notify.godeleted−63

@@ -1,63 +0,0 @@

1−// Package notify defines domain notification types for agent events.

2−// These types are decoupled from UI concerns so the agent can publish

3−// events without importing UI packages.

4−package notify

5−

6−// Type identifies the kind of agent notification.

7−type Type string

8−

9−const (

10− // TypeAgentFinished indicates the agent has completed its turn.

11− TypeAgentFinished Type = "agent_finished"

12− // TypeReAuthenticate indicates the agent encountered an

13− // authentication error and the user needs to re-authenticate.

14− TypeReAuthenticate Type = "re_authenticate"

15− // TypeAgentError indicates the agent's turn terminated with an

16− // error. The error text is carried in Notification.Message.

17− TypeAgentError Type = "error"

18−)

19−

20−// Notification represents a domain event published by the agent.

21−type Notification struct {

22− SessionID string

23− SessionTitle string

24− Type Type

25− ProviderID string

26− // RunID, when non-empty, is the caller-supplied correlator

27− // (proto.AgentMessage.RunID) for the run that produced this

28− // notification. It lets observers attribute a TypeAgentError to a

29− // specific request rather than to any in-flight run on the

30− // session. Empty when no caller set one.

31− RunID string

32− // Message carries the error text for TypeAgentError. Other

33− // notification types ignore it.

34− Message string

35−}

36−

37−// RunComplete is the authoritative end-of-run signal for a session.

38−// It is published exactly once per top-level agent run (per

39−// [sessionAgent.Run] invocation that actually executed) after all

40−// message updates for the turn have been flushed via

41−// message.Service.FlushAll. Carries the final assistant text and

42−// message ID so non-interactive clients can reconcile stdout even if

43−// SSE events arrive out of order or are dropped by the broker. Error

44−// is non-empty when the run terminated with an error; Cancelled is

45−// true when the run terminated due to context cancellation. The two

46−// are mutually exclusive in the success case but may overlap when a

47−// cancel triggers a downstream error.

48−//

49−// RunID identifies the specific request that produced this event.

50−// It is the value the caller set on `proto.AgentMessage.RunID` (or

51−// equivalently propagated via agent.WithRunID on the context that

52−// reaches the coordinator); empty when no caller set one. Filtering

53−// by RunID lets a client correlate a SendMessage call with its

54−// terminal event even when the session is busy and other turns are

55−// finishing on the same session.

56−type RunComplete struct {

57− SessionID string

58− RunID string

59− MessageID string

60− Text string

61− Error string

62− Cancelled bool

63−}

internal/agent/prompt/prompt.go+37−68

@@ -4,7 +4,6 @@ import (

4 "cmp"

5 "context"

6 "fmt"

7− "log/slog"

8 "os"

7 "os"

9 "path/filepath"

8 "path/filepath"

10 "runtime"

9 "runtime"

@@ -13,7 +12,6 @@ import (

13 "time"

12 "time"

15 "github.com/charmbracelet/crush/internal/config"

14 "github.com/charmbracelet/crush/internal/config"

16− "github.com/charmbracelet/crush/internal/filepathext"

17 "github.com/charmbracelet/crush/internal/home"

15 "github.com/charmbracelet/crush/internal/home"

18 "github.com/charmbracelet/crush/internal/shell"

16 "github.com/charmbracelet/crush/internal/shell"

19 "github.com/charmbracelet/crush/internal/skills"

17 "github.com/charmbracelet/crush/internal/skills"

@@ -29,17 +27,16 @@ type Prompt struct {

29}

27}

31type PromptDat struct {

29type PromptDat struct {

32− Provider string

30+ Provider string

33− Model string

31+ Model string

34− Config config.Config

32+ Config config.Config

35− WorkingDir string

33+ WorkingDir string

36− IsGitRepo bool

34+ IsGitRepo bool

37− Platform string

35+ Platform string

38− Date string

36+ Date string

39− GitStatus string

37+ GitStatus string

40− ContextFiles []ContextFile

38+ ContextFiles []ContextFile

41− GlobalContextFiles []ContextFile

39+ AvailSkillXML string

42− AvailSkillXML string

43}

40}

45type ContextFile struct {

42type ContextFile struct {

@@ -79,13 +76,13 @@ func NewPrompt(name, promptTemplate string, opts ...Option) (*Prompt, error) {

79 return p, nil

76 return p, nil

80}

77}

82−func (p *Prompt) Build(ctx context.Context, provider, model string, store *config.ConfigStore) (string, error) {

79+func (p *Prompt) Build(ctx context.Context, provider, model string, cfg config.Config) (string, error) {

83 t, err := template.New(p.name).Parse(p.template)

80 t, err := template.New(p.name).Parse(p.template)

84 if err != nil {

81 if err != nil {

85 return "", fmt.Errorf("parsing template: %w", err)

82 return "", fmt.Errorf("parsing template: %w", err)

86 }

83 }

87 var sb strings.Builder

84 var sb strings.Builder

88− d, err := p.promptData(ctx, provider, model, store)

85+ d, err := p.promptData(ctx, provider, model, cfg)

89 if err != nil {

86 if err != nil {

90 return "", err

87 return "", err

91 }

88 }

@@ -107,15 +104,18 @@ func processFile(filePath string) *ContextFile {

107 }

104 }

108}

105}

109

106

110−func processContextPath(p string, store *config.ConfigStore) []ContextFile {

107+func processContextPath(p string, cfg config.Config) []ContextFile {

111 var contexts []ContextFile

108 var contexts []ContextFile

112− fullPath := filepathext.SmartJoin(store.WorkingDir(), p)

109+ fullPath := p

110+ if !filepath.IsAbs(p) {

111+ fullPath = filepath.Join(cfg.WorkingDir(), p)

112+ }

113 info, err := os.Stat(fullPath)

114 if err != nil {

115 return contexts

116 }

117 if info.IsDir() {

118− _ = filepath.WalkDir(fullPath, func(path string, d os.DirEntry, err error) error {

118+ filepath.WalkDir(fullPath, func(path string, d os.DirEntry, err error) error {

119 if err != nil {

120 return err

121 }

@@ -136,11 +136,11 @@ func processContextPath(p string, store *config.ConfigStore) []ContextFile {

136}

137

138// expandPath expands ~ and environment variables in file paths

139−func expandPath(path string, store *config.ConfigStore) string {

139+func expandPath(path string, cfg config.Config) string {

140 path = home.Long(path)

141 // Handle environment variable expansion using the same pattern as config

142 if strings.HasPrefix(path, "$") {

143− if expanded, err := store.Resolver().ResolveValue(path); err == nil {

143+ if expanded, err := cfg.Resolver().ResolveValue(path); err == nil {

144 path = expanded

145 }

146 }

@@ -148,67 +148,39 @@ func expandPath(path string, store *config.ConfigStore) string {

148 return path

149}

150

151−// loadContextFiles loads and deduplicates context files from a list of paths.

151+func (p *Prompt) promptData(ctx context.Context, provider, model string, cfg config.Config) (PromptDat, error) {

152−func loadContextFiles(paths []string, store *config.ConfigStore) map[string][]ContextFile {

152+ workingDir := cmp.Or(p.workingDir, cfg.WorkingDir())

153+ platform := cmp.Or(p.platform, runtime.GOOS)

154+

153 files := map[string][]ContextFile{}

155 files := map[string][]ContextFile{}

156+

154− for _, pth := range paths {

157+ for _, pth := range cfg.Options.ContextPaths {

155− expanded := expandPath(pth, store)

158+ expanded := expandPath(pth, cfg)

156 pathKey := strings.ToLower(expanded)

159 pathKey := strings.ToLower(expanded)

157 if _, ok := files[pathKey]; ok {

160 if _, ok := files[pathKey]; ok {

158 continue

161 continue

159 }

162 }

160− files[pathKey] = processContextPath(expanded, store)

163+ content := processContextPath(expanded, cfg)

164+ files[pathKey] = content

161 }

165 }

162− return files

163−}

164−

165−func (p *Prompt) promptData(ctx context.Context, provider, model string, store *config.ConfigStore) (PromptDat, error) {

166− workingDir := cmp.Or(p.workingDir, store.WorkingDir())

167− platform := cmp.Or(p.platform, runtime.GOOS)

168−

169− cfg := store.Config()

170− contextFiles := loadContextFiles(cfg.Options.ContextPaths, store)

171− globalContextFiles := loadContextFiles(cfg.Options.GlobalContextPaths, store)

172

166

173 // Discover and load skills metadata.

167 // Discover and load skills metadata.

174 var availSkillXML string

168 var availSkillXML string

175−

176− // Start with builtin skills.

177− allSkills := skills.DiscoverBuiltin()

178− builtinNames := make(map[string]bool, len(allSkills))

179− for _, s := range allSkills {

180− builtinNames[s.Name] = true

181− }

182−

183− // Discover user skills from configured paths.

184 if len(cfg.Options.SkillsPaths) > 0 {

169 if len(cfg.Options.SkillsPaths) > 0 {

185 expandedPaths := make([]string, 0, len(cfg.Options.SkillsPaths))

170 expandedPaths := make([]string, 0, len(cfg.Options.SkillsPaths))

186 for _, pth := range cfg.Options.SkillsPaths {

171 for _, pth := range cfg.Options.SkillsPaths {

187− expandedPaths = append(expandedPaths, expandPath(pth, store))

172+ expandedPaths = append(expandedPaths, expandPath(pth, cfg))

188 }

173 }

189− for _, userSkill := range skills.Discover(expandedPaths) {

174+ if discoveredSkills := skills.Discover(expandedPaths); len(discoveredSkills) > 0 {

190− if builtinNames[userSkill.Name] {

175+ availSkillXML = skills.ToPromptXML(discoveredSkills)

191− slog.Warn("User skill overrides builtin skill", "name", userSkill.Name)

192− }

193− allSkills = append(allSkills, userSkill)

194 }

176 }

195 }

177 }

196

178

197− // Deduplicate: user skills override builtins with the same name.

198− allSkills = skills.Deduplicate(allSkills)

199−

200− // Filter out disabled skills.

201− allSkills = skills.Filter(allSkills, cfg.Options.DisabledSkills)

202−

203− if len(allSkills) > 0 {

204− availSkillXML = skills.ToPromptXML(allSkills)

205− }

206−

207− isGit := isGitRepo(store.WorkingDir())

179+ isGit := isGitRepo(cfg.WorkingDir())

208 data := PromptDat{

180 data := PromptDat{

209 Provider: provider,

181 Provider: provider,

210 Model: model,

182 Model: model,

211− Config: *cfg,

183+ Config: cfg,

212 WorkingDir: filepath.ToSlash(workingDir),

184 WorkingDir: filepath.ToSlash(workingDir),

213 IsGitRepo: isGit,

185 IsGitRepo: isGit,

214 Platform: platform,

186 Platform: platform,

@@ -217,17 +189,14 @@ func (p *Prompt) promptData(ctx context.Context, provider, model string, store *

217 }

189 }

218 if isGit {

190 if isGit {

219 var err error

191 var err error

220− data.GitStatus, err = getGitStatus(ctx, store.WorkingDir())

192+ data.GitStatus, err = getGitStatus(ctx, cfg.WorkingDir())

221 if err != nil {

193 if err != nil {

222 return PromptDat{}, err

194 return PromptDat{}, err

223 }

195 }

224 }

196 }

225

197

226− for _, files := range contextFiles {

198+ for _, contextFiles := range files {

227− data.ContextFiles = append(data.ContextFiles, files...)

199+ data.ContextFiles = append(data.ContextFiles, contextFiles...)

228− }

229− for _, files := range globalContextFiles {

230− data.GlobalContextFiles = append(data.GlobalContextFiles, files...)

231 }

200 }

232 return data, nil

201 return data, nil

233}

202}

internal/agent/prompts.go+1−1

@@ -33,7 +33,7 @@ func taskPrompt(opts ...prompt.Option) (*prompt.Prompt, error) {

33 return systemPrompt, nil

34}

36−func InitializePrompt(cfg *config.ConfigStore) (string, error) {

36+func InitializePrompt(cfg config.Config) (string, error) {

37 systemPrompt, err := prompt.NewPrompt("initialize", string(initializePromptTmpl))

38 if err != nil {

39 return "", err

internal/agent/queued_runid_test.godeleted−181

@@ -1,181 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "errors"

6− "sync/atomic"

7− "testing"

8− "time"

9−

10− "charm.land/catwalk/pkg/catwalk"

11− "charm.land/fantasy"

12− "github.com/charmbracelet/crush/internal/agent/notify"

13− "github.com/charmbracelet/crush/internal/message"

14− "github.com/charmbracelet/crush/internal/pubsub"

15− "github.com/stretchr/testify/require"

16−)

17−

18−// gatedStreamModel streams a single text part followed by a clean finish,

19−// but blocks the very first Stream call until its gate is released. That

20−// lets a test hold a run "active" (past PrepareStep, inside Stream) just

21−// long enough to enqueue a follow-up prompt behind the busy session.

22−// Subsequent Stream calls (e.g. the recursive run draining the queue)

23−// proceed immediately.

24−type gatedStreamModel struct {

25− text string

26− gate chan struct{}

27− entered chan struct{}

28− calls atomic.Int64

29−}

30−

31−func (m *gatedStreamModel) Provider() string { return "fake" }

32−func (m *gatedStreamModel) Model() string { return "fake-model" }

33−

34−func (m *gatedStreamModel) Generate(ctx context.Context, call fantasy.Call) (*fantasy.Response, error) {

35− return &fantasy.Response{

36− Content: fantasy.ResponseContent{fantasy.TextContent{Text: m.text}},

37− FinishReason: fantasy.FinishReasonStop,

38− }, nil

39−}

40−

41−func (m *gatedStreamModel) Stream(ctx context.Context, call fantasy.Call) (fantasy.StreamResponse, error) {

42− if m.calls.Add(1) == 1 {

43− close(m.entered)

44− select {

45− case <-m.gate:

46− case <-ctx.Done():

47− }

48− }

49− text := m.text

50− return func(yield func(fantasy.StreamPart) bool) {

51− if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextStart, ID: "1"}) {

52− return

53− }

54− if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextDelta, ID: "1", Delta: text}) {

55− return

56− }

57− if !yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeTextEnd, ID: "1"}) {

58− return

59− }

60− yield(fantasy.StreamPart{Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonStop})

61− }, nil

62−}

63−

64−func (m *gatedStreamModel) GenerateObject(ctx context.Context, call fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {

65− return nil, errors.New("not implemented")

66−}

67−

68−func (m *gatedStreamModel) StreamObject(ctx context.Context, call fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {

69− return nil, errors.New("not implemented")

70−}

71−

72−// TestRun_QueuedRunIDPromptRunsRecursivelyAndPublishesRunComplete is the

73−// end-to-end proof of fix 2: a prompt carrying a RunID that is queued

74−// behind a busy session must NOT be silently folded into the active turn.

75−// It runs as its own turn via the recursive run path and publishes its

76−// own terminal RunComplete, so a `crush run` caller blocking on that

77−// RunID does not hang. The active turn keeps its own RunComplete too.

78−func TestRun_QueuedRunIDPromptRunsRecursivelyAndPublishesRunComplete(t *testing.T) {

79− t.Parallel()

80−

81− env := testEnv(t)

82− broker := pubsub.NewBroker[notify.RunComplete]()

83− t.Cleanup(broker.Shutdown)

84−

85− large := &gatedStreamModel{

86− text: "done",

87− gate: make(chan struct{}),

88− entered: make(chan struct{}),

89− }

90− small := &finishStreamModel{text: "title"}

91−

92− sa := NewSessionAgent(SessionAgentOptions{

93− LargeModel: Model{Model: large, CatwalkCfg: catwalk.Model{ContextWindow: 200000, DefaultMaxTokens: 10000}},

94− SmallModel: Model{Model: small, CatwalkCfg: catwalk.Model{ContextWindow: 200000, DefaultMaxTokens: 10000}},

95− IsYolo: true,

96− Sessions: env.sessions,

97− Messages: env.messages,

98− RunComplete: broker,

99− }).(*sessionAgent)

100−

101− sess, err := env.sessions.Create(t.Context(), "session")

102− require.NoError(t, err)

103−

104− subCtx, subCancel := context.WithCancel(t.Context())

105− defer subCancel()

106− ch := broker.Subscribe(subCtx)

107−

108− // Start the main turn; it blocks inside Stream once active.

109− mainDone := make(chan error, 1)

110− go func() {

111− _, runErr := sa.Run(t.Context(), SessionAgentCall{

112− SessionID: sess.ID,

113− RunID: "run-main",

114− Prompt: "main",

115− })

116− mainDone <- runErr

117− }()

118−

119− // Wait until the main turn is active (inside Stream).

120− select {

121− case <-large.entered:

122− case <-time.After(5 * time.Second):

123− t.Fatal("main run never entered Stream")

124− }

125− require.True(t, sa.IsSessionBusy(sess.ID), "main run must be active before enqueueing the follow-up")

126−

127− // Enqueue a RunID-bearing follow-up behind the busy session.

128− res, err := sa.Run(t.Context(), SessionAgentCall{

129− SessionID: sess.ID,

130− RunID: "run-follow",

131− Prompt: "follow",

132− })

133− require.NoError(t, err)

134− require.Nil(t, res, "a busy-session follow-up must enqueue and return (nil, nil)")

135− require.Equal(t, 1, sa.QueuedPrompts(sess.ID), "the follow-up must be queued, not folded")

136−

137− // Release the main turn so it completes and hands off to the queue.

138− close(large.gate)

139− require.NoError(t, <-mainDone)

140−

141− // Both turns must publish their own terminal RunComplete.

142− got := map[string]notify.RunComplete{}

143− deadline := time.After(5 * time.Second)

144− for len(got) < 2 {

145− select {

146− case ev := <-ch:

147− got[ev.Payload.RunID] = ev.Payload

148− case <-deadline:

149− t.Fatalf("timed out waiting for both RunCompletes; got %v", got)

150− }

151− }

152−

153− main, ok := got["run-main"]

154− require.True(t, ok, "the active turn must publish its own RunComplete")

155− require.Empty(t, main.Error)

156− require.False(t, main.Cancelled)

157−

158− follow, ok := got["run-follow"]

159− require.True(t, ok,

160− "the queued RunID prompt must publish its own RunComplete instead of being folded silently")

161− require.Empty(t, follow.Error)

162− require.False(t, follow.Cancelled)

163− require.Equal(t, "done", follow.Text, "the queued prompt ran as its own turn")

164−

165− // Two distinct assistant turns prove the follow-up was not folded.

166− msgs, err := env.messages.List(t.Context(), sess.ID)

167− require.NoError(t, err)

168− var assistants, follows int

169− for _, m := range msgs {

170− switch m.Role {

171− case message.Assistant:

172− assistants++

173− case message.User:

174− if m.Content().String() == "follow" {

175− follows++

176− }

177− }

178− }

179− require.Equal(t, 2, assistants, "the active turn and the recursive turn each produce one assistant message")

180− require.Equal(t, 1, follows, "the follow-up prompt is its own user turn")

181−}

internal/agent/run_complete_test.godeleted−322

@@ -1,322 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "testing"

6− "time"

7−

8− "github.com/charmbracelet/crush/internal/agent/notify"

9− "github.com/charmbracelet/crush/internal/pubsub"

10− "github.com/stretchr/testify/require"

11−)

12−

13−// TestSessionAgentRun_QueueStripsOnComplete verifies that when a Run

14−// call is enqueued (because the session is already busy), the

15−// OnComplete hook is NOT propagated onto the queued copy. The hook

16−// belongs to the caller's retry/coalesce scope (typically

17−// coordinator.Run) which has already returned by the time the queue

18−// drains; carrying it forward would silently funnel the terminal

19−// event into a closure nobody reads, and subscribers (`crush run`)

20−// would hang waiting for a RunComplete that never publishes.

21−func TestSessionAgentRun_QueueStripsOnComplete(t *testing.T) {

22− t.Parallel()

23−

24− env := testEnv(t)

25− a := NewSessionAgent(SessionAgentOptions{

26− Sessions: env.sessions,

27− Messages: env.messages,

28− }).(*sessionAgent)

29−

30− const sessionID = "queued-session"

31− // Mark the session as busy so Run takes the queue branch

32− // without needing a real model.

33− a.activeRequests.Set(sessionID, func() {})

34−

35− var called bool

36− hook := func(notify.RunComplete) { called = true }

37−

38− res, err := a.Run(t.Context(), SessionAgentCall{

39− SessionID: sessionID,

40− RunID: "run-xyz",

41− Prompt: "queued prompt",

42− OnComplete: hook,

43− })

44− require.NoError(t, err)

45− require.Nil(t, res, "queued Run must return (nil, nil)")

46− require.False(t, called,

47− "OnComplete must not fire on the enqueue path; the caller's scope is still live")

48−

49− queued, ok := a.messageQueue.Get(sessionID)

50− require.True(t, ok)

51− require.Len(t, queued, 1)

52− require.Nil(t, queued[0].OnComplete,

53− "queued SessionAgentCall must have OnComplete stripped so the drain falls back to the default broker publish")

54− require.Equal(t, "queued prompt", queued[0].Prompt,

55− "all other fields must be preserved on the queued copy")

56− require.Equal(t, "run-xyz", queued[0].RunID,

57− "RunID must be preserved on the queued copy so the drained turn's "+

58− "RunComplete still correlates with the originating SendMessage")

59−}

60−

61−// TestDrainQueueForStep_FiltersUnderDispatchLock verifies that the queue

62−// drain evaluates the per-session cancel mark while holding the dispatch

63−// mutex (canceledBySeq's documented precondition). Queued calls at or

64−// below the cancel high-water mark are dropped, calls queued after the

65−// cancel (higher seq) are folded, untracked enqueues (seq == 0) are

66−// dropped whenever any mark is present, and the queue is cleared. These

67−// calls carry no RunID, so all foldable survivors are returned for

68−// folding (the existing follow-up behavior).

69−func TestDrainQueueForStep_FiltersUnderDispatchLock(t *testing.T) {

70− t.Parallel()

71−

72− env := testEnv(t)

73− a := NewSessionAgent(SessionAgentOptions{

74− Sessions: env.sessions,

75− Messages: env.messages,

76− }).(*sessionAgent)

77−

78− const sessionID = "drain-session"

79− a.messageQueue.Set(sessionID, []SessionAgentCall{

80− {SessionID: sessionID, Prompt: "below", acceptSeq: 1},

81− {SessionID: sessionID, Prompt: "at-mark", acceptSeq: 2},

82− {SessionID: sessionID, Prompt: "after", acceptSeq: 3},

83− {SessionID: sessionID, Prompt: "untracked", acceptSeq: 0},

84− })

85− // Cancel high-water mark at seq 2: seq <= 2 and seq == 0 are covered.

86− a.cancelMark.Set(sessionID, 2)

87−

88− fold, canceledWithRunID := a.drainQueueForStep(sessionID)

89−

90− require.Len(t, fold, 1,

91− "only the follow-up queued after the cancel (seq > mark) must be folded")

92− require.Equal(t, "after", fold[0].Prompt)

93− require.Empty(t, canceledWithRunID,

94− "no dropped call carried a RunID, so none need a terminal RunComplete")

95−

96− _, ok := a.messageQueue.Get(sessionID)

97− require.False(t, ok, "drain must clear the session message queue when nothing is kept")

98−}

99−

100−// TestDrainQueueForStep_NoMarkFoldsAllNonRunID verifies that with no

101−// cancel mark recorded, every queued call without a RunID is folded.

102−func TestDrainQueueForStep_NoMarkFoldsAllNonRunID(t *testing.T) {

103− t.Parallel()

104−

105− env := testEnv(t)

106− a := NewSessionAgent(SessionAgentOptions{

107− Sessions: env.sessions,

108− Messages: env.messages,

109− }).(*sessionAgent)

110−

111− const sessionID = "drain-nomark"

112− a.messageQueue.Set(sessionID, []SessionAgentCall{

113− {SessionID: sessionID, Prompt: "a", acceptSeq: 0},

114− {SessionID: sessionID, Prompt: "b", acceptSeq: 5},

115− })

116−

117− fold, canceledWithRunID := a.drainQueueForStep(sessionID)

118− require.Len(t, fold, 2, "no cancel mark means all non-RunID queued calls are folded")

119− require.Empty(t, canceledWithRunID)

120−}

121−

122−// TestDrainQueueForStep_KeepsRunIDPromptsQueued is the core of fix 2: a

123−// queued prompt that carries a RunID must NOT be folded into the active

124−// turn. Folding it would silently absorb it into another turn and never

125−// publish a RunComplete for its RunID, hanging a `crush run` caller that

126−// blocks on that event. Such prompts are left in the queue so the

127−// recursive run path gives each its own turn and its own RunComplete.

128−// Non-RunID prompts are still folded.

129−func TestDrainQueueForStep_KeepsRunIDPromptsQueued(t *testing.T) {

130− t.Parallel()

131−

132− env := testEnv(t)

133− a := NewSessionAgent(SessionAgentOptions{

134− Sessions: env.sessions,

135− Messages: env.messages,

136− }).(*sessionAgent)

137−

138− const sessionID = "drain-runid"

139− a.messageQueue.Set(sessionID, []SessionAgentCall{

140− {SessionID: sessionID, Prompt: "fold-me", acceptSeq: 1},

141− {SessionID: sessionID, RunID: "run-a", Prompt: "keep-me", acceptSeq: 2},

142− {SessionID: sessionID, RunID: "run-b", Prompt: "keep-me-too", acceptSeq: 3},

143− })

144−

145− fold, canceledWithRunID := a.drainQueueForStep(sessionID)

146−

147− require.Len(t, fold, 1, "only the non-RunID prompt is folded into the active turn")

148− require.Equal(t, "fold-me", fold[0].Prompt)

149− require.Empty(t, canceledWithRunID)

150−

151− kept, ok := a.messageQueue.Get(sessionID)

152− require.True(t, ok, "RunID-bearing prompts must remain queued for the recursive run path")

153− require.Len(t, kept, 2)

154− require.Equal(t, "run-a", kept[0].RunID)

155− require.Equal(t, "run-b", kept[1].RunID)

156−}

157−

158−// TestDrainQueueForStep_ReportsCanceledRunIDDrops verifies that a queued

159−// prompt carrying a RunID that is dropped because a cancel covers it is

160−// reported in canceledWithRunID so the caller can publish its terminal

161−// cancelled RunComplete. A canceled prompt without a RunID is dropped

162−// silently as before.

163−func TestDrainQueueForStep_ReportsCanceledRunIDDrops(t *testing.T) {

164− t.Parallel()

165−

166− env := testEnv(t)

167− a := NewSessionAgent(SessionAgentOptions{

168− Sessions: env.sessions,

169− Messages: env.messages,

170− }).(*sessionAgent)

171−

172− const sessionID = "drain-cancel-runid"

173− a.messageQueue.Set(sessionID, []SessionAgentCall{

174− {SessionID: sessionID, RunID: "run-canceled", Prompt: "canceled", acceptSeq: 1},

175− {SessionID: sessionID, Prompt: "canceled-no-runid", acceptSeq: 1},

176− {SessionID: sessionID, RunID: "run-survives", Prompt: "survives", acceptSeq: 5},

177− })

178− a.cancelMark.Set(sessionID, 2)

179−

180− fold, canceledWithRunID := a.drainQueueForStep(sessionID)

181−

182− require.Empty(t, fold, "no uncanceled non-RunID prompts to fold")

183− require.Len(t, canceledWithRunID, 1,

184− "only the dropped RunID-bearing prompt needs a terminal RunComplete")

185− require.Equal(t, "run-canceled", canceledWithRunID[0].RunID)

186−

187− kept, ok := a.messageQueue.Get(sessionID)

188− require.True(t, ok)

189− require.Len(t, kept, 1, "the uncanceled RunID prompt stays queued")

190− require.Equal(t, "run-survives", kept[0].RunID)

191−}

192−

193−// TestRunCompletePublisher_MustDeliverOverTakesPublish exercises the

194−// pubsub.Publisher interface change end-to-end: a Broker is the only

195−// concrete Publisher implementation and must satisfy both Publish and

196−// PublishMustDeliver. The coordinator's final RunComplete emit relies

197−// on PublishMustDeliver to apply bounded-blocking semantics so a

198−// momentarily-full subscriber buffer can't silently drop the

199−// authoritative end-of-run event.

200−func TestRunCompletePublisher_MustDeliverOverTakesPublish(t *testing.T) {

201− t.Parallel()

202−

203− broker := pubsub.NewBroker[notify.RunComplete]()

204− t.Cleanup(broker.Shutdown)

205−

206− // Subscribe before publishing so the event is delivered.

207− ctx, cancel := context.WithCancel(t.Context())

208− defer cancel()

209− ch := broker.Subscribe(ctx)

210−

211− rc := notify.RunComplete{SessionID: "S", MessageID: "m", Text: "ok"}

212− var pub pubsub.Publisher[notify.RunComplete] = broker

213− pub.PublishMustDeliver(t.Context(), pubsub.UpdatedEvent, rc)

214−

215− select {

216− case got := <-ch:

217− require.Equal(t, rc, got.Payload)

218− case <-time.After(time.Second):

219− t.Fatal("PublishMustDeliver did not deliver event")

220− }

221−}

222−

223−// requireSingleCancelledRunComplete reads exactly one RunComplete from ch,

224−// asserts it is the cancelled terminal event for runID, and verifies no

225−// second event arrives. This observes the published pubsub event rather

226−// than internal bookkeeping, which is the contract a `crush run` caller

227−// blocking on the broker actually relies on.

228−func requireSingleCancelledRunComplete(t *testing.T, ch <-chan pubsub.Event[notify.RunComplete], sessionID, runID string) {

229− t.Helper()

230− select {

231− case ev := <-ch:

232− require.Equal(t, runID, ev.Payload.RunID,

233− "the published RunComplete must carry the dropped queued prompt's RunID")

234− require.Equal(t, sessionID, ev.Payload.SessionID)

235− require.True(t, ev.Payload.Cancelled,

236− "a dropped queued prompt must publish a cancelled RunComplete")

237− case <-time.After(5 * time.Second):

238− t.Fatal("timed out waiting for the cancelled RunComplete")

239− }

240− select {

241− case extra := <-ch:

242− t.Fatalf("expected exactly one RunComplete, got a second: %+v", extra.Payload)

243− case <-time.After(100 * time.Millisecond):

244− }

245−}

246−

247−// TestCancel_QueuedRunIDPromptPublishesCancelledRunComplete proves the

248−// terminal-event behavior end-to-end: a RunID-bearing prompt sitting in

249−// the queue that is canceled while queued (via the public Cancel path,

250−// which routes through clearQueueAndNotify -> publishCanceledQueueDrops)

251−// must emit exactly one cancelled RunComplete on the broker for its

252−// RunID. A queued prompt without a RunID is dropped silently. This is the

253−// coverage the earlier drain test lacked: it asserted the returned

254−// bookkeeping slice, not the published event a `crush run` caller awaits.

255−func TestCancel_QueuedRunIDPromptPublishesCancelledRunComplete(t *testing.T) {

256− t.Parallel()

257−

258− env := testEnv(t)

259− broker := pubsub.NewBroker[notify.RunComplete]()

260− t.Cleanup(broker.Shutdown)

261−

262− a := NewSessionAgent(SessionAgentOptions{

263− Sessions: env.sessions,

264− Messages: env.messages,

265− RunComplete: broker,

266− }).(*sessionAgent)

267−

268− subCtx, subCancel := context.WithCancel(t.Context())

269− defer subCancel()

270− ch := broker.Subscribe(subCtx)

271−

272− const sessionID = "cancel-queued-runid"

273− a.messageQueue.Set(sessionID, []SessionAgentCall{

274− {SessionID: sessionID, Prompt: "no-runid", acceptSeq: 1},

275− {SessionID: sessionID, RunID: "run-queued", Prompt: "queued", acceptSeq: 2},

276− })

277−

278− a.Cancel(sessionID)

279−

280− requireSingleCancelledRunComplete(t, ch, sessionID, "run-queued")

281−

282− _, ok := a.messageQueue.Get(sessionID)

283− require.False(t, ok, "Cancel must clear the queue")

284−}

285−

286−// TestDrainQueueForStep_DroppedRunIDPublishesCancelledRunComplete drives

287−// the production drain sequence (drainQueueForStep then

288−// publishCanceledQueueDrops, mirroring the PrepareStep handoff) and

289−// asserts the dropped RunID-bearing prompt actually publishes exactly one

290−// cancelled RunComplete on the broker. The companion bookkeeping test

291−// covers the returned slice; this one covers the observable terminal

292−// event.

293−func TestDrainQueueForStep_DroppedRunIDPublishesCancelledRunComplete(t *testing.T) {

294− t.Parallel()

295−

296− env := testEnv(t)

297− broker := pubsub.NewBroker[notify.RunComplete]()

298− t.Cleanup(broker.Shutdown)

299−

300− a := NewSessionAgent(SessionAgentOptions{

301− Sessions: env.sessions,

302− Messages: env.messages,

303− RunComplete: broker,

304− }).(*sessionAgent)

305−

306− subCtx, subCancel := context.WithCancel(t.Context())

307− defer subCancel()

308− ch := broker.Subscribe(subCtx)

309−

310− const sessionID = "drain-drop-runid"

311− a.messageQueue.Set(sessionID, []SessionAgentCall{

312− {SessionID: sessionID, RunID: "run-dropped", Prompt: "dropped", acceptSeq: 1},

313− {SessionID: sessionID, Prompt: "dropped-no-runid", acceptSeq: 1},

314− })

315− a.cancelMark.Set(sessionID, 2)

316−

317− _, canceledWithRunID := a.drainQueueForStep(sessionID)

318− require.Len(t, canceledWithRunID, 1)

319− a.publishCanceledQueueDrops(canceledWithRunID)

320−

321− requireSingleCancelledRunComplete(t, ch, sessionID, "run-dropped")

322−}

internal/agent/run_marker.godeleted−52

@@ -1,52 +0,0 @@

1−package agent

2−

3−import (

4− "context"

5− "sync/atomic"

6−)

7−

8−// runCompleteMarkerKey is the unexported context key carrying a

9−// [runCompleteMarker] from the dispatch boundary (backend.runAgent)

10−// down into the coordinator. It lets the dispatcher learn whether the

11−// coordinator already published the authoritative terminal

12−// notify.RunComplete for the run, so a fallback terminal event is only

13−// emitted when one is actually missing (e.g. an error returned before

14−// sessionAgent.Run ever executed). It avoids a breaking change to the

15−// Coordinator interface.

16−type runCompleteMarkerKey struct{}

17−

18−// runCompleteMarker records whether a terminal RunComplete has been

19−// published for a run. It is shared by pointer through the context so

20−// a publish deep in the call stack is observable by the dispatcher

21−// after the call returns.

22−type runCompleteMarker struct {

23− published atomic.Bool

24−}

25−

26−// WithRunCompleteMarker returns ctx carrying a fresh marker the

27−// coordinator can flag via [MarkRunCompletePublished] once it emits the

28−// run's terminal RunComplete. Callers read the result with

29−// [RunCompletePublished]. Attaching the marker is optional: code paths

30−// without one simply skip the dedup signal.

31−func WithRunCompleteMarker(ctx context.Context) context.Context {

32− return context.WithValue(ctx, runCompleteMarkerKey{}, &runCompleteMarker{})

33−}

34−

35−// MarkRunCompletePublished records that the authoritative terminal

36−// RunComplete has been published for the run carried by ctx. It is a

37−// no-op when no marker is present (e.g. the in-process/local Run path,

38−// which is not dispatched through backend.runAgent).

39−func MarkRunCompletePublished(ctx context.Context) {

40− if m, ok := ctx.Value(runCompleteMarkerKey{}).(*runCompleteMarker); ok {

41− m.published.Store(true)

42− }

43−}

44−

45−// RunCompletePublished reports whether [MarkRunCompletePublished] was

46−// called on ctx's marker. It returns false when no marker is present.

47−func RunCompletePublished(ctx context.Context) bool {

48− if m, ok := ctx.Value(runCompleteMarkerKey{}).(*runCompleteMarker); ok {

49− return m.published.Load()

50− }

51− return false

52−}

internal/agent/runid.godeleted−33

@@ -1,33 +0,0 @@

1−package agent

2−

3−import "context"

4−

5−// runIDContextKey is the unexported context key used to carry a

6−// caller-supplied RunID from the workspace HTTP boundary

7−// (backend.SendMessage) down into coordinator.Run without forcing a

8−// breaking change to the Coordinator.Run signature. The value is

9−// then copied onto SessionAgentCall.RunID by the coordinator so the

10−// agent's terminal RunComplete event can echo it back to the

11−// originating caller.

12−type runIDContextKey struct{}

13−

14−// WithRunID returns ctx tagged with a per-request RunID. It is the

15−// boundary helper for callers that need their SendMessage→Run

16−// terminal event to be uniquely correlatable (e.g. `crush run`

17−// against a session that may be busy). Empty runIDs are stored

18−// as-is; downstream code treats an empty RunID as "caller did not

19−// supply one" and falls back to SessionID-only correlation.

20−func WithRunID(ctx context.Context, runID string) context.Context {

21− return context.WithValue(ctx, runIDContextKey{}, runID)

22−}

23−

24−// RunIDFromContext returns the RunID set by [WithRunID], or "" if

25−// none was set or the value is not a string. Exported because the

26−// coordinator and tests in other packages need to read it; safe to

27−// call on any context.

28−func RunIDFromContext(ctx context.Context) string {

29− if v, ok := ctx.Value(runIDContextKey{}).(string); ok {

30− return v

31− }

32− return ""

33−}

internal/agent/templates/agent_tool.md+15−1

@@ -1 +1,15 @@

1−Launch a new agent that has access to the following tools: glob, grep, ls, view. When you are searching for a keyword or file and are not confident that you will find the right match on the first try, use the agent tool to perform the search for you.

1+Launch a new agent that has access to the following tools: GlobTool, GrepTool, LS, View. When you are searching for a keyword or file and are not confident that you will find the right match on the first try, use the Agent tool to perform the search for you.

