perf(chat): throttle streaming markdown render and defer chart mounts

- add useStreamingMarkdown: cap mid-stream markdown re-render to ~140ms,
  full-fidelity render once the segment completes
- skip code-block language auto-detection while streaming (escaped plain
  text), restore full highlighting on the final render
- defer echarts/mermaid blocks to a lightweight loading placeholder while
  streaming so their parsers never run on truncated source
- bypass the render cache for streaming-mode output
- wire into ContentSegment and MessageBubble (content + thinking)

mateclaw-ui/src/assets/main.css

@@ -616,6 +616,29 @@ html.dark body::before {
   font-size: 13px;
 }
 
+/* Streaming placeholder for echarts/mermaid blocks (shown until the fenced
+   source finishes streaming, then replaced by the real chart). */
+.markdown-body .chart-loading {
+  display: flex;
+  align-items: center;
+  justify-content: center;
+  gap: 8px;
+  margin: 14px 0;
+  min-height: 120px;
+  border-radius: 12px;
+  background: var(--mc-bg-elevated);
+  border: 1px solid var(--mc-border-light);
+}
+.markdown-body .chart-loading__dot {
+  width: 9px;
+  height: 9px;
+  border-radius: 50%;
+  background: var(--mc-border, #cbd5e1);
+  animation: mc-product-pulse 1.4s ease-in-out infinite;
+}
+.markdown-body .chart-loading__dot:nth-child(2) { animation-delay: 0.2s; }
+.markdown-body .chart-loading__dot:nth-child(3) { animation-delay: 0.4s; }
+
 /* ================================================================
    Product cards (from ```product-cards fenced blocks — shopping /
    price-comparison results rendered as a clickable card grid)

mateclaw-ui/src/components/chat/ContentSegment.vue

@@ -1,6 +1,6 @@
 <script setup lang="ts">
 import { computed } from 'vue'
-import { useMarkdownRenderer } from '@/composables/useMarkdownRenderer'
+import { useStreamingMarkdown } from '@/composables/useStreamingMarkdown'
 import TypingCursor from './TypingCursor.vue'
 import type { MessageSegment } from '@/types'
 
@@ -11,10 +11,14 @@ const props = withDefaults(defineProps<{
   showCursor: false,
 })
 
-const { renderMarkdown } = useMarkdownRenderer()
-
-const renderedContent = computed(() => renderMarkdown(props.segment.text || ''))
 const isRunning = computed(() => props.segment.status === 'running')
+
+// Throttle markdown rendering while the segment streams; render once at full
+// fidelity the moment it completes.
+const { html: renderedContent } = useStreamingMarkdown(
+  () => props.segment.text || '',
+  () => isRunning.value,
+)
 </script>
 
 <template>

mateclaw-ui/src/components/chat/MessageBubble.vue

@@ -440,7 +440,7 @@ import {
   VideoPause,
   WarningFilled,
 } from '@element-plus/icons-vue'
-import { useMarkdownRenderer } from '@/composables/useMarkdownRenderer'
+import { useStreamingMarkdown } from '@/composables/useStreamingMarkdown'
 import { useAuthenticatedAttachment } from '@/composables/useAuthenticatedAttachment'
 import { useToolLabel } from '@/composables/useToolLabel'
 import { http } from '@/api'
@@ -460,7 +460,6 @@ import type { BrowserAction } from './BrowserTimeline.vue'
 import type { Message, MessageSegment, ChatAttachment, ToolCallMeta, PlanMeta } from '@/types'
 import type { ChatErrorInfo } from '@/types/chatError'
 
-const { renderMarkdown } = useMarkdownRenderer()
 const { t, locale } = useI18n()
 const { getToolLabel } = useToolLabel()
 const { blobUrls, loadAllImages, loadAllVideos, loadAllAudios, loadAllModels, downloadFile, openImage, getDisplayUrl, revokeAll } = useAuthenticatedAttachment()
@@ -614,10 +613,12 @@ const toggleThinking = () => {
   emit('toggle-thinking', localThinkingExpanded.value)
 }
 
