3295a08 · crush · git.iko.soy

git.iko.soy/crush

on main help sign in

fix: cache streaming thinking renders to avoid CPU burn during long reasoning traces

3295a08Kieran Klukas4 months ago

internal/ui/chat/assistant.go+15−7

compare with main

unifiedsplit

@@ -144,6 +144,13 @@ type AssistantMessageItem struct {

144 // docs/notes/2026-05-12-chat-rendering-perf.md. See

145 // streaming_markdown.go for the full algorithm.

146 streamingContent streamingMarkdown

147+

148+ // streamingThinking applies the same stable-prefix caching to

149+ // the thinking/reasoning section. Without this, every streaming

150+ // delta forces a full glamour re-render of the entire accumulated

151+ // thinking text, which burns CPU and starves the terminal emulator

152+ // during long reasoning traces.

153+ streamingThinking streamingMarkdown

147}

154}

148

155

149var _ Expandable = (*AssistantMessageItem)(nil)

156var _ Expandable = (*AssistantMessageItem)(nil)

@@ -447,13 +454,7 @@ func (a *AssistantMessageItem) cachedError(width int) string {

447// lines so the visual box matches what the user sees today.

454// lines so the visual box matches what the user sees today.

448func (a *AssistantMessageItem) renderThinking(thinking string, width int) string {

455func (a *AssistantMessageItem) renderThinking(thinking string, width int) string {

449 renderer := common.QuietMarkdownRenderer(a.sty, width)

456 renderer := common.QuietMarkdownRenderer(a.sty, width)

450− mu := common.LockMarkdownRenderer(renderer)

457+ rendered := a.streamingThinking.Render(thinking, width, renderer)

451− mu.Lock()

452− rendered, err := renderer.Render(thinking)

453− mu.Unlock()

454− if err != nil {

455− rendered = thinking

456− }

457 rendered = strings.TrimSpace(rendered)

458 rendered = strings.TrimSpace(rendered)

458

459

459 lines := strings.Split(rendered, "\n")

460 lines := strings.Split(rendered, "\n")

@@ -587,6 +588,7 @@ func (a *AssistantMessageItem) clearCache() {

587 a.contentSec.reset()

588 a.contentSec.reset()

588 a.errorSec.reset()

589 a.errorSec.reset()

589 a.streamingContent.Reset()

590 a.streamingContent.Reset()

591+ a.streamingThinking.Reset()

590}

592}

591

593

592// ToggleExpanded advances the F5 thinking view-mode cycle and returns

594// ToggleExpanded advances the F5 thinking view-mode cycle and returns

@@ -617,6 +619,12 @@ func (a *AssistantMessageItem) ToggleExpanded() bool {

617 case thinkingFullExpanded:

619 case thinkingFullExpanded:

618 a.thinkingViewMode = thinkingCollapsed

620 a.thinkingViewMode = thinkingCollapsed

619 }

621 }

622+ // View-mode changes alter the windowing slice applied after

623+ // glamour render. The streaming prefix cache may have been

624+ // seeded under a different slice regime, and glued renders are

625+ // not byte-identical to monolithic ones. Drop the prefix cache

626+ // so the next render is clean.

627+ a.streamingThinking.Reset()

620 a.Bump()

628 a.Bump()

621 return a.thinkingViewMode != thinkingCollapsed

629 return a.thinkingViewMode != thinkingCollapsed

622}

630}