3+<usage>

4+- If you are searching for a keyword like "config" or "logger", or for questions like "which file does X?", the Agent tool is strongly recommended

5+- If you want to read a specific file path, use the View or GlobTool tool instead of the Agent tool, to find the match more quickly

6+- If you are searching for a specific class definition like "class Foo", use the GlobTool tool instead, to find the match more quickly

7+</usage>

9+<usage_notes>

10+1. Launch multiple agents concurrently whenever possible, to maximize performance; to do that, use a single message with multiple tool uses

11+2. When the agent is done, it will return a single message back to you. The result returned by the agent is not visible to the user. To show the user the result, you should send a text message back to the user with a concise summary of the result.

12+3. Each agent invocation is stateless. You will not be able to send additional messages to the agent, nor will the agent be able to communicate with you outside of its final report. Therefore, your prompt should contain a highly detailed task description for the agent to perform autonomously and you should specify exactly what information the agent should return back to you in its final and only message to you.

13+4. The agent's outputs should generally be trusted

14+5. IMPORTANT: The agent can not use Bash, Replace, Edit, so can not modify files. If you want to use these tools, use them directly instead of going through the agent.

15+</usage_notes>

internal/agent/templates/agentic_fetch.md+64−1

@@ -1 +1,64 @@

1−Fetch a URL or search the web using an AI sub-agent that can extract, summarize, and answer questions. Slower and costlier than fetch; use fetch for raw content or API responses.

1+Fetches content from a URL or searches the web, then processes it using an AI model to extract information or answer questions.

3+<when_to_use>

4+Use this tool when you need to:

5+- Search the web for information (omit the url parameter)

6+- Extract specific information from a webpage (provide a url)

7+- Answer questions about web content

8+- Summarize or analyze web pages

9+- Research topics by searching and following links

10+

11+DO NOT use this tool when:

12+- You just need raw content without analysis (use fetch instead - faster and cheaper)

13+- You want direct access to API responses or JSON (use fetch instead)

14+- You don't need the content processed or interpreted (use fetch instead)

15+</when_to_use>

16+

17+<usage>

18+- Provide a prompt describing what information you want to find or extract (required)

19+- Optionally provide a URL to fetch and analyze specific content

20+- If no URL is provided, the agent will search the web to find relevant information

21+- The tool spawns a sub-agent with web_search, web_fetch, and analysis tools

22+- Returns the agent's response about the content

23+</usage>

24+

25+<parameters>

26+- prompt: What information you want to find or extract (required)

27+- url: The URL to fetch content from (optional - if not provided, agent will search the web)

28+</parameters>

29+

30+<usage_notes>

31+- IMPORTANT: If an MCP-provided web fetch tool is available, prefer using that tool instead of this one, as it may have fewer restrictions. All MCP-provided tools start with "mcp_".

32+- When using URL mode: The URL must be a fully-formed valid URL. HTTP URLs will be automatically upgraded to HTTPS.

33+- When searching: Just provide the prompt describing what you want to find - the agent will search and fetch relevant pages.

34+- The sub-agent can perform multiple searches and fetch multiple pages to gather comprehensive information.

35+- This tool is read-only and does not modify any files.

36+- Results will be summarized if the content is very large.

37+- This tool uses AI processing and costs more tokens than the simple fetch tool.

38+</usage_notes>

39+

40+<limitations>

41+- Max response size: 5MB per page

42+- Only supports HTTP and HTTPS protocols

43+- Cannot handle authentication or cookies

44+- Some websites may block automated requests

45+- Uses additional tokens for AI processing

46+- Search results depend on DuckDuckGo availability

47+</limitations>

48+

49+<tips>

50+- Be specific in your prompt about what information you want

51+- For research tasks, omit the URL and let the agent search and follow relevant links

52+- For complex pages, ask the agent to focus on specific sections

53+- The agent has access to web_search, web_fetch, grep, and view tools

54+- If you just need raw content, use the fetch tool instead to save tokens

55+</tips>

56+

57+<examples>

58+Search for information:

59+- prompt: "What are the main new features in the latest Python release?"

60+

61+Fetch and analyze a URL:

62+- url: "https://docs.python.org/3/whatsnew/3.12.html"

63+- prompt: "Summarize the key changes in Python 3.12"

64+</examples>

internal/agent/templates/coder.md.tpl+10−38

@@ -3,12 +3,12 @@ You are Crush, a powerful AI Assistant that runs in the CLI.

3<critical_rules>

4These rules override everything else. Follow them strictly:

6−1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.

6+1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.

72. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.

83. **TEST AFTER CHANGES**: Run tests immediately after each modification.

94. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.

105. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.

11−6. **NEVER COMMIT**: Unless user explicitly says "commit". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.

11+6. **NEVER COMMIT**: Unless user explicitly says "commit".

127. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.

138. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.

149. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.

@@ -16,13 +16,10 @@ These rules override everything else. Follow them strictly:

1611. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.

1712. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.

1813. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.

19−14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.

20−15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.

21</critical_rules>

19</critical_rules>

23<communication_style>

21<communication_style>

24Keep responses minimal:

22Keep responses minimal:

25−- ALWAYS think and respond in the same spoken language the prompt was written in.

26- Under 4 lines of text (tool use doesn't count)

23- Under 4 lines of text (tool use doesn't count)

27- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.

24- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.

28- No preamble ("Here's...", "I'll...")

25- No preamble ("Here's...", "I'll...")

@@ -141,10 +138,10 @@ Examples of autonomous decisions:

141

138

142Never use `apply_patch` or similar - those tools don't exist.

139Never use `apply_patch` or similar - those tools don't exist.

143

140

144−Critical: ALWAYS read the relevant context of files before editing them in this conversation.

141+Critical: ALWAYS read files before editing them in this conversation.

145

142

146When using edit tools:

143When using edit tools:

147−1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)

144+1. Read the file first - note the EXACT indentation (spaces vs tabs, count)

1482. Copy the exact text including ALL whitespace, newlines, and indentation

1452. Copy the exact text including ALL whitespace, newlines, and indentation

1493. Include 3-5 lines of context before and after the target

1463. Include 3-5 lines of context before and after the target

1504. Verify your old_string would appear exactly once in the file

1474. Verify your old_string would appear exactly once in the file

@@ -252,7 +249,7 @@ Common errors:

252<memory_instructions>

249<memory_instructions>

253Memory files store commands, preferences, and codebase info. Update them when you discover:

250Memory files store commands, preferences, and codebase info. Update them when you discover:

254- Build/test/lint commands

251- Build/test/lint commands

255−- Code style preferences

252+- Code style preferences

256- Important codebase patterns

253- Important codebase patterns

257- Useful project information

254- Useful project information

258</memory_instructions>

255</memory_instructions>

@@ -378,43 +375,18 @@ Diagnostics (lint/typecheck) included in tool output.

378{{.AvailSkillXML}}

375{{.AvailSkillXML}}

379

376

380<skills_usage>

377<skills_usage>

381−The `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.

378+When a user task matches a skill's description, read the skill's SKILL.md file to get full instructions.

382−

379+Skills are activated by reading their location path. Follow the skill's instructions to complete the task.

383−MANDATORY activation flow:

384−1. Scan `<available_skills>` against the current user task.

385−2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.

386−3. Read the entire SKILL.md and follow its instructions.

387−4. Only then execute the task, using the skill's prescribed commands/tools.

388−

389−Do NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.

390−

391−Builtin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The "crush://" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.

392−

393−Do not use MCP tools (including read_mcp_resource) to load skills.

394−If a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).

380+If a skill mentions scripts, references, or assets, they are placed in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).

395</skills_usage>

381</skills_usage>

396{{end}}

382{{end}}

397

383

398{{if .ContextFiles}}

384{{if .ContextFiles}}

399−# Project-Specific Context

385+<memory>

400−Make sure to follow the instructions in the context below.

401−<project_context>

402{{range .ContextFiles}}

386{{range .ContextFiles}}

403<file path="{{.Path}}">

387<file path="{{.Path}}">

404{{.Content}}

388{{.Content}}

405</file>

389</file>

406{{end}}

390{{end}}

407−</project_context>

391+</memory>

408−{{end}}

409−{{if .GlobalContextFiles}}

410−

411−# User context

412−The following is personal content added by the user that they'd like you to follow no matter what project you're working in.

413−<user_preferences>

414−{{range .GlobalContextFiles}}

415−<file path="{{.Path}}">

416−{{.Content}}

417−</file>

418−{{end}}

419−</user_preferences>

420{{end}}

392{{end}}

internal/agent/templates/initialize.md.tpl+3−5

@@ -2,7 +2,7 @@ Analyze this codebase and create/update **{{.Config.Options.InitializeAs}}** to

3**First**: Check if directory is empty or contains only config files. If so, stop and say "Directory appears empty or only contains config. Add source code first, then run this command to generate {{.Config.Options.InitializeAs}}."

5−**Goal**: Document what an agent needs to know to work in this codebase - commands, patterns, conventions, gotchas, overall architecture, how components fit together

5+**Goal**: Document what an agent needs to know to work in this codebase - commands, patterns, conventions, gotchas.

7**Discovery process**:

@@ -10,20 +10,18 @@ Analyze this codebase and create/update **{{.Config.Options.InitializeAs}}** to

102. Look for existing rule files (`.cursor/rules/*.md`, `.cursorrules`, `.github/copilot-instructions.md`, `claude.md`, `agents.md`) - only read if they exist

113. Identify project type from config files and directory structure

124. Find build/test/lint commands from config files, scripts, Makefiles, or CI configs

13−5. Read representative source files to understand code patterns, architecture, control/data flow

13+5. Read representative source files to understand code patterns

146. If {{.Config.Options.InitializeAs}} exists, read and improve it

16**Content to include**:

18- Essential commands (build, test, run, deploy, etc.) - whatever is relevant for this project

19−- Code organization and structure, application architecture and control/data flow

19+- Code organization and structure

20- Naming conventions and style patterns

21- Testing approach and patterns

22- Important gotchas or non-obvious patterns

23- Any project-specific context from existing rule files

25−**Note:** LLM agents learn and adapt to their context as they obtain it, so mentioning obvious details they would immediately pick up from reading a file or two is actively detrimental. Keep the principles of progressive disclosure in mind and focus primarily on non-obvious knowledge that saves the agent from trial-and-error discovery: gotchas, implicit conventions, commands with surprising flags, and context that isn't self-evident from the code in a single file.

26−

27**Format**: Clear markdown sections. Use your judgment on structure based on what you find. Aim for completeness over brevity - include everything an agent would need to know.

25**Format**: Clear markdown sections. Use your judgment on structure based on what you find. Aim for completeness over brevity - include everything an agent would need to know.

29**Critical**: Only document what you actually observe. Never invent commands, patterns, or conventions. If you can't find something, don't include it.

27**Critical**: Only document what you actually observe. Never invent commands, patterns, or conventions. If you can't find something, don't include it.

internal/agent/templates/title.md+7−8

@@ -1,11 +1,10 @@

1−You will generate a short title based on the first message a user begins a conversation with.

1+you will generate a short title based on the first message a user begins a conversation with

3<rules>

4−- Keep the title in the same language that the user wrote their message in.

5−- Ensure it is not more than 50 characters long.

4+- ensure it is not more than 50 characters long

6−- The title should be a summary of the user's message.

5+- the title should be a summary of the user's message

7−- It should be one line long.

6+- it should be one line long

8−- Do not use quotes or colons.

7+- do not use quotes or colons

9−- The entire text you return will be used as the title.

8+- the entire text you return will be used as the title

10−- Never return anything that is more than one sentence (one line) long.

9+- never return anything that is more than one sentence (one line) long

11</rules>

10</rules>

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/bash_tool.yamlnew file+213

@@ -0,0 +1,213 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 802

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse bash to create a file named test.txt with content ''hello bash''. do not print its timestamp\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01Lij7hAdrqXtCCiddzXdoGK","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":152,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":2,"service_tier":"standard"}}}

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Bash"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" File"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Creation"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Comman"} }

47+

48+ event: content_block_delta

49+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d"} }

50+

51+ event: content_block_stop

52+ data: {"type":"content_block_stop","index":0 }

53+

54+ event: message_delta

55+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":152,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":8} }

56+

57+ event: message_stop

58+ data: {"type":"message_stop" }

59+

60+ headers:

61+ Content-Type:

62+ - text/event-stream; charset=utf-8

63+ status: 200 OK

64+ code: 200

65+ duration: 811.339375ms

66+- id: 1

67+ request:

68+ proto: HTTP/1.1

69+ proto_major: 1

70+ proto_minor: 1

71+ content_length: 52393

72+ host: ""

73+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use bash to create a file named test.txt with content 'hello bash'. do not print its timestamp\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/bash_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

74+ headers:

75+ Accept:

76+ - application/json

77+ Content-Type:

78+ - application/json

79+ User-Agent:

80+ - Anthropic/Go 1.14.0

81+ url: https://api.anthropic.com/v1/messages

82+ method: POST

83+ response:

84+ proto: HTTP/2.0

85+ proto_major: 2

86+ proto_minor: 0

87+ content_length: -1

88+ body: |+

89+ event: message_start

90+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_018EZSLb8cAiYosSKFfVfGjP","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4543,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4543,"ephemeral_1h_input_tokens":0},"output_tokens":8,"service_tier":"standard"}} }

91+

92+ event: content_block_start

93+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_013uhtHrcioLU9dEwS6BpjWD","name":"bash","input":{}} }

94+

95+ event: ping

96+ data: {"type": "ping"}

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"command\""} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":": \"echo 'h"}}

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ell"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"o "} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"bash"} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"' >"} }

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":" te"} }

121+

122+ event: content_block_delta

123+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"st.txt\""} }

124+

125+ event: content_block_delta

126+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", "} }

127+

128+ event: content_block_delta

129+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"descri"} }

130+

131+ event: content_block_delta

132+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ption\""} }

133+

134+ event: content_block_delta

135+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":": \"Crea"} }

136+

137+ event: content_block_delta

138+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"te "} }

139+

140+ event: content_block_delta

141+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"test.txt w"} }

142+

143+ event: content_block_delta

144+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ith"} }

145+

146+ event: content_block_delta

147+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":" content\"}"} }

148+

149+ event: content_block_stop

150+ data: {"type":"content_block_stop","index":0 }

151+

152+ event: message_delta

153+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4543,"cache_read_input_tokens":9490,"output_tokens":82} }

154+

155+ event: message_stop

156+ data: {"type":"message_stop" }

157+

158+ headers:

159+ Content-Type:

160+ - text/event-stream; charset=utf-8

161+ status: 200 OK

162+ code: 200

163+ duration: 1.380791709s

164+- id: 2

165+ request:

166+ proto: HTTP/1.1

167+ proto_major: 1

168+ proto_minor: 1

169+ content_length: 52744

170+ host: ""

171+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use bash to create a file named test.txt with content 'hello bash'. do not print its timestamp\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_013uhtHrcioLU9dEwS6BpjWD\",\"input\":{\"command\":\"echo 'hello bash' \\u003e test.txt\",\"description\":\"Create test.txt with content\"},\"name\":\"bash\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_013uhtHrcioLU9dEwS6BpjWD\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"no output\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/bash_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

172+ headers:

173+ Accept:

174+ - application/json

175+ Content-Type:

176+ - application/json

177+ User-Agent:

178+ - Anthropic/Go 1.14.0

179+ url: https://api.anthropic.com/v1/messages

180+ method: POST

181+ response:

182+ proto: HTTP/2.0

183+ proto_major: 2

184+ proto_minor: 0

185+ content_length: -1

186+ body: |+

187+ event: message_start

188+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_019z3iNk2iq5DAfodREhCCzC","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":94,"cache_read_input_tokens":14033,"cache_creation":{"ephemeral_5m_input_tokens":94,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

189+

190+ event: content_block_start

191+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

192+

193+ event: ping

194+ data: {"type": "ping"}

195+

196+ event: content_block_delta

197+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Done"} }

198+

199+ event: content_block_stop

200+ data: {"type":"content_block_stop","index":0 }

201+

202+ event: message_delta

203+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":94,"cache_read_input_tokens":14033,"output_tokens":4} }

204+

205+ event: message_stop

206+ data: {"type":"message_stop" }

207+

208+ headers:

209+ Content-Type:

210+ - text/event-stream; charset=utf-8

211+ status: 200 OK

212+ code: 200

213+ duration: 2.965123583s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/download_tool.yamlnew file+256

@@ -0,0 +1,256 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 823

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\ndownload the file from https://example-files.online-convert.com/document/txt/example.txt and save it as example.txt\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_011YAinXEYrTqbpsq4sBCjxA","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":160,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Downloa"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d File"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" from Example"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" URL"} }

47+

48+ event: content_block_stop

49+ data: {"type":"content_block_stop","index":0 }

50+

51+ event: message_delta

52+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":160,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":8} }

53+

54+ event: message_stop

55+ data: {"type":"message_stop" }

56+

57+ headers:

58+ Content-Type:

59+ - text/event-stream; charset=utf-8

60+ status: 200 OK

61+ code: 200

62+ duration: 640.901ms

63+- id: 1

64+ request:

65+ proto: HTTP/1.1

66+ proto_major: 1

67+ proto_minor: 1

68+ content_length: 52418

69+ host: ""

70+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"download the file from https://example-files.online-convert.com/document/txt/example.txt and save it as example.txt\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/download_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

71+ headers:

72+ Accept:

73+ - application/json

74+ Content-Type:

75+ - application/json

76+ User-Agent:

77+ - Anthropic/Go 1.14.0

78+ url: https://api.anthropic.com/v1/messages

79+ method: POST

80+ response:

81+ proto: HTTP/2.0

82+ proto_major: 2

83+ proto_minor: 0

84+ content_length: -1

85+ body: |+

86+ event: message_start

87+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01Ubo3m8oVn1kppZgkSCrixN","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4551,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4551,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

88+

89+ event: content_block_start

90+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01Nrp1Tn9bF743SfWymTxUkC","name":"download","input":{}} }

91+

92+ event: ping

93+ data: {"type": "ping"}

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\""}}

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ur"}}

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"l\": \"h"} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ttps://ex"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ample-file"} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"s.online-co"} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"nvert.com/do"} }

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"cumen"} }

121+

122+ event: content_block_delta

123+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"t/txt/exa"} }

124+

125+ event: content_block_delta

126+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"mple.txt\""} }

127+

128+ event: content_block_delta

129+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \"f"} }

130+

131+ event: content_block_delta

132+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ile_pa"} }

133+

134+ event: content_block_delta

135+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"th\": \""}}

136+

137+ event: content_block_delta

138+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"example"} }

139+

140+ event: content_block_delta

141+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":".txt\"}"} }

142+

143+ event: content_block_stop

144+ data: {"type":"content_block_stop","index":0 }

145+

146+ event: message_delta

147+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4551,"cache_read_input_tokens":9490,"output_tokens":90} }

148+

149+ event: message_stop

150+ data: {"type":"message_stop" }

151+

152+ headers:

153+ Content-Type:

154+ - text/event-stream; charset=utf-8

155+ status: 200 OK

156+ code: 200

157+ duration: 1.208203792s

158+- id: 2

159+ request:

160+ proto: HTTP/1.1

161+ proto_major: 1

162+ proto_minor: 1

163+ content_length: 0

164+ host: example-files.online-convert.com

165+ headers:

166+ User-Agent:

167+ - crush/1.0

168+ url: https://example-files.online-convert.com/document/txt/example.txt

169+ method: GET

170+ response:

171+ proto: HTTP/2.0

172+ proto_major: 2

173+ proto_minor: 0

174+ content_length: -1

175+ uncompressed: true

176+ body: |-

177+ TXT test file

178+ Purpose: Provide example of this file type

179+ Document file type: TXT

180+ Version: 1.0

181+ Remark:

182+

183+ Example content:

184+ The names "John Doe" for males, "Jane Doe" or "Jane Roe" for females, or "Jonnie Doe" and "Janie Doe" for children, or just "Doe" non-gender-specifically are used as placeholder names for a party whose true identity is unknown or must be withheld in a legal action, case, or discussion. The names are also used to refer to acorpse or hospital patient whose identity is unknown. This practice is widely used in the United States and Canada, but is rarely used in other English-speaking countries including the United Kingdom itself, from where the use of "John Doe" in a legal context originates. The names Joe Bloggs or John Smith are used in the UK instead, as well as in Australia and New Zealand.

185+

186+ John Doe is sometimes used to refer to a typical male in other contexts as well, in a similar manner to John Q. Public, known in Great Britain as Joe Public, John Smith or Joe Bloggs. For example, the first name listed on a form is often John Doe, along with a fictional address or other fictional information to provide an example of how to fill in the form. The name is also used frequently in popular culture, for example in the Frank Capra film Meet John Doe. John Doe was also the name of a 2002 American television series.

187+

188+ Similarly, a child or baby whose identity is unknown may be referred to as Baby Doe. A notorious murder case in Kansas City, Missouri, referred to the baby victim as Precious Doe. Other unidentified female murder victims are Cali Doe and Princess Doe. Additional persons may be called James Doe, Judy Doe, etc. However, to avoid possible confusion, if two anonymous or unknown parties are cited in a specific case or action, the surnames Doe and Roe may be used simultaneously; for example, "John Doe v. Jane Roe". If several anonymous parties are referenced, they may simply be labelled John Doe #1, John Doe #2, etc. (the U.S. Operation Delego cited 21 (numbered) "John Doe"s) or labelled with other variants of Doe / Roe / Poe / etc. Other early alternatives such as John Stiles and Richard Miles are now rarely used, and Mary Major has been used in some American federal cases.

189+

190+

191+

192+ File created by https://www.online-convert.com

193+ More example files: https://www.online-convert.com/file-type

194+ Text of Example content: Wikipedia (https://en.wikipedia.org/wiki/John_Doe)

195+ License: Attribution-ShareAlike 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)

196+

197+ Feel free to use and share the file according to the license above.

198+ headers:

199+ Content-Type:

200+ - text/plain; charset=UTF-8

201+ status: 200 OK

202+ code: 200

203+ duration: 194.328542ms

204+- id: 3

205+ request:

206+ proto: HTTP/1.1

207+ proto_major: 1

208+ proto_minor: 1

209+ content_length: 52864

210+ host: ""

211+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"download the file from https://example-files.online-convert.com/document/txt/example.txt and save it as example.txt\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01Nrp1Tn9bF743SfWymTxUkC\",\"input\":{\"file_path\":\"example.txt\",\"url\":\"https://example-files.online-convert.com/document/txt/example.txt\"},\"name\":\"download\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01Nrp1Tn9bF743SfWymTxUkC\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"Successfully downloaded 2574 bytes to example.txt (Content-Type: text/plain; charset=UTF-8)\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/download_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

212+ headers:

213+ Accept:

214+ - application/json

215+ Content-Type:

216+ - application/json

217+ User-Agent:

218+ - Anthropic/Go 1.14.0

219+ url: https://api.anthropic.com/v1/messages

220+ method: POST

221+ response:

222+ proto: HTTP/2.0

223+ proto_major: 2

224+ proto_minor: 0

225+ content_length: -1

226+ body: |+

227+ event: message_start

228+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01DqZpyfqZVVermCgv37vJd3","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":4,"cache_creation_input_tokens":126,"cache_read_input_tokens":14041,"cache_creation":{"ephemeral_5m_input_tokens":126,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

229+

230+ event: content_block_start

231+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

232+

233+ event: ping

234+ data: {"type": "ping"}

235+

236+ event: content_block_delta

237+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Downloade"} }

238+

239+ event: content_block_delta

240+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d"} }

241+

242+ event: content_block_stop

243+ data: {"type":"content_block_stop","index":0 }

244+

245+ event: message_delta

246+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":4,"cache_creation_input_tokens":126,"cache_read_input_tokens":14041,"output_tokens":4} }

247+

248+ event: message_stop

249+ data: {"type":"message_stop" }

250+

251+ headers:

252+ Content-Type:

253+ - text/event-stream; charset=utf-8

254+ status: 200 OK

255+ code: 200

256+ duration: 2.392586375s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/fetch_tool.yamlnew file+302

@@ -0,0 +1,302 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 844

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nfetch the content from https://example-files.online-convert.com/website/html/example.html and tell me if it contains the word ''John Doe''\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01QHs42eeZPJbxjrFBiaSQu3","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":167,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Website"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Content"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Check"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" for"} }

47+

48+ event: content_block_delta

49+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" 'John Doe"} }

50+

51+ event: content_block_delta

52+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"'"} }

53+

54+ event: content_block_stop

55+ data: {"type":"content_block_stop","index":0 }

56+

57+ event: message_delta

58+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":167,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":13} }

59+

60+ event: message_stop

61+ data: {"type":"message_stop" }

62+

63+ headers:

64+ Content-Type:

65+ - text/event-stream; charset=utf-8

66+ status: 200 OK

67+ code: 200

68+ duration: 601.885125ms

69+- id: 1

70+ request:

71+ proto: HTTP/1.1

72+ proto_major: 1

73+ proto_minor: 1

74+ content_length: 52436

75+ host: ""

76+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"fetch the content from https://example-files.online-convert.com/website/html/example.html and tell me if it contains the word 'John Doe'\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/fetch_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

77+ headers:

78+ Accept:

79+ - application/json

80+ Content-Type:

81+ - application/json

82+ User-Agent:

83+ - Anthropic/Go 1.14.0

84+ url: https://api.anthropic.com/v1/messages

85+ method: POST

86+ response:

87+ proto: HTTP/2.0

88+ proto_major: 2

89+ proto_minor: 0

90+ content_length: -1

91+ body: |+

92+ event: message_start

93+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_012kt7M1ooLoNTTfqAY4XYHC","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4558,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4558,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

94+

95+ event: content_block_start

96+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01FZaJdfZV7w7UVgkJHHPJnK","name":"fetch","input":{}} }

97+

98+ event: ping

99+ data: {"type": "ping"}

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"u"} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"rl\": \"https:"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"//example"} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"-files.o"} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"nlin"}}

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"e-co"} }

121+

122+ event: content_block_delta

123+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"nvert."} }

124+

125+ event: content_block_delta

126+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"com/w"}}

127+

128+ event: content_block_delta

129+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ebsite"} }

130+

131+ event: content_block_delta

132+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"/html/"} }

133+

134+ event: content_block_delta

135+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ex"} }

136+

137+ event: content_block_delta

138+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ample"} }

139+

140+ event: content_block_delta

141+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":".html\""} }

142+

143+ event: content_block_delta

144+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \"format\":"} }

145+

146+ event: content_block_delta

147+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":" \"text\"}"} }

148+

149+ event: content_block_stop

150+ data: {"type":"content_block_stop","index":0 }

151+

152+ event: message_delta

153+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4558,"cache_read_input_tokens":9490,"output_tokens":86} }

154+

155+ event: message_stop

156+ data: {"type":"message_stop" }

157+

158+ headers:

159+ Content-Type:

160+ - text/event-stream; charset=utf-8

161+ status: 200 OK

162+ code: 200

163+ duration: 2.293590416s

164+- id: 2

165+ request:

166+ proto: HTTP/1.1

167+ proto_major: 1

168+ proto_minor: 1

169+ content_length: 0

170+ host: example-files.online-convert.com

171+ headers:

172+ User-Agent:

173+ - crush/1.0

174+ url: https://example-files.online-convert.com/website/html/example.html

175+ method: GET

176+ response:

177+ proto: HTTP/2.0

178+ proto_major: 2

179+ proto_minor: 0

180+ content_length: -1

181+ uncompressed: true

182+ body: |-

183+ <!DOCTYPE html>

184+

185+ <html>

186+ <body>

187+

188+ <h1>HTML test file</h1>

189+

190+ Purpose: Provide example of this file type

191+ Document file type: HTML

192+ Version: 1.0

193+ Remark:

194+

195+ <img align="right" style="height:5.0551cm;width:7.62cm;" src="data:image/*;base64,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

196+ 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

197+ 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

198+ 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

199+ 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

200+ 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

201+ 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

202+ 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

203+ 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

204+ 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

205+ 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

206+ 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"/>

207+

208+ Example content:

209+ The names "John Doe" for males, "Jane Doe" or "Jane Roe" for females, or "Jonnie Doe" and "Janie Doe" for children, or just "Doe" non-gender-specifically are used as placeholder names for a party whose true identity is unknown or must be withheld in a legal action, case, or discussion. The names are also used to refer to acorpse or hospital patient whose identity is unknown. This practice is widely used in the United States and Canada, but is rarely used in other English-speaking countries including the United Kingdom itself, from where the use of "John Doe" in a legal context originates. The names Joe Bloggs or John Smith are used in the UK instead, as well as in Australia and New Zealand.

210+

211+ John Doe is sometimes used to refer to a typical male in other contexts as well, in a similar manner to John Q. Public, known in Great Britain as Joe Public, John Smith or Joe Bloggs. For example, the first name listed on a form is often John Doe, along with a fictional address or other fictional information to provide an example of how to fill in the form. The name is also used frequently in popular culture, for example in the Frank Capra film Meet John Doe. John Doe was also the name of a 2002 American television series.

212+

213+ Similarly, a child or baby whose identity is unknown may be referred to as Baby Doe. A notorious murder case in Kansas City, Missouri, referred to the baby victim as Precious Doe. Other unidentified female murder victims are Cali Doe and Princess Doe. Additional persons may be called James Doe, Judy Doe, etc. However, to avoid possible confusion, if two anonymous or unknown parties are cited in a specific case or action, the surnames Doe and Roe may be used simultaneously; for example, "John Doe v. Jane Roe". If several anonymous parties are referenced, they may simply be labelled John Doe #1, John Doe #2, etc. (the U.S. Operation Delego cited 21 (numbered) "John Doe"s) or labelled with other variants of Doe / Roe / Poe / etc. Other early alternatives such as John Stiles and Richard Miles are now rarely used, and Mary Major has been used in some American federal cases.

214+

215+ <img style="height:1.6881cm;width:1.6881cm;" align="left" src="data:image/*;base64,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

216+ B1/bMVO7nYqfTdlIhfMUvxUd51lhy67Tyq1UqJI9Jew1tnpUt0B6+/rwB89+fbtcyp1Q+gRk7z/Gy9tfzMHyrrfpKG8eqVDetsaC8o5Juz6zqvKuz6GyvKtkeXn3fbDzOauVdwpGynvzd+wv64P1QRgMBZhAGAwFmEAYDAWYQBgMBZhAGAwFmEAYDAWYQBgMBZhAGAwFmEAYDAWYQBgMBZhAGAwFmEAYDAWYQBgMBf4/fQGdtWECNsgAAAAASUVORK5CYII="/>

217+

218+ File created by <a href="https://www.online-convert.com" target="_blank">https://www.online-convert.com</a>

219+ More example files: <a href="https://www.online-convert.com/file-type" target="_blank">https://www.online-convert.com/file-type</a>

220+ Text of "Example content": <a href="https://en.wikipedia.org/wiki/John_Doe" target="_blank">Wikipedia</a>

221+ License: <a href="https://creativecommons.org/licenses/by-sa/3.0/" target="_blank">Attribution-ShareAlike 3.0 Unported</a>

222+

223+ <img style="height:0.6949cm;width:1.9741cm;" align="right" src="data:image/*;base64,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"/>

224+

225+ Feel free to use and share the file according to license above.

226+

227+ </body>

228+ </html>

229+ headers:

230+ Content-Type:

231+ - text/html; charset=UTF-8

232+ status: 200 OK

233+ code: 200

234+ duration: 50.850459ms

235+- id: 3

236+ request:

237+ proto: HTTP/1.1

238+ proto_major: 1

239+ proto_minor: 1

240+ content_length: 55284

241+ host: ""

242+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"fetch the content from https://example-files.online-convert.com/website/html/example.html and tell me if it contains the word 'John Doe'\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01FZaJdfZV7w7UVgkJHHPJnK\",\"input\":{\"format\":\"text\",\"url\":\"https://example-files.online-convert.com/website/html/example.html\"},\"name\":\"fetch\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01FZaJdfZV7w7UVgkJHHPJnK\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"HTML test file Purpose: Provide example of this file type Document file type: HTML Version: 1.0 Remark: Example content: The names \\\"John Doe\\\" for males, \\\"Jane Doe\\\" or \\\"Jane Roe\\\" for females, or \\\"Jonnie Doe\\\" and \\\"Janie Doe\\\" for children, or just \\\"Doe\\\" non-gender-specifically are used as placeholder names for a party whose true identity is unknown or must be withheld in a legal action, case, or discussion. The names are also used to refer to acorpse or hospital patient whose identity is unknown. This practice is widely used in the United States and Canada, but is rarely used in other English-speaking countries including the United Kingdom itself, from where the use of \\\"John Doe\\\" in a legal context originates. The names Joe Bloggs or John Smith are used in the UK instead, as well as in Australia and New Zealand. John Doe is sometimes used to refer to a typical male in other contexts as well, in a similar manner to John Q. Public, known in Great Britain as Joe Public, John Smith or Joe Bloggs. For example, the first name listed on a form is often John Doe, along with a fictional address or other fictional information to provide an example of how to fill in the form. The name is also used frequently in popular culture, for example in the Frank Capra film Meet John Doe. John Doe was also the name of a 2002 American television series. Similarly, a child or baby whose identity is unknown may be referred to as Baby Doe. A notorious murder case in Kansas City, Missouri, referred to the baby victim as Precious Doe. Other unidentified female murder victims are Cali Doe and Princess Doe. Additional persons may be called James Doe, Judy Doe, etc. However, to avoid possible confusion, if two anonymous or unknown parties are cited in a specific case or action, the surnames Doe and Roe may be used simultaneously; for example, \\\"John Doe v. Jane Roe\\\". If several anonymous parties are referenced, they may simply be labelled John Doe #1, John Doe #2, etc. (the U.S. Operation Delego cited 21 (numbered) \\\"John Doe\\\"s) or labelled with other variants of Doe / Roe / Poe / etc. Other early alternatives such as John Stiles and Richard Miles are now rarely used, and Mary Major has been used in some American federal cases. File created by https://www.online-convert.com More example files: https://www.online-convert.com/file-type Text of \\\"Example content\\\": Wikipedia License: Attribution-ShareAlike 3.0 Unported Feel free to use and share the file according to license above.\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/fetch_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

243+ headers:

244+ Accept:

245+ - application/json

246+ Content-Type:

247+ - application/json

248+ User-Agent:

249+ - Anthropic/Go 1.14.0

250+ url: https://api.anthropic.com/v1/messages

251+ method: POST

252+ response:

253+ proto: HTTP/2.0

254+ proto_major: 2

255+ proto_minor: 0

256+ content_length: -1

257+ body: |+

258+ event: message_start

259+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01D9K2wp4oAovQ8hreUHQB5H","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":4,"cache_creation_input_tokens":710,"cache_read_input_tokens":14048,"cache_creation":{"ephemeral_5m_input_tokens":710,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

260+

261+ event: content_block_start

262+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

263+

264+ event: ping

265+ data: {"type": "ping"}

266+

267+ event: content_block_delta

268+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Yes"} }

269+

270+ event: content_block_delta

271+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":", the content"} }

272+

273+ event: content_block_delta

274+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" contains \""} }

275+

276+ event: content_block_delta

277+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"John Doe\" multiple"} }

278+

279+ event: content_block_delta

280+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" times throughout"} }

281+

282+ event: content_block_delta

283+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" the HTML "} }

284+

285+ event: content_block_delta

286+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"document."} }

287+

288+ event: content_block_stop

289+ data: {"type":"content_block_stop","index":0 }

290+

291+ event: message_delta

292+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":4,"cache_creation_input_tokens":710,"cache_read_input_tokens":14048,"output_tokens":20} }

293+

294+ event: message_stop

295+ data: {"type":"message_stop"}

296+

297+ headers:

298+ Content-Type:

299+ - text/event-stream; charset=utf-8

300+ status: 200 OK

301+ code: 200

302+ duration: 1.997422583s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/glob_tool.yamlnew file+180

@@ -0,0 +1,180 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 763

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse glob to find all .go files in the current directory\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_016WcJGfCcpgVqfScPMNdmBp","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":142,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Fin"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d Go"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Files"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" with"} }

47+

48+ event: content_block_delta

49+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Glob"} }

50+

51+ event: content_block_stop

52+ data: {"type":"content_block_stop","index":0 }

53+

54+ event: message_delta

55+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":142,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":9} }

56+

57+ event: message_stop

58+ data: {"type":"message_stop" }

59+

60+ headers:

61+ Content-Type:

62+ - text/event-stream; charset=utf-8

63+ status: 200 OK

64+ code: 200

65+ duration: 683.862917ms

66+- id: 1

67+ request:

68+ proto: HTTP/1.1

69+ proto_major: 1

70+ proto_minor: 1

71+ content_length: 52354

72+ host: ""

73+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use glob to find all .go files in the current directory\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/glob_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

74+ headers:

75+ Accept:

76+ - application/json

77+ Content-Type:

78+ - application/json

79+ User-Agent:

80+ - Anthropic/Go 1.14.0

81+ url: https://api.anthropic.com/v1/messages

82+ method: POST

83+ response:

84+ proto: HTTP/2.0

85+ proto_major: 2

86+ proto_minor: 0

87+ content_length: -1

88+ body: |+

89+ event: message_start

90+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01N81tb9PX4XZneahamhjL1e","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4533,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4533,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}}}

91+

92+ event: content_block_start

93+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_018v77nmaX7HkuieMesYWGn8","name":"glob","input":{}} }

94+

95+ event: ping

96+ data: {"type": "ping"}

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"pattern"} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\": "} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"*.go\"}"} }

109+

110+ event: content_block_stop

111+ data: {"type":"content_block_stop","index":0 }

112+

113+ event: message_delta

114+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4533,"cache_read_input_tokens":9490,"output_tokens":52}}

115+

116+ event: message_stop

117+ data: {"type":"message_stop" }

118+

119+ headers:

120+ Content-Type:

121+ - text/event-stream; charset=utf-8

122+ status: 200 OK

123+ code: 200

124+ duration: 2.771247709s

125+- id: 2

126+ request:

127+ proto: HTTP/1.1

128+ proto_major: 1

129+ proto_minor: 1

130+ content_length: 52687

131+ host: ""

132+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use glob to find all .go files in the current directory\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_018v77nmaX7HkuieMesYWGn8\",\"input\":{\"pattern\":\"*.go\"},\"name\":\"glob\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_018v77nmaX7HkuieMesYWGn8\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"/tmp/crush-test/TestCoderAgent/anthropic-sonnet/glob_tool/main.go\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/glob_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

133+ headers:

134+ Accept:

135+ - application/json

136+ Content-Type:

137+ - application/json

138+ User-Agent:

139+ - Anthropic/Go 1.14.0

140+ url: https://api.anthropic.com/v1/messages

141+ method: POST

142+ response:

143+ proto: HTTP/2.0

144+ proto_major: 2

145+ proto_minor: 0

146+ content_length: -1

147+ body: |+

148+ event: message_start

149+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01ShF6YZDVaziREi6YGbu26Q","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":86,"cache_read_input_tokens":14023,"cache_creation":{"ephemeral_5m_input_tokens":86,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

150+

151+ event: content_block_start

152+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

153+

154+ event: ping

155+ data: {"type": "ping"}

156+

157+ event: content_block_delta

158+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"`"} }

159+

160+ event: content_block_delta

161+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"main"}}

162+

163+ event: content_block_delta

164+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".go`"} }

165+

166+ event: content_block_stop

167+ data: {"type":"content_block_stop","index":0 }

168+

169+ event: message_delta

170+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":86,"cache_read_input_tokens":14023,"output_tokens":8} }

171+

172+ event: message_stop

173+ data: {"type":"message_stop" }

174+

175+ headers:

176+ Content-Type:

177+ - text/event-stream; charset=utf-8

178+ status: 200 OK

179+ code: 200

180+ duration: 1.889802958s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/grep_tool.yamlnew file+198

@@ -0,0 +1,198 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 761

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse grep to search for the word ''package'' in go files\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_0181LpjAHqcVep2LP8kMttkT","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":144,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Go"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Files"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Package"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Search with"} }

47+

48+ event: content_block_delta

49+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Grep"} }

50+

51+ event: content_block_stop

52+ data: {"type":"content_block_stop","index":0 }

53+

54+ event: message_delta

55+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":144,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":10} }

56+

57+ event: message_stop

58+ data: {"type":"message_stop" }

59+

60+ headers:

61+ Content-Type:

62+ - text/event-stream; charset=utf-8

63+ status: 200 OK

64+ code: 200

65+ duration: 647.254041ms

66+- id: 1

67+ request:

68+ proto: HTTP/1.1

69+ proto_major: 1

70+ proto_minor: 1

71+ content_length: 52352

72+ host: ""

73+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use grep to search for the word 'package' in go files\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/grep_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

74+ headers:

75+ Accept:

76+ - application/json

77+ Content-Type:

78+ - application/json

79+ User-Agent:

80+ - Anthropic/Go 1.14.0

81+ url: https://api.anthropic.com/v1/messages

82+ method: POST

83+ response:

84+ proto: HTTP/2.0

85+ proto_major: 2

86+ proto_minor: 0

87+ content_length: -1

88+ body: |+

89+ event: message_start

90+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01FvrRdvQTegpjqBeGJYt2h7","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4535,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4535,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

91+

92+ event: content_block_start

93+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01TkHn7cSXHd7iPVfdsTyjPQ","name":"grep","input":{}}}

94+

95+ event: ping

96+ data: {"type": "ping"}

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"pa"} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ttern\": "} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"package\""}}

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \"includ"} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"e\": \"*.go\"}"} }

115+

116+ event: content_block_stop

117+ data: {"type":"content_block_stop","index":0}

118+

119+ event: message_delta

120+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4535,"cache_read_input_tokens":9490,"output_tokens":69} }

121+

122+ event: message_stop

123+ data: {"type":"message_stop" }

124+

125+ headers:

126+ Content-Type:

127+ - text/event-stream; charset=utf-8

128+ status: 200 OK

129+ code: 200

130+ duration: 2.177929292s

131+- id: 2

132+ request:

133+ proto: HTTP/1.1

134+ proto_major: 1

135+ proto_minor: 1

136+ content_length: 52757

137+ host: ""

138+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use grep to search for the word 'package' in go files\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01TkHn7cSXHd7iPVfdsTyjPQ\",\"input\":{\"include\":\"*.go\",\"pattern\":\"package\"},\"name\":\"grep\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01TkHn7cSXHd7iPVfdsTyjPQ\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"Found 1 matches\\n/tmp/crush-test/TestCoderAgent/anthropic-sonnet/grep_tool/main.go:\\n Line 1, Char 1: package main\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/grep_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

139+ headers:

140+ Accept:

141+ - application/json

142+ Content-Type:

143+ - application/json

144+ User-Agent:

145+ - Anthropic/Go 1.14.0

146+ url: https://api.anthropic.com/v1/messages

147+ method: POST

148+ response:

149+ proto: HTTP/2.0

150+ proto_major: 2

151+ proto_minor: 0

152+ content_length: -1

153+ body: |+

154+ event: message_start

155+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01FDcw9tmjXhqMW72BPVHHAd","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":125,"cache_read_input_tokens":14025,"cache_creation":{"ephemeral_5m_input_tokens":125,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

156+

157+ event: content_block_start

158+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

159+

160+ event: ping

161+ data: {"type": "ping"}

162+

163+ event: content_block_delta

164+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Foun"} }

165+

166+ event: content_block_delta

167+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d 1 match:"} }

168+

169+ event: content_block_delta

170+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\n- main"} }

171+

172+ event: content_block_delta

173+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".go:1 "}}

174+

175+ event: content_block_delta

176+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"-"}}

177+

178+ event: content_block_delta

179+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" `"}}

180+

181+ event: content_block_delta

182+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"package main`"} }

183+

184+ event: content_block_stop

185+ data: {"type":"content_block_stop","index":0 }

186+

187+ event: message_delta

188+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":125,"cache_read_input_tokens":14025,"output_tokens":22} }

189+

190+ event: message_stop

191+ data: {"type":"message_stop" }

192+

193+ headers:

194+ Content-Type:

195+ - text/event-stream; charset=utf-8

196+ status: 200 OK

197+ code: 200

198+ duration: 2.618235084s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/ls_tool.yamlnew file+171

@@ -0,0 +1,171 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 757

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse ls to list the files in the current directory\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01MFyJCeDt5XVr7qQrdbfQ3Z","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":140,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"List"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Files"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" in Current Directory"} }

44+

45+ event: content_block_stop

46+ data: {"type":"content_block_stop","index":0 }

47+

48+ event: message_delta

49+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":140,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":8} }

50+

51+ event: message_stop

52+ data: {"type":"message_stop" }

53+

54+ headers:

55+ Content-Type:

56+ - text/event-stream; charset=utf-8

57+ status: 200 OK

58+ code: 200

59+ duration: 539.973208ms

60+- id: 1

61+ request:

62+ proto: HTTP/1.1

63+ proto_major: 1

64+ proto_minor: 1

65+ content_length: 52346

66+ host: ""

67+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use ls to list the files in the current directory\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/ls_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

68+ headers:

69+ Accept:

70+ - application/json

71+ Content-Type:

72+ - application/json

73+ User-Agent:

74+ - Anthropic/Go 1.14.0

75+ url: https://api.anthropic.com/v1/messages

76+ method: POST

77+ response:

78+ proto: HTTP/2.0

79+ proto_major: 2

80+ proto_minor: 0

81+ content_length: -1

82+ body: |+

83+ event: message_start

84+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_017yVYyY1diqcpzgTunk2Ye9","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4531,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4531,"ephemeral_1h_input_tokens":0},"output_tokens":2,"service_tier":"standard"}} }

85+

86+ event: content_block_start

87+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_018FMgJN6RBKgVVepJjmxo81","name":"ls","input":{}} }

88+

89+ event: ping

90+ data: {"type": "ping"}

91+

92+ event: content_block_delta

93+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"path\": \".\"}"} }

97+

98+ event: content_block_stop

99+ data: {"type":"content_block_stop","index":0 }

100+

101+ event: message_delta

102+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4531,"cache_read_input_tokens":9490,"output_tokens":50} }

103+

104+ event: message_stop

105+ data: {"type":"message_stop" }

106+

107+ headers:

108+ Content-Type:

109+ - text/event-stream; charset=utf-8

110+ status: 200 OK

111+ code: 200

112+ duration: 1.345310041s

113+- id: 2

114+ request:

115+ proto: HTTP/1.1

116+ proto_major: 1

117+ proto_minor: 1

118+ content_length: 52693

119+ host: ""

120+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use ls to list the files in the current directory\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_018FMgJN6RBKgVVepJjmxo81\",\"input\":{\"path\":\".\"},\"name\":\"ls\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_018FMgJN6RBKgVVepJjmxo81\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\n- /tmp/crush-test/TestCoderAgent/anthropic-sonnet/ls_tool/\\n - go.mod\\n - main.go\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/ls_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

121+ headers:

122+ Accept:

123+ - application/json

124+ Content-Type:

125+ - application/json

126+ User-Agent:

127+ - Anthropic/Go 1.14.0

128+ url: https://api.anthropic.com/v1/messages

129+ method: POST

130+ response:

131+ proto: HTTP/2.0

132+ proto_major: 2

133+ proto_minor: 0

134+ content_length: -1

135+ body: |+

136+ event: message_start

137+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01Qj3zitYX9Y43328gMRzR7s","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":97,"cache_read_input_tokens":14021,"cache_creation":{"ephemeral_5m_input_tokens":97,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

138+

139+ event: content_block_start

140+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

141+

142+ event: ping

143+ data: {"type": "ping"}

144+

145+ event: content_block_delta

146+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"**"} }

147+

148+ event: content_block_delta

149+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"go"} }

150+

151+ event: content_block_delta

152+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".mod**,"} }

153+

154+ event: content_block_delta

155+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" **main.go**"} }

156+

157+ event: content_block_stop

158+ data: {"type":"content_block_stop","index":0}

159+

160+ event: message_delta

161+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":97,"cache_read_input_tokens":14021,"output_tokens":13} }

162+

163+ event: message_stop

164+ data: {"type":"message_stop" }

165+

166+ headers:

167+ Content-Type:

168+ - text/event-stream; charset=utf-8

169+ status: 200 OK

170+ code: 200

171+ duration: 1.081939s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/multiedit_tool.yamlnew file+374

@@ -0,0 +1,374 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 836

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse multiedit to change ''Hello, World!'' to ''Hello, Crush!'' and add a comment ''// Greeting'' above the fmt.Println line in main.go\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01MWKFC4YJpdfdLd6h1kfcfd","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":170,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":2,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Modify"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Go"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Hello"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Worl"} }

47+

48+ event: content_block_delta

49+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d Greeting"} }

50+

51+ event: content_block_stop

52+ data: {"type":"content_block_stop","index":0 }

53+

54+ event: message_delta

55+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":170,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":10} }

56+

57+ event: message_stop

58+ data: {"type":"message_stop" }

59+

60+ headers:

61+ Content-Type:

62+ - text/event-stream; charset=utf-8

63+ status: 200 OK

64+ code: 200

65+ duration: 664.226709ms

66+- id: 1

67+ request:

68+ proto: HTTP/1.1

69+ proto_major: 1

70+ proto_minor: 1

71+ content_length: 52432

72+ host: ""

73+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use multiedit to change 'Hello, World!' to 'Hello, Crush!' and add a comment '// Greeting' above the fmt.Println line in main.go\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

74+ headers:

75+ Accept:

76+ - application/json

77+ Content-Type:

78+ - application/json

79+ User-Agent:

80+ - Anthropic/Go 1.14.0

81+ url: https://api.anthropic.com/v1/messages

82+ method: POST

83+ response:

84+ proto: HTTP/2.0

85+ proto_major: 2

86+ proto_minor: 0

87+ content_length: -1

88+ body: |+

89+ event: message_start

90+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01GN3npLhkirRANxfHbitvEC","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4563,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4563,"ephemeral_1h_input_tokens":0},"output_tokens":3,"service_tier":"standard"}} }

91+

92+ event: content_block_start

93+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

94+

95+ event: ping

96+ data: {"type": "ping"}

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"I'll rea"} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d the file"} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" first"} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":","} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" then use mult"} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"iedit to make both"} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" changes."} }

118+

119+ event: content_block_stop

120+ data: {"type":"content_block_stop","index":0 }

121+

122+ event: content_block_start

123+ data: {"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"toolu_01F2DF1unat7CCccKgnyga2E","name":"view","input":{}} }

124+

125+ event: content_block_delta

126+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":""} }

127+

128+ event: content_block_delta

129+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"{\"file_p"}}

130+

131+ event: content_block_delta

132+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"ath\": \""} }

133+

134+ event: content_block_delta

135+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"/tmp/cru"} }

136+

137+ event: content_block_delta

138+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"sh-test/T"}}

139+

140+ event: content_block_delta

141+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"estCoderAg"} }

142+

143+ event: content_block_delta

144+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"ent"} }

145+

146+ event: content_block_delta

147+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"/anthropic"} }

148+

149+ event: content_block_delta

150+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"-sonnet/mul"} }

151+

152+ event: content_block_delta

153+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"ti"} }

154+

155+ event: content_block_delta

156+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"edit_tool/m"} }

157+

158+ event: content_block_delta

159+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"ain.go\"}"} }

160+

161+ event: content_block_stop

162+ data: {"type":"content_block_stop","index":1 }

163+

164+ event: message_delta

165+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4563,"cache_read_input_tokens":9490,"output_tokens":97} }

166+

167+ event: message_stop

168+ data: {"type":"message_stop" }

169+

170+ headers:

171+ Content-Type:

172+ - text/event-stream; charset=utf-8

173+ status: 200 OK

174+ code: 200

175+ duration: 2.400389333s

176+- id: 2

177+ request:

178+ proto: HTTP/1.1

179+ proto_major: 1

180+ proto_minor: 1

181+ content_length: 53032

182+ host: ""

183+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use multiedit to change 'Hello, World!' to 'Hello, Crush!' and add a comment '// Greeting' above the fmt.Println line in main.go\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"text\":\"I'll read the file first, then use multiedit to make both changes.\",\"type\":\"text\"},{\"id\":\"toolu_01F2DF1unat7CCccKgnyga2E\",\"input\":{\"file_path\":\"/tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool/main.go\"},\"name\":\"view\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01F2DF1unat7CCccKgnyga2E\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\u003cfile\\u003e\\n 1|package main\\n 2|\\n 3|import \\\"fmt\\\"\\n 4|\\n 5|func main() {\\n 6|\\tfmt.Println(\\\"Hello, World!\\\")\\n 7|}\\n\\u003c/file\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

184+ headers:

185+ Accept:

186+ - application/json

187+ Content-Type:

188+ - application/json

189+ User-Agent:

190+ - Anthropic/Go 1.14.0

191+ url: https://api.anthropic.com/v1/messages

192+ method: POST

193+ response:

194+ proto: HTTP/2.0

195+ proto_major: 2

196+ proto_minor: 0

197+ content_length: -1

198+ body: |+

199+ event: message_start

200+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_0159aNnXatZgFaD28pxojeM7","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":161,"cache_read_input_tokens":14053,"cache_creation":{"ephemeral_5m_input_tokens":161,"ephemeral_1h_input_tokens":0},"output_tokens":25,"service_tier":"standard"}} }

201+

202+ event: content_block_start

203+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_0127xZ5cYcrHZt7yeWXucfzJ","name":"multiedit","input":{}} }

204+

205+ event: ping

206+ data: {"type": "ping"}

207+

208+ event: content_block_delta

209+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

210+

211+ event: content_block_delta

212+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"fi"} }

213+

214+ event: content_block_delta

215+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"le_path\": "} }

216+

217+ event: content_block_delta

218+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"/tmp/cru"} }

219+

220+ event: content_block_delta

221+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"sh-test/Tes"} }

222+

223+ event: content_block_delta

224+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"tCoderA"} }

225+

226+ event: content_block_delta

227+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"gent"} }

228+

229+ event: content_block_delta

230+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"/anthropic-"} }

231+

232+ event: content_block_delta

233+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"sonnet/"} }

234+

235+ event: content_block_delta

236+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"multiedit_to"} }

237+

238+ event: content_block_delta

239+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ol/main.go\""} }

240+

241+ event: content_block_delta

242+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \"edits"} }

243+

244+ event: content_block_delta

245+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\": [{\"old_"} }

246+

247+ event: content_block_delta

248+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"stri"} }

249+

250+ event: content_block_delta

251+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ng\":\""} }

252+

253+ event: content_block_delta

254+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"func "}}

255+

256+ event: content_block_delta

257+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"mai"} }

258+

259+ event: content_block_delta

260+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"n() {\\n"} }

261+

262+ event: content_block_delta

263+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\\tfmt.Prin"} }

264+

265+ event: content_block_delta

266+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"tln(\\\"Hello,"} }

267+

268+ event: content_block_delta

269+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":" World"} }

270+

271+ event: content_block_delta

272+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"!\\\")\\n}\",\"n"} }

273+

274+ event: content_block_delta

275+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ew_string\":"} }

276+

277+ event: content_block_delta

278+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"func mai"} }

279+

280+ event: content_block_delta

281+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"n() {\\n\\t// "} }

282+

283+ event: content_block_delta

284+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"Greeting\\"} }

285+

286+ event: content_block_delta

287+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"n\\tf"} }

288+

289+ event: content_block_delta

290+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"mt"} }

291+

292+ event: content_block_delta

293+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":".Print"} }

294+

295+ event: content_block_delta

296+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ln(\\\"H"} }

297+

298+ event: content_block_delta

299+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ello, C"} }

300+

301+ event: content_block_delta

302+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"rush"} }

303+

304+ event: content_block_delta

305+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"!\\\")"} }

306+

307+ event: content_block_delta

308+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\\n}\"}]}"} }

309+

310+ event: content_block_stop

311+ data: {"type":"content_block_stop","index":0 }

312+

313+ event: message_delta

314+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":161,"cache_read_input_tokens":14053,"output_tokens":176} }

315+

316+ event: message_stop

317+ data: {"type":"message_stop" }

318+

319+ headers:

320+ Content-Type:

321+ - text/event-stream; charset=utf-8

322+ status: 200 OK

323+ code: 200

324+ duration: 1.529795125s

325+- id: 3

326+ request:

327+ proto: HTTP/1.1

328+ proto_major: 1

329+ proto_minor: 1

330+ content_length: 53670

331+ host: ""

332+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use multiedit to change 'Hello, World!' to 'Hello, Crush!' and add a comment '// Greeting' above the fmt.Println line in main.go\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"text\":\"I'll read the file first, then use multiedit to make both changes.\",\"type\":\"text\"},{\"id\":\"toolu_01F2DF1unat7CCccKgnyga2E\",\"input\":{\"file_path\":\"/tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool/main.go\"},\"name\":\"view\",\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01F2DF1unat7CCccKgnyga2E\",\"content\":[{\"text\":\"\\u003cfile\\u003e\\n 1|package main\\n 2|\\n 3|import \\\"fmt\\\"\\n 4|\\n 5|func main() {\\n 6|\\tfmt.Println(\\\"Hello, World!\\\")\\n 7|}\\n\\u003c/file\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_0127xZ5cYcrHZt7yeWXucfzJ\",\"input\":{\"edits\":[{\"new_string\":\"func main() {\\n\\t// Greeting\\n\\tfmt.Println(\\\"Hello, Crush!\\\")\\n}\",\"old_string\":\"func main() {\\n\\tfmt.Println(\\\"Hello, World!\\\")\\n}\"}],\"file_path\":\"/tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool/main.go\"},\"name\":\"multiedit\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_0127xZ5cYcrHZt7yeWXucfzJ\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\u003cresult\\u003e\\nApplied 1 edits to file: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool/main.go\\n\\u003c/result\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/multiedit_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

333+ headers:

334+ Accept:

335+ - application/json

336+ Content-Type:

337+ - application/json

338+ User-Agent:

339+ - Anthropic/Go 1.14.0

340+ url: https://api.anthropic.com/v1/messages

341+ method: POST

342+ response:

343+ proto: HTTP/2.0

344+ proto_major: 2

345+ proto_minor: 0

346+ content_length: -1

347+ body: |+

348+ event: message_start

349+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01J5DJLpRcfec3fCncNWVbiZ","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":221,"cache_read_input_tokens":14214,"cache_creation":{"ephemeral_5m_input_tokens":221,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}}}

350+

351+ event: content_block_start

352+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

353+

354+ event: ping

355+ data: {"type": "ping"}

356+

357+ event: content_block_delta

358+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Done"} }

359+

360+ event: content_block_stop

361+ data: {"type":"content_block_stop","index":0 }

362+

363+ event: message_delta

364+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":221,"cache_read_input_tokens":14214,"output_tokens":4} }

365+

366+ event: message_stop

367+ data: {"type":"message_stop" }

368+

369+ headers:

370+ Content-Type:

371+ - text/event-stream; charset=utf-8

372+ status: 200 OK

373+ code: 200

374+ duration: 5.723296875s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/parallel_tool_calls.yamlnew file+204

@@ -0,0 +1,204 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 842

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse glob to find all .go files and use ls to list the current directory, it is very important that you run both tool calls in parallel\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_011AxTk6fRGrP3MaJPbdM1kh","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":159,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":2,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Parallel"}}

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Go File"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" an"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d Directory Listing"} }

47+

48+ event: content_block_stop

49+ data: {"type":"content_block_stop","index":0 }

50+

51+ event: message_delta

52+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":159,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":11} }

53+

54+ event: message_stop

55+ data: {"type":"message_stop" }

56+

57+ headers:

58+ Content-Type:

59+ - text/event-stream; charset=utf-8

60+ status: 200 OK

61+ code: 200

62+ duration: 636.447167ms

63+- id: 1

64+ request:

65+ proto: HTTP/1.1

66+ proto_major: 1

67+ proto_minor: 1

68+ content_length: 52443

69+ host: ""

70+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use glob to find all .go files and use ls to list the current directory, it is very important that you run both tool calls in parallel\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/parallel_tool_calls\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

71+ headers:

72+ Accept:

73+ - application/json

74+ Content-Type:

75+ - application/json

76+ User-Agent:

77+ - Anthropic/Go 1.14.0

78+ url: https://api.anthropic.com/v1/messages

79+ method: POST

80+ response:

81+ proto: HTTP/2.0

82+ proto_major: 2

83+ proto_minor: 0

84+ content_length: -1

85+ body: |+

86+ event: message_start

87+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_019QtBed29oZzEdoeDnYjK3p","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4552,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4552,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

88+

89+ event: content_block_start

90+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01HksmvDoHPMYWEPjPnG84Uc","name":"glob","input":{}} }

91+

92+ event: ping

93+ data: {"type": "ping"}

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"pattern\""} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":": \"**/*.go\"}"} }

103+

104+ event: content_block_stop

105+ data: {"type":"content_block_stop","index":0 }

106+

107+ event: content_block_start

108+ data: {"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"toolu_01JpofSJ9YTdmSKLzx81RmLL","name":"ls","input":{}} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":""} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"{\"path\": \".\"}"}}

115+

116+ event: content_block_stop

117+ data: {"type":"content_block_stop","index":1 }

118+

119+ event: message_delta

120+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4552,"cache_read_input_tokens":9490,"output_tokens":86} }

121+

122+ event: message_stop

123+ data: {"type":"message_stop"}

124+

125+ headers:

126+ Content-Type:

127+ - text/event-stream; charset=utf-8

128+ status: 200 OK

129+ code: 200

130+ duration: 2.092154708s

131+- id: 2

132+ request:

133+ proto: HTTP/1.1

134+ proto_major: 1

135+ proto_minor: 1

136+ content_length: 53087

137+ host: ""

138+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use glob to find all .go files and use ls to list the current directory, it is very important that you run both tool calls in parallel\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01HksmvDoHPMYWEPjPnG84Uc\",\"input\":{\"pattern\":\"**/*.go\"},\"name\":\"glob\",\"type\":\"tool_use\"},{\"id\":\"toolu_01JpofSJ9YTdmSKLzx81RmLL\",\"input\":{\"path\":\".\"},\"name\":\"ls\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01HksmvDoHPMYWEPjPnG84Uc\",\"content\":[{\"text\":\"/tmp/crush-test/TestCoderAgent/anthropic-sonnet/parallel_tool_calls/main.go\",\"type\":\"text\"}],\"type\":\"tool_result\"},{\"tool_use_id\":\"toolu_01JpofSJ9YTdmSKLzx81RmLL\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\n- /tmp/crush-test/TestCoderAgent/anthropic-sonnet/parallel_tool_calls/\\n - go.mod\\n - main.go\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/parallel_tool_calls\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

139+ headers:

140+ Accept:

141+ - application/json

142+ Content-Type:

143+ - application/json

144+ User-Agent:

145+ - Anthropic/Go 1.14.0

146+ url: https://api.anthropic.com/v1/messages

147+ method: POST

148+ response:

149+ proto: HTTP/2.0

150+ proto_major: 2

151+ proto_minor: 0

152+ content_length: -1

153+ body: |+

154+ event: message_start

155+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01LH4CcQvVEVQhnn6Tqe2SrN","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":6,"cache_creation_input_tokens":207,"cache_read_input_tokens":14042,"cache_creation":{"ephemeral_5m_input_tokens":207,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

156+

157+ event: content_block_start

158+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

159+

160+ event: ping

161+ data: {"type": "ping"}

162+

163+ event: content_block_delta

164+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"**"} }

165+

166+ event: content_block_delta

167+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Go"} }

168+

169+ event: content_block_delta

170+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" files found**: main"} }

171+

172+ event: content_block_delta

173+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".go"} }

174+

175+ event: content_block_delta

176+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\n\n**Directory"} }

177+

178+ event: content_block_delta

179+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" contents"} }

180+

181+ event: content_block_delta

182+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"**:"} }

183+

184+ event: content_block_delta

185+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" go"} }

186+

187+ event: content_block_delta

188+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".mod, main.go"} }

189+

190+ event: content_block_stop

191+ data: {"type":"content_block_stop","index":0 }

192+

193+ event: message_delta

194+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":6,"cache_creation_input_tokens":207,"cache_read_input_tokens":14042,"output_tokens":23} }

195+

196+ event: message_stop

197+ data: {"type":"message_stop"}

198+

199+ headers:

200+ Content-Type:

201+ - text/event-stream; charset=utf-8

202+ status: 200 OK

203+ code: 200

204+ duration: 1.624386875s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/read_a_file.yamlnew file+192

@@ -0,0 +1,192 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 723

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nRead the go mod\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01BS8RX55UAXGbe4XdL5VxK5","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":134,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Rea"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d Go"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Module"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Details"} }

47+

48+ event: content_block_stop

49+ data: {"type":"content_block_stop","index":0 }

50+

51+ event: message_delta

52+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":134,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":7} }

53+

54+ event: message_stop

55+ data: {"type":"message_stop" }

56+

57+ headers:

58+ Content-Type:

59+ - text/event-stream; charset=utf-8

60+ status: 200 OK

61+ code: 200

62+ duration: 589.305792ms

63+- id: 1

64+ request:

65+ proto: HTTP/1.1

66+ proto_major: 1

67+ proto_minor: 1

68+ content_length: 52316

69+ host: ""

70+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"Read the go mod\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/read_a_file\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

71+ headers:

72+ Accept:

73+ - application/json

74+ Content-Type:

75+ - application/json

76+ User-Agent:

77+ - Anthropic/Go 1.14.0

78+ url: https://api.anthropic.com/v1/messages

79+ method: POST

80+ response:

81+ proto: HTTP/2.0

82+ proto_major: 2

83+ proto_minor: 0

84+ content_length: -1

85+ body: |+

86+ event: message_start

87+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01JLKef6JCmfXXnQuhuCdQd9","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4527,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4527,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

88+

89+ event: content_block_start

90+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01F7YcxHnHkoFTwRJt2WcEPE","name":"view","input":{}} }

91+

92+ event: ping

93+ data: {"type": "ping"}

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"file_p"} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"at"} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"h\""} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":": \"go.m"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"od\"}"} }

112+

113+ event: content_block_stop

114+ data: {"type":"content_block_stop","index":0 }

115+

116+ event: message_delta

117+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4527,"cache_read_input_tokens":9490,"output_tokens":55} }

118+

119+ event: message_stop

120+ data: {"type":"message_stop" }

121+

122+ headers:

123+ Content-Type:

124+ - text/event-stream; charset=utf-8

125+ status: 200 OK

126+ code: 200

127+ duration: 2.071498208s

128+- id: 2

129+ request:

130+ proto: HTTP/1.1

131+ proto_major: 1

132+ proto_minor: 1

133+ content_length: 52689

134+ host: ""

135+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"Read the go mod\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01F7YcxHnHkoFTwRJt2WcEPE\",\"input\":{\"file_path\":\"go.mod\"},\"name\":\"view\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01F7YcxHnHkoFTwRJt2WcEPE\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\u003cfile\\u003e\\n 1|module example.com/testproject\\n 2|\\n 3|go 1.23\\n\\u003c/file\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/read_a_file\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

136+ headers:

137+ Accept:

138+ - application/json

139+ Content-Type:

140+ - application/json

141+ User-Agent:

142+ - Anthropic/Go 1.14.0

143+ url: https://api.anthropic.com/v1/messages

144+ method: POST

145+ response:

146+ proto: HTTP/2.0

147+ proto_major: 2

148+ proto_minor: 0

149+ content_length: -1

150+ body: |+

151+ event: message_start

152+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01DqkRgmKqtdzVozWkQSSTay","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":96,"cache_read_input_tokens":14017,"cache_creation":{"ephemeral_5m_input_tokens":96,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

153+

154+ event: content_block_start

155+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

156+

157+ event: ping

158+ data: {"type": "ping"}

159+

160+ event: content_block_delta

161+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"```"} }

162+

163+ event: content_block_delta

164+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"go"} }

165+

166+ event: content_block_delta

167+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\nmodule"} }

168+

169+ event: content_block_delta

170+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" example.com/testproject\n\ngo"} }

171+

172+ event: content_block_delta

173+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" 1.23"} }

174+

175+ event: content_block_delta

176+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\n```"}}

177+

178+ event: content_block_stop

179+ data: {"type":"content_block_stop","index":0 }

180+

181+ event: message_delta

182+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":96,"cache_read_input_tokens":14017,"output_tokens":21} }

183+

184+ event: message_stop

185+ data: {"type":"message_stop" }

186+

187+ headers:

188+ Content-Type:

189+ - text/event-stream; charset=utf-8

190+ status: 200 OK

191+ code: 200

192+ duration: 3.342119125s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/simple_test.yamlnew file+136

@@ -0,0 +1,136 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 713

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nHello\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01RhwrPwqgU78kv4XsoesFPF","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":131,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}}}

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Quick"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Chat"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Opener"} }

44+

45+ event: content_block_stop

46+ data: {"type":"content_block_stop","index":0 }

47+

48+ event: message_delta

49+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":131,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":7}}

50+

51+ event: message_stop

52+ data: {"type":"message_stop" }

53+

54+ headers:

55+ Content-Type:

56+ - text/event-stream; charset=utf-8

57+ status: 200 OK

58+ code: 200

59+ duration: 705.524375ms

60+- id: 1

61+ request:

62+ proto: HTTP/1.1

63+ proto_major: 1

64+ proto_minor: 1

65+ content_length: 52306

66+ host: ""

67+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"Hello\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/simple_test\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

68+ headers:

69+ Accept:

70+ - application/json

71+ Content-Type:

72+ - application/json

73+ User-Agent:

74+ - Anthropic/Go 1.14.0

75+ url: https://api.anthropic.com/v1/messages

76+ method: POST

77+ response:

78+ proto: HTTP/2.0

79+ proto_major: 2

80+ proto_minor: 0

81+ content_length: -1

82+ body: |+

83+ event: message_start

84+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01Pcqx7Mq3aWnT6VUDVGAs71","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":14012,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":14012,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

85+

86+ event: content_block_start

87+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

88+

89+ event: ping

90+ data: {"type": "ping"}

91+

92+ event: content_block_delta

93+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Hello"} }

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"! I'm ready to help you with"}}

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" any"} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" coding"} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" tasks,"} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" file"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" operations, or questions"} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" you have. What"} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" would you like to work"} }

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" on?"} }

121+

122+ event: content_block_stop

123+ data: {"type":"content_block_stop","index":0 }

124+

125+ event: message_delta

126+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":14012,"cache_read_input_tokens":0,"output_tokens":32} }

127+

128+ event: message_stop

129+ data: {"type":"message_stop" }

130+

131+ headers:

132+ Content-Type:

133+ - text/event-stream; charset=utf-8

134+ status: 200 OK

135+ code: 200

136+ duration: 2.2130525s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/sourcegraph_tool.yamlnew file+454

@@ -0,0 +1,454 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 768

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse sourcegraph to search for ''func main'' in Go repositories\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_01FkLqqw1crC8WC9dAUtwxqu","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":145,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":3,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Searching Go"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Repos"}}

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" for Main"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Function"} }

47+

48+ event: content_block_stop

49+ data: {"type":"content_block_stop","index":0 }

50+

51+ event: message_delta

52+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":145,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":11} }

53+

54+ event: message_stop

55+ data: {"type":"message_stop" }

56+

57+ headers:

58+ Content-Type:

59+ - text/event-stream; charset=utf-8

60+ status: 200 OK

61+ code: 200

62+ duration: 640.628291ms

63+- id: 1

64+ request:

65+ proto: HTTP/1.1

66+ proto_major: 1

67+ proto_minor: 1

68+ content_length: 52366

69+ host: ""

70+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use sourcegraph to search for 'func main' in Go repositories\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/sourcegraph_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

71+ headers:

72+ Accept:

73+ - application/json

74+ Content-Type:

75+ - application/json

76+ User-Agent:

77+ - Anthropic/Go 1.14.0

78+ url: https://api.anthropic.com/v1/messages

79+ method: POST

80+ response:

81+ proto: HTTP/2.0

82+ proto_major: 2

83+ proto_minor: 0

84+ content_length: -1

85+ body: |+

86+ event: message_start

87+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01Rf2rLFJUs5ePRyK3fB6DAt","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":14028,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":14028,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

88+

89+ event: content_block_start

90+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01DtbQComcz9jYF4xhrFfpEd","name":"sourcegraph","input":{}} }

91+

92+ event: ping

93+ data: {"type": "ping"}

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"que"} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ry\": "} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"func ma"} }

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"in lang:"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"go\""} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", "} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"c"} }

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"oun"} }

121+

122+ event: content_block_delta

123+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"t\": 10}"} }

124+

125+ event: content_block_stop

126+ data: {"type":"content_block_stop","index":0 }

127+

128+ event: message_delta

129+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":14028,"cache_read_input_tokens":0,"output_tokens":74} }

130+

131+ event: message_stop

132+ data: {"type":"message_stop" }

133+

134+ headers:

135+ Content-Type:

136+ - text/event-stream; charset=utf-8

137+ status: 200 OK

138+ code: 200

139+ duration: 2.687317s

140+- id: 2

141+ request:

142+ proto: HTTP/1.1

143+ proto_major: 1

144+ proto_minor: 1

145+ content_length: 424

146+ host: sourcegraph.com

147+ body: '{"query":"query Search($query: String!) { search(query: $query, version: V2, patternType: keyword ) { results { matchCount, limitHit, resultCount, approximateResultCount, missing { name }, timedout { name }, indexUnavailable, results { __typename, ... on FileMatch { repository { name }, file { path, url, content }, lineMatches { preview, lineNumber, offsetAndLengths } } } } } }","variables":{"query":"func main lang:go"}}'

148+ headers:

149+ Content-Type:

150+ - application/json

151+ User-Agent:

152+ - crush/1.0

153+ url: https://sourcegraph.com/.api/graphql

154+ method: POST

155+ response:

156+ proto: HTTP/2.0

157+ proto_major: 2

158+ proto_minor: 0

159+ content_length: -1

160+ uncompressed: true

161+ body: '{"data":{"search":{"results":{"matchCount":30,"limitHit":true,"resultCount":30,"approximateResultCount":"30+","missing":[],"timedout":[],"indexUnavailable":false,"results":[{"__typename":"FileMatch","repository":{"name":"github.com/Code-Hex/funcy-mock"},"file":{"path":"cmd/funcygen/main.go","url":"/github.com/Code-Hex/funcy-mock/-/blob/cmd/funcygen/main.go","content":"package main\n\nimport (\n\t\"os\"\n\n\tfuncy \"github.com/Code-Hex/funcy-mock\"\n)\n\nfunc main() {\n\tos.Exit(funcy.Run())\n}\n"},"lineMatches":[{"preview":"func main() {","lineNumber":8,"offsetAndLengths":[[0,9]]},{"preview":"\tos.Exit(funcy.Run())","lineNumber":9,"offsetAndLengths":[[9,4]]},{"preview":"package main","lineNumber":0,"offsetAndLengths":[[8,4]]},{"preview":"\tfuncy \"github.com/Code-Hex/funcy-mock\"","lineNumber":5,"offsetAndLengths":[[1,4],[28,4]]}]},{"__typename":"FileMatch","repository":{"name":"github.com/callicoder/golang-tutorials"},"file":{"path":"06-functions/01-function-declaration/main.go","url":"/github.com/callicoder/golang-tutorials/-/blob/06-functions/01-function-declaration/main.go","content":"package main\n\nimport \"fmt\"\n\nfunc avg(x float64, y float64) float64 {\n\treturn (x + y) / 2\n}\n\nfunc main() {\n\tx := 5.75\n\ty := 6.25\n\n\tresult := avg(x, y)\n\n\tfmt.Printf(\"Average of %.2f and %.2f = %.2f\\n\", x, y, result)\n}\n"},"lineMatches":[{"preview":"func main() {","lineNumber":8,"offsetAndLengths":[[0,9]]},{"preview":"package main","lineNumber":0,"offsetAndLengths":[[8,4]]},{"preview":"func avg(x float64, y float64) float64 {","lineNumber":4,"offsetAndLengths":[[0,4]]}]},{"__typename":"FileMatch","repository":{"name":"github.com/callicoder/golang-tutorials"},"file":{"path":"13-methods/02-methods-are-functions/main.go","url":"/github.com/callicoder/golang-tutorials/-/blob/13-methods/02-methods-are-functions/main.go","content":"package main\n\nimport (\n\t\"fmt\"\n)\n\n// Struct type - `Point`\ntype Point struct {\n\tX, Y float64\n}\n\nfunc IsAboveFunc(p Point, y float64) bool {\n\treturn p.Y \u003e y\n}\n\n/*\nCompare the above function with the corresponding method -\nfunc (p Point) IsAbove(y float64) bool {\n\treturn p.Y \u003e y\n}\n*/\n\nfunc main() {\n\tp := Point{2.5, -3.0}\n\n\tfmt.Println(\"Point : \", p)\n\n\tfmt.Println(\"Is Point p located above the line y = 1.0 ? : \", IsAboveFunc(p, 1))\n}\n"},"lineMatches":[{"preview":"func main() {","lineNumber":22,"offsetAndLengths":[[0,9]]},{"preview":"package main","lineNumber":0,"offsetAndLengths":[[8,4]]},{"preview":"func IsAboveFunc(p Point, y float64) bool {","lineNumber":11,"offsetAndLengths":[[0,4],[12,4]]},{"preview":"Compare the above function with the corresponding method -","lineNumber":16,"offsetAndLengths":[[18,4]]},{"preview":"func (p Point) IsAbove(y float64) bool {","lineNumber":17,"offsetAndLengths":[[0,4]]},{"preview":"\tfmt.Println(\"Is Point p located above the line y = 1.0 ? : \", IsAboveFunc(p, 1))","lineNumber":27,"offsetAndLengths":[[70,4]]}]},{"__typename":"FileMatch","repository":{"name":"github.com/whitesquirrell/C0deVari4nt"},"file":{"path":"api-server/vscode-codeql-starter/codeql-go/ql/test/library-tests/semmle/go/Function/main.go","url":"/github.com/whitesquirrell/C0deVari4nt/-/blob/api-server/vscode-codeql-starter/codeql-go/ql/test/library-tests/semmle/go/Function/main.go","content":"package main\n\ntype t int\n\nfunc main() {}\n\nfunc f1(x int) {}\n\nfunc (t t) f2(x, y int) {}\n\nfunc (_ t) f3(x int, y string) {}\n\nfunc f4(x int, y ...string) {}\n\nfunc f5(x int) int { return 0 }\n\nfunc f6(x int) (y int) { return 0 }\n\nfunc f7() (x, y int) { return 0, 0 }\n\nfunc f8() (x int, _ string) { return 0, \"\" }\n\nfunc f9() (_ int, y string) { return 0, \"\" }\n"},"lineMatches":[{"preview":"func main() {}","lineNumber":4,"offsetAndLengths":[[0,9]]},{"preview":"func f1(x int) {}","lineNumber":6,"offsetAndLengths":[[0,4]]},{"preview":"func (t t) f2(x, y int) {}","lineNumber":8,"offsetAndLengths":[[0,4]]},{"preview":"func (_ t) f3(x int, y string) {}","lineNumber":10,"offsetAndLengths":[[0,4]]},{"preview":"func f4(x int, y ...string) {}","lineNumber":12,"offsetAndLengths":[[0,4]]},{"preview":"func f5(x int) int { return 0 }","lineNumber":14,"offsetAndLengths":[[0,4]]},{"preview":"func f6(x int) (y int) { return 0 }","lineNumber":16,"offsetAndLengths":[[0,4]]},{"preview":"func f7() (x, y int) { return 0, 0 }","lineNumber":18,"offsetAndLengths":[[0,4]]},{"preview":"func f8() (x int, _ string) { return 0, \"\" }","lineNumber":20,"offsetAndLengths":[[0,4]]},{"preview":"func f9() (_ int, y string) { return 0, \"\" }","lineNumber":22,"offsetAndLengths":[[0,4]]},{"preview":"package main","lineNumber":0,"offsetAndLengths":[[8,4]]}]},{"__typename":"FileMatch","repository":{"name":"github.com/callicoder/golang-tutorials"},"file":{"path":"13-methods/07-method-with-value-receiver-vs-function-with-value-argument/main.go","url":"/github.com/callicoder/golang-tutorials/-/blob/13-methods/07-method-with-value-receiver-vs-function-with-value-argument/main.go","content":"package main\n\nimport (\n\t\"fmt\"\n)\n\n// Struct type - `Point`\ntype Point struct {\n\tX, Y float64\n}\n\nfunc (p Point) IsAbove(y float64) bool {\n\treturn p.Y \u003e y\n}\n\nfunc IsAboveFunc(p Point, y float64) bool {\n\treturn p.Y \u003e y\n}\n\nfunc main() {\n\tp := Point{0, 4}\n\tptr := \u0026p\n\n\tfmt.Println(\"Point p = \", p)\n\n\t// Calling a Method with Value receiver\n\tp.IsAbove(1) // Valid\n\tptr.IsAbove(1) // Valid\n\n\t// Calling a Function with a Value argument\n\tIsAboveFunc(p, 1) // Valid\n\t// IsAboveFunc(ptr, 1) // Not Valid\n}\n"},"lineMatches":[{"preview":"func main() {","lineNumber":19,"offsetAndLengths":[[0,9]]},{"preview":"package main","lineNumber":0,"offsetAndLengths":[[8,4]]},{"preview":"func IsAboveFunc(p Point, y float64) bool {","lineNumber":15,"offsetAndLengths":[[0,4],[12,4]]}]}]}}}}'

162+ headers:

163+ Content-Type:

164+ - application/json

165+ status: 200 OK

166+ code: 200

167+ duration: 1.505534709s

168+- id: 3

169+ request:

170+ proto: HTTP/1.1

171+ proto_major: 1

172+ proto_minor: 1

173+ content_length: 63051

174+ host: ""

175+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use sourcegraph to search for 'func main' in Go repositories\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01DtbQComcz9jYF4xhrFfpEd\",\"input\":{\"count\":10,\"query\":\"func main lang:go\"},\"name\":\"sourcegraph\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01DtbQComcz9jYF4xhrFfpEd\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"# Sourcegraph Search Results\\n\\nFound 30 matches across 30 results\\n(Result limit reached, try a more specific query)\\n\\n## Result 1: github.com/Code-Hex/funcy-mock/cmd/funcygen/main.go\\n\\nURL: /github.com/Code-Hex/funcy-mock/-/blob/cmd/funcygen/main.go\\n\\n```\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"os\\\"\\n5| \\n6| \\tfuncy \\\"github.com/Code-Hex/funcy-mock\\\"\\n7| )\\n8| func main() {\\n9| func main() {\\n10| \\tos.Exit(funcy.Run())\\n11| }\\n12| \\n```\\n\\n```\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"os\\\"\\n5| \\n6| \\tfuncy \\\"github.com/Code-Hex/funcy-mock\\\"\\n7| )\\n8| \\n9| \\tos.Exit(funcy.Run())\\n10| \\tos.Exit(funcy.Run())\\n11| }\\n12| \\n```\\n\\n```\\n0| package main\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"os\\\"\\n5| \\n6| \\tfuncy \\\"github.com/Code-Hex/funcy-mock\\\"\\n7| )\\n8| \\n9| func main() {\\n10| \\tos.Exit(funcy.Run())\\n```\\n\\n```\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"os\\\"\\n5| \\tfuncy \\\"github.com/Code-Hex/funcy-mock\\\"\\n6| \\tfuncy \\\"github.com/Code-Hex/funcy-mock\\\"\\n7| )\\n8| \\n9| func main() {\\n10| \\tos.Exit(funcy.Run())\\n11| }\\n12| \\n```\\n\\n## Result 2: github.com/callicoder/golang-tutorials/06-functions/01-function-declaration/main.go\\n\\nURL: /github.com/callicoder/golang-tutorials/-/blob/06-functions/01-function-declaration/main.go\\n\\n```\\n1| package main\\n2| \\n3| import \\\"fmt\\\"\\n4| \\n5| func avg(x float64, y float64) float64 {\\n6| \\treturn (x + y) / 2\\n7| }\\n8| func main() {\\n9| func main() {\\n10| \\tx := 5.75\\n11| \\ty := 6.25\\n12| \\n13| \\tresult := avg(x, y)\\n14| \\n15| \\tfmt.Printf(\\\"Average of %.2f and %.2f = %.2f\\\\n\\\", x, y, result)\\n16| }\\n17| \\n```\\n\\n```\\n0| package main\\n1| package main\\n2| \\n3| import \\\"fmt\\\"\\n4| \\n5| func avg(x float64, y float64) float64 {\\n6| \\treturn (x + y) / 2\\n7| }\\n8| \\n9| func main() {\\n10| \\tx := 5.75\\n```\\n\\n```\\n1| package main\\n2| \\n3| import \\\"fmt\\\"\\n4| func avg(x float64, y float64) float64 {\\n5| func avg(x float64, y float64) float64 {\\n6| \\treturn (x + y) / 2\\n7| }\\n8| \\n9| func main() {\\n10| \\tx := 5.75\\n11| \\ty := 6.25\\n12| \\n13| \\tresult := avg(x, y)\\n14| \\n```\\n\\n## Result 3: github.com/callicoder/golang-tutorials/13-methods/02-methods-are-functions/main.go\\n\\nURL: /github.com/callicoder/golang-tutorials/-/blob/13-methods/02-methods-are-functions/main.go\\n\\n```\\n12| func IsAboveFunc(p Point, y float64) bool {\\n13| \\treturn p.Y \\u003e y\\n14| }\\n15| \\n16| /*\\n17| Compare the above function with the corresponding method -\\n18| func (p Point) IsAbove(y float64) bool {\\n19| \\treturn p.Y \\u003e y\\n20| }\\n21| */\\n22| func main() {\\n23| func main() {\\n24| \\tp := Point{2.5, -3.0}\\n25| \\n26| \\tfmt.Println(\\\"Point : \\\", p)\\n27| \\n28| \\tfmt.Println(\\\"Is Point p located above the line y = 1.0 ? : \\\", IsAboveFunc(p, 1))\\n29| }\\n30| \\n```\\n\\n```\\n0| package main\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"fmt\\\"\\n5| )\\n6| \\n7| // Struct type - `Point`\\n8| type Point struct {\\n9| \\tX, Y float64\\n10| }\\n```\\n\\n```\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"fmt\\\"\\n5| )\\n6| \\n7| // Struct type - `Point`\\n8| type Point struct {\\n9| \\tX, Y float64\\n10| }\\n11| func IsAboveFunc(p Point, y float64) bool {\\n12| func IsAboveFunc(p Point, y float64) bool {\\n13| \\treturn p.Y \\u003e y\\n14| }\\n15| \\n16| /*\\n17| Compare the above function with the corresponding method -\\n18| func (p Point) IsAbove(y float64) bool {\\n19| \\treturn p.Y \\u003e y\\n20| }\\n21| */\\n```\\n\\n```\\n6| \\n7| // Struct type - `Point`\\n8| type Point struct {\\n9| \\tX, Y float64\\n10| }\\n11| \\n12| func IsAboveFunc(p Point, y float64) bool {\\n13| \\treturn p.Y \\u003e y\\n14| }\\n15| \\n16| Compare the above function with the corresponding method -\\n17| Compare the above function with the corresponding method -\\n18| func (p Point) IsAbove(y float64) bool {\\n19| \\treturn p.Y \\u003e y\\n20| }\\n21| */\\n22| \\n23| func main() {\\n24| \\tp := Point{2.5, -3.0}\\n25| \\n26| \\tfmt.Println(\\\"Point : \\\", p)\\n```\\n\\n```\\n7| // Struct type - `Point`\\n8| type Point struct {\\n9| \\tX, Y float64\\n10| }\\n11| \\n12| func IsAboveFunc(p Point, y float64) bool {\\n13| \\treturn p.Y \\u003e y\\n14| }\\n15| \\n16| /*\\n17| func (p Point) IsAbove(y float64) bool {\\n18| func (p Point) IsAbove(y float64) bool {\\n19| \\treturn p.Y \\u003e y\\n20| }\\n21| */\\n22| \\n23| func main() {\\n24| \\tp := Point{2.5, -3.0}\\n25| \\n26| \\tfmt.Println(\\\"Point : \\\", p)\\n27| \\n```\\n\\n```\\n17| Compare the above function with the corresponding method -\\n18| func (p Point) IsAbove(y float64) bool {\\n19| \\treturn p.Y \\u003e y\\n20| }\\n21| */\\n22| \\n23| func main() {\\n24| \\tp := Point{2.5, -3.0}\\n25| \\n26| \\tfmt.Println(\\\"Point : \\\", p)\\n27| \\tfmt.Println(\\\"Is Point p located above the line y = 1.0 ? : \\\", IsAboveFunc(p, 1))\\n28| \\tfmt.Println(\\\"Is Point p located above the line y = 1.0 ? : \\\", IsAboveFunc(p, 1))\\n29| }\\n30| \\n```\\n\\n## Result 4: github.com/whitesquirrell/C0deVari4nt/api-server/vscode-codeql-starter/codeql-go/ql/test/library-tests/semmle/go/Function/main.go\\n\\nURL: /github.com/whitesquirrell/C0deVari4nt/-/blob/api-server/vscode-codeql-starter/codeql-go/ql/test/library-tests/semmle/go/Function/main.go\\n\\n```\\n1| package main\\n2| \\n3| type t int\\n4| func main() {}\\n5| func main() {}\\n6| \\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n```\\n\\n```\\n1| package main\\n2| \\n3| type t int\\n4| \\n5| func main() {}\\n6| func f1(x int) {}\\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n```\\n\\n```\\n1| package main\\n2| \\n3| type t int\\n4| \\n5| func main() {}\\n6| \\n7| func f1(x int) {}\\n8| func (t t) f2(x, y int) {}\\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| \\n```\\n\\n```\\n1| package main\\n2| \\n3| type t int\\n4| \\n5| func main() {}\\n6| \\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| func (_ t) f3(x int, y string) {}\\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| \\n19| func f7() (x, y int) { return 0, 0 }\\n20| \\n```\\n\\n```\\n2| \\n3| type t int\\n4| \\n5| func main() {}\\n6| \\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| func f4(x int, y ...string) {}\\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| \\n19| func f7() (x, y int) { return 0, 0 }\\n20| \\n21| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n22| \\n```\\n\\n```\\n4| \\n5| func main() {}\\n6| \\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| func f5(x int) int { return 0 }\\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| \\n19| func f7() (x, y int) { return 0, 0 }\\n20| \\n21| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n22| \\n23| func f9() (_ int, y string) { return 0, \\\"\\\" }\\n24| \\n```\\n\\n```\\n6| \\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| func f6(x int) (y int) { return 0 }\\n17| func f6(x int) (y int) { return 0 }\\n18| \\n19| func f7() (x, y int) { return 0, 0 }\\n20| \\n21| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n22| \\n23| func f9() (_ int, y string) { return 0, \\\"\\\" }\\n24| \\n```\\n\\n```\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| func f7() (x, y int) { return 0, 0 }\\n19| func f7() (x, y int) { return 0, 0 }\\n20| \\n21| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n22| \\n23| func f9() (_ int, y string) { return 0, \\\"\\\" }\\n24| \\n```\\n\\n```\\n10| \\n11| func (_ t) f3(x int, y string) {}\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| \\n19| func f7() (x, y int) { return 0, 0 }\\n20| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n21| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n22| \\n23| func f9() (_ int, y string) { return 0, \\\"\\\" }\\n24| \\n```\\n\\n```\\n12| \\n13| func f4(x int, y ...string) {}\\n14| \\n15| func f5(x int) int { return 0 }\\n16| \\n17| func f6(x int) (y int) { return 0 }\\n18| \\n19| func f7() (x, y int) { return 0, 0 }\\n20| \\n21| func f8() (x int, _ string) { return 0, \\\"\\\" }\\n22| func f9() (_ int, y string) { return 0, \\\"\\\" }\\n23| func f9() (_ int, y string) { return 0, \\\"\\\" }\\n24| \\n```\\n\\n```\\n0| package main\\n1| package main\\n2| \\n3| type t int\\n4| \\n5| func main() {}\\n6| \\n7| func f1(x int) {}\\n8| \\n9| func (t t) f2(x, y int) {}\\n10| \\n```\\n\\n## Result 5: github.com/callicoder/golang-tutorials/13-methods/07-method-with-value-receiver-vs-function-with-value-argument/main.go\\n\\nURL: /github.com/callicoder/golang-tutorials/-/blob/13-methods/07-method-with-value-receiver-vs-function-with-value-argument/main.go\\n\\n```\\n9| \\tX, Y float64\\n10| }\\n11| \\n12| func (p Point) IsAbove(y float64) bool {\\n13| \\treturn p.Y \\u003e y\\n14| }\\n15| \\n16| func IsAboveFunc(p Point, y float64) bool {\\n17| \\treturn p.Y \\u003e y\\n18| }\\n19| func main() {\\n20| func main() {\\n21| \\tp := Point{0, 4}\\n22| \\tptr := \\u0026p\\n23| \\n24| \\tfmt.Println(\\\"Point p = \\\", p)\\n25| \\n26| \\t// Calling a Method with Value receiver\\n27| \\tp.IsAbove(1) // Valid\\n28| \\tptr.IsAbove(1) // Valid\\n29| \\n```\\n\\n```\\n0| package main\\n1| package main\\n2| \\n3| import (\\n4| \\t\\\"fmt\\\"\\n5| )\\n6| \\n7| // Struct type - `Point`\\n8| type Point struct {\\n9| \\tX, Y float64\\n10| }\\n```\\n\\n```\\n5| )\\n6| \\n7| // Struct type - `Point`\\n8| type Point struct {\\n9| \\tX, Y float64\\n10| }\\n11| \\n12| func (p Point) IsAbove(y float64) bool {\\n13| \\treturn p.Y \\u003e y\\n14| }\\n15| func IsAboveFunc(p Point, y float64) bool {\\n16| func IsAboveFunc(p Point, y float64) bool {\\n17| \\treturn p.Y \\u003e y\\n18| }\\n19| \\n20| func main() {\\n21| \\tp := Point{0, 4}\\n22| \\tptr := \\u0026p\\n23| \\n24| \\tfmt.Println(\\\"Point p = \\\", p)\\n25| \\n```\\n\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/sourcegraph_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

176+ headers:

177+ Accept:

178+ - application/json

179+ Content-Type:

180+ - application/json

181+ User-Agent:

182+ - Anthropic/Go 1.14.0

183+ url: https://api.anthropic.com/v1/messages

184+ method: POST

185+ response:

186+ proto: HTTP/2.0

187+ proto_major: 2

188+ proto_minor: 0

189+ content_length: -1

190+ body: |+

191+ event: message_start

192+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_014y6FKU3R7sSQVReuxECULe","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":4678,"cache_read_input_tokens":14028,"cache_creation":{"ephemeral_5m_input_tokens":4678,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

193+

194+ event: content_block_start

195+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

196+

197+ event: ping

198+ data: {"type": "ping"}

199+

200+ event: content_block_delta

201+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Foun"} }

202+

203+ event: content_block_delta

204+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d "} }

205+

206+ event: content_block_delta

207+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"30"} }

208+

209+ event: content_block_delta

210+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" results for"} }

211+

212+ event: content_block_delta

213+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" `"} }

214+

215+ event: content_block_delta

216+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"func"} }

217+

218+ event: content_block_delta

219+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" main"} }

220+

221+ event: content_block_delta

222+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"`"} }

223+

224+ event: content_block_delta

225+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" in Go"} }

226+

227+ event: content_block_delta

228+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" repositories"} }

229+

230+ event: content_block_delta

231+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":". Here"} }

232+

233+ event: content_block_delta

234+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" are"} }

235+

236+ event: content_block_delta

237+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" some"} }

238+

239+ event: content_block_delta

240+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" examples"} }

241+

242+ event: content_block_delta

243+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":":\n\n**"} }

244+

245+ event: content_block_delta

246+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Simple"} }

247+

248+ event: content_block_delta

249+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" main"} }

250+

251+ event: content_block_delta

252+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" functions"} }

253+

254+ event: content_block_delta

255+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":":**\n- `"} }

256+

257+ event: content_block_delta

258+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"github"} }

259+

260+ event: content_block_delta

261+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".com/Code"} }

262+

263+ event: content_block_delta

264+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"-Hex/fun"} }

265+

266+ event: content_block_delta

267+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"cy-"} }

268+

269+ event: content_block_delta

270+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"mock"} }

271+

272+ event: content_block_delta

273+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"/"} }

274+

275+ event: content_block_delta

276+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"cm"} }

277+

278+ event: content_block_delta

279+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d/func"} }

280+

281+ event: content_block_delta

282+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"ygen"} }

283+

284+ event: content_block_delta

285+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"/main.go:"} }

286+

287+ event: content_block_delta

288+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"9"} }

289+

290+ event: content_block_delta

291+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"`"} }

292+

293+ event: content_block_delta

294+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" -"} }

295+

296+ event: content_block_delta

297+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Minimal"} }

298+

299+ event: content_block_delta

300+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" main"} }

301+

302+ event: content_block_delta

303+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" that"}}

304+

305+ event: content_block_delta

306+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" calls"} }

307+

308+ event: content_block_delta

309+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" `os"} }

310+

311+ event: content_block_delta

312+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".Exit(fun"} }

313+

314+ event: content_block_delta

315+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"cy.Run()"}}

316+

317+ event: content_block_delta

318+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":")`\n- `github.com/"} }

319+

320+ event: content_block_delta

321+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"callicoder/golang-tutorials/"}}

322+

323+ event: content_block_delta

324+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"06"} }

325+

326+ event: content_block_delta

327+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"-functions"} }

328+

329+ event: content_block_delta

330+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"/01-function"} }

331+

332+ event: content_block_delta

333+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"-declaration"} }

334+

335+ event: content_block_delta

336+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"/main.go:9` -"} }

337+

338+ event: content_block_delta

339+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Basic"} }

340+

341+ event: content_block_delta

342+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" tutorial"} }

343+

344+ event: content_block_delta

345+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" showing"} }

346+

347+ event: content_block_delta

348+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" function declaration an"} }

349+

350+ event: content_block_delta

351+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d usage"} }

352+

353+ event: content_block_delta

354+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\n\n**Test"} }

355+

356+ event: content_block_delta

357+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" files:**\n- `github.com"} }

358+

359+ event: content_block_delta

360+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"/whit"} }

361+

362+ event: content_block_delta

363+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"esquirrell/C0de"} }

364+

365+ event: content_block_delta

366+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Vari4nt/."} }

367+

368+ event: content_block_delta

369+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"../Function"} }

370+

371+ event: content_block_delta

372+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"/main.go:5"} }

373+

374+ event: content_block_delta

375+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"` - Empty"} }

376+

377+ event: content_block_delta

378+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" main"}}

379+

380+ event: content_block_delta

381+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" function in test library"} }

382+

383+ event: content_block_delta

384+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\n\n**Metho"} }

385+

386+ event: content_block_delta

387+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d demonstration"} }

388+

389+ event: content_block_delta

390+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":":**\n- `github.com/"} }

391+

392+ event: content_block_delta

393+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"callicoder/golang-tutorials/"} }

394+

395+ event: content_block_delta

396+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"13-methods/`"} }

397+

398+ event: content_block_delta

399+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" - Multiple"} }

400+

401+ event: content_block_delta

402+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" examples"} }

403+

404+ event: content_block_delta

405+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" showing"} }

406+

407+ event: content_block_delta

408+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" methods"} }

409+

410+ event: content_block_delta

411+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" vs"} }

412+

413+ event: content_block_delta

414+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" functions in"} }

415+

416+ event: content_block_delta

417+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" Go\n\nThe"} }

418+

419+ event: content_block_delta

420+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" search"} }

421+

422+ event: content_block_delta

423+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" hit"} }

424+

425+ event: content_block_delta

426+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" the limit (30 matches"} }

427+

428+ event: content_block_delta

429+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"), indicating"} }

430+

431+ event: content_block_delta

432+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" `"} }

433+

434+ event: content_block_delta

435+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"func main` is very"} }

436+

437+ event: content_block_delta

438+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" common in Go repositories."} }

439+

440+ event: content_block_stop

441+ data: {"type":"content_block_stop","index":0 }

442+

443+ event: message_delta

444+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":4678,"cache_read_input_tokens":14028,"output_tokens":212} }

445+

446+ event: message_stop

447+ data: {"type":"message_stop" }

448+

449+ headers:

450+ Content-Type:

451+ - text/event-stream; charset=utf-8

452+ status: 200 OK

453+ code: 200

454+ duration: 2.250806s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/update_a_file.yamlnew file+314

@@ -0,0 +1,314 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 777

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nupdate the main.go file by changing the print to say hello from crush\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_0198CQ2JnP8dYvrKr27saGVX","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":145,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Update"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" main"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".go print"} }

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" message"} }

47+

48+ event: content_block_stop

49+ data: {"type":"content_block_stop","index":0 }

50+

51+ event: message_delta

52+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":145,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":9} }

53+

54+ event: message_stop

55+ data: {"type":"message_stop" }

56+

57+ headers:

58+ Content-Type:

59+ - text/event-stream; charset=utf-8

60+ status: 200 OK

61+ code: 200

62+ duration: 557.108833ms

63+- id: 1

64+ request:

65+ proto: HTTP/1.1

66+ proto_major: 1

67+ proto_minor: 1

68+ content_length: 52372

69+ host: ""

70+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"update the main.go file by changing the print to say hello from crush\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/update_a_file\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

71+ headers:

72+ Accept:

73+ - application/json

74+ Content-Type:

75+ - application/json

76+ User-Agent:

77+ - Anthropic/Go 1.14.0

78+ url: https://api.anthropic.com/v1/messages

79+ method: POST

80+ response:

81+ proto: HTTP/2.0

82+ proto_major: 2

83+ proto_minor: 0

84+ content_length: -1

85+ body: |+

86+ event: message_start

87+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_012phMFzjQWf6vHiicCKhg8G","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4538,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4538,"ephemeral_1h_input_tokens":0},"output_tokens":3,"service_tier":"standard"}} }

88+

89+ event: content_block_start

90+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

91+

92+ event: ping

93+ data: {"type": "ping"}

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"I'll rea"} }

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"d the main.go file first"} }

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":","} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" then update it."} }

106+

107+ event: content_block_stop

108+ data: {"type":"content_block_stop","index":0 }

109+

110+ event: content_block_start

111+ data: {"type":"content_block_start","index":1,"content_block":{"type":"tool_use","id":"toolu_01Wo6wBTug9S8KGxDe7uGt6o","name":"view","input":{}} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":""} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"{\"file_"} }

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":"path\": \"main"} }

121+

122+ event: content_block_delta

123+ data: {"type":"content_block_delta","index":1,"delta":{"type":"input_json_delta","partial_json":".go\"}"} }

124+

125+ event: content_block_stop

126+ data: {"type":"content_block_stop","index":1 }

127+

128+ event: message_delta

129+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4538,"cache_read_input_tokens":9490,"output_tokens":70} }

130+

131+ event: message_stop

132+ data: {"type":"message_stop" }

133+

134+ headers:

135+ Content-Type:

136+ - text/event-stream; charset=utf-8

137+ status: 200 OK

138+ code: 200

139+ duration: 2.076319s

140+- id: 2

141+ request:

142+ proto: HTTP/1.1

143+ proto_major: 1

144+ proto_minor: 1

145+ content_length: 52892

146+ host: ""

147+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"update the main.go file by changing the print to say hello from crush\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"text\":\"I'll read the main.go file first, then update it.\",\"type\":\"text\"},{\"id\":\"toolu_01Wo6wBTug9S8KGxDe7uGt6o\",\"input\":{\"file_path\":\"main.go\"},\"name\":\"view\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01Wo6wBTug9S8KGxDe7uGt6o\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\u003cfile\\u003e\\n 1|package main\\n 2|\\n 3|import \\\"fmt\\\"\\n 4|\\n 5|func main() {\\n 6|\\tfmt.Println(\\\"Hello, World!\\\")\\n 7|}\\n\\u003c/file\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/update_a_file\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

148+ headers:

149+ Accept:

150+ - application/json

151+ Content-Type:

152+ - application/json

153+ User-Agent:

154+ - Anthropic/Go 1.14.0

155+ url: https://api.anthropic.com/v1/messages

156+ method: POST

157+ response:

158+ proto: HTTP/2.0

159+ proto_major: 2

160+ proto_minor: 0

161+ content_length: -1

162+ body: |+

163+ event: message_start

164+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01E3AUv7t6VcuJTmnXVsCoHb","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":134,"cache_read_input_tokens":14028,"cache_creation":{"ephemeral_5m_input_tokens":134,"ephemeral_1h_input_tokens":0},"output_tokens":18,"service_tier":"standard"}} }

165+

166+ event: content_block_start

167+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01UhVFWneNPWYNZRcLBofsnJ","name":"edit","input":{}} }

168+

169+ event: ping

170+ data: {"type": "ping"}

171+

172+ event: content_block_delta

173+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

174+

175+ event: content_block_delta

176+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"file_path\""} }

177+

178+ event: content_block_delta

179+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":": \"main.g"} }

180+

181+ event: content_block_delta

182+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"o\""} }

183+

184+ event: content_block_delta

185+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \"ol"} }

186+

187+ event: content_block_delta

188+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"d_string\": "} }

189+

190+ event: content_block_delta

191+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\"func main()"} }

192+

193+ event: content_block_delta

194+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":" {\\n\\t"} }

195+

196+ event: content_block_delta

197+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"fmt."} }

198+

199+ event: content_block_delta

200+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"Println(\\\"He"} }

201+

202+ event: content_block_delta

203+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ll"} }

204+

205+ event: content_block_delta

206+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"o, World!"} }

207+

208+ event: content_block_delta

209+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\\\")\\n}\""} }

210+

211+ event: content_block_delta

212+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \"new_"}}

213+

214+ event: content_block_delta

215+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"str"} }

216+

217+ event: content_block_delta

218+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ing\":"} }

219+

220+ event: content_block_delta

221+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":" \"func "} }

222+

223+ event: content_block_delta

224+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ma"} }

225+

226+ event: content_block_delta

227+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"in("} }

228+

229+ event: content_block_delta

230+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":") {\\n"} }

231+

232+ event: content_block_delta

233+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\\tfmt.P"} }

234+

235+ event: content_block_delta

236+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"rintln(\\\"He"} }

237+

238+ event: content_block_delta

239+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"llo "} }

240+

241+ event: content_block_delta

242+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"fro"} }

243+

244+ event: content_block_delta

245+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"m Crus"} }

246+

247+ event: content_block_delta

248+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"h\\\")\\n}\"}"} }

249+

250+ event: content_block_stop

251+ data: {"type":"content_block_stop","index":0 }

252+

253+ event: message_delta

254+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":134,"cache_read_input_tokens":14028,"output_tokens":124} }

255+

256+ event: message_stop

257+ data: {"type":"message_stop" }

258+

259+ headers:

260+ Content-Type:

261+ - text/event-stream; charset=utf-8

262+ status: 200 OK

263+ code: 200

264+ duration: 2.838877416s

265+- id: 3

266+ request:

267+ proto: HTTP/1.1

268+ proto_major: 1

269+ proto_minor: 1

270+ content_length: 53438

271+ host: ""

272+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"update the main.go file by changing the print to say hello from crush\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"text\":\"I'll read the main.go file first, then update it.\",\"type\":\"text\"},{\"id\":\"toolu_01Wo6wBTug9S8KGxDe7uGt6o\",\"input\":{\"file_path\":\"main.go\"},\"name\":\"view\",\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01Wo6wBTug9S8KGxDe7uGt6o\",\"content\":[{\"text\":\"\\u003cfile\\u003e\\n 1|package main\\n 2|\\n 3|import \\\"fmt\\\"\\n 4|\\n 5|func main() {\\n 6|\\tfmt.Println(\\\"Hello, World!\\\")\\n 7|}\\n\\u003c/file\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01UhVFWneNPWYNZRcLBofsnJ\",\"input\":{\"file_path\":\"main.go\",\"new_string\":\"func main() {\\n\\tfmt.Println(\\\"Hello from Crush\\\")\\n}\",\"old_string\":\"func main() {\\n\\tfmt.Println(\\\"Hello, World!\\\")\\n}\"},\"name\":\"edit\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01UhVFWneNPWYNZRcLBofsnJ\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\u003cresult\\u003e\\nContent replaced in file: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/update_a_file/main.go\\n\\u003c/result\\u003e\\n\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/update_a_file\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

273+ headers:

274+ Accept:

275+ - application/json

276+ Content-Type:

277+ - application/json

278+ User-Agent:

279+ - Anthropic/Go 1.14.0

280+ url: https://api.anthropic.com/v1/messages

281+ method: POST

282+ response:

283+ proto: HTTP/2.0

284+ proto_major: 2

285+ proto_minor: 0

286+ content_length: -1

287+ body: |+

288+ event: message_start

289+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_016C3xzWvix2X7Hr9Y6X2xWf","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":5,"cache_creation_input_tokens":176,"cache_read_input_tokens":14162,"cache_creation":{"ephemeral_5m_input_tokens":176,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

290+

291+ event: content_block_start

292+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

293+

294+ event: ping

295+ data: {"type": "ping"}

296+

297+ event: content_block_delta

298+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Done"} }

299+

300+ event: content_block_stop

301+ data: {"type":"content_block_stop","index":0 }

302+

303+ event: message_delta

304+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":5,"cache_creation_input_tokens":176,"cache_read_input_tokens":14162,"output_tokens":4} }

305+

306+ event: message_stop

307+ data: {"type":"message_stop" }

308+

309+ headers:

310+ Content-Type:

311+ - text/event-stream; charset=utf-8

312+ status: 200 OK

313+ code: 200

314+ duration: 2.1880785s

internal/agent/testdata/TestCoderAgent/anthropic-sonnet/write_tool.yamlnew file+228

@@ -0,0 +1,228 @@

1+---

2+version: 2

3+interactions:

4+- id: 0

5+ request:

6+ proto: HTTP/1.1

7+ proto_major: 1

8+ proto_minor: 1

9+ content_length: 817

10+ host: ""

11+ body: '{"max_tokens":40,"messages":[{"content":[{"text":"Generate a concise title for the following content:\n\nuse write to create a new file called config.json with content ''{\"name\": \"test\", \"version\": \"1.0.0\"}''\n \u003cthink\u003e\n\n\u003c/think\u003e","type":"text"}],"role":"user"}],"model":"claude-3-5-haiku-20241022","system":[{"text":"you will generate a short title based on the first message a user begins a conversation with\n\n\u003crules\u003e\n- ensure it is not more than 50 characters long\n- the title should be a summary of the user''s message\n- it should be one line long\n- do not use quotes or colons\n- the entire text you return will be used as the title\n- never return anything that is more than one sentence (one line) long\n\u003c/rules\u003e\n\n /no_think","type":"text"}],"stream":true}'

12+ headers:

13+ Accept:

14+ - application/json

15+ Content-Type:

16+ - application/json

17+ User-Agent:

18+ - Anthropic/Go 1.14.0

19+ url: https://api.anthropic.com/v1/messages

20+ method: POST

21+ response:

22+ proto: HTTP/2.0

23+ proto_major: 2

24+ proto_minor: 0

25+ content_length: -1

26+ body: |+

27+ event: message_start

28+ data: {"type":"message_start","message":{"model":"claude-3-5-haiku-20241022","id":"msg_016VUcmtZyoNJiJysc1oxRtZ","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":161,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

29+

30+ event: content_block_start

31+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

32+

33+ event: ping

34+ data: {"type": "ping"}

35+

36+ event: content_block_delta

37+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Creating"} }

38+

39+ event: content_block_delta

40+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" config"} }

41+

42+ event: content_block_delta

43+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":".json with JSON"}}

44+

45+ event: content_block_delta

46+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":" data"} }

47+

48+ event: content_block_stop

49+ data: {"type":"content_block_stop","index":0 }

50+

51+ event: message_delta

52+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":161,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":10} }

53+

54+ event: message_stop

55+ data: {"type":"message_stop" }

56+

57+ headers:

58+ Content-Type:

59+ - text/event-stream; charset=utf-8

60+ status: 200 OK

61+ code: 200

62+ duration: 555.094125ms

63+- id: 1

64+ request:

65+ proto: HTTP/1.1

66+ proto_major: 1

67+ proto_minor: 1

68+ content_length: 52409

69+ host: ""

70+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"},{\"text\":\"use write to create a new file called config.json with content '{\\\"name\\\": \\\"test\\\", \\\"version\\\": \\\"1.0.0\\\"}'\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/write_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

71+ headers:

72+ Accept:

73+ - application/json

74+ Content-Type:

75+ - application/json

76+ User-Agent:

77+ - Anthropic/Go 1.14.0

78+ url: https://api.anthropic.com/v1/messages

79+ method: POST

80+ response:

81+ proto: HTTP/2.0

82+ proto_major: 2

83+ proto_minor: 0

84+ content_length: -1

85+ body: |+

86+ event: message_start

87+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01Dgqq36mPzsNyVqkBHtnYua","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":2,"cache_creation_input_tokens":4552,"cache_read_input_tokens":9490,"cache_creation":{"ephemeral_5m_input_tokens":4552,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

88+

89+ event: content_block_start

90+ data: {"type":"content_block_start","index":0,"content_block":{"type":"tool_use","id":"toolu_01DBdhpH1BHBEzKcLwfNYDJh","name":"write","input":{}} }

91+

92+ event: ping

93+ data: {"type": "ping"}

94+

95+ event: content_block_delta

96+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":""} }

97+

98+ event: content_block_delta

99+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\"file_"}}

100+

101+ event: content_block_delta

102+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"path\": \"/"} }

103+

104+ event: content_block_delta

105+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"tm"}}

106+

107+ event: content_block_delta

108+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"p/crush-te"} }

109+

110+ event: content_block_delta

111+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"st"} }

112+

113+ event: content_block_delta

114+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"/TestCoderA"} }

115+

116+ event: content_block_delta

117+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"gen"} }

118+

119+ event: content_block_delta

120+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"t/anthropic"} }

121+

122+ event: content_block_delta

123+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"-sonnet/wr"} }

124+

125+ event: content_block_delta

126+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ite_tool/con"} }

127+

128+ event: content_block_delta

129+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"fi"} }

130+

131+ event: content_block_delta

132+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"g.json\""}}

133+

134+ event: content_block_delta

135+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \""} }

136+

137+ event: content_block_delta

138+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"conte"} }

139+

140+ event: content_block_delta

141+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"nt\": \""} }

142+

143+ event: content_block_delta

144+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"{\\\""} }

145+

146+ event: content_block_delta

147+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"name\\"} }

148+

149+ event: content_block_delta

150+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"\": \\\"test\\\""} }

151+

152+ event: content_block_delta

153+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":", \\\"v"} }

154+

155+ event: content_block_delta

156+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"ersion\\\""} }

157+

158+ event: content_block_delta

159+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":": \\\"1.0."}}

160+

161+ event: content_block_delta

162+ data: {"type":"content_block_delta","index":0,"delta":{"type":"input_json_delta","partial_json":"0\\\"}\"}"} }

163+

164+ event: content_block_stop

165+ data: {"type":"content_block_stop","index":0 }

166+

167+ event: message_delta

168+ data: {"type":"message_delta","delta":{"stop_reason":"tool_use","stop_sequence":null},"usage":{"input_tokens":2,"cache_creation_input_tokens":4552,"cache_read_input_tokens":9490,"output_tokens":109} }

169+

170+ event: message_stop

171+ data: {"type":"message_stop" }

172+

173+ headers:

174+ Content-Type:

175+ - text/event-stream; charset=utf-8

176+ status: 200 OK

177+ code: 200

178+ duration: 1.323930875s

179+- id: 2

180+ request:

181+ proto: HTTP/1.1

182+ proto_major: 1

183+ proto_minor: 1

184+ content_length: 52941

185+ host: ""

186+ body: "{\"max_tokens\":10000,\"messages\":[{\"content\":[{\"text\":\"\\u003csystem_reminder\\u003eThis is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.\\u003c/system_reminder\\u003e\",\"type\":\"text\"},{\"text\":\"use write to create a new file called config.json with content '{\\\"name\\\": \\\"test\\\", \\\"version\\\": \\\"1.0.0\\\"}'\",\"type\":\"text\"}],\"role\":\"user\"},{\"content\":[{\"id\":\"toolu_01DBdhpH1BHBEzKcLwfNYDJh\",\"input\":{\"content\":\"{\\\"name\\\": \\\"test\\\", \\\"version\\\": \\\"1.0.0\\\"}\",\"file_path\":\"/tmp/crush-test/TestCoderAgent/anthropic-sonnet/write_tool/config.json\"},\"name\":\"write\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"tool_use\"}],\"role\":\"assistant\"},{\"content\":[{\"tool_use_id\":\"toolu_01DBdhpH1BHBEzKcLwfNYDJh\",\"cache_control\":{\"type\":\"ephemeral\"},\"content\":[{\"text\":\"\\u003cresult\\u003e\\nFile successfully written: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/write_tool/config.json\\n\\u003c/result\\u003e\",\"type\":\"text\"}],\"type\":\"tool_result\"}],\"role\":\"user\"}],\"model\":\"claude-sonnet-4-5-20250929\",\"system\":[{\"text\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n\\u003ccritical_rules\\u003e\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ BEFORE EDITING**: Never edit a file you haven't already read in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default \\u003c4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\".\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n\\u003c/critical_rules\\u003e\\n\\n\\u003ccommunication_style\\u003e\\nKeep responses minimal:\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n\\u003c/communication_style\\u003e\\n\\n\\u003ccode_references\\u003e\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n\\u003c/code_references\\u003e\\n\\n\\u003cworkflow\\u003e\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n\\u003c/workflow\\u003e\\n\\n\\u003cdecision_making\\u003e\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n\\u003c/decision_making\\u003e\\n\\n\\u003cediting_files\\u003e\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the file first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n\\u003c/editing_files\\u003e\\n\\n\\u003cwhitespace_and_exact_matching\\u003e\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n\\u003c/whitespace_and_exact_matching\\u003e\\n\\n\\u003ctask_completion\\u003e\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n\\u003c/task_completion\\u003e\\n\\n\\u003cerror_handling\\u003e\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n\\u003c/error_handling\\u003e\\n\\n\\u003cmemory_instructions\\u003e\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences \\n- Important codebase patterns\\n- Useful project information\\n\\u003c/memory_instructions\\u003e\\n\\n\\u003ccode_conventions\\u003e\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n\\u003c/code_conventions\\u003e\\n\\n\\u003ctesting\\u003e\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n\\u003c/testing\\u003e\\n\\n\\u003ctool_usage\\u003e\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n\\u003cbash_commands\\u003e\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `\\u0026` for background processes that won't stop on their own (e.g., `node server.js \\u0026`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status \\u0026\\u0026 git diff HEAD \\u0026\\u0026 git log -n 3`)\\n\\u003c/bash_commands\\u003e\\n\\u003c/tool_usage\\u003e\\n\\n\\u003cproactiveness\\u003e\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n\\u003c/proactiveness\\u003e\\n\\n\\u003cfinal_answers\\u003e\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n\\u003c/final_answers\\u003e\\n\\n\\u003cenv\\u003e\\nWorking directory: /tmp/crush-test/TestCoderAgent/anthropic-sonnet/write_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n\\u003c/env\\u003e\\n\\n\\n\\n\\n\",\"cache_control\":{\"type\":\"ephemeral\"},\"type\":\"text\"}],\"tool_choice\":{\"disable_parallel_tool_use\":false,\"type\":\"auto\"},\"tools\":[{\"input_schema\":{\"properties\":{\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"},\"name\":\"bash\",\"description\":\"Executes bash commands with automatic background conversion for long-running tasks.\\n\\n\\u003ccross_platform\\u003e\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n\\u003c/cross_platform\\u003e\\n\\n\\u003cexecution_steps\\u003e\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with \\u003ccwd\\u003e\\u003c/cwd\\u003e tags\\n\\u003c/execution_steps\\u003e\\n\\n\\u003cusage_notes\\u003e\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '\\u0026\\u0026', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n\\u003c/usage_notes\\u003e\\n\\n\\u003cbackground_execution\\u003e\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `\\u0026` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n\\u003c/background_execution\\u003e\\n\\n\\u003cgit_commits\\u003e\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in \\u003ccommit_analysis\\u003e tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond git context\\n - Draft concise (1-2 sentences) message focusing on \\\"why\\\" not \\\"what\\\"\\n - Use clear language, accurate reflection (\\\"add\\\"=new feature, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix)\\n - Avoid generic messages, review draft\\n\\n4. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n Commit message here.\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n\\n Co-Authored-By: Crush \\u003ccrush@charm.land\\u003e\\n\\n\\n EOF\\n )\\\"\\n\\n5. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n6. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response.\\n\\u003c/git_commits\\u003e\\n\\n\\u003cpull_requests\\u003e\\nUse gh command for ALL GitHub tasks. When user asks to create PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n3. Commit changes if needed\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in \\u003cpr_analysis\\u003e tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Clear, concise language\\n - Accurate reflection of changes and purpose\\n - Avoid generic summaries\\n - Review draft\\n\\n6. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat \\u0026lt;\\u0026lt;'EOF'\\n\\n ## Summary\\n\\n \\u0026lt;1-3 bullet points\\u003e\\n\\n ## Test plan\\n\\n [Checklist of TODOs...]\\n\\n\\n \U0001F498 Generated with Crush\\n\\n\\n EOF\\n )\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n\\u003c/pull_requests\\u003e\\n\\n\\u003cexamples\\u003e\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar \\u0026\\u0026 pytest tests\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"},\"name\":\"download\",\"description\":\"Downloads binary data from URL and saves to local file.\\n\\n\\u003cusage\\u003e\\n- Provide URL to download from\\n- Specify local file path where content should be saved\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Downloads any file type (binary or text)\\n- Auto-creates parent directories if missing\\n- Handles large files efficiently with streaming\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 100MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Will overwrite existing files without warning\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use absolute paths or paths relative to working directory\\n- Set appropriate timeouts for large files or slow connections\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"},\"name\":\"edit\",\"description\":\"Edits files by replacing text, creating new files, or deleting content. For moving/renaming use Bash 'mv'. For large edits use Write tool.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. For new files: Use LS tool to verify parent directory exists\\n3. **CRITICAL**: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. old_string: Text to replace (must match exactly including whitespace/indentation)\\n3. new_string: Replacement text\\n4. replace_all: Replace all occurrences (default false)\\n\\u003c/parameters\\u003e\\n\\n\\u003cspecial_cases\\u003e\\n\\n- Create file: provide file_path + new_string, leave old_string empty\\n- Delete content: provide file_path + old_string, leave new_string empty\\n \\u003c/special_cases\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\nEXACT MATCHING: The tool is extremely literal. Text must match **EXACTLY**\\n\\n- Every space and tab character\\n- Every blank line\\n- Every newline character\\n- Indentation level (count the spaces/tabs)\\n- Comment spacing (`// comment` vs `//comment`)\\n- Brace positioning (`func() {` vs `func(){`)\\n\\nCommon failures:\\n\\n```\\nExpected: \\\" func foo() {\\\" (4 spaces)\\nProvided: \\\" func foo() {\\\" (2 spaces) ❌ FAILS\\n\\nExpected: \\\"}\\\\n\\\\nfunc bar() {\\\" (2 newlines)\\nProvided: \\\"}\\\\nfunc bar() {\\\" (1 newline) ❌ FAILS\\n\\nExpected: \\\"// Comment\\\" (space after //)\\nProvided: \\\"//Comment\\\" (no space) ❌ FAILS\\n```\\n\\nUNIQUENESS (when replace_all=false): old_string MUST uniquely identify target instance\\n\\n- Include 3-5 lines context BEFORE and AFTER change point\\n- Include exact whitespace, indentation, surrounding code\\n- If text appears multiple times, add more context to make it unique\\n\\nSINGLE INSTANCE: Tool changes ONE instance when replace_all=false\\n\\n- For multiple instances: set replace_all=true OR make separate calls with unique context\\n- Plan calls carefully to avoid conflicts\\n\\nVERIFICATION BEFORE USING: Before every edit\\n\\n1. View the file and locate exact target location\\n2. Check how many instances of target text exist\\n3. Copy the EXACT text including all whitespace\\n4. Verify you have enough context for unique identification\\n5. Double-check indentation matches (count spaces/tabs)\\n6. Plan separate calls or use replace_all for multiple changes\\n \\u003c/critical_requirements\\u003e\\n\\n\\u003cwarnings\\u003e\\nTool fails if:\\n- old_string matches multiple locations and replace_all=false\\n- old_string doesn't match exactly (including whitespace)\\n- Insufficient context causes wrong instance change\\n- Indentation is off by even one space\\n- Missing or extra blank lines\\n- Wrong tabs vs spaces\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf you get \\\"old_string not found in file\\\":\\n\\n1. **View the file again** at the specific location\\n2. **Copy more context** - include entire function if needed\\n3. **Check whitespace**:\\n - Count indentation spaces/tabs\\n - Look for blank lines\\n - Check for trailing spaces\\n4. **Verify character-by-character** that your old_string matches\\n5. **Never guess** - always View the file to get exact text\\n \\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n\\n- Ensure edits result in correct, idiomatic code\\n- Don't leave code in broken state\\n- Use absolute file paths (starting with /)\\n- Use forward slashes (/) for cross-platform compatibility\\n- Multiple edits to same file: send all in single message with multiple tool calls\\n- **When in doubt, include MORE context rather than less**\\n- Match the existing code style exactly (spaces, tabs, blank lines)\\n \\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nBefore submitting an edit, verify:\\n\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if they exist\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3-5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all)\\n- [ ] Copied text character-for-character, not approximated\\n \\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Exact match with context\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n\\nnew_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n // New validation\\\\n if len(input) \\u003e 1000 {\\\\n return errors.New(\\\\\\\"input too long\\\\\\\")\\\\n }\\\\n return nil\\\\n}\\\"\\n```\\n\\n❌ Incorrect: Not enough context\\n\\n```\\nold_string: \\\"return nil\\\" // Appears many times!\\n```\\n\\n❌ Incorrect: Wrong indentation\\n\\n```\\nold_string: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 2 spaces\\n// But file actually has: \\\" if input == \\\\\\\"\\\\\\\" {\\\" // 4 spaces\\n```\\n\\n✅ Correct: Including context to make unique\\n\\n```\\nold_string: \\\"func ProcessData(input string) error {\\\\n if input == \\\\\\\"\\\\\\\" {\\\\n return errors.New(\\\\\\\"empty input\\\\\\\")\\\\n }\\\\n return nil\\\"\\n```\\n\\n\\u003c/examples\\u003e\\n\\n\\u003cwindows_notes\\u003e\\n\\n- Forward slashes work throughout (C:/path/file)\\n- File permissions handled automatically\\n- Line endings converted automatically (\\\\n ↔ \\\\r\\\\n)\\n \\u003c/windows_notes\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"},\"name\":\"multiedit\",\"description\":\"Makes multiple edits to a single file in one operation. Built on Edit tool for efficient multiple find-and-replace operations. Prefer over Edit tool for multiple edits to same file.\\n\\n\\u003cprerequisites\\u003e\\n1. Use View tool to understand file contents and context\\n2. Verify directory path is correct\\n3. CRITICAL: Note exact whitespace, indentation, and formatting from View output\\n\\u003c/prerequisites\\u003e\\n\\n\\u003cparameters\\u003e\\n1. file_path: Absolute path to file (required)\\n2. edits: Array of edit operations, each containing:\\n - old_string: Text to replace (must match exactly including whitespace/indentation)\\n - new_string: Replacement text\\n - replace_all: Replace all occurrences (optional, defaults to false)\\n\\u003c/parameters\\u003e\\n\\n\\u003coperation\\u003e\\n- Edits applied sequentially in provided order.\\n- Each edit operates on result of previous edit.\\n- PARTIAL SUCCESS: If some edits fail, successful edits are still applied. Failed edits are returned in the response.\\n- File is modified if at least one edit succeeds.\\n- Ideal for several changes to different parts of same file.\\n\\u003c/operation\\u003e\\n\\n\\u003cinherited_rules\\u003e\\nAll instructions from the Edit tool documentation apply verbatim to every edit item:\\n- Critical requirements for exact matching and uniqueness\\n- Warnings and common failures (tabs vs spaces, blank lines, brace placement, etc.)\\n- Verification steps before using, recovery steps, best practices, and whitespace checklist\\nUse the same level of precision as Edit. Multiedit often fails due to formatting mismatches—double-check whitespace for every edit.\\n\\u003c/inherited_rules\\u003e\\n\\n\\u003ccritical_requirements\\u003e\\n1. Apply Edit tool rules to EACH edit (see edit.md).\\n2. Edits are applied in order; successful edits are kept even if later edits fail.\\n3. Plan sequence carefully: earlier edits change the file content that later edits must match.\\n4. Ensure each old_string is unique at its application time (after prior edits).\\n5. Check the response for failed edits and retry them if needed.\\n\\u003c/critical_requirements\\u003e\\n\\n\\u003cverification_before_using\\u003e\\n1. View the file and copy exact text (including whitespace) for each target.\\n2. Check how many instances each old_string has BEFORE the sequence starts.\\n3. Dry-run mentally: after applying edit #N, will edit #N+1 still match? Adjust old_string/new_string accordingly.\\n4. Prefer fewer, larger context blocks over many tiny fragments that are easy to misalign.\\n5. If edits are independent, consider separate multiedit batches per logical region.\\n\\u003c/verification_before_using\\u003e\\n\\n\\u003cwarnings\\u003e\\n- Operation continues even if some edits fail; check response for failed edits.\\n- Earlier edits can invalidate later matches (added/removed spaces, lines, or reordered text).\\n- Mixed tabs/spaces, trailing spaces, or missing blank lines commonly cause failures.\\n- replace_all may affect unintended regions—use carefully or provide more context.\\n\\u003c/warnings\\u003e\\n\\n\\u003crecovery_steps\\u003e\\nIf some edits fail:\\n1. Check the response metadata for the list of failed edits with their error messages.\\n2. View the file again to see the current state after successful edits.\\n3. Adjust the failed edits based on the new file content.\\n4. Retry the failed edits with corrected old_string values.\\n5. Consider breaking complex batches into smaller, independent operations.\\n\\u003c/recovery_steps\\u003e\\n\\n\\u003cbest_practices\\u003e\\n- Ensure all edits result in correct, idiomatic code; don't leave code broken.\\n- Use absolute file paths (starting with /).\\n- Use replace_all only when you're certain; otherwise provide unique context.\\n- Match existing style exactly (spaces, tabs, blank lines).\\n- Review failed edits in the response and retry with corrections.\\n\\u003c/best_practices\\u003e\\n\\n\\u003cwhitespace_checklist\\u003e\\nFor EACH edit, verify:\\n- [ ] Viewed the file first\\n- [ ] Counted indentation spaces/tabs\\n- [ ] Included blank lines if present\\n- [ ] Matched brace/bracket positioning\\n- [ ] Included 3–5 lines of surrounding context\\n- [ ] Verified text appears exactly once (or using replace_all deliberately)\\n- [ ] Copied text character-for-character, not approximated\\n\\u003c/whitespace_checklist\\u003e\\n\\n\\u003cexamples\\u003e\\n✅ Correct: Sequential edits where the second match accounts for the first change\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n doNew()\\\\n}\\\",\\n },\\n {\\n // Uses context that still exists AFTER the first replacement\\n old_string: \\\"func B() {\\\\n callA()\\\\n}\\\",\\n new_string: \\\"func B() {\\\\n callA()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n❌ Incorrect: Second old_string no longer matches due to whitespace change introduced by the first edit\\n\\n```\\nedits: [\\n {\\n old_string: \\\"func A() {\\\\n doOld()\\\\n}\\\",\\n new_string: \\\"func A() {\\\\n\\\\n doNew()\\\\n}\\\", // Added extra blank line\\n },\\n {\\n old_string: \\\"func A() {\\\\n doNew()\\\\n}\\\", // Missing the new blank line, will FAIL\\n new_string: \\\"func A() {\\\\n doNew()\\\\n logChange()\\\\n}\\\",\\n },\\n]\\n```\\n\\n✅ Correct: Handling partial success\\n\\n```\\n// If edit 2 fails, edit 1 is still applied\\n// Response will indicate:\\n// - edits_applied: 1\\n// - edits_failed: [{index: 2, error: \\\"...\\\", edit: {...}}]\\n// You can then retry edit 2 with corrected context\\n```\\n\\u003c/examples\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"},\"name\":\"fetch\",\"description\":\"Fetches raw content from URL and returns it in specified format without any AI processing.\\n\\n\\u003cwhen_to_use\\u003e\\nUse this tool when you need:\\n- Raw, unprocessed content from a URL\\n- Direct access to API responses or JSON data\\n- HTML/text/markdown content without interpretation\\n- Simple, fast content retrieval without analysis\\n- To save tokens by avoiding AI processing\\n\\nDO NOT use this tool when you need to:\\n- Extract specific information from a webpage (use agentic_fetch instead)\\n- Answer questions about web content (use agentic_fetch instead)\\n- Analyze or summarize web pages (use agentic_fetch instead)\\n\\u003c/when_to_use\\u003e\\n\\n\\u003cusage\\u003e\\n- Provide URL to fetch content from\\n- Specify desired output format (text, markdown, or html)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Supports three output formats: text, markdown, html\\n- Auto-handles HTTP redirects\\n- Fast and lightweight - no AI processing\\n- Sets reasonable timeouts to prevent hanging\\n- Validates input parameters before requests\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max response size: 5MB\\n- Only supports HTTP and HTTPS protocols\\n- Cannot handle authentication or cookies\\n- Some websites may block automated requests\\n- Returns raw content only - no analysis or extraction\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use text format for plain text content or simple API responses\\n- Use markdown format for content that should be rendered with formatting\\n- Use html format when you need raw HTML structure\\n- Set appropriate timeouts for potentially slow websites\\n- If the user asks to analyze or extract from a page, use agentic_fetch instead\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"glob\",\"description\":\"Fast file pattern matching tool that finds files by name/pattern, returning paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide glob pattern to match against file paths\\n- Optional starting directory (defaults to current working directory)\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cpattern_syntax\\u003e\\n- '\\\\*' matches any sequence of non-separator characters\\n- '\\\\*\\\\*' matches any sequence including separators\\n- '?' matches any single non-separator character\\n- '[...]' matches any character in brackets\\n- '[!...]' matches any character not in brackets\\n\\u003c/pattern_syntax\\u003e\\n\\n\\u003cexamples\\u003e\\n- '*.js' - JavaScript files in current directory\\n- '**/*.js' - JavaScript files in any subdirectory\\n- 'src/**/*.{ts,tsx}' - TypeScript files in src directory\\n- '*.{html,css,js}' - HTML, CSS, and JS files\\n\\u003c/examples\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Does not search file contents (use Grep for that)\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Path separators handled automatically (/ and \\\\ work)\\n- Uses ripgrep (rg) if available, otherwise Go implementation\\n- Patterns should use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Combine with Grep: find files with Glob, search contents with Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine pattern if needed\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"},\"name\":\"grep\",\"description\":\"Fast content search tool that finds files containing specific text/patterns, returning matching paths sorted by modification time (newest first).\\n\\n\\u003cusage\\u003e\\n- Provide regex pattern to search within file contents\\n- Set literal_text=true for exact text with special characters (recommended for non-regex users)\\n- Optional starting directory (defaults to current working directory)\\n- Optional include pattern to filter which files to search\\n- Results sorted with most recently modified files first\\n\\u003c/usage\\u003e\\n\\n\\u003cregex_syntax\\u003e\\nWhen literal_text=false (supports standard regex):\\n\\n- 'function' searches for literal text \\\"function\\\"\\n- 'log\\\\..\\\\*Error' finds text starting with \\\"log.\\\" and ending with \\\"Error\\\"\\n- 'import\\\\s+.\\\\*\\\\s+from' finds import statements in JavaScript/TypeScript\\n\\u003c/regex_syntax\\u003e\\n\\n\\u003cinclude_patterns\\u003e\\n- '\\\\*.js' - Only search JavaScript files\\n- '\\\\*.{ts,tsx}' - Only search TypeScript files\\n- '\\\\*.go' - Only search Go files\\n\\u003c/include_patterns\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 100 files (newest first)\\n- Performance depends on number of files searched\\n- Very large binary files may be skipped\\n- Hidden files (starting with '.') skipped\\n\\u003c/limitations\\u003e\\n\\n\\u003cignore_support\\u003e\\n- Respects .gitignore patterns to skip ignored files/directories\\n- Respects .crushignore patterns for additional ignore rules\\n- Both ignore files auto-detected in search root directory\\n\\u003c/ignore_support\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Uses ripgrep (rg) if available for better performance\\n- Falls back to Go implementation if ripgrep unavailable\\n- File paths normalized automatically for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- For faster searches: use Glob to find relevant files first, then Grep\\n- For iterative exploration requiring multiple searches, consider Agent tool\\n- Check if results truncated and refine search pattern if needed\\n- Use literal_text=true for exact text with special characters (dots, parentheses, etc.)\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"},\"name\":\"ls\",\"description\":\"Shows files and subdirectories in tree structure for exploring project organization.\\n\\n\\u003cusage\\u003e\\n- Provide path to list (defaults to current working directory)\\n- Optional glob patterns to ignore\\n- Results displayed in tree structure\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Hierarchical view of files and directories\\n- Auto-skips hidden files/directories (starting with '.')\\n- Skips common system directories like __pycache__\\n- Can filter files matching specific patterns\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Results limited to 1000 files\\n- Large directories truncated\\n- No file sizes or permissions shown\\n- Cannot recursively list all directories in large projects\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Hidden file detection uses Unix convention (files starting with '.')\\n- Windows hidden files (with hidden attribute) not auto-skipped\\n- Common Windows directories (System32, Program Files) not in default ignore\\n- Path separators handled automatically (/ and \\\\ work)\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use Glob for finding files by name patterns instead of browsing\\n- Use Grep for searching file contents\\n- Combine with other tools for effective exploration\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"},\"name\":\"sourcegraph\",\"description\":\"Search code across public repositories using Sourcegraph's GraphQL API.\\n\\n\\u003cusage\\u003e\\n- Provide search query using Sourcegraph syntax\\n- Optional result count (default: 10, max: 20)\\n- Optional timeout for request\\n\\u003c/usage\\u003e\\n\\n\\u003cbasic_syntax\\u003e\\n- \\\"fmt.Println\\\" - exact matches\\n- \\\"file:.go fmt.Println\\\" - limit to Go files\\n- \\\"repo:^github\\\\.com/golang/go$ fmt.Println\\\" - specific repos\\n- \\\"lang:go fmt.Println\\\" - limit to Go code\\n- \\\"fmt.Println AND log.Fatal\\\" - combined terms\\n- \\\"fmt\\\\.(Print|Printf|Println)\\\" - regex patterns\\n- \\\"\\\\\\\"exact phrase\\\\\\\"\\\" - exact phrase matching\\n- \\\"-file:test\\\" or \\\"-repo:forks\\\" - exclude matches\\n\\u003c/basic_syntax\\u003e\\n\\n\\u003ckey_filters\\u003e\\nRepository: repo:name, repo:^exact$, repo:org/repo@branch, -repo:exclude, fork:yes, archived:yes, visibility:public\\nFile: file:\\\\.js$, file:internal/, -file:test, file:has.content(text)\\nContent: content:\\\"exact\\\", -content:\\\"unwanted\\\", case:yes\\nType: type:symbol, type:file, type:path, type:diff, type:commit\\nTime: after:\\\"1 month ago\\\", before:\\\"2023-01-01\\\", author:name, message:\\\"fix\\\"\\nResult: select:repo, select:file, select:content, count:100, timeout:30s\\n\\u003c/key_filters\\u003e\\n\\n\\u003cexamples\\u003e\\n- \\\"file:.go context.WithTimeout\\\" - Go code using context.WithTimeout\\n- \\\"lang:typescript useState type:symbol\\\" - TypeScript React useState hooks\\n- \\\"repo:^github\\\\.com/kubernetes/kubernetes$ pod list type:file\\\" - Kubernetes pod files\\n- \\\"file:Dockerfile (alpine OR ubuntu) -content:alpine:latest\\\" - Dockerfiles with base images\\n\\u003c/examples\\u003e\\n\\n\\u003cboolean_operators\\u003e\\n- \\\"term1 AND term2\\\" - both terms\\n- \\\"term1 OR term2\\\" - either term\\n- \\\"term1 NOT term2\\\" - term1 but not term2\\n- \\\"term1 and (term2 or term3)\\\" - grouping with parentheses\\n\\u003c/boolean_operators\\u003e\\n\\n\\u003climitations\\u003e\\n- Only searches public repositories\\n- Rate limits may apply\\n- Complex queries take longer\\n- Max 20 results per query\\n\\u003c/limitations\\u003e\\n\\n\\u003ctips\\u003e\\n- Use specific file extensions to narrow results\\n- Add repo: filters for targeted searches\\n- Use type:symbol for function/method definitions\\n- Use type:file to find relevant files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 2000)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"},\"name\":\"view\",\"description\":\"Reads and displays file contents with line numbers for examining code, logs, or text data.\\n\\n\\u003cusage\\u003e\\n- Provide file path to read\\n- Optional offset: start reading from specific line (0-based)\\n- Optional limit: control lines read (default 2000)\\n- Don't use for directories (use LS tool instead)\\n- Supports image files (PNG, JPEG, GIF, BMP, SVG, WebP)\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Displays contents with line numbers\\n- Can read from any file position using offset\\n- Handles large files by limiting lines read\\n- Auto-truncates very long lines for display\\n- Suggests similar filenames when file not found\\n- Renders image files directly in terminal\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Max file size: 5MB\\n- Default limit: 2000 lines\\n- Lines \\u003e2000 chars truncated\\n- Binary files (except images) cannot be displayed\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Handles Windows (CRLF) and Unix (LF) line endings\\n- Works with forward slashes (/) and backslashes (\\\\)\\n- Auto-detects text encoding for common formats\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use with Glob to find files first\\n- For code exploration: Grep to find relevant files, then View to examine\\n- For large files: use offset parameter for specific sections\\n- View tool automatically detects and renders image files\\n\\u003c/tips\\u003e\\n\"},{\"input_schema\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"},\"name\":\"write\",\"description\":\"Creates or updates files in filesystem for saving/modifying text content.\\n\\n\\u003cusage\\u003e\\n- Provide file path to write\\n- Include content to write to file\\n- Tool creates necessary parent directories automatically\\n\\u003c/usage\\u003e\\n\\n\\u003cfeatures\\u003e\\n- Creates new files or overwrites existing ones\\n- Auto-creates parent directories if missing\\n- Checks if file modified since last read for safety\\n- Avoids unnecessary writes when content unchanged\\n\\u003c/features\\u003e\\n\\n\\u003climitations\\u003e\\n- Read file before writing to avoid conflicts\\n- Cannot append (rewrites entire file)\\n\\u003c/limitations\\u003e\\n\\n\\u003ccross_platform\\u003e\\n- Use forward slashes (/) for compatibility\\n\\u003c/cross_platform\\u003e\\n\\n\\u003ctips\\u003e\\n- Use View tool first to examine existing files before modifying\\n- Use LS tool to verify location when creating new files\\n- Combine with Glob/Grep to find and modify multiple files\\n- Include descriptive comments when changing existing code\\n\\u003c/tips\\u003e\\n\",\"cache_control\":{\"type\":\"ephemeral\"}}],\"stream\":true}"

187+ headers:

188+ Accept:

189+ - application/json

190+ Content-Type:

191+ - application/json

192+ User-Agent:

193+ - Anthropic/Go 1.14.0

194+ url: https://api.anthropic.com/v1/messages

195+ method: POST

196+ response:

197+ proto: HTTP/2.0

198+ proto_major: 2

199+ proto_minor: 0

200+ content_length: -1

201+ body: |+

202+ event: message_start

203+ data: {"type":"message_start","message":{"model":"claude-sonnet-4-5-20250929","id":"msg_01Pwt34S6JT6HPnxxemHPT1P","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":4,"cache_creation_input_tokens":155,"cache_read_input_tokens":14042,"cache_creation":{"ephemeral_5m_input_tokens":155,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard"}} }

204+

205+ event: content_block_start

206+ data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""} }

207+

208+ event: ping

209+ data: {"type": "ping"}

210+

211+ event: content_block_delta

212+ data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Done"} }

213+

214+ event: content_block_stop

215+ data: {"type":"content_block_stop","index":0 }

216+

217+ event: message_delta

218+ data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":4,"cache_creation_input_tokens":155,"cache_read_input_tokens":14042,"output_tokens":4} }

219+

220+ event: message_stop

221+ data: {"type":"message_stop" }

222+

223+ headers:

224+ Content-Type:

225+ - text/event-stream; charset=utf-8

226+ status: 200 OK

227+ code: 200

228+ duration: 1.744400375s

internal/agent/testdata/TestCoderAgent/glm-5.1/bash_tool.yamldeleted−425

@@ -1,425 +0,0 @@

1−---

2−version: 2

3−interactions:

4−- id: 0

5− request:

6− proto: HTTP/1.1

7− proto_major: 1

8− proto_minor: 1

9− content_length: 853

10− host: ""

11− body: '{"messages":[{"content":"You will generate a short title based on the first message a user begins a conversation with.\n\n<rules>\n- Keep the title in the same language that the user wrote their message in.\n- Ensure it is not more than 50 characters long.\n- The title should be a summary of the user''s message.\n- It should be one line long.\n- Do not use quotes or colons.\n- The entire text you return will be used as the title.\n- Never return anything that is more than one sentence (one line) long.\n</rules>\n\n /no_think","role":"system"},{"content":"Generate a concise title for the following content:\n\nuse bash to create a file named test.txt with content ''hello bash''. do not print its timestamp\n <think>\n\n</think>","role":"user"}],"model":"gpt-oss-120b","max_completion_tokens":40,"stream_options":{"include_usage":true},"stream":true}'

12− headers:

13− Accept:

14− - application/json

15− Content-Type:

16− - application/json

17− User-Agent:

18− - Charm-Crush/devel (https://charm.land/crush)

19− url: https://hyper.charm.land/v1/chat/completions

20− method: POST

21− response:

22− proto: HTTP/2.0

23− proto_major: 2

24− proto_minor: 0

25− content_length: -1

26− uncompressed: true

27− body: |+

28− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

29−

30− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"We"},"finish_reason":null}]}

31−

32− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" need"},"finish_reason":null}]}

33−

34− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

35−

36− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" generate"},"finish_reason":null}]}

37−

38− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

39−

40− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" short"},"finish_reason":null}]}

41−

42− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

43−

44− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" based"},"finish_reason":null}]}

45−

46− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" on"},"finish_reason":null}]}

47−

48− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

49−

50− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" user's"},"finish_reason":null}]}

51−

52− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" first"},"finish_reason":null}]}

53−

54− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" message"},"finish_reason":null}]}

55−

56− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

57−

58− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" The"},"finish_reason":null}]}

59−

60− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

61−

62− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" wrote"},"finish_reason":null}]}

63−

64− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

65−

66− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" English"},"finish_reason":null}]}

67−

68− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

69−

70− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Title"},"finish_reason":null}]}

71−

72− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" must"},"finish_reason":null}]}

73−

74− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" be"},"finish_reason":null}]}

75−

76− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \u003c="},"finish_reason":null}]}

77−

78− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

79−

80− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

81−

82− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

83−

84− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

85−

86− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" quotes"},"finish_reason":null}]}

87−

88− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"/"},"finish_reason":null}]}

89−

90− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"col"},"finish_reason":null}]}

91−

92− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ons"},"finish_reason":null}]}

93−

94− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

95−

96− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

97−

98− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" line"},"finish_reason":null}]}

99−

100− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

101−

102− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

103−

104− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" more"},"finish_reason":null}]}

105−

106− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" than"},"finish_reason":null}]}

107−

108− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

109−

110− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" sentence"},"finish_reason":null}]}

111−

112− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

113−

114− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Summ"},"finish_reason":null}]}

115−

116− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ar"},"finish_reason":null}]}

117−

118− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ize"},"finish_reason":null}]}

119−

120− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

121−

122− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

123−

124− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"bash"},"finish_reason":null}]}

125−

126− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" create"},"finish_reason":null}]}

127−

128− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" test"},"finish_reason":null}]}

129−

130− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

131−

132− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

133−

134− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" hello"},"finish_reason":null}]}

135−

136− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" bash"},"finish_reason":null}]}

137−

138− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" without"},"finish_reason":null}]}

139−

140− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" timestamp"},"finish_reason":null}]}

141−

142− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

143−

144− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Let's"},"finish_reason":null}]}

145−

146− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" craft"},"finish_reason":null}]}

147−

148− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

149−

150− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

151−

152− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Create"},"finish_reason":null}]}

153−

154− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" test"},"finish_reason":null}]}

155−

156− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

157−

158− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

159−

160− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" hello"},"finish_reason":null}]}

161−

162− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" bash"},"finish_reason":null}]}

163−

164− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" using"},"finish_reason":null}]}

165−

166− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Bash"},"finish_reason":null}]}

167−

168− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

169−

170− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" That's"},"finish_reason":null}]}

171−

172− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" under"},"finish_reason":null}]}

173−

174− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

175−

176− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

177−

178− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

179−

180− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

181−

182− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

183−

184− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

185−

186− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Create"},"finish_reason":null}]}

187−

188− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" test"},"finish_reason":null}]}

189−

190− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

191−

192− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

193−

194− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" hello"},"finish_reason":null}]}

195−

196− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" bash"},"finish_reason":null}]}

197−

198− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" using"},"finish_reason":null}]}

199−

200− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Bash"},"finish_reason":null}]}

201−

202− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\""},"finish_reason":null}]}

203−

204− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

205−

206− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

207−

208− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"38"},"finish_reason":null}]}

209−

210− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

211−

212− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

213−

214− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" No"},"finish_reason":null}]}

215−

216− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" quotes"},"finish_reason":null}]}

217−

218− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

219−

220− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Good"},"finish_reason":null}]}

221−

222− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

223−

224− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"Create"},"finish_reason":null}]}

225−

226− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" test"},"finish_reason":null}]}

227−

228− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":".txt"},"finish_reason":null}]}

229−

230− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" with"},"finish_reason":null}]}

231−

232− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" hello"},"finish_reason":null}]}

233−

234− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" bash"},"finish_reason":null}]}

235−

236− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" using"},"finish_reason":null}]}

237−

238− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" Bash"},"finish_reason":null}]}

239−

240− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

241−

242− data: {"id":"chatcmpl-9b4429c7811fb2c7070cc503","object":"chat.completion.chunk","created":1782413792,"model":"gpt-oss-120b","choices":[],"usage":{"prompt_tokens":224,"completion_tokens":115,"total_tokens":339,"completion_tokens_details":{"reasoning_tokens":98},"cost":{"usd":0.00015733800000000003,"hypercredits":0.0031467600000000006},"remaining":{"hypercredits":0}}}

243−

244− data: [DONE]

245−

246− headers:

247− Content-Type:

248− - text/event-stream

249− status: 200 OK

250− code: 200

251− duration: 700.546792ms

252−- id: 1

253− request:

254− proto: HTTP/1.1

255− proto_major: 1

256− proto_minor: 1

257− content_length: 35672

258− host: ""

259− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/bash_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use bash to create a file named test.txt with content 'hello bash'. do not print its timestamp\",\"role\":\"user\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

260− headers:

261− Accept:

262− - application/json

263− Content-Type:

264− - application/json

265− User-Agent:

266− - Charm-Crush/devel (https://charm.land/crush)

267− url: https://hyper.charm.land/v1/chat/completions

268− method: POST

269− response:

270− proto: HTTP/2.0

271− proto_major: 2

272− proto_minor: 0

273− content_length: -1

274− uncompressed: true

275− body: |+

276− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

277−

278− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"The"},"finish_reason":null}]}

279−

280− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

281−

282− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" wants"},"finish_reason":null}]}

283−

284− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" me"},"finish_reason":null}]}

285−

286− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

287−

288− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" create"},"finish_reason":null}]}

289−

290− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

291−

292− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" file"},"finish_reason":null}]}

293−

294− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" named"},"finish_reason":null}]}

295−

296− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" test"},"finish_reason":null}]}

297−

298− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

299−

300− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

301−

302− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" content"},"finish_reason":null}]}

303−

304− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

305−

306− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"hello"},"finish_reason":null}]}

307−

308− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" bash"},"finish_reason":null}]}

309−

310− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

311−

312− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" using"},"finish_reason":null}]}

313−

314− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

315−

316− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" bash"},"finish_reason":null}]}

317−

318− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" tool"},"finish_reason":null}]}

319−

320− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

321−

322− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" and"},"finish_reason":null}]}

323−

324− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" not"},"finish_reason":null}]}

325−

326− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" print"},"finish_reason":null}]}

327−

328− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" its"},"finish_reason":null}]}

329−

330− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" timestamp"},"finish_reason":null}]}

331−

332− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

333−

334− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_51c682620b31061c","type":"function","index":0,"function":{"name":"bash","arguments":""}}]},"finish_reason":null}]}

335−

336− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"command\": \""}}]},"finish_reason":null}]}

337−

338− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"echo"}}]},"finish_reason":null}]}

339−

340− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" '"}}]},"finish_reason":null}]}

341−

342− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"hello"}}]},"finish_reason":null}]}

343−

344− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" bash"}}]},"finish_reason":null}]}

345−

346− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"'"}}]},"finish_reason":null}]}

347−

348− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" \u003e"}}]},"finish_reason":null}]}

349−

350− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" test"}}]},"finish_reason":null}]}

351−

352− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".txt"}}]},"finish_reason":null}]}

353−

354− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

355−

356− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":", \"description\": \""}}]},"finish_reason":null}]}

357−

358− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Create"}}]},"finish_reason":null}]}

359−

360− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" test"}}]},"finish_reason":null}]}

361−

362− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".txt"}}]},"finish_reason":null}]}

363−

364− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" with"}}]},"finish_reason":null}]}

365−

366− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":" content"}}]},"finish_reason":null}]}

367−

368− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

369−

370− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"}"}}]},"finish_reason":null}]}

371−

372− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

373−

374− data: {"id":"chatcmpl-5baef5f0d9ce99f46bc74840","object":"chat.completion.chunk","created":1782413793,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":686,"completion_tokens":56,"total_tokens":8221,"prompt_tokens_details":{"cached_tokens":7479},"completion_tokens_details":{"reasoning_tokens":32},"cost":{"usd":0.0013476,"hypercredits":0.026952},"remaining":{"hypercredits":0}}}

375−

376− data: [DONE]

377−

378− headers:

379− Content-Type:

380− - text/event-stream

381− status: 200 OK

382− code: 200

383− duration: 1.461643167s

384−- id: 2

385− request:

386− proto: HTTP/1.1

387− proto_major: 1

388− proto_minor: 1

389− content_length: 36122

390− host: ""

391− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/bash_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use bash to create a file named test.txt with content 'hello bash'. do not print its timestamp\",\"role\":\"user\"},{\"tool_calls\":[{\"id\":\"call_51c682620b31061c\",\"function\":{\"arguments\":\"{\\\"command\\\": \\\"echo 'hello bash' > test.txt\\\", \\\"description\\\": \\\"Create test.txt with content\\\"}\",\"name\":\"bash\"},\"type\":\"function\"}],\"role\":\"assistant\",\"reasoning_content\":\"The user wants me to create a file named test.txt with content 'hello bash' using the bash tool, and not print its timestamp.\"},{\"content\":\"no output\",\"tool_call_id\":\"call_51c682620b31061c\",\"role\":\"tool\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

392− headers:

393− Accept:

394− - application/json

395− Content-Type:

396− - application/json

397− User-Agent:

398− - Charm-Crush/devel (https://charm.land/crush)

399− url: https://hyper.charm.land/v1/chat/completions

400− method: POST

401− response:

402− proto: HTTP/2.0

403− proto_major: 2

404− proto_minor: 0

405− content_length: -1

406− uncompressed: true

407− body: |+

408− data: {"id":"chatcmpl-985f3df336cbf13a5d358129","object":"chat.completion.chunk","created":1782413794,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

409−

410− data: {"id":"chatcmpl-985f3df336cbf13a5d358129","object":"chat.completion.chunk","created":1782413794,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"Done"},"finish_reason":null}]}

411−

412− data: {"id":"chatcmpl-985f3df336cbf13a5d358129","object":"chat.completion.chunk","created":1782413794,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"."},"finish_reason":null}]}

413−

414− data: {"id":"chatcmpl-985f3df336cbf13a5d358129","object":"chat.completion.chunk","created":1782413794,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

415−

416− data: {"id":"chatcmpl-985f3df336cbf13a5d358129","object":"chat.completion.chunk","created":1782413794,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":720,"completion_tokens":4,"total_tokens":8203,"prompt_tokens_details":{"cached_tokens":7479},"cost":{"usd":0.0010255999999999998,"hypercredits":0.020511999999999996},"remaining":{"hypercredits":0}}}

417−

418− data: [DONE]

419−

420− headers:

421− Content-Type:

422− - text/event-stream

423− status: 200 OK

424− code: 200

425− duration: 1.122182833s

internal/agent/testdata/TestCoderAgent/glm-5.1/download_tool.yamldeleted−483

@@ -1,483 +0,0 @@

1−---

2−version: 2

3−interactions:

4−- id: 0

5− request:

6− proto: HTTP/1.1

7− proto_major: 1

8− proto_minor: 1

9− content_length: 35697

10− host: ""

11− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/download_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"download the file from https://example-files.online-convert.com/document/txt/example.txt and save it as example.txt\",\"role\":\"user\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

12− headers:

13− Accept:

14− - application/json

15− Content-Type:

16− - application/json

17− User-Agent:

18− - Charm-Crush/devel (https://charm.land/crush)

19− url: https://hyper.charm.land/v1/chat/completions

20− method: POST

21− response:

22− proto: HTTP/2.0

23− proto_major: 2

24− proto_minor: 0

25− content_length: -1

26− uncompressed: true

27− body: |+

28− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

29−

30− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"The"},"finish_reason":null}]}

31−

32− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

33−

34− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" wants"},"finish_reason":null}]}

35−

36− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

37−

38− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" download"},"finish_reason":null}]}

39−

40− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

41−

42− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" file"},"finish_reason":null}]}

43−

44− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" from"},"finish_reason":null}]}

45−

46− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

47−

48− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" URL"},"finish_reason":null}]}

49−

50− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" and"},"finish_reason":null}]}

51−

52− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" save"},"finish_reason":null}]}

53−

54− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" it"},"finish_reason":null}]}

55−

56− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" as"},"finish_reason":null}]}

57−

58− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

59−

60− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

61−

62− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

63−

64− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_4cbf337c01e5b47c","type":"function","index":0,"function":{"name":"download","arguments":""}}]},"finish_reason":null}]}

65−

66− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"file_path\": \""}}]},"finish_reason":null}]}

67−

68− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/tmp"}}]},"finish_reason":null}]}

69−

70− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/c"}}]},"finish_reason":null}]}

71−

72− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"rush"}}]},"finish_reason":null}]}

73−

74− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-test"}}]},"finish_reason":null}]}

75−

76− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/Test"}}]},"finish_reason":null}]}

77−

78− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Coder"}}]},"finish_reason":null}]}

79−

80− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Agent"}}]},"finish_reason":null}]}

81−

82− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/g"}}]},"finish_reason":null}]}

83−

84− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"lm"}}]},"finish_reason":null}]}

85−

86− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-"}}]},"finish_reason":null}]}

87−

88− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"5"}}]},"finish_reason":null}]}

89−

90− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"."}}]},"finish_reason":null}]}

91−

92− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"1"}}]},"finish_reason":null}]}

93−

94− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/download"}}]},"finish_reason":null}]}

95−

96− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"_tool"}}]},"finish_reason":null}]}

97−

98− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/example"}}]},"finish_reason":null}]}

99−

100− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".txt"}}]},"finish_reason":null}]}

101−

102− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

103−

104− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":", \"url\": \""}}]},"finish_reason":null}]}

105−

106− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"https"}}]},"finish_reason":null}]}

107−

108− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"://"}}]},"finish_reason":null}]}

109−

110− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"example"}}]},"finish_reason":null}]}

111−

112− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-files"}}]},"finish_reason":null}]}

113−

114− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".online"}}]},"finish_reason":null}]}

115−

116− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-"}}]},"finish_reason":null}]}

117−

118− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"convert"}}]},"finish_reason":null}]}

119−

120− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".com"}}]},"finish_reason":null}]}

121−

122− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/document"}}]},"finish_reason":null}]}

123−

124− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/t"}}]},"finish_reason":null}]}

125−

126− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"xt"}}]},"finish_reason":null}]}

127−

128− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/example"}}]},"finish_reason":null}]}

129−

130− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".txt"}}]},"finish_reason":null}]}

131−

132− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

133−

134− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"}"}}]},"finish_reason":null}]}

135−

136− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

137−

138− data: {"id":"chatcmpl-cc0c9ef25b702fedc808d605","object":"chat.completion.chunk","created":1782413796,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":689,"completion_tokens":63,"total_tokens":8231,"prompt_tokens_details":{"cached_tokens":7479},"completion_tokens_details":{"reasoning_tokens":18},"cost":{"usd":0.001321,"hypercredits":0.026419999999999996},"remaining":{"hypercredits":0}}}

139−

140− data: [DONE]

141−

142− headers:

143− Content-Type:

144− - text/event-stream

145− status: 200 OK

146− code: 200

147− duration: 936.027666ms

148−- id: 1

149− request:

150− proto: HTTP/1.1

151− proto_major: 1

152− proto_minor: 1

153− content_length: 874

154− host: ""

155− body: '{"messages":[{"content":"You will generate a short title based on the first message a user begins a conversation with.\n\n<rules>\n- Keep the title in the same language that the user wrote their message in.\n- Ensure it is not more than 50 characters long.\n- The title should be a summary of the user''s message.\n- It should be one line long.\n- Do not use quotes or colons.\n- The entire text you return will be used as the title.\n- Never return anything that is more than one sentence (one line) long.\n</rules>\n\n /no_think","role":"system"},{"content":"Generate a concise title for the following content:\n\ndownload the file from https://example-files.online-convert.com/document/txt/example.txt and save it as example.txt\n <think>\n\n</think>","role":"user"}],"model":"gpt-oss-120b","max_completion_tokens":40,"stream_options":{"include_usage":true},"stream":true}'

156− headers:

157− Accept:

158− - application/json

159− Content-Type:

160− - application/json

161− User-Agent:

162− - Charm-Crush/devel (https://charm.land/crush)

163− url: https://hyper.charm.land/v1/chat/completions

164− method: POST

165− response:

166− proto: HTTP/2.0

167− proto_major: 2

168− proto_minor: 0

169− content_length: -1

170− uncompressed: true

171− body: |+

172− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

173−

174− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"The"},"finish_reason":null}]}

175−

176− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

177−

178− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" asks"},"finish_reason":null}]}

179−

180− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

181−

182− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

183−

184− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Generate"},"finish_reason":null}]}

185−

186− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

187−

188− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" concise"},"finish_reason":null}]}

189−

190− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

191−

192− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

193−

194− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

195−

196− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" following"},"finish_reason":null}]}

197−

198− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" content"},"finish_reason":null}]}

199−

200− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

201−

202− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" download"},"finish_reason":null}]}

203−

204− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

205−

206− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" file"},"finish_reason":null}]}

207−

208− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" from"},"finish_reason":null}]}

209−

210− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" https"},"finish_reason":null}]}

211−

212− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"://"},"finish_reason":null}]}

213−

214− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"example"},"finish_reason":null}]}

215−

216− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"-files"},"finish_reason":null}]}

217−

218− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".online"},"finish_reason":null}]}

219−

220− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"-con"},"finish_reason":null}]}

221−

222− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"vert"},"finish_reason":null}]}

223−

224− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".com"},"finish_reason":null}]}

225−

226− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"/document"},"finish_reason":null}]}

227−

228− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"/t"},"finish_reason":null}]}

229−

230− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"xt"},"finish_reason":null}]}

231−

232− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"/example"},"finish_reason":null}]}

233−

234− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

235−

236− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" and"},"finish_reason":null}]}

237−

238− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" save"},"finish_reason":null}]}

239−

240− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" it"},"finish_reason":null}]}

241−

242− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" as"},"finish_reason":null}]}

243−

244− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

245−

246− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

247−

248− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

249−

250− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Need"},"finish_reason":null}]}

251−

252− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

253−

254− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" short"},"finish_reason":null}]}

255−

256− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

257−

258− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

259−

260− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" same"},"finish_reason":null}]}

261−

262− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" language"},"finish_reason":null}]}

263−

264− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

265−

266− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"English"},"finish_reason":null}]}

267−

268− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"),"},"finish_reason":null}]}

269−

270− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \u003c="},"finish_reason":null}]}

271−

272− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

273−

274− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" chars"},"finish_reason":null}]}

275−

276− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

277−

278− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

279−

280− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" line"},"finish_reason":null}]}

281−

282− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

283−

284− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

285−

286− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" quotes"},"finish_reason":null}]}

287−

288− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

289−

290− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

291−

292− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" colon"},"finish_reason":null}]}

293−

294− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

295−

296− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Something"},"finish_reason":null}]}

297−

298− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" like"},"finish_reason":null}]}

299−

300− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

301−

302− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Download"},"finish_reason":null}]}

303−

304− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" and"},"finish_reason":null}]}

305−

306− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Save"},"finish_reason":null}]}

307−

308− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

309−

310− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

311−

312− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" from"},"finish_reason":null}]}

313−

314− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" URL"},"finish_reason":null}]}

315−

316− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

317−

318− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

319−

320− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

321−

322− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

323−

324− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

325−

326− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Download"},"finish_reason":null}]}

327−

328− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" and"},"finish_reason":null}]}

329−

330− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Save"},"finish_reason":null}]}

331−

332− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

333−

334− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".txt"},"finish_reason":null}]}

335−

336− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" from"},"finish_reason":null}]}

337−

338− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" URL"},"finish_reason":null}]}

339−

340− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\""},"finish_reason":null}]}

341−

342− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

343−

344− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

345−

346− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"38"},"finish_reason":null}]}

347−

348− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

349−

350− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

351−

352− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Good"},"finish_reason":null}]}

353−

354− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

355−

356− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"Download"},"finish_reason":null}]}

357−

358− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" and"},"finish_reason":null}]}

359−

360− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" Save"},"finish_reason":null}]}

361−

362− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" example"},"finish_reason":null}]}

363−

364− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":".txt"},"finish_reason":null}]}

365−

366− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" from"},"finish_reason":null}]}

367−

368− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" URL"},"finish_reason":null}]}

369−

370− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

371−

372− data: {"id":"chatcmpl-cacd42b690dbfa0c757a15f0","object":"chat.completion.chunk","created":1782413796,"model":"gpt-oss-120b","choices":[],"usage":{"prompt_tokens":160,"completion_tokens":108,"total_tokens":336,"prompt_tokens_details":{"cached_tokens":68},"completion_tokens_details":{"reasoning_tokens":103},"cost":{"usd":0.000145462,"hypercredits":0.00290924},"remaining":{"hypercredits":0}}}

373−

374− data: [DONE]

375−

376− headers:

377− Content-Type:

378− - text/event-stream

379− status: 200 OK

380− code: 200

381− duration: 1.326897958s

382−- id: 2

383− request:

384− proto: HTTP/1.1

385− proto_major: 1

386− proto_minor: 1

387− content_length: 0

388− host: example-files.online-convert.com

389− headers:

390− User-Agent:

391− - crush/1.0

392− url: https://example-files.online-convert.com/document/txt/example.txt

393− method: GET

394− response:

395− proto: HTTP/2.0

396− proto_major: 2

397− proto_minor: 0

398− content_length: -1

399− uncompressed: true

400− body: |-

401− TXT test file

402− Purpose: Provide example of this file type

403− Document file type: TXT

404− Version: 1.0

405− Remark:

406−

407− Example content:

408− The names "John Doe" for males, "Jane Doe" or "Jane Roe" for females, or "Jonnie Doe" and "Janie Doe" for children, or just "Doe" non-gender-specifically are used as placeholder names for a party whose true identity is unknown or must be withheld in a legal action, case, or discussion. The names are also used to refer to acorpse or hospital patient whose identity is unknown. This practice is widely used in the United States and Canada, but is rarely used in other English-speaking countries including the United Kingdom itself, from where the use of "John Doe" in a legal context originates. The names Joe Bloggs or John Smith are used in the UK instead, as well as in Australia and New Zealand.

409−

410− John Doe is sometimes used to refer to a typical male in other contexts as well, in a similar manner to John Q. Public, known in Great Britain as Joe Public, John Smith or Joe Bloggs. For example, the first name listed on a form is often John Doe, along with a fictional address or other fictional information to provide an example of how to fill in the form. The name is also used frequently in popular culture, for example in the Frank Capra film Meet John Doe. John Doe was also the name of a 2002 American television series.

411−

412− Similarly, a child or baby whose identity is unknown may be referred to as Baby Doe. A notorious murder case in Kansas City, Missouri, referred to the baby victim as Precious Doe. Other unidentified female murder victims are Cali Doe and Princess Doe. Additional persons may be called James Doe, Judy Doe, etc. However, to avoid possible confusion, if two anonymous or unknown parties are cited in a specific case or action, the surnames Doe and Roe may be used simultaneously; for example, "John Doe v. Jane Roe". If several anonymous parties are referenced, they may simply be labelled John Doe #1, John Doe #2, etc. (the U.S. Operation Delego cited 21 (numbered) "John Doe"s) or labelled with other variants of Doe / Roe / Poe / etc. Other early alternatives such as John Stiles and Richard Miles are now rarely used, and Mary Major has been used in some American federal cases.

413−

414−

415−

416− File created by https://www.online-convert.com

417− More example files: https://www.online-convert.com/file-type

418− Text of Example content: Wikipedia (https://en.wikipedia.org/wiki/John_Doe)

419− License: Attribution-ShareAlike 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)

420−

421− Feel free to use and share the file according to the license above.

422− headers:

423− Content-Type:

424− - text/plain; charset=UTF-8

425− status: 200 OK

426− code: 200

427− duration: 893.137459ms

428−- id: 3

429− request:

430− proto: HTTP/1.1

431− proto_major: 1

432− proto_minor: 1

433− content_length: 36247

434− host: ""

435− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/download_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"download the file from https://example-files.online-convert.com/document/txt/example.txt and save it as example.txt\",\"role\":\"user\"},{\"tool_calls\":[{\"id\":\"call_4cbf337c01e5b47c\",\"function\":{\"arguments\":\"{\\\"file_path\\\": \\\"/tmp/crush-test/TestCoderAgent/glm-5.1/download_tool/example.txt\\\", \\\"url\\\": \\\"https://example-files.online-convert.com/document/txt/example.txt\\\"}\",\"name\":\"download\"},\"type\":\"function\"}],\"role\":\"assistant\",\"reasoning_content\":\"The user wants to download a file from a URL and save it as example.txt.\"},{\"content\":\"Successfully downloaded 2574 bytes to example.txt (Content-Type: text/plain; charset=UTF-8)\",\"tool_call_id\":\"call_4cbf337c01e5b47c\",\"role\":\"tool\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

436− headers:

437− Accept:

438− - application/json

439− Content-Type:

440− - application/json

441− User-Agent:

442− - Charm-Crush/devel (https://charm.land/crush)

443− url: https://hyper.charm.land/v1/chat/completions

444− method: POST

445− response:

446− proto: HTTP/2.0

447− proto_major: 2

448− proto_minor: 0

449− content_length: -1

450− uncompressed: true

451− body: |+

452− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

453−

454− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"Download"},"finish_reason":null}]}

455−

456− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"ed"},"finish_reason":null}]}

457−

458− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" and"},"finish_reason":null}]}

459−

460− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" saved"},"finish_reason":null}]}

461−

462− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" as"},"finish_reason":null}]}

463−

464− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `"},"finish_reason":null}]}

465−

466− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"example"},"finish_reason":null}]}

467−

468− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":".txt"},"finish_reason":null}]}

469−

470− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`."},"finish_reason":null}]}

471−

472− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

473−

474− data: {"id":"chatcmpl-4bda680be45220e2a1b071f0","object":"chat.completion.chunk","created":1782413799,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":73,"completion_tokens":11,"total_tokens":8251,"prompt_tokens_details":{"cached_tokens":8167},"cost":{"usd":0.0001506,"hypercredits":0.003012},"remaining":{"hypercredits":0}}}

475−

476− data: [DONE]

477−

478− headers:

479− Content-Type:

480− - text/event-stream

481− status: 200 OK

482− code: 200

483− duration: 1.486211917s

internal/agent/testdata/TestCoderAgent/glm-5.1/fetch_tool.yamldeleted−650

@@ -1,650 +0,0 @@

1−---

2−version: 2

3−interactions:

4−- id: 0

5− request:

6− proto: HTTP/1.1

7− proto_major: 1

8− proto_minor: 1

9− content_length: 35715

10− host: ""

11− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/fetch_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"fetch the content from https://example-files.online-convert.com/website/html/example.html and tell me if it contains the word 'John Doe'\",\"role\":\"user\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

12− headers:

13− Accept:

14− - application/json

15− Content-Type:

16− - application/json

17− User-Agent:

18− - Charm-Crush/devel (https://charm.land/crush)

19− url: https://hyper.charm.land/v1/chat/completions

20− method: POST

21− response:

22− proto: HTTP/2.0

23− proto_major: 2

24− proto_minor: 0

25− content_length: -1

26− uncompressed: true

27− body: |+

28− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

29−

30− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"The"},"finish_reason":null}]}

31−

32− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

33−

34− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" wants"},"finish_reason":null}]}

35−

36− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" me"},"finish_reason":null}]}

37−

38− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

39−

40− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" fetch"},"finish_reason":null}]}

41−

42− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" content"},"finish_reason":null}]}

43−

44− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" from"},"finish_reason":null}]}

45−

46− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

47−

48− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" URL"},"finish_reason":null}]}

49−

50− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" and"},"finish_reason":null}]}

51−

52− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" check"},"finish_reason":null}]}

53−

54− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" if"},"finish_reason":null}]}

55−

56− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" it"},"finish_reason":null}]}

57−

58− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" contains"},"finish_reason":null}]}

59−

60− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

61−

62− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" word"},"finish_reason":null}]}

63−

64− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

65−

66− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"John"},"finish_reason":null}]}

67−

68− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":" Doe"},"finish_reason":null}]}

69−

70− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"reasoning_content":"'."},"finish_reason":null}]}

71−

72− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_60557666b1c79950","type":"function","index":0,"function":{"name":"fetch","arguments":""}}]},"finish_reason":null}]}

73−

74− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"format\": \""}}]},"finish_reason":null}]}

75−

76− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"text"}}]},"finish_reason":null}]}

77−

78− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

79−

80− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":", \"url\": \""}}]},"finish_reason":null}]}

81−

82− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"https"}}]},"finish_reason":null}]}

83−

84− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"://"}}]},"finish_reason":null}]}

85−

86− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"example"}}]},"finish_reason":null}]}

87−

88− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-files"}}]},"finish_reason":null}]}

89−

90− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".online"}}]},"finish_reason":null}]}

91−

92− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"-"}}]},"finish_reason":null}]}

93−

94− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"convert"}}]},"finish_reason":null}]}

95−

96− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".com"}}]},"finish_reason":null}]}

97−

98− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/"}}]},"finish_reason":null}]}

99−

100− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"website"}}]},"finish_reason":null}]}

101−

102− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/html"}}]},"finish_reason":null}]}

103−

104− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/example"}}]},"finish_reason":null}]}

105−

106− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":".html"}}]},"finish_reason":null}]}

107−

108− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

109−

110− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"}"}}]},"finish_reason":null}]}

111−

112− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

113−

114− data: {"id":"chatcmpl-adf239d098d5e63d542c1773","object":"chat.completion.chunk","created":1782413801,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":696,"completion_tokens":50,"total_tokens":8225,"prompt_tokens_details":{"cached_tokens":7479},"completion_tokens_details":{"reasoning_tokens":23},"cost":{"usd":0.0012956,"hypercredits":0.025912},"remaining":{"hypercredits":0}}}

115−

116− data: [DONE]

117−

118− headers:

119− Content-Type:

120− - text/event-stream

121− status: 200 OK

122− code: 200

123− duration: 2.230852292s

124−- id: 1

125− request:

126− proto: HTTP/1.1

127− proto_major: 1

128− proto_minor: 1

129− content_length: 895

130− host: ""

131− body: '{"messages":[{"content":"You will generate a short title based on the first message a user begins a conversation with.\n\n<rules>\n- Keep the title in the same language that the user wrote their message in.\n- Ensure it is not more than 50 characters long.\n- The title should be a summary of the user''s message.\n- It should be one line long.\n- Do not use quotes or colons.\n- The entire text you return will be used as the title.\n- Never return anything that is more than one sentence (one line) long.\n</rules>\n\n /no_think","role":"system"},{"content":"Generate a concise title for the following content:\n\nfetch the content from https://example-files.online-convert.com/website/html/example.html and tell me if it contains the word ''John Doe''\n <think>\n\n</think>","role":"user"}],"model":"gpt-oss-120b","max_completion_tokens":40,"stream_options":{"include_usage":true},"stream":true}'

132− headers:

133− Accept:

134− - application/json

135− Content-Type:

136− - application/json

137− User-Agent:

138− - Charm-Crush/devel (https://charm.land/crush)

139− url: https://hyper.charm.land/v1/chat/completions

140− method: POST

141− response:

142− proto: HTTP/2.0

143− proto_major: 2

144− proto_minor: 0

145− content_length: -1

146− uncompressed: true

147− body: |+

148− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

149−

150− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" wants"},"finish_reason":null}]}

151−

152− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

153−

154− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" short"},"finish_reason":null}]}

155−

156− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

157−

158− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

159−

160− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" their"},"finish_reason":null}]}

161−

162− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" request"},"finish_reason":null}]}

163−

164− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" content"},"finish_reason":null}]}

165−

166− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

167−

168− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Title"},"finish_reason":null}]}

169−

170− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

171−

172− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" English"},"finish_reason":null}]}

173−

174− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

175−

176− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Summ"},"finish_reason":null}]}

177−

178− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ar"},"finish_reason":null}]}

179−

180− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ize"},"finish_reason":null}]}

181−

182− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

183−

184− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

185−

186− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Check"},"finish_reason":null}]}

187−

188− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

189−

190− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".html"},"finish_reason":null}]}

191−

192− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

193−

194− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" word"},"finish_reason":null}]}

195−

196− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

197−

198− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"John"},"finish_reason":null}]}

199−

200− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Doe"},"finish_reason":null}]}

201−

202− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

203−

204− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

205−

206− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Must"},"finish_reason":null}]}

207−

208− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" be"},"finish_reason":null}]}

209−

210− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \u003c="},"finish_reason":null}]}

211−

212− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

213−

214− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" chars"},"finish_reason":null}]}

215−

216− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

217−

218− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Let's"},"finish_reason":null}]}

219−

220− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" craft"},"finish_reason":null}]}

221−

222− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

223−

224− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

225−

226− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Check"},"finish_reason":null}]}

227−

228− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" if"},"finish_reason":null}]}

229−

230− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

231−

232− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".html"},"finish_reason":null}]}

233−

234− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" contains"},"finish_reason":null}]}

235−

236− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

237−

238− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"John"},"finish_reason":null}]}

239−

240− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Doe"},"finish_reason":null}]}

241−

242− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

243−

244− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

245−

246− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

247−

248− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

249−

250− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

251−

252− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

253−

254− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

255−

256− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

257−

258− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Check"},"finish_reason":null}]}

259−

260− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" if"},"finish_reason":null}]}

261−

262− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" example"},"finish_reason":null}]}

263−

264− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".html"},"finish_reason":null}]}

265−

266− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" contains"},"finish_reason":null}]}

267−

268− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

269−

270− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"John"},"finish_reason":null}]}

271−

272− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Doe"},"finish_reason":null}]}

273−

274− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'\""},"finish_reason":null}]}

275−

276− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" that's"},"finish_reason":null}]}

277−

278− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

279−

280− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"41"},"finish_reason":null}]}

281−

282− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" maybe"},"finish_reason":null}]}

283−

284− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

285−

286− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Let's"},"finish_reason":null}]}

287−

288− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" count"},"finish_reason":null}]}

289−

290− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

291−

292− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" C"},"finish_reason":null}]}

293−

294− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

295−

296− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

297−

298− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

299−

300− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"h"},"finish_reason":null}]}

301−

302− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

303−

304− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"e"},"finish_reason":null}]}

305−

306− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"3"},"finish_reason":null}]}

307−

308− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"c"},"finish_reason":null}]}

309−

310− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

311−

312− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"k"},"finish_reason":null}]}

313−

314− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

315−

316− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

317−

318− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"6"},"finish_reason":null}]}

319−

320− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" i"},"finish_reason":null}]}

321−

322− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

323−

324− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"f"},"finish_reason":null}]}

325−

326− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"8"},"finish_reason":null}]}

327−

328− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

329−

330− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"9"},"finish_reason":null}]}

331−

332− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" e"},"finish_reason":null}]}

333−

334− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"10"},"finish_reason":null}]}

335−

336− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"x"},"finish_reason":null}]}

337−

338− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"11"},"finish_reason":null}]}

339−

340− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"a"},"finish_reason":null}]}

341−

342− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"12"},"finish_reason":null}]}

343−

344− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"m"},"finish_reason":null}]}

345−

346− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"13"},"finish_reason":null}]}

347−

348− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"p"},"finish_reason":null}]}

349−

350− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"14"},"finish_reason":null}]}

351−

352− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"l"},"finish_reason":null}]}

353−

354− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"15"},"finish_reason":null}]}

355−

356− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"e"},"finish_reason":null}]}

357−

358− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"16"},"finish_reason":null}]}

359−

360− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

361−

362− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"17"},"finish_reason":null}]}

363−

364− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"h"},"finish_reason":null}]}

365−

366− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"18"},"finish_reason":null}]}

367−

368− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"t"},"finish_reason":null}]}

369−

370− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"19"},"finish_reason":null}]}

371−

372− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"m"},"finish_reason":null}]}

373−

374− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"20"},"finish_reason":null}]}

375−

376− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"l"},"finish_reason":null}]}

377−

378− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"21"},"finish_reason":null}]}

379−

380− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

381−

382− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"22"},"finish_reason":null}]}

383−

384− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" c"},"finish_reason":null}]}

385−

386− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"23"},"finish_reason":null}]}

387−

388− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"o"},"finish_reason":null}]}

389−

390− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"24"},"finish_reason":null}]}

391−

392− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"n"},"finish_reason":null}]}

393−

394− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"25"},"finish_reason":null}]}

395−

396− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"t"},"finish_reason":null}]}

397−

398− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"26"},"finish_reason":null}]}

399−

400− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"a"},"finish_reason":null}]}

401−

402− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"27"},"finish_reason":null}]}

403−

404− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"i"},"finish_reason":null}]}

405−

406− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"28"},"finish_reason":null}]}

407−

408− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"n"},"finish_reason":null}]}

409−

410− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"29"},"finish_reason":null}]}

411−

412− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"s"},"finish_reason":null}]}

413−

414− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"30"},"finish_reason":null}]}

415−

416− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

417−

418− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"31"},"finish_reason":null}]}

419−

420− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

421−

422− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"32"},"finish_reason":null}]}

423−

424− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"J"},"finish_reason":null}]}

425−

426− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"33"},"finish_reason":null}]}

427−

428− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"o"},"finish_reason":null}]}

429−

430− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"34"},"finish_reason":null}]}

431−

432− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"h"},"finish_reason":null}]}

433−

434− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"35"},"finish_reason":null}]}

435−

436− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"n"},"finish_reason":null}]}

437−

438− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"36"},"finish_reason":null}]}

439−

440− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

441−

442− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"37"},"finish_reason":null}]}

443−

444− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"D"},"finish_reason":null}]}

445−

446− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"38"},"finish_reason":null}]}

447−

448− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"o"},"finish_reason":null}]}

449−

450− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"39"},"finish_reason":null}]}

451−

452− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"e"},"finish_reason":null}]}

453−

454− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"40"},"finish_reason":null}]}

455−

456− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

457−

458− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"41"},"finish_reason":null}]}

459−

460− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

461−

462− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" That's"},"finish_reason":null}]}

463−

464− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

465−

466− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"41"},"finish_reason":null}]}

467−

468− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

469−

470− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

471−

472− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" within"},"finish_reason":null}]}

473−

474− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

475−

476− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

477−

478− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

479−

480− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Good"},"finish_reason":null}]}

481−

482− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

483−

484− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"Check"},"finish_reason":null}]}

485−

486− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" if"},"finish_reason":null}]}

487−

488− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" example"},"finish_reason":null}]}

489−

490− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":".html"},"finish_reason":null}]}

491−

492− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" contains"},"finish_reason":null}]}

493−

494− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" '"},"finish_reason":null}]}

495−

496− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"John"},"finish_reason":null}]}

497−

498− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" Doe"},"finish_reason":null}]}

499−

500− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"'"},"finish_reason":null}]}

501−

502− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

503−

504− data: {"id":"chatcmpl-2422745116026947d8ea8426","object":"chat.completion.chunk","created":1782413800,"model":"gpt-oss-120b","choices":[],"usage":{"prompt_tokens":173,"completion_tokens":183,"total_tokens":420,"prompt_tokens_details":{"cached_tokens":64},"cost":{"usd":0.00013191,"hypercredits":0.0026382},"remaining":{"hypercredits":0}}}

505−

506− data: [DONE]

507−

508− headers:

509− Content-Type:

510− - text/event-stream

511− status: 200 OK

512− code: 200

513− duration: 1.005238292s

514−- id: 2

515− request:

516− proto: HTTP/1.1

517− proto_major: 1

518− proto_minor: 1

519− content_length: 0

520− host: example-files.online-convert.com

521− headers:

522− User-Agent:

523− - crush/1.0

524− url: https://example-files.online-convert.com/website/html/example.html

525− method: GET

526− response:

527− proto: HTTP/2.0

528− proto_major: 2

529− proto_minor: 0

530− content_length: -1

531− uncompressed: true

532− body: |-

533− <!DOCTYPE html>

534−

535− <html>

536− <body>

537−

538− <h1>HTML test file</h1>

539−

540− Purpose: Provide example of this file type

541− Document file type: HTML

542− Version: 1.0

543− Remark:

544−

545− <img align="right" style="height:5.0551cm;width:7.62cm;" src="data:image/*;base64,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

546− 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

547− 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

548− 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

549− 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

550− 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

551− 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

552− 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

553− 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

554− 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

555− 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

556− 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"/>

557−

558− Example content:

559− The names "John Doe" for males, "Jane Doe" or "Jane Roe" for females, or "Jonnie Doe" and "Janie Doe" for children, or just "Doe" non-gender-specifically are used as placeholder names for a party whose true identity is unknown or must be withheld in a legal action, case, or discussion. The names are also used to refer to acorpse or hospital patient whose identity is unknown. This practice is widely used in the United States and Canada, but is rarely used in other English-speaking countries including the United Kingdom itself, from where the use of "John Doe" in a legal context originates. The names Joe Bloggs or John Smith are used in the UK instead, as well as in Australia and New Zealand.

560−

561− John Doe is sometimes used to refer to a typical male in other contexts as well, in a similar manner to John Q. Public, known in Great Britain as Joe Public, John Smith or Joe Bloggs. For example, the first name listed on a form is often John Doe, along with a fictional address or other fictional information to provide an example of how to fill in the form. The name is also used frequently in popular culture, for example in the Frank Capra film Meet John Doe. John Doe was also the name of a 2002 American television series.

562−

563− Similarly, a child or baby whose identity is unknown may be referred to as Baby Doe. A notorious murder case in Kansas City, Missouri, referred to the baby victim as Precious Doe. Other unidentified female murder victims are Cali Doe and Princess Doe. Additional persons may be called James Doe, Judy Doe, etc. However, to avoid possible confusion, if two anonymous or unknown parties are cited in a specific case or action, the surnames Doe and Roe may be used simultaneously; for example, "John Doe v. Jane Roe". If several anonymous parties are referenced, they may simply be labelled John Doe #1, John Doe #2, etc. (the U.S. Operation Delego cited 21 (numbered) "John Doe"s) or labelled with other variants of Doe / Roe / Poe / etc. Other early alternatives such as John Stiles and Richard Miles are now rarely used, and Mary Major has been used in some American federal cases.

564−

565− <img style="height:1.6881cm;width:1.6881cm;" align="left" src="data:image/*;base64,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

566− B1/bMVO7nYqfTdlIhfMUvxUd51lhy67Tyq1UqJI9Jew1tnpUt0B6+/rwB89+fbtcyp1Q+gRk7z/Gy9tfzMHyrrfpKG8eqVDetsaC8o5Juz6zqvKuz6GyvKtkeXn3fbDzOauVdwpGynvzd+wv64P1QRgMBZhAGAwFmEAYDAWYQBgMBZhAGAwFmEAYDAWYQBgMBZhAGAwFmEAYDAWYQBgMBZhAGAwFmEAYDAWYQBgMBf4/fQGdtWECNsgAAAAASUVORK5CYII="/>

567−

568− File created by <a href="https://www.online-convert.com" target="_blank">https://www.online-convert.com</a>

569− More example files: <a href="https://www.online-convert.com/file-type" target="_blank">https://www.online-convert.com/file-type</a>

570− Text of "Example content": <a href="https://en.wikipedia.org/wiki/John_Doe" target="_blank">Wikipedia</a>

571− License: <a href="https://creativecommons.org/licenses/by-sa/3.0/" target="_blank">Attribution-ShareAlike 3.0 Unported</a>

572−

573− <img style="height:0.6949cm;width:1.9741cm;" align="right" src="data:image/*;base64,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"/>

574−

575− Feel free to use and share the file according to license above.

576−

577− </body>

578− </html>

579− headers:

580− Content-Type:

581− - text/html; charset=UTF-8

582− status: 200 OK

583− code: 200

584− duration: 147.645375ms

585−- id: 3

586− request:

587− proto: HTTP/1.1

588− proto_major: 1

589− proto_minor: 1

590− content_length: 36231

591− host: ""

592− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/fetch_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"fetch the content from https://example-files.online-convert.com/website/html/example.html and tell me if it contains the word 'John Doe'\",\"role\":\"user\"},{\"tool_calls\":[{\"id\":\"call_60557666b1c79950\",\"function\":{\"arguments\":\"{\\\"format\\\": \\\"text\\\", \\\"url\\\": \\\"https://example-files.online-convert.com/website/html/example.html\\\"}\",\"name\":\"fetch\"},\"type\":\"function\"}],\"role\":\"assistant\",\"reasoning_content\":\"The user wants me to fetch content from a URL and check if it contains the word 'John Doe'.\"},{\"content\":\"HTML test file Purpose: Provide example of this file type Document file type: HTML Version: 1.0 Remark:\",\"tool_call_id\":\"call_60557666b1c79950\",\"role\":\"tool\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

593− headers:

594− Accept:

595− - application/json

596− Content-Type:

597− - application/json

598− User-Agent:

599− - Charm-Crush/devel (https://charm.land/crush)

600− url: https://hyper.charm.land/v1/chat/completions

601− method: POST

602− response:

603− proto: HTTP/2.0

604− proto_major: 2

605− proto_minor: 0

606− content_length: -1

607− uncompressed: true

608− body: |+

609− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

610−

611− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"No"},"finish_reason":null}]}

612−

613− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":","},"finish_reason":null}]}

614−

615− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" the"},"finish_reason":null}]}

616−

617− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" fetched"},"finish_reason":null}]}

618−

619− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" content"},"finish_reason":null}]}

620−

621− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" does"},"finish_reason":null}]}

622−

623− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" not"},"finish_reason":null}]}

624−

625− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" contain"},"finish_reason":null}]}

626−

627− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" the"},"finish_reason":null}]}

628−

629− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" word"},"finish_reason":null}]}

630−

631− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" \""},"finish_reason":null}]}

632−

633− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"John"},"finish_reason":null}]}

634−

635− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" Doe"},"finish_reason":null}]}

636−

637− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"\"."},"finish_reason":null}]}

638−

639− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

640−

641− data: {"id":"chatcmpl-cd7a3a1eb813288fd5e12324","object":"chat.completion.chunk","created":1782413803,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":753,"completion_tokens":16,"total_tokens":8248,"prompt_tokens_details":{"cached_tokens":7479},"cost":{"usd":0.0011246000000000001,"hypercredits":0.022492},"remaining":{"hypercredits":0}}}

642−

643− data: [DONE]

644−

645− headers:

646− Content-Type:

647− - text/event-stream

648− status: 200 OK

649− code: 200

650− duration: 977.248916ms

internal/agent/testdata/TestCoderAgent/glm-5.1/glob_tool.yamldeleted−737

@@ -1,737 +0,0 @@

1−---

2−version: 2

3−interactions:

4−- id: 0

5− request:

6− proto: HTTP/1.1

7− proto_major: 1

8− proto_minor: 1

9− content_length: 35633

10− host: ""

11− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/glob_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use glob to find all .go files in the current directory\",\"role\":\"user\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

12− headers:

13− Accept:

14− - application/json

15− Content-Type:

16− - application/json

17− User-Agent:

18− - Charm-Crush/devel (https://charm.land/crush)

19− url: https://hyper.charm.land/v1/chat/completions

20− method: POST

21− response:

22− proto: HTTP/2.0

23− proto_major: 2

24− proto_minor: 0

25− content_length: -1

26− uncompressed: true

27− body: |+

28− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

29−

30− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_ad74b69d4b477328","type":"function","index":0,"function":{"name":"glob","arguments":""}}]},"finish_reason":null}]}

31−

32− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"pattern\": \""}}]},"finish_reason":null}]}

33−

34− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"**"}}]},"finish_reason":null}]}

35−

36− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"/*."}}]},"finish_reason":null}]}

37−

38− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"go"}}]},"finish_reason":null}]}

39−

40− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

41−

42− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"}"}}]},"finish_reason":null}]}

43−

44− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

45−

46− data: {"id":"chatcmpl-680977ab8e48dc961b5a65b4","object":"chat.completion.chunk","created":1782413805,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":679,"completion_tokens":13,"total_tokens":8171,"prompt_tokens_details":{"cached_tokens":7479},"cost":{"usd":0.0010077999999999999,"hypercredits":0.020155999999999997},"remaining":{"hypercredits":0}}}

47−

48− data: [DONE]

49−

50− headers:

51− Content-Type:

52− - text/event-stream

53− status: 200 OK

54− code: 200

55− duration: 2.191829958s

56−- id: 1

57− request:

58− proto: HTTP/1.1

59− proto_major: 1

60− proto_minor: 1

61− content_length: 35910

62− host: ""

63− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/glob_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use glob to find all .go files in the current directory\",\"role\":\"user\"},{\"tool_calls\":[{\"id\":\"call_ad74b69d4b477328\",\"function\":{\"arguments\":\"{\\\"pattern\\\": \\\"**/*.go\\\"}\",\"name\":\"glob\"},\"type\":\"function\"}],\"role\":\"assistant\"},{\"content\":\"/tmp/crush-test/TestCoderAgent/glm-5.1/glob_tool/main.go\",\"tool_call_id\":\"call_ad74b69d4b477328\",\"role\":\"tool\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

64− headers:

65− Accept:

66− - application/json

67− Content-Type:

68− - application/json

69− User-Agent:

70− - Charm-Crush/devel (https://charm.land/crush)

71− url: https://hyper.charm.land/v1/chat/completions

72− method: POST

73− response:

74− proto: HTTP/2.0

75− proto_major: 2

76− proto_minor: 0

77− content_length: -1

78− uncompressed: true

79− body: |+

80− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

81−

82− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"Found"},"finish_reason":null}]}

83−

84− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" "},"finish_reason":null}]}

85−

86− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"1"},"finish_reason":null}]}

87−

88− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `."},"finish_reason":null}]}

89−

90− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"go"},"finish_reason":null}]}

91−

92− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`"},"finish_reason":null}]}

93−

94− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" file"},"finish_reason":null}]}

95−

96− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":":"},"finish_reason":null}]}

97−

98− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `"},"finish_reason":null}]}

99−

100− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"main"},"finish_reason":null}]}

101−

102− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":".go"},"finish_reason":null}]}

103−

104− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`"},"finish_reason":null}]}

105−

106− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

107−

108− data: {"id":"chatcmpl-85e120b7a3fdf21599aee1ba","object":"chat.completion.chunk","created":1782413807,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":36,"completion_tokens":14,"total_tokens":8207,"prompt_tokens_details":{"cached_tokens":8157},"cost":{"usd":0.00011200000000000001,"hypercredits":0.0022400000000000002},"remaining":{"hypercredits":0}}}

109−

110− data: [DONE]

111−

112− headers:

113− Content-Type:

114− - text/event-stream

115− status: 200 OK

116− code: 200

117− duration: 1.449032542s

118−- id: 2

119− request:

120− proto: HTTP/1.1

121− proto_major: 1

122− proto_minor: 1

123− content_length: 814

124− host: ""

125− body: '{"messages":[{"content":"You will generate a short title based on the first message a user begins a conversation with.\n\n<rules>\n- Keep the title in the same language that the user wrote their message in.\n- Ensure it is not more than 50 characters long.\n- The title should be a summary of the user''s message.\n- It should be one line long.\n- Do not use quotes or colons.\n- The entire text you return will be used as the title.\n- Never return anything that is more than one sentence (one line) long.\n</rules>\n\n /no_think","role":"system"},{"content":"Generate a concise title for the following content:\n\nuse glob to find all .go files in the current directory\n <think>\n\n</think>","role":"user"}],"model":"gpt-oss-120b","max_completion_tokens":40,"stream_options":{"include_usage":true},"stream":true}'

126− headers:

127− Accept:

128− - application/json

129− Content-Type:

130− - application/json

131− User-Agent:

132− - Charm-Crush/devel (https://charm.land/crush)

133− url: https://hyper.charm.land/v1/chat/completions

134− method: POST

135− response:

136− proto: HTTP/2.0

137− proto_major: 2

138− proto_minor: 0

139− content_length: -1

140− uncompressed: true

141− body: |+

142− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

143−

144− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

145−

146− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" wants"},"finish_reason":null}]}

147−

148− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

149−

150− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" concise"},"finish_reason":null}]}

151−

152− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

153−

154− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

155−

156− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

157−

158− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" content"},"finish_reason":null}]}

159−

160− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

161−

162− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

163−

164− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"use"},"finish_reason":null}]}

165−

166− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" glob"},"finish_reason":null}]}

167−

168− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

169−

170− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" find"},"finish_reason":null}]}

171−

172− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" all"},"finish_reason":null}]}

173−

174− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ."},"finish_reason":null}]}

175−

176− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"go"},"finish_reason":null}]}

177−

178− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

179−

180− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

181−

182− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

183−

184− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" current"},"finish_reason":null}]}

185−

186− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" directory"},"finish_reason":null}]}

187−

188− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

189−

190− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Title"},"finish_reason":null}]}

191−

192− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

193−

194− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" English"},"finish_reason":null}]}

195−

196− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

197−

198− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \u003c="},"finish_reason":null}]}

199−

200− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

201−

202− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

203−

204− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

205−

206− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

207−

208− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" line"},"finish_reason":null}]}

209−

210− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

211−

212− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

213−

214− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" quotes"},"finish_reason":null}]}

215−

216− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" or"},"finish_reason":null}]}

217−

218− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" col"},"finish_reason":null}]}

219−

220− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ons"},"finish_reason":null}]}

221−

222− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

223−

224− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Something"},"finish_reason":null}]}

225−

226− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" like"},"finish_reason":null}]}

227−

228− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

229−

230− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Finding"},"finish_reason":null}]}

231−

232− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ."},"finish_reason":null}]}

233−

234− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"go"},"finish_reason":null}]}

235−

236− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Files"},"finish_reason":null}]}

237−

238− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

239−

240− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Glob"},"finish_reason":null}]}

241−

242− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

243−

244− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Current"},"finish_reason":null}]}

245−

246− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Directory"},"finish_reason":null}]}

247−

248− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

249−

250− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" That's"},"finish_reason":null}]}

251−

252− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" under"},"finish_reason":null}]}

253−

254− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

255−

256− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

257−

258− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

259−

260− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

261−

262− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

263−

264− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

265−

266− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

267−

268− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Finding"},"finish_reason":null}]}

269−

270− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ."},"finish_reason":null}]}

271−

272− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"go"},"finish_reason":null}]}

273−

274− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Files"},"finish_reason":null}]}

275−

276− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

277−

278− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Glob"},"finish_reason":null}]}

279−

280− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

281−

282− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Current"},"finish_reason":null}]}

283−

284− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Directory"},"finish_reason":null}]}

285−

286− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\""},"finish_reason":null}]}

287−

288− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

289−

290− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" let's"},"finish_reason":null}]}

291−

292− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" count"},"finish_reason":null}]}

293−

294− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

295−

296− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Finding"},"finish_reason":null}]}

297−

298− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

299−

300− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

301−

302− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

303−

304− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" +"},"finish_reason":null}]}

305−

306− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

307−

308− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

309−

310− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

311−

312− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

313−

314− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

315−

316− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"8"},"finish_reason":null}]}

317−

318− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

319−

320− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ."},"finish_reason":null}]}

321−

322− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"go"},"finish_reason":null}]}

323−

324− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

325−

326− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"3"},"finish_reason":null}]}

327−

328− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

329−

330− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" =\u003e"},"finish_reason":null}]}

331−

332− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" total"},"finish_reason":null}]}

333−

334− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

335−

336− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"11"},"finish_reason":null}]}

337−

338− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

339−

340− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

341−

342− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

343−

344− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"12"},"finish_reason":null}]}

345−

346− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

347−

348− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Files"},"finish_reason":null}]}

349−

350− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

351−

352− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

353−

354− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

355−

356− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"17"},"finish_reason":null}]}

357−

358− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

359−

360− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

361−

362− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

363−

364− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"18"},"finish_reason":null}]}

365−

366− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

367−

368− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" with"},"finish_reason":null}]}

369−

370− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

371−

372− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

373−

374− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

375−

376− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"22"},"finish_reason":null}]}

377−

378− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

379−

380− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

381−

382− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

383−

384− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"23"},"finish_reason":null}]}

385−

386− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

387−

388− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Glob"},"finish_reason":null}]}

389−

390− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

391−

392− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

393−

394− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

395−

396− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"27"},"finish_reason":null}]}

397−

398− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

399−

400− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

401−

402− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

403−

404− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"28"},"finish_reason":null}]}

405−

406− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

407−

408− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

409−

410− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

411−

412− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

413−

414− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

415−

416− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"30"},"finish_reason":null}]}

417−

418− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

419−

420− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

421−

422− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

423−

424− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"31"},"finish_reason":null}]}

425−

426− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

427−

428− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Current"},"finish_reason":null}]}

429−

430− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

431−

432− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

433−

434− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

435−

436− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"38"},"finish_reason":null}]}

437−

438− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

439−

440− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

441−

442− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

443−

444− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"39"},"finish_reason":null}]}

445−

446− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

447−

448− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Directory"},"finish_reason":null}]}

449−

450− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

451−

452− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"9"},"finish_reason":null}]}

453−

454− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

455−

456− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"48"},"finish_reason":null}]}

457−

458− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

459−

460− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" So"},"finish_reason":null}]}

461−

462− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

463−

464− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"48"},"finish_reason":null}]}

465−

466− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

467−

468− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

469−

470− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" okay"},"finish_reason":null}]}

471−

472− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

473−

474− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Or"},"finish_reason":null}]}

475−

476− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

477−

478− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Use"},"finish_reason":null}]}

479−

480− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" glob"},"finish_reason":null}]}

481−

482− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

483−

484− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" locate"},"finish_reason":null}]}

485−

486− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ."},"finish_reason":null}]}

487−

488− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"go"},"finish_reason":null}]}

489−

490− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

491−

492− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

493−

494− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" current"},"finish_reason":null}]}

495−

496− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" dir"},"finish_reason":null}]}

497−

498− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

499−

500− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Let's"},"finish_reason":null}]}

501−

502− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" count"},"finish_reason":null}]}

503−

504− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

505−

506− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Use"},"finish_reason":null}]}

507−

508− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

509−

510− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"3"},"finish_reason":null}]}

511−

512− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")+"},"finish_reason":null}]}

513−

514− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"space"},"finish_reason":null}]}

515−

516− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

517−

518− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

519−

520− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

521−

522− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" glob"},"finish_reason":null}]}

523−

524− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

525−

526− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

527−

528− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

529−

530− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"8"},"finish_reason":null}]}

531−

532− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

533−

534− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

535−

536− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

537−

538− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"9"},"finish_reason":null}]}

539−

540− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

541−

542− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

543−

544− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

545−

546− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

547−

548− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

549−

550− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"11"},"finish_reason":null}]}

551−

552− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

553−

554− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

555−

556− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

557−

558− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"12"},"finish_reason":null}]}

559−

560− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

561−

562− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" locate"},"finish_reason":null}]}

563−

564− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

565−

566− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"6"},"finish_reason":null}]}

567−

568− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

569−

570− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"18"},"finish_reason":null}]}

571−

572− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

573−

574− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

575−

576− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

577−

578− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"19"},"finish_reason":null}]}

579−

580− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

581−

582− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ."},"finish_reason":null}]}

583−

584− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"go"},"finish_reason":null}]}

585−

586− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

587−

588− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"3"},"finish_reason":null}]}

589−

590− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

591−

592− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"22"},"finish_reason":null}]}

593−

594− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

595−

596− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

597−

598− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

599−

600− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"23"},"finish_reason":null}]}

601−

602− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

603−

604− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

605−

606− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

607−

608− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

609−

610− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

611−

612− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"28"},"finish_reason":null}]}

613−

614− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

615−

616− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

617−

618− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

619−

620− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"29"},"finish_reason":null}]}

621−

622− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

623−

624− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

625−

626− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

627−

628− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

629−

630− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

631−

632− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"31"},"finish_reason":null}]}

633−

634− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

635−

636− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

637−

638− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

639−

640− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"32"},"finish_reason":null}]}

641−

642− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

643−

644− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" current"},"finish_reason":null}]}

645−

646− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

647−

648− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

649−

650− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

651−

652− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"39"},"finish_reason":null}]}

653−

654− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

655−

656− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

657−

658− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"="},"finish_reason":null}]}

659−

660− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"40"},"finish_reason":null}]}

661−

662− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

663−

664− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" dir"},"finish_reason":null}]}

665−

666− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

667−

668− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"3"},"finish_reason":null}]}

669−

670− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")="},"finish_reason":null}]}

671−

672− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"43"},"finish_reason":null}]}

673−

674− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

675−

676− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" That's"},"finish_reason":null}]}

677−

678− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

679−

680− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"43"},"finish_reason":null}]}

681−

682− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

683−

684− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Both"},"finish_reason":null}]}

685−

686− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" fine"},"finish_reason":null}]}

687−

688− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

689−

690− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Let's"},"finish_reason":null}]}

691−

692− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" output"},"finish_reason":null}]}

693−

694− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

695−

696− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" line"},"finish_reason":null}]}

697−

698− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

699−

700− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

701−

702− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" quotes"},"finish_reason":null}]}

703−

704− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

705−

706− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"Use"},"finish_reason":null}]}

707−

708− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" glob"},"finish_reason":null}]}

709−

710− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" to"},"finish_reason":null}]}

711−

712− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" locate"},"finish_reason":null}]}

713−

714− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" ."},"finish_reason":null}]}

715−

716− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"go"},"finish_reason":null}]}

717−

718− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" files"},"finish_reason":null}]}

719−

720− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" in"},"finish_reason":null}]}

721−

722− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" current"},"finish_reason":null}]}

723−

724− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" directory"},"finish_reason":null}]}

725−

726− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

727−

728− data: {"id":"chatcmpl-eb1b020a610e9fa7da88e894","object":"chat.completion.chunk","created":1782413804,"model":"gpt-oss-120b","choices":[],"usage":{"prompt_tokens":157,"completion_tokens":298,"total_tokens":519,"prompt_tokens_details":{"cached_tokens":64},"cost":{"usd":0.00019385200000000003,"hypercredits":0.0038770400000000004},"remaining":{"hypercredits":0}}}

729−

730− data: [DONE]

731−

732− headers:

733− Content-Type:

734− - text/event-stream

735− status: 200 OK

736− code: 200

737− duration: 648.718041ms

internal/agent/testdata/TestCoderAgent/glm-5.1/grep_tool.yamldeleted−499

@@ -1,499 +0,0 @@

1−---

2−version: 2

3−interactions:

4−- id: 0

5− request:

6− proto: HTTP/1.1

7− proto_major: 1

8− proto_minor: 1

9− content_length: 35631

10− host: ""

11− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/grep_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use grep to search for the word 'package' in go files\",\"role\":\"user\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

12− headers:

13− Accept:

14− - application/json

15− Content-Type:

16− - application/json

17− User-Agent:

18− - Charm-Crush/devel (https://charm.land/crush)

19− url: https://hyper.charm.land/v1/chat/completions

20− method: POST

21− response:

22− proto: HTTP/2.0

23− proto_major: 2

24− proto_minor: 0

25− content_length: -1

26− uncompressed: true

27− body: |+

28− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

29−

30− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_a57bfe141bf85225","type":"function","index":0,"function":{"name":"grep","arguments":""}}]},"finish_reason":null}]}

31−

32− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"include\": \""}}]},"finish_reason":null}]}

33−

34− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"*."}}]},"finish_reason":null}]}

35−

36− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"go"}}]},"finish_reason":null}]}

37−

38− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

39−

40− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":", \"pattern\": \""}}]},"finish_reason":null}]}

41−

42− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"package"}}]},"finish_reason":null}]}

43−

44− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\""}}]},"finish_reason":null}]}

45−

46− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"}"}}]},"finish_reason":null}]}

47−

48− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

49−

50− data: {"id":"chatcmpl-8110189342807f502f916f95","object":"chat.completion.chunk","created":1782413815,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":680,"completion_tokens":18,"total_tokens":8177,"prompt_tokens_details":{"cached_tokens":7479},"cost":{"usd":0.0010312,"hypercredits":0.020623999999999997},"remaining":{"hypercredits":0}}}

51−

52− data: [DONE]

53−

54− headers:

55− Content-Type:

56− - text/event-stream

57− status: 200 OK

58− code: 200

59− duration: 1.505379s

60−- id: 1

61− request:

62− proto: HTTP/1.1

63− proto_major: 1

64− proto_minor: 1

65− content_length: 812

66− host: ""

67− body: '{"messages":[{"content":"You will generate a short title based on the first message a user begins a conversation with.\n\n<rules>\n- Keep the title in the same language that the user wrote their message in.\n- Ensure it is not more than 50 characters long.\n- The title should be a summary of the user''s message.\n- It should be one line long.\n- Do not use quotes or colons.\n- The entire text you return will be used as the title.\n- Never return anything that is more than one sentence (one line) long.\n</rules>\n\n /no_think","role":"system"},{"content":"Generate a concise title for the following content:\n\nuse grep to search for the word ''package'' in go files\n <think>\n\n</think>","role":"user"}],"model":"gpt-oss-120b","max_completion_tokens":40,"stream_options":{"include_usage":true},"stream":true}'

68− headers:

69− Accept:

70− - application/json

71− Content-Type:

72− - application/json

73− User-Agent:

74− - Charm-Crush/devel (https://charm.land/crush)

75− url: https://hyper.charm.land/v1/chat/completions

76− method: POST

77− response:

78− proto: HTTP/2.0

79− proto_major: 2

80− proto_minor: 0

81− content_length: -1

82− uncompressed: true

83− body: |+

84− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

85−

86− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"The"},"finish_reason":null}]}

87−

88− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

89−

90− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" wants"},"finish_reason":null}]}

91−

92− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

93−

94− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" concise"},"finish_reason":null}]}

95−

96− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

97−

98− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

99−

100− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" They"},"finish_reason":null}]}

101−

102− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" gave"},"finish_reason":null}]}

103−

104− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

105−

106− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" phrase"},"finish_reason":null}]}

107−

108− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

109−

110− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

111−

112− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"use"},"finish_reason":null}]}

113−

114− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" grep"},"finish_reason":null}]}

115−

116− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

117−

118− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" search"},"finish_reason":null}]}

119−

120− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

121−

122− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

123−

124− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" word"},"finish_reason":null}]}

125−

126− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

127−

128− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"package"},"finish_reason":null}]}

129−

130− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

131−

132− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

133−

134− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" go"},"finish_reason":null}]}

135−

136− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

137−

138− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

139−

140− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Need"},"finish_reason":null}]}

141−

142− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

143−

144− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" produce"},"finish_reason":null}]}

145−

146− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

147−

148− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

149−

150− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

151−

152− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" same"},"finish_reason":null}]}

153−

154− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" language"},"finish_reason":null}]}

155−

156− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

157−

158− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"English"},"finish_reason":null}]}

159−

160− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")."},"finish_reason":null}]}

161−

162− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Under"},"finish_reason":null}]}

163−

164− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

165−

166− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

167−

168− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" chars"},"finish_reason":null}]}

169−

170− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

171−

172− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

173−

174− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" line"},"finish_reason":null}]}

175−

176− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

177−

178− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Something"},"finish_reason":null}]}

179−

180− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" like"},"finish_reason":null}]}

181−

182− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

183−

184− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

185−

186− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"G"},"finish_reason":null}]}

187−

188− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"rep"},"finish_reason":null}]}

189−

190− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

191−

192− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

193−

194− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"package"},"finish_reason":null}]}

195−

196− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

197−

198− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

199−

200− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Go"},"finish_reason":null}]}

201−

202− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

203−

204− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

205−

206− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" That's"},"finish_reason":null}]}

207−

208− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" within"},"finish_reason":null}]}

209−

210− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" limit"},"finish_reason":null}]}

211−

212− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

213−

214− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Title"},"finish_reason":null}]}

215−

216− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" length"},"finish_reason":null}]}

217−

218− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"?"},"finish_reason":null}]}

219−

220− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

221−

222− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

223−

224− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

225−

226− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

227−

228− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"G"},"finish_reason":null}]}

229−

230− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"rep"},"finish_reason":null}]}

231−

232− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

233−

234− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

235−

236− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"package"},"finish_reason":null}]}

237−

238− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"'"},"finish_reason":null}]}

239−

240− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

241−

242− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Go"},"finish_reason":null}]}

243−

244− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

245−

246− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\""},"finish_reason":null}]}

247−

248− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

249−

250− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" let's"},"finish_reason":null}]}

251−

252− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" count"},"finish_reason":null}]}

253−

254− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

255−

256− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" G"},"finish_reason":null}]}

257−

258− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"("},"finish_reason":null}]}

259−

260− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

261−

262− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

263−

264− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" r"},"finish_reason":null}]}

265−

266− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

267−

268− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" e"},"finish_reason":null}]}

269−

270− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"3"},"finish_reason":null}]}

271−

272− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" p"},"finish_reason":null}]}

273−

274− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

275−

276− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

277−

278− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

279−

280− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" f"},"finish_reason":null}]}

281−

282− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"6"},"finish_reason":null}]}

283−

284− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" o"},"finish_reason":null}]}

285−

286− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

287−

288− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" r"},"finish_reason":null}]}

289−

290− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"8"},"finish_reason":null}]}

291−

292− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

293−

294− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"9"},"finish_reason":null}]}

295−

296− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

297−

298− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"10"},"finish_reason":null}]}

299−

300− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" p"},"finish_reason":null}]}

301−

302− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"11"},"finish_reason":null}]}

303−

304− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

305−

306− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"12"},"finish_reason":null}]}

307−

308− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" c"},"finish_reason":null}]}

309−

310− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"13"},"finish_reason":null}]}

311−

312− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" k"},"finish_reason":null}]}

313−

314− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"14"},"finish_reason":null}]}

315−

316− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

317−

318− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"15"},"finish_reason":null}]}

319−

320− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" g"},"finish_reason":null}]}

321−

322− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"16"},"finish_reason":null}]}

323−

324− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" e"},"finish_reason":null}]}

325−

326− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"17"},"finish_reason":null}]}

327−

328− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" '"},"finish_reason":null}]}

329−

330− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"18"},"finish_reason":null}]}

331−

332− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

333−

334− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"19"},"finish_reason":null}]}

335−

336− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" i"},"finish_reason":null}]}

337−

338− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"20"},"finish_reason":null}]}

339−

340− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" n"},"finish_reason":null}]}

341−

342− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"21"},"finish_reason":null}]}

343−

344− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

345−

346− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"22"},"finish_reason":null}]}

347−

348− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" G"},"finish_reason":null}]}

349−

350− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"23"},"finish_reason":null}]}

351−

352− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" o"},"finish_reason":null}]}

353−

354− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"24"},"finish_reason":null}]}

355−

356− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

357−

358− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"25"},"finish_reason":null}]}

359−

360− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" f"},"finish_reason":null}]}

361−

362− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"26"},"finish_reason":null}]}

363−

364− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" i"},"finish_reason":null}]}

365−

366− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"27"},"finish_reason":null}]}

367−

368− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" l"},"finish_reason":null}]}

369−

370− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"28"},"finish_reason":null}]}

371−

372− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" e"},"finish_reason":null}]}

373−

374− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"29"},"finish_reason":null}]}

375−

376− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" s"},"finish_reason":null}]}

377−

378− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"30"},"finish_reason":null}]}

379−

380− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

381−

382− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" So"},"finish_reason":null}]}

383−

384− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

385−

386− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"30"},"finish_reason":null}]}

387−

388− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

389−

390− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

391−

392− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" under"},"finish_reason":null}]}

393−

394− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

395−

396− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

397−

398− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

399−

400− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Provide"},"finish_reason":null}]}

401−

402− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" that"},"finish_reason":null}]}

403−

404− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

405−

406− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"G"},"finish_reason":null}]}

407−

408− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"rep"},"finish_reason":null}]}

409−

410− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" for"},"finish_reason":null}]}

411−

412− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" '"},"finish_reason":null}]}

413−

414− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"package"},"finish_reason":null}]}

415−

416− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"'"},"finish_reason":null}]}

417−

418− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" in"},"finish_reason":null}]}

419−

420− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

421−

422− data: {"id":"chatcmpl-249484b499306e8f4f6345cf","object":"chat.completion.chunk","created":1782413814,"model":"gpt-oss-120b","choices":[],"usage":{"prompt_tokens":201,"completion_tokens":179,"total_tokens":380,"completion_tokens_details":{"reasoning_tokens":129},"cost":{"usd":0.00020686400000000002,"hypercredits":0.00413728},"remaining":{"hypercredits":0}}}

423−

424− data: [DONE]

425−

426− headers:

427− Content-Type:

428− - text/event-stream

429− status: 200 OK

430− code: 200

431− duration: 541.974834ms

432−- id: 2

433− request:

434− proto: HTTP/1.1

435− proto_major: 1

436− proto_minor: 1

437− content_length: 35983

438− host: ""

439− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/grep_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use grep to search for the word 'package' in go files\",\"role\":\"user\"},{\"tool_calls\":[{\"id\":\"call_a57bfe141bf85225\",\"function\":{\"arguments\":\"{\\\"include\\\": \\\"*.go\\\", \\\"pattern\\\": \\\"package\\\"}\",\"name\":\"grep\"},\"type\":\"function\"}],\"role\":\"assistant\"},{\"content\":\"Found 1 matches\\n/tmp/crush-test/TestCoderAgent/glm-5.1/grep_tool/main.go:\\n Line 1, Char 1: package main\\n\",\"tool_call_id\":\"call_a57bfe141bf85225\",\"role\":\"tool\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

440− headers:

441− Accept:

442− - application/json

443− Content-Type:

444− - application/json

445− User-Agent:

446− - Charm-Crush/devel (https://charm.land/crush)

447− url: https://hyper.charm.land/v1/chat/completions

448− method: POST

449− response:

450− proto: HTTP/2.0

451− proto_major: 2

452− proto_minor: 0

453− content_length: -1

454− uncompressed: true

455− body: |+

456− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

457−

458− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"1"},"finish_reason":null}]}

459−

460− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" match"},"finish_reason":null}]}

461−

462− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" found"},"finish_reason":null}]}

463−

464− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":":"},"finish_reason":null}]}

465−

466− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `"},"finish_reason":null}]}

467−

468− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"main"},"finish_reason":null}]}

469−

470− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":".go"},"finish_reason":null}]}

471−

472− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":":"},"finish_reason":null}]}

473−

474− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"1"},"finish_reason":null}]}

475−

476− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`"},"finish_reason":null}]}

477−

478− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" —"},"finish_reason":null}]}

479−

480− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `"},"finish_reason":null}]}

481−

482− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"package"},"finish_reason":null}]}

483−

484− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" main"},"finish_reason":null}]}

485−

486− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`"},"finish_reason":null}]}

487−

488− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

489−

490− data: {"id":"chatcmpl-09dddfb926db7fdea7e49d0d","object":"chat.completion.chunk","created":1782413816,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":59,"completion_tokens":17,"total_tokens":8234,"prompt_tokens_details":{"cached_tokens":8158},"cost":{"usd":0.0001574,"hypercredits":0.0031479999999999998},"remaining":{"hypercredits":0}}}

491−

492− data: [DONE]

493−

494− headers:

495− Content-Type:

496− - text/event-stream

497− status: 200 OK

498− code: 200

499− duration: 1.500804s

internal/agent/testdata/TestCoderAgent/glm-5.1/ls_tool.yamldeleted−485

@@ -1,485 +0,0 @@

1−---

2−version: 2

3−interactions:

4−- id: 0

5− request:

6− proto: HTTP/1.1

7− proto_major: 1

8− proto_minor: 1

9− content_length: 35625

10− host: ""

11− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/ls_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use ls to list the files in the current directory\",\"role\":\"user\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

12− headers:

13− Accept:

14− - application/json

15− Content-Type:

16− - application/json

17− User-Agent:

18− - Charm-Crush/devel (https://charm.land/crush)

19− url: https://hyper.charm.land/v1/chat/completions

20− method: POST

21− response:

22− proto: HTTP/2.0

23− proto_major: 2

24− proto_minor: 0

25− content_length: -1

26− uncompressed: true

27− body: |+

28− data: {"id":"chatcmpl-edc9f684f6145a18378c95dd","object":"chat.completion.chunk","created":1782413818,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

29−

30− data: {"id":"chatcmpl-edc9f684f6145a18378c95dd","object":"chat.completion.chunk","created":1782413818,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_9fc14d4b22da05b9","type":"function","index":0,"function":{"name":"ls","arguments":""}}]},"finish_reason":null}]}

31−

32− data: {"id":"chatcmpl-edc9f684f6145a18378c95dd","object":"chat.completion.chunk","created":1782413818,"model":"glm-5.1","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{}"}}]},"finish_reason":null}]}

33−

34− data: {"id":"chatcmpl-edc9f684f6145a18378c95dd","object":"chat.completion.chunk","created":1782413818,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}

35−

36− data: {"id":"chatcmpl-edc9f684f6145a18378c95dd","object":"chat.completion.chunk","created":1782413818,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":677,"completion_tokens":5,"total_tokens":8161,"prompt_tokens_details":{"cached_tokens":7479},"cost":{"usd":0.0009697999999999999,"hypercredits":0.019395999999999997},"remaining":{"hypercredits":0}}}

37−

38− data: [DONE]

39−

40− headers:

41− Content-Type:

42− - text/event-stream

43− status: 200 OK

44− code: 200

45− duration: 1.511164s

46−- id: 1

47− request:

48− proto: HTTP/1.1

49− proto_major: 1

50− proto_minor: 1

51− content_length: 808

52− host: ""

53− body: '{"messages":[{"content":"You will generate a short title based on the first message a user begins a conversation with.\n\n<rules>\n- Keep the title in the same language that the user wrote their message in.\n- Ensure it is not more than 50 characters long.\n- The title should be a summary of the user''s message.\n- It should be one line long.\n- Do not use quotes or colons.\n- The entire text you return will be used as the title.\n- Never return anything that is more than one sentence (one line) long.\n</rules>\n\n /no_think","role":"system"},{"content":"Generate a concise title for the following content:\n\nuse ls to list the files in the current directory\n <think>\n\n</think>","role":"user"}],"model":"gpt-oss-120b","max_completion_tokens":40,"stream_options":{"include_usage":true},"stream":true}'

54− headers:

55− Accept:

56− - application/json

57− Content-Type:

58− - application/json

59− User-Agent:

60− - Charm-Crush/devel (https://charm.land/crush)

61− url: https://hyper.charm.land/v1/chat/completions

62− method: POST

63− response:

64− proto: HTTP/2.0

65− proto_major: 2

66− proto_minor: 0

67− content_length: -1

68− uncompressed: true

69− body: |+

70− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

71−

72− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" user"},"finish_reason":null}]}

73−

74− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" asks"},"finish_reason":null}]}

75−

76− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

77−

78− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

79−

80− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Generate"},"finish_reason":null}]}

81−

82− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

83−

84− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" concise"},"finish_reason":null}]}

85−

86− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

87−

88− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" for"},"finish_reason":null}]}

89−

90− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

91−

92− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" following"},"finish_reason":null}]}

93−

94− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" content"},"finish_reason":null}]}

95−

96− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":\n\n"},"finish_reason":null}]}

97−

98− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"use"},"finish_reason":null}]}

99−

100− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ls"},"finish_reason":null}]}

101−

102− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" to"},"finish_reason":null}]}

103−

104− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" list"},"finish_reason":null}]}

105−

106− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

107−

108− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

109−

110− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

111−

112− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" the"},"finish_reason":null}]}

113−

114− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" current"},"finish_reason":null}]}

115−

116− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" directory"},"finish_reason":null}]}

117−

118− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"\n\n"},"finish_reason":null}]}

119−

120− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"We"},"finish_reason":null}]}

121−

122− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" need"},"finish_reason":null}]}

123−

124− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" a"},"finish_reason":null}]}

125−

126− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" short"},"finish_reason":null}]}

127−

128− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" title"},"finish_reason":null}]}

129−

130− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

131−

132− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" same"},"finish_reason":null}]}

133−

134− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" language"},"finish_reason":null}]}

135−

136− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

137−

138− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"English"},"finish_reason":null}]}

139−

140− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"),"},"finish_reason":null}]}

141−

142− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" max"},"finish_reason":null}]}

143−

144− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

145−

146− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

147−

148− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

149−

150− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

151−

152− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" one"},"finish_reason":null}]}

153−

154− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" line"},"finish_reason":null}]}

155−

156− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":","},"finish_reason":null}]}

157−

158− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" no"},"finish_reason":null}]}

159−

160− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" quotes"},"finish_reason":null}]}

161−

162− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" or"},"finish_reason":null}]}

163−

164− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" col"},"finish_reason":null}]}

165−

166− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"ons"},"finish_reason":null}]}

167−

168− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

169−

170− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Something"},"finish_reason":null}]}

171−

172− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" like"},"finish_reason":null}]}

173−

174− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" \""},"finish_reason":null}]}

175−

176− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"List"},"finish_reason":null}]}

177−

178− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

179−

180− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

181−

182− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" current"},"finish_reason":null}]}

183−

184− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" directory"},"finish_reason":null}]}

185−

186− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" using"},"finish_reason":null}]}

187−

188− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ls"},"finish_reason":null}]}

189−

190− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"\"."},"finish_reason":null}]}

191−

192− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" That's"},"finish_reason":null}]}

193−

194− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" within"},"finish_reason":null}]}

195−

196− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

197−

198− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"50"},"finish_reason":null}]}

199−

200− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" chars"},"finish_reason":null}]}

201−

202− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"?"},"finish_reason":null}]}

203−

204− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" Count"},"finish_reason":null}]}

205−

206− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":":"},"finish_reason":null}]}

207−

208− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" List"},"finish_reason":null}]}

209−

210− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

211−

212− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

213−

214− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

215−

216− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

217−

218− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

219−

220− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

221−

222− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

223−

224− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" files"},"finish_reason":null}]}

225−

226− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

227−

228− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

229−

230− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

231−

232− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

233−

234− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

235−

236− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

237−

238− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

239−

240− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" in"},"finish_reason":null}]}

241−

242− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

243−

244− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

245−

246− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

247−

248− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

249−

250− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

251−

252− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

253−

254− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

255−

256− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" current"},"finish_reason":null}]}

257−

258− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

259−

260− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

261−

262− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

263−

264− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

265−

266− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

267−

268− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

269−

270− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

271−

272− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" directory"},"finish_reason":null}]}

273−

274− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

275−

276− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"9"},"finish_reason":null}]}

277−

278− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

279−

280− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

281−

282− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

283−

284− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

285−

286− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

287−

288− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" using"},"finish_reason":null}]}

289−

290− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

291−

292− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

293−

294− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

295−

296− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" space"},"finish_reason":null}]}

297−

298− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

299−

300− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

301−

302− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

303−

304− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ls"},"finish_reason":null}]}

305−

306− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ("},"finish_reason":null}]}

307−

308− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

309−

310− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":")"},"finish_reason":null}]}

311−

312− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

313−

314− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" total"},"finish_reason":null}]}

315−

316− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

317−

318− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"4"},"finish_reason":null}]}

319−

320− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

321−

322− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

323−

324− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

325−

326− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

327−

328− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

329−

330− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

331−

332− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

333−

334− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

335−

336− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

337−

338− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

339−

340− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

341−

342− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"7"},"finish_reason":null}]}

343−

344− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

345−

346− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

347−

348− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

349−

350− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"9"},"finish_reason":null}]}

351−

352− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

353−

354− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

355−

356− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

357−

358− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"5"},"finish_reason":null}]}

359−

360− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

361−

362− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"1"},"finish_reason":null}]}

363−

364− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"+"},"finish_reason":null}]}

365−

366− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"2"},"finish_reason":null}]}

367−

368− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" ="},"finish_reason":null}]}

369−

370− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" "},"finish_reason":null}]}

371−

372− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"41"},"finish_reason":null}]}

373−

374− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" characters"},"finish_reason":null}]}

375−

376− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

377−

378− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" OK"},"finish_reason":null}]}

379−

380− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":".\n\n"},"finish_reason":null}]}

381−

382− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"Return"},"finish_reason":null}]}

383−

384− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":" that"},"finish_reason":null}]}

385−

386− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"reasoning_content":"."},"finish_reason":null}]}

387−

388− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":"List"},"finish_reason":null}]}

389−

390− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" files"},"finish_reason":null}]}

391−

392− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" in"},"finish_reason":null}]}

393−

394− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" current"},"finish_reason":null}]}

395−

396− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" directory"},"finish_reason":null}]}

397−

398− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" using"},"finish_reason":null}]}

399−

400− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{"content":" ls"},"finish_reason":null}]}

401−

402− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

403−

404− data: {"id":"chatcmpl-a1787c3e58fc5f4d633cd7f2","object":"chat.completion.chunk","created":1782413817,"model":"gpt-oss-120b","choices":[],"usage":{"prompt_tokens":155,"completion_tokens":172,"total_tokens":391,"prompt_tokens_details":{"cached_tokens":64},"cost":{"usd":0.000122704,"hypercredits":0.00245408},"remaining":{"hypercredits":0}}}

405−

406− data: [DONE]

407−

408− headers:

409− Content-Type:

410− - text/event-stream

411− status: 200 OK

412− code: 200

413− duration: 531.140917ms

414−- id: 2

415− request:

416− proto: HTTP/1.1

417− proto_major: 1

418− proto_minor: 1

419− content_length: 35898

420− host: ""

421− body: "{\"messages\":[{\"content\":\"You are Crush, a powerful AI Assistant that runs in the CLI.\\n\\n<critical_rules>\\nThese rules override everything else. Follow them strictly:\\n\\n1. **READ THE RELEVANT CONTEXT BEFORE EDITING**: Never edit a file you haven't already read the relevant context for in this conversation. Once read, you don't need to re-read unless it changed. Pay close attention to exact formatting, indentation, and whitespace - these must match exactly in your edits.\\n2. **BE AUTONOMOUS**: Don't ask questions - search, read, think, decide, act. Break complex tasks into steps and complete them all. Systematically try alternative strategies (different commands, search terms, tools, refactors, or scopes) until either the task is complete or you hit a hard external limit (missing credentials, permissions, files, or network access you cannot change). Only stop for actual blocking errors, not perceived difficulty.\\n3. **TEST AFTER CHANGES**: Run tests immediately after each modification.\\n4. **BE CONCISE**: Keep output concise (default <4 lines), unless explaining complex changes or asked for detail. Conciseness applies to output only, not to thoroughness of work.\\n5. **USE EXACT MATCHES**: When editing, match text exactly including whitespace, indentation, and line breaks.\\n6. **NEVER COMMIT**: Unless user explicitly says \\\"commit\\\". When committing, follow the `<git_commits>` format from the bash tool description exactly, including any configured attribution lines.\\n7. **FOLLOW MEMORY FILE INSTRUCTIONS**: If memory files contain specific instructions, preferences, or commands, you MUST follow them.\\n8. **NEVER ADD COMMENTS**: Only add comments if the user asked you to do so. Focus on *why* not *what*. NEVER communicate with the user through code comments.\\n9. **SECURITY FIRST**: Only assist with defensive security tasks. Refuse to create, modify, or improve code that may be used maliciously.\\n10. **NO URL GUESSING**: Only use URLs provided by the user or found in local files.\\n11. **NEVER PUSH TO REMOTE**: Don't push changes to remote repositories unless explicitly asked.\\n12. **DON'T REVERT CHANGES**: Don't revert changes unless they caused errors or the user explicitly asks.\\n13. **TOOL CONSTRAINTS**: Only use documented tools. Never attempt 'apply_patch' or 'apply_diff' - they don't exist. Use 'edit' or 'multiedit' instead.\\n14. **LOAD MATCHING SKILLS**: If any entry in `<available_skills>` matches the current task, you MUST call `view` on its `<location>` before taking any other action for that task. The `<description>` is only a trigger — the actual procedure, scripts, and references live in SKILL.md. Do NOT infer a skill's behavior from its description or skip loading it because you think you already know how to do the task.\\n15. **LIMIT FILE READS**: Avoid reading entire files, as they can be very large. Read only the sections you need using 'offset' and 'limit' parameters.\\n</critical_rules>\\n\\n<communication_style>\\nKeep responses minimal:\\n- ALWAYS think and respond in the same spoken language the prompt was written in.\\n- Under 4 lines of text (tool use doesn't count)\\n- Conciseness is about **text only**: always fully implement the requested feature, tests, and wiring even if that requires many tool calls.\\n- No preamble (\\\"Here's...\\\", \\\"I'll...\\\")\\n- No postamble (\\\"Let me know...\\\", \\\"Hope this helps...\\\")\\n- One-word answers when possible\\n- No emojis ever\\n- No explanations unless user asks\\n- Never send acknowledgement-only responses; after receiving new context or instructions, immediately continue the task or state the concrete next action you will take.\\n- Use rich Markdown formatting (headings, bullet lists, tables, code fences) for any multi-sentence or explanatory answer; only use plain unformatted text if the user explicitly asks.\\n\\nExamples:\\nuser: what is 2+2?\\nassistant: 4\\n\\nuser: list files in src/\\nassistant: [uses ls tool]\\nfoo.c, bar.c, baz.c\\n\\nuser: which file has the foo implementation?\\nassistant: src/foo.c\\n\\nuser: add error handling to the login function\\nassistant: [searches for login, reads file, edits with exact match, runs tests]\\nDone\\n\\nuser: Where are errors from the client handled?\\nassistant: Clients are marked as failed in the `connectToServer` function in src/services/process.go:712.\\n</communication_style>\\n\\n<code_references>\\nWhen referencing specific functions or code locations, use the pattern `file_path:line_number` to help users navigate:\\n- Example: \\\"The error is handled in src/main.go:45\\\"\\n- Example: \\\"See the implementation in pkg/utils/helper.go:123-145\\\"\\n</code_references>\\n\\n<workflow>\\nFor every task, follow this sequence internally (don't narrate it):\\n\\n**Before acting**:\\n- Search codebase for relevant files\\n- Read files to understand current state\\n- Check memory for stored commands\\n- Identify what needs to change\\n- Use `git log` and `git blame` for additional context when needed\\n\\n**While acting**:\\n- Read entire file before editing it\\n- Before editing: verify exact whitespace and indentation from View output\\n- Use exact text for find/replace (include whitespace)\\n- Make one logical change at a time\\n- After each change: run tests\\n- If tests fail: fix immediately\\n- If edit fails: read more context, don't guess - the text must match exactly\\n- Keep going until query is completely resolved before yielding to user\\n- For longer tasks, send brief progress updates (under 10 words) BUT IMMEDIATELY CONTINUE WORKING - progress updates are not stopping points\\n\\n**Before finishing**:\\n- Verify ENTIRE query is resolved (not just first step)\\n- All described next steps must be completed\\n- Cross-check the original prompt and your own mental checklist; if any feasible part remains undone, continue working instead of responding.\\n- Run lint/typecheck if in memory\\n- Verify all changes work\\n- Keep response under 4 lines\\n\\n**Key behaviors**:\\n- Use find_references before changing shared code\\n- Follow existing patterns (check similar files)\\n- If stuck, try different approach (don't repeat failures)\\n- Make decisions yourself (search first, don't ask)\\n- Fix problems at root cause, not surface-level patches\\n- Don't fix unrelated bugs or broken tests (mention them in final message if relevant)\\n</workflow>\\n\\n<decision_making>\\n**Make decisions autonomously** - don't ask when you can:\\n- Search to find the answer\\n- Read files to see patterns\\n- Check similar code\\n- Infer from context\\n- Try most likely approach\\n- When requirements are underspecified but not obviously dangerous, make the most reasonable assumptions based on project patterns and memory files, briefly state them if needed, and proceed instead of waiting for clarification.\\n\\n**Only stop/ask user if**:\\n- Truly ambiguous business requirement\\n- Multiple valid approaches with big tradeoffs\\n- Could cause data loss\\n- Exhausted all attempts and hit actual blocking errors\\n\\n**When requesting information/access**:\\n- Exhaust all available tools, searches, and reasonable assumptions first.\\n- Never say \\\"Need more info\\\" without detail.\\n- In the same message, list each missing item, why it is required, acceptable substitutes, and what you already attempted.\\n- State exactly what you will do once the information arrives so the user knows the next step.\\n\\nWhen you must stop, first finish all unblocked parts of the request, then clearly report: (a) what you tried, (b) exactly why you are blocked, and (c) the minimal external action required. Don't stop just because one path failed—exhaust multiple plausible approaches first.\\n\\n**Never stop for**:\\n- Task seems too large (break it down)\\n- Multiple files to change (change them)\\n- Concerns about \\\"session limits\\\" (no such limits exist)\\n- Work will take many steps (do all the steps)\\n\\nExamples of autonomous decisions:\\n- File location → search for similar files\\n- Test command → check package.json/memory\\n- Code style → read existing code\\n- Library choice → check what's used\\n- Naming → follow existing names\\n</decision_making>\\n\\n<editing_files>\\n**Available edit tools:**\\n- `edit` - Single find/replace in a file\\n- `multiedit` - Multiple find/replace operations in one file\\n- `write` - Create/overwrite entire file\\n\\nNever use `apply_patch` or similar - those tools don't exist.\\n\\nCritical: ALWAYS read the relevant context of files before editing them in this conversation.\\n\\nWhen using edit tools:\\n1. Read the relevant context first - note the EXACT indentation (spaces vs tabs, count)\\n2. Copy the exact text including ALL whitespace, newlines, and indentation\\n3. Include 3-5 lines of context before and after the target\\n4. Verify your old_string would appear exactly once in the file\\n5. If uncertain about whitespace, include more surrounding context\\n6. Verify edit succeeded\\n7. Run tests\\n\\n**Whitespace matters**:\\n- Count spaces/tabs carefully (use View tool line numbers as reference)\\n- Include blank lines if they exist\\n- Match line endings exactly\\n- When in doubt, include MORE context rather than less\\n\\nEfficiency tips:\\n- Don't re-read files after successful edits (tool will fail if it didn't work)\\n- Same applies for making folders, deleting files, etc.\\n\\nCommon mistakes to avoid:\\n- Editing without reading first\\n- Approximate text matches\\n- Wrong indentation (spaces vs tabs, wrong count)\\n- Missing or extra blank lines\\n- Not enough context (text appears multiple times)\\n- Trimming whitespace that exists in the original\\n- Not testing after changes\\n</editing_files>\\n\\n<whitespace_and_exact_matching>\\nThe Edit tool is extremely literal. \\\"Close enough\\\" will fail.\\n\\n**Before every edit**:\\n1. View the file and locate the exact lines to change\\n2. Copy the text EXACTLY including:\\n - Every space and tab\\n - Every blank line\\n - Opening/closing braces position\\n - Comment formatting\\n3. Include enough surrounding lines (3-5) to make it unique\\n4. Double-check indentation level matches\\n\\n**Common failures**:\\n- `func foo() {` vs `func foo(){` (space before brace)\\n- Tab vs 4 spaces vs 2 spaces\\n- Missing blank line before/after\\n- `// comment` vs `//comment` (space after //)\\n- Different number of spaces in indentation\\n\\n**If edit fails**:\\n- View the file again at the specific location\\n- Copy even more context\\n- Check for tabs vs spaces\\n- Verify line endings\\n- Try including the entire function/block if needed\\n- Never retry with guessed changes - get the exact text first\\n</whitespace_and_exact_matching>\\n\\n<task_completion>\\nEnsure every task is implemented completely, not partially or sketched.\\n\\n1. **Think before acting** (for non-trivial tasks)\\n - Identify all components that need changes (models, logic, routes, config, tests, docs)\\n - Consider edge cases and error paths upfront\\n - Form a mental checklist of requirements before making the first edit\\n - This planning happens internally - don't narrate it to the user\\n\\n2. **Implement end-to-end**\\n - Treat every request as complete work: if adding a feature, wire it fully\\n - Update all affected files (callers, configs, tests, docs)\\n - Don't leave TODOs or \\\"you'll also need to...\\\" - do it yourself\\n - No task is too large - break it down and complete all parts\\n - For multi-part prompts, treat each bullet/question as a checklist item and ensure every item is implemented or answered. Partial completion is not an acceptable final state.\\n\\n3. **Verify before finishing**\\n - Re-read the original request and verify each requirement is met\\n - Check for missing error handling, edge cases, or unwired code\\n - Run tests to confirm the implementation works\\n - Only say \\\"Done\\\" when truly done - never stop mid-task\\n</task_completion>\\n\\n<error_handling>\\nWhen errors occur:\\n1. Read complete error message\\n2. Understand root cause (isolate with debug logs or minimal reproduction if needed)\\n3. Try different approach (don't repeat same action)\\n4. Search for similar code that works\\n5. Make targeted fix\\n6. Test to verify\\n7. For each error, attempt at least two or three distinct remediation strategies (search similar code, adjust commands, narrow or widen scope, change approach) before concluding the problem is externally blocked.\\n\\nCommon errors:\\n- Import/Module → check paths, spelling, what exists\\n- Syntax → check brackets, indentation, typos\\n- Tests fail → read test, see what it expects\\n- File not found → use ls, check exact path\\n\\n**Edit tool \\\"old_string not found\\\"**:\\n- View the file again at the target location\\n- Copy the EXACT text including all whitespace\\n- Include more surrounding context (full function if needed)\\n- Check for tabs vs spaces, extra/missing blank lines\\n- Count indentation spaces carefully\\n- Don't retry with approximate matches - get the exact text\\n</error_handling>\\n\\n<memory_instructions>\\nMemory files store commands, preferences, and codebase info. Update them when you discover:\\n- Build/test/lint commands\\n- Code style preferences\\n- Important codebase patterns\\n- Useful project information\\n</memory_instructions>\\n\\n<code_conventions>\\nBefore writing code:\\n1. Check if library exists (look at imports, package.json)\\n2. Read similar code for patterns\\n3. Match existing style\\n4. Use same libraries/frameworks\\n5. Follow security best practices (never log secrets)\\n6. Don't use one-letter variable names unless requested\\n\\nNever assume libraries are available - verify first.\\n\\n**Ambition vs. precision**:\\n- New projects → be creative and ambitious with implementation\\n- Existing codebases → be surgical and precise, respect surrounding code\\n- Don't change filenames or variables unnecessarily\\n- Don't add formatters/linters/tests to codebases that don't have them\\n</code_conventions>\\n\\n<testing>\\nAfter significant changes:\\n- Start testing as specific as possible to code changed, then broaden to build confidence\\n- Use self-verification: write unit tests, add output logs, or use debug statements to verify your solutions\\n- Run relevant test suite\\n- If tests fail, fix before continuing\\n- Check memory for test commands\\n- Run lint/typecheck if available (on precise targets when possible)\\n- For formatters: iterate max 3 times to get it right; if still failing, present correct solution and note formatting issue\\n- Suggest adding commands to memory if not found\\n- Don't fix unrelated bugs or test failures (not your responsibility)\\n</testing>\\n\\n<tool_usage>\\n- Default to using tools (ls, grep, view, agent, tests, web_fetch, etc.) rather than speculation whenever they can reduce uncertainty or unlock progress, even if it takes multiple tool calls.\\n- Search before assuming\\n- Read files before editing\\n- Always use absolute paths for file operations (editing, reading, writing)\\n- Use Agent tool for complex searches\\n- Run tools in parallel when safe (no dependencies)\\n- When making multiple independent bash calls, send them in a single message with multiple tool calls for parallel execution\\n- Summarize tool output for user (they don't see it)\\n- Never use `curl` through the bash tool it is not allowed use the fetch tool instead.\\n- Only use the tools you know exist.\\n\\n<bash_commands>\\n**CRITICAL**: The `description` parameter is REQUIRED for all bash tool calls. Always provide it.\\n\\nWhen running non-trivial bash commands (especially those that modify the system):\\n- Briefly explain what the command does and why you're running it\\n- This ensures the user understands potentially dangerous operations\\n- Simple read-only commands (ls, cat, etc.) don't need explanation\\n- Use `&` for background processes that won't stop on their own (e.g., `node server.js &`)\\n- Avoid interactive commands - use non-interactive versions (e.g., `npm init -y` not `npm init`)\\n- Combine related commands to save time (e.g., `git status && git diff HEAD && git log -n 3`)\\n</bash_commands>\\n</tool_usage>\\n\\n<proactiveness>\\nBalance autonomy with user intent:\\n- When asked to do something → do it fully (including ALL follow-ups and \\\"next steps\\\")\\n- Never describe what you'll do next - just do it\\n- When the user provides new information or clarification, incorporate it immediately and keep executing instead of stopping with an acknowledgement.\\n- Responding with only a plan, outline, or TODO list (or any other purely verbal response) is failure; you must execute the plan via tools whenever execution is possible.\\n- When asked how to approach → explain first, don't auto-implement\\n- After completing work → stop, don't explain (unless asked)\\n- Don't surprise user with unexpected actions\\n</proactiveness>\\n\\n<final_answers>\\nAdapt verbosity to match the work completed:\\n\\n**Default (under 4 lines)**:\\n- Simple questions or single-file changes\\n- Casual conversation, greetings, acknowledgements\\n- One-word answers when possible\\n\\n**More detail allowed (up to 10-15 lines)**:\\n- Large multi-file changes that need walkthrough\\n- Complex refactoring where rationale adds value\\n- Tasks where understanding the approach is important\\n- When mentioning unrelated bugs/issues found\\n- Suggesting logical next steps user might want\\n- Structure longer answers with Markdown sections and lists, and put all code, commands, and config in fenced code blocks.\\n\\n**What to include in verbose answers**:\\n- Brief summary of what was done and why\\n- Key files/functions changed (with `file:line` references)\\n- Any important decisions or tradeoffs made\\n- Next steps or things user should verify\\n- Issues found but not fixed\\n\\n**What to avoid**:\\n- Don't show full file contents unless explicitly asked\\n- Don't explain how to save files or copy code (user has access to your work)\\n- Don't use \\\"Here's what I did\\\" or \\\"Let me know if...\\\" style preambles/postambles\\n- Keep tone direct and factual, like handing off work to a teammate\\n</final_answers>\\n\\n<env>\\nWorking directory: /tmp/crush-test/TestCoderAgent/glm-5.1/ls_tool\\nIs directory a git repo: no\\nPlatform: linux\\nToday's date: 1/1/2025\\n\\n</env>\\n\\n\\n\\n<available_skills>\\n <skill>\\n <name>crush-hooks</name>\\n <description>Use when the user wants to add, write, debug, or configure a Crush hook — gating or blocking tool calls, approving or rewriting tool input before execution, injecting context into tool results, or troubleshooting hook behavior in crush.json.</description>\\n <location>crush://skills/crush-hooks/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n <skill>\\n <name>jq</name>\\n <description>Use when the user needs to query, filter, reshape, extract, create, or construct JSON data — including API responses, config files, log output, or any structured data — or when helping the user write or debug JSON transformations.</description>\\n <location>crush://skills/jq/SKILL.md</location>\\n <type>builtin</type>\\n </skill>\\n</available_skills>\\n\\n<skills_usage>\\nThe `<description>` of each skill is a TRIGGER — it tells you *when* a skill applies. It is NOT a specification of what the skill does or how to do it. The procedure, scripts, commands, references, and required flags live only in the SKILL.md body. You do not know what a skill actually does until you have read its SKILL.md.\\n\\nMANDATORY activation flow:\\n1. Scan `<available_skills>` against the current user task.\\n2. If any skill's `<description>` matches, call the View tool with its `<location>` EXACTLY as shown — before any other tool call that performs the task.\\n3. Read the entire SKILL.md and follow its instructions.\\n4. Only then execute the task, using the skill's prescribed commands/tools.\\n\\nDo NOT skip step 2 because you think you already know how to do the task. Do NOT infer a skill's behavior from its name or description. If you find yourself about to run `bash`, `edit`, or any task-doing tool for a skill-eligible request without having just viewed the SKILL.md, stop and load the skill first.\\n\\nBuiltin skills (type=builtin) use virtual `crush://skills/...` location identifiers. The \\\"crush://\\\" prefix is NOT a URL, network address, or MCP resource — it is a special internal identifier the View tool understands natively. Pass the `<location>` verbatim to View.\\n\\nDo not use MCP tools (including read_mcp_resource) to load skills.\\nIf a skill mentions scripts, references, or assets, they live in the same folder as the skill itself (e.g., scripts/, references/, assets/ subdirectories within the skill's folder).\\n</skills_usage>\\n\\n\\n\\n\\n\",\"role\":\"system\"},{\"content\":\"<system_reminder>This is a reminder that your todo list is currently empty. DO NOT mention this to the user explicitly because they are already aware.\\nIf you are working on tasks that would benefit from a todo list please use the \\\"todos\\\" tool to create one.\\nIf not, please feel free to ignore. Again do not mention this message to the user.</system_reminder>\",\"role\":\"user\"},{\"content\":\"use ls to list the files in the current directory\",\"role\":\"user\"},{\"tool_calls\":[{\"id\":\"call_9fc14d4b22da05b9\",\"function\":{\"arguments\":\"{}\",\"name\":\"ls\"},\"type\":\"function\"}],\"role\":\"assistant\"},{\"content\":\"\\n- /tmp/crush-test/TestCoderAgent/glm-5.1/ls_tool/\\n - go.mod\\n - main.go\\n\",\"tool_call_id\":\"call_9fc14d4b22da05b9\",\"role\":\"tool\"}],\"model\":\"glm-5.1\",\"max_tokens\":10000,\"stream_options\":{\"include_usage\":true},\"tool_choice\":\"auto\",\"tools\":[{\"function\":{\"name\":\"bash\",\"strict\":false,\"description\":\"Execute shell commands; long-running commands automatically move to background and return a shell ID.\\n\\n<cross_platform>\\nUses mvdan/sh interpreter (Bash-compatible on all platforms including Windows).\\nUse forward slashes for paths: \\\"ls C:/foo/bar\\\" not \\\"ls C:\\\\foo\\\\bar\\\".\\nCommon shell builtins and core utils available on Windows.\\n</cross_platform>\\n\\n<execution_steps>\\n1. Directory Verification: If creating directories/files, use LS tool to verify parent exists\\n2. Security Check: Banned commands (alias, aria2c, axel, chrome, curl, curlie, firefox, http-prompt, httpie, links, lynx, nc, safari, scp, ssh, telnet, w3m, wget, xh, doas, su, sudo, apk, apt, apt-cache, apt-get, dnf, dpkg, emerge, home-manager, makepkg, opkg, pacman, paru, pkg, pkg_add, pkg_delete, portage, rpm, yay, yum, zypper, at, batch, chkconfig, crontab, fdisk, mkfs, mount, parted, service, systemctl, umount, firewall-cmd, ifconfig, ip, iptables, netstat, pfctl, route, ufw) return error - explain to user. Safe read-only commands execute without prompts\\n3. Command Execution: Execute with proper quoting, capture output\\n4. Auto-Background: Commands exceeding 1 minute (default, configurable via `auto_background_after`) automatically move to background and return shell ID\\n5. Output Processing: Truncate if exceeds 30000 characters\\n6. Return Result: Include errors, metadata with <cwd></cwd> tags\\n</execution_steps>\\n\\n<usage_notes>\\n- Command required, working_dir optional (defaults to current directory)\\n- IMPORTANT: Use Grep/Glob/Agent tools instead of 'find'/'grep'. Use View/LS tools instead of 'cat'/'head'/'tail'/'ls'\\n- Chain with ';' or '&&', avoid newlines except in quoted strings\\n- Each command runs in independent shell (no state persistence between calls)\\n- Prefer absolute paths over 'cd' (use 'cd' only if user explicitly requests)\\n</usage_notes>\\n\\n<background_execution>\\n- Set run_in_background=true to run commands in a separate background shell\\n- Returns a shell ID for managing the background process\\n- Use job_output tool to view current output from background shell\\n- Use job_kill tool to terminate a background shell\\n- IMPORTANT: NEVER use `&` at the end of commands to run in background - use run_in_background parameter instead\\n- Commands that should run in background:\\n * Long-running servers (e.g., `npm start`, `python -m http.server`, `node server.js`)\\n * Watch/monitoring tasks (e.g., `npm run watch`, `tail -f logfile`)\\n * Continuous processes that don't exit on their own\\n * Any command expected to run indefinitely\\n- Commands that should NOT run in background:\\n * Build commands (e.g., `npm run build`, `go build`)\\n * Test suites (e.g., `npm test`, `pytest`)\\n * Git operations\\n * File operations\\n * Short-lived scripts\\n</background_execution>\\n\\n<git_message_quality>\\nThese rules apply whenever creating or updating commit messages, PR titles, or PR bodies:\\n\\n- Messages MUST be understandable to someone unfamiliar with the codebase.\\n- Before creating or updating a message, verify this litmus test: a new contributor reading only the commit message or PR title/body should understand what problem this solves, why it matters, and the impact without opening files, reading the diff, or knowing internal code names.\\n- Avoid code identifiers, filenames, function names, and implementation details unless they are necessary for understanding the user-facing impact.\\n- Bad: \\\"Add NameFromHex with sync.Once lazy init\\\"\\n- Good: \\\"Improve color name lookup performance while keeping startup fast\\\"\\n</git_message_quality>\\n\\n<commit_messages>\\nCommit messages are for future readers scanning history. Before committing:\\n\\n- Follow <git_message_quality>.\\n- Draft a concise 1-2 sentence message focusing on why the change exists and what outcome it enables, not a list of files or implementation details.\\n- Use clear, accurate verbs (\\\"add\\\"=new capability, \\\"update\\\"=enhancement, \\\"fix\\\"=bug fix) and avoid generic messages.\\n- The first line MUST be under 72 characters.\\n- Add a body only when it is needed to explain the reasoning, tradeoffs, or important context; wrap body lines at 72 characters.\\n- If the change is internal-only, still describe the benefit or maintenance outcome rather than naming private code.\\n- Bad: \\\"fix: nil pointer in session.go\\\"\\n- Good: \\\"fix: prevent session loading from crashing on missing metadata\\\"\\n- Bad: \\\"refactor: move PromptBuilder into internal/agent\\\"\\n- Good: \\\"refactor: make prompt assembly easier to maintain\\\"\\n</commit_messages>\\n\\n<git_commits>\\nWhen user asks to create git commit:\\n\\n1. Single message with three tool_use blocks (IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - git log (recent commit message style)\\n\\n2. Add relevant untracked files to staging. Don't commit files already modified at conversation start unless relevant.\\n\\n3. Analyze staged changes in <commit_analysis> tags:\\n - List changed/added files, summarize nature (feature/enhancement/bug fix/refactoring/test/docs)\\n - Brainstorm purpose/motivation, assess project impact, check for sensitive info\\n - Don't use tools beyond the context of git\\n\\n4. Draft a commit message:\\n - Follow <commit_messages>\\n - Review draft against the litmus test before committing\\n\\n5. Create commit with attribution using HEREDOC:\\n git commit -m \\\"$(cat <<'EOF'\\nCommit message here.\\n\\n\\n\U0001F498 Generated with Crush\\n\\n\\n\\nCo-Authored-By: Crush <crush@charm.land>\\n\\nEOF\\n)\\\"\\n\\n6. If pre-commit hook fails, retry ONCE. If fails again, hook preventing commit. If succeeds but files modified, MUST amend.\\n\\n7. Run git status to verify.\\n\\nNotes: Use \\\"git commit -am\\\" when possible, don't stage unrelated files, NEVER update config, don't push, no -i flags, no empty commits, return empty response, when rebasing always use -m.\\n</git_commits>\\n\\n<pull_requests>\\n\\n\\nWhen user asks you to create or update a PR:\\n\\n1. Single message with multiple tool_use blocks (VERY IMPORTANT for speed):\\n - git status (untracked files)\\n - git diff (staged/unstaged changes)\\n - Check if branch tracks remote and is up to date\\n - git log and 'git diff main...HEAD' (full commit history from main divergence)\\n\\n2. Create new branch if needed\\n\\n3. Commit changes if needed\\n\\n4. Push to remote with -u flag if needed\\n\\n5. Analyze changes in <pr_analysis> tags:\\n - List commits since diverging from main\\n - Summarize nature of changes\\n - Brainstorm purpose/motivation\\n - Assess project impact\\n - Don't use tools beyond git context\\n - Check for sensitive information\\n\\n6. Draft a PR message:\\n - Follow <git_message_quality>\\n - Draft concise (1-2 bullet points) PR summary focusing on \\\"why\\\"\\n - Ensure summary reflects ALL changes since main divergence\\n - Use clear, concise language\\n - Provide an accurate reflection of changes and purpose\\n - Avoid generic summaries; messages should be thoughtful\\n - Review draft against the litmus test before creating or updating the PR\\n\\n7. Create PR with gh pr create using HEREDOC:\\n gh pr create --title \\\"title\\\" --body \\\"$(cat <<'EOF'\\n\\n<summary>\\n\\n\U0001F498 Generated with Crush\\n\\nEOF\\n)\\\"\\n\\nImportant:\\n\\n- Return empty response - user sees gh output\\n- Never update git config\\n</pull_requests>\\n\\n<examples>\\nGood: pytest /foo/bar/tests\\nBad: cd /foo/bar && pytest tests\\n</examples>\\n\",\"parameters\":{\"properties\":{\"auto_background_after\":{\"description\":\"Seconds to wait before automatically moving the command to a background job (default: 60)\",\"type\":\"integer\"},\"command\":{\"description\":\"The command to execute\",\"type\":\"string\"},\"description\":{\"description\":\"A brief description of what the command does, try to keep it under 30 characters or so\",\"type\":\"string\"},\"run_in_background\":{\"description\":\"Set to true (boolean) to run this command in the background. Use job_output to read the output later.\",\"type\":\"boolean\"},\"working_dir\":{\"description\":\"The working directory to execute the command in (defaults to current directory)\",\"type\":\"string\"}},\"required\":[\"description\",\"command\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"download\",\"strict\":false,\"description\":\"Download a URL directly to a local file (binary-safe, streaming, max 600s timeout); overwrites without warning. For reading content into context use fetch.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The local file path where the downloaded content should be saved\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 600)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to download from\",\"type\":\"string\"}},\"required\":[\"url\",\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"edit\",\"strict\":false,\"description\":\"Edit a file by exact find-and-replace; can also create or delete content. For renames/moves use bash. For large edits use write.\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"},\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false)\",\"type\":\"boolean\"}},\"required\":[\"file_path\",\"old_string\",\"new_string\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"multiedit\",\"strict\":false,\"description\":\"Apply multiple find-and-replace edits to a single file in one operation; edits run sequentially. Prefer over edit for multiple changes to the same file. Same exact-match rules as edit apply.\",\"parameters\":{\"properties\":{\"edits\":{\"description\":\"Array of edit operations to perform sequentially on the file\",\"items\":{\"properties\":{\"new_string\":{\"description\":\"The text to replace it with\",\"type\":\"string\"},\"old_string\":{\"description\":\"The text to replace\",\"type\":\"string\"},\"replace_all\":{\"description\":\"Replace all occurrences of old_string (default false).\",\"type\":\"boolean\"}},\"required\":[\"old_string\",\"new_string\"],\"type\":\"object\"},\"type\":\"array\"},\"file_path\":{\"description\":\"The absolute path to the file to modify\",\"type\":\"string\"}},\"required\":[\"file_path\",\"edits\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"fetch\",\"strict\":false,\"description\":\"Fetch raw content from a URL as text, markdown, or html (max 100KB); no AI processing. For analysis or extraction use agentic_fetch.\\n\",\"parameters\":{\"properties\":{\"format\":{\"description\":\"The format to return the content in (text, markdown, or html)\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"},\"url\":{\"description\":\"The URL to fetch content from\",\"type\":\"string\"}},\"required\":[\"url\",\"format\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"glob\",\"strict\":false,\"description\":\"Find files by name/pattern (glob syntax), sorted by modification time; max 100 results; skips hidden files. Use grep to search file contents.\\n\",\"parameters\":{\"properties\":{\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The glob pattern to match files against\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"grep\",\"strict\":false,\"description\":\"Search file contents by regex or literal text; returns matching file paths sorted by modification time (max 100); respects .gitignore. Use glob to filter by filename, not contents.\\n\",\"parameters\":{\"properties\":{\"include\":{\"description\":\"File pattern to include in the search (e.g. \\\"*.js\\\", \\\"*.{ts,tsx}\\\")\",\"type\":\"string\"},\"literal_text\":{\"description\":\"If true, the pattern will be treated as literal text with special regex characters escaped. Default is false.\",\"type\":\"boolean\"},\"path\":{\"description\":\"The directory to search in. Defaults to the current working directory.\",\"type\":\"string\"},\"pattern\":{\"description\":\"The regex pattern to search for in file contents\",\"type\":\"string\"}},\"required\":[\"pattern\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"ls\",\"strict\":false,\"description\":\"List files and directories as a tree; skips hidden files and common system dirs; max 1000 files. Use glob to find files by pattern, grep to search contents.\\n\",\"parameters\":{\"properties\":{\"depth\":{\"description\":\"The maximum depth to traverse\",\"type\":\"integer\"},\"ignore\":{\"description\":\"List of glob patterns to ignore\",\"items\":{\"type\":\"string\"},\"type\":\"array\"},\"path\":{\"description\":\"The path to the directory to list (defaults to current working directory)\",\"type\":\"string\"}},\"required\":[],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"sourcegraph\",\"strict\":false,\"description\":\"Search code across public GitHub repositories via Sourcegraph; supports regex, language/repo/file filters, and symbol search (max 20 results). Only searches public repos.\\n\",\"parameters\":{\"properties\":{\"context_window\":{\"description\":\"The context around the match to return (default: 10 lines)\",\"type\":\"integer\"},\"count\":{\"description\":\"Optional number of results to return (default: 10, max: 20)\",\"type\":\"integer\"},\"query\":{\"description\":\"The Sourcegraph search query\",\"type\":\"string\"},\"timeout\":{\"description\":\"Optional timeout in seconds (max 120)\",\"type\":\"integer\"}},\"required\":[\"query\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"view\",\"strict\":false,\"description\":\"Read a file by path with line numbers; supports offset and line limit (default 200, max 200KB returned file content section); renders images (PNG, JPEG, GIF, WebP); use ls for directories.\\n\",\"parameters\":{\"properties\":{\"file_path\":{\"description\":\"The path to the file to read\",\"type\":\"string\"},\"limit\":{\"description\":\"The number of lines to read (defaults to 200)\",\"type\":\"integer\"},\"offset\":{\"description\":\"The line number to start reading from (0-based)\",\"type\":\"integer\"}},\"required\":[\"file_path\"],\"type\":\"object\"}},\"type\":\"function\"},{\"function\":{\"name\":\"write\",\"strict\":false,\"description\":\"Create or overwrite a file with given content; auto-creates parent dirs. Cannot append. Read the file first to avoid conflicts. For surgical changes use edit or multiedit.\",\"parameters\":{\"properties\":{\"content\":{\"description\":\"The content to write to the file\",\"type\":\"string\"},\"file_path\":{\"description\":\"The path to the file to write\",\"type\":\"string\"}},\"required\":[\"file_path\",\"content\"],\"type\":\"object\"}},\"type\":\"function\"}],\"stream\":true}"

422− headers:

423− Accept:

424− - application/json

425− Content-Type:

426− - application/json

427− User-Agent:

428− - Charm-Crush/devel (https://charm.land/crush)

429− url: https://hyper.charm.land/v1/chat/completions

430− method: POST

431− response:

432− proto: HTTP/2.0

433− proto_major: 2

434− proto_minor: 0

435− content_length: -1

436− uncompressed: true

437− body: |+

438− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]}

439−

440− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"Two"},"finish_reason":null}]}

441−

442− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" files"},"finish_reason":null}]}

443−

444− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" in"},"finish_reason":null}]}

445−

446− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" the"},"finish_reason":null}]}

447−

448− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" current"},"finish_reason":null}]}

449−

450− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" directory"},"finish_reason":null}]}

451−

452− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":":\n"},"finish_reason":null}]}

453−

454− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"-"},"finish_reason":null}]}

455−

456− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `"},"finish_reason":null}]}

457−

458− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"go"},"finish_reason":null}]}

459−

460− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":".mod"},"finish_reason":null}]}

461−

462− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`\n"},"finish_reason":null}]}

463−

464− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"-"},"finish_reason":null}]}

465−

466− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":" `"},"finish_reason":null}]}

467−

468− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"main"},"finish_reason":null}]}

469−

470− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":".go"},"finish_reason":null}]}

471−

472− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{"content":"`"},"finish_reason":null}]}

473−

474− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

475−

476− data: {"id":"chatcmpl-0df4ed5d1b2533d9447fa033","object":"chat.completion.chunk","created":1782413819,"model":"glm-5.1","choices":[],"usage":{"prompt_tokens":8195,"completion_tokens":19,"total_tokens":8214,"cost":{"usd":0.0115566,"hypercredits":0.231132},"remaining":{"hypercredits":0}}}

477−

478− data: [DONE]

479−

480− headers:

481− Content-Type:

482− - text/event-stream

483− status: 200 OK

484− code: 200

485− duration: 1.065962584s

internal/agent/testdata/TestCoderAgent/glm-5.1/multiedit_tool.yamldeleted−1439