-const renderedThinkingContent = computed(() => {
-  if (!thinkingContent.value) return ''
-  return renderMarkdown(thinkingContent.value)
-})
+// Throttle thinking + main-content markdown while the turn streams; both render
+// once at full fidelity when generation stops.
+const { html: renderedThinkingContent } = useStreamingMarkdown(
+  () => thinkingContent.value,
+  () => isGenerating.value,
+)
 
 // --- 主内容 ---
 const isApprovalPlaceholder = (text: string) => {
@@ -643,10 +644,10 @@ const parseErrorText = computed(() => {
   return errorPart?.text || ''
 })
 
-const renderedContent = computed(() => {
-  if (!displayContent.value) return ''
-  return renderMarkdown(displayContent.value)
-})
+const { html: renderedContent } = useStreamingMarkdown(
+  () => displayContent.value,
+  () => isGenerating.value,
+)
 
 const showLoadingIndicator = computed(() => {
   return isGenerating.value && !displayContent.value

mateclaw-ui/src/composables/__tests__/streaming-render.test.ts

@@ -0,0 +1,72 @@
+// @vitest-environment happy-dom
+import { describe, it, expect } from 'vitest'
+import { useMarkdownRenderer } from '../useMarkdownRenderer'
+
+const { renderMarkdown } = useMarkdownRenderer()
+
+// An unlabeled fenced code block. The default (final) render auto-detects the
+// language via hljs.highlightAuto — the single most expensive step. The
+// streaming render must skip it and emit escaped plain text instead.
+const UNLABELED_CODE = `\`\`\`
+function add(a, b) {
+  return a + b
+}
+\`\`\``
+
+describe('streaming markdown render mode', () => {
+  it('skips code auto-highlight while streaming', () => {
+    const streamed = renderMarkdown(UNLABELED_CODE, { streaming: true })
+    // No highlight.js token spans in the streamed (cheap) render. The
+    // structural `hljs-lines` gutter wrapper is always present, so we assert on
+    // the token spans (hljs-keyword / hljs-string / …) specifically.
+    expect(streamed).not.toContain('<span class="hljs-')
+    // The code still shows, just as escaped plain text inside the gutter list.
+    expect(streamed).toContain('function add')
+  })
+
+  it('auto-highlights the same block on the final (non-streaming) render', () => {
+    const final = renderMarkdown(UNLABELED_CODE, { streaming: false })
+    // highlight.js emits token classes once auto-detection runs.
+    expect(final).toContain('<span class="hljs-')
+  })
+
+  it('respects an explicit language even while streaming', () => {
+    const labeled = '```js\nconst x = 1\n```'
+    const streamed = renderMarkdown(labeled, { streaming: true })
+    // Explicit language uses single-grammar hljs.highlight (cheap), kept on.
+    expect(streamed).toContain('<span class="hljs-')
+  })
+
+  it('does not serve a streaming render back to a later final render', () => {
+    // Same source text rendered streaming-first then final: the final must not
+    // be the cached low-fidelity streaming output.
+    const md = '```\nlet y = 2\n```'
+    const streamed = renderMarkdown(md, { streaming: true })
+    const final = renderMarkdown(md, { streaming: false })
+    expect(streamed).not.toContain('<span class="hljs-')
+    expect(final).toContain('<span class="hljs-')
+  })
+
+  it('defers echarts to a placeholder while streaming, real block on final', () => {
+    const md = '```echarts\n{"series":[{"type":"bar","data":[1,2,3]}]}\n```'
+    const streamed = renderMarkdown(md, { streaming: true })
+    expect(streamed).toContain('chart-loading')
+    expect(streamed).not.toContain('echarts-block')
+    expect(streamed).not.toContain('data-echarts-option')
+
+    const final = renderMarkdown(md, { streaming: false })
+    expect(final).toContain('class="echarts-block"')
+    expect(final).toContain('data-echarts-option')
+  })
+
+  it('defers mermaid to a placeholder while streaming, real block on final', () => {
+    const md = '```mermaid\ngraph TD; A-->B;\n```'
+    const streamed = renderMarkdown(md, { streaming: true })
+    expect(streamed).toContain('chart-loading')
+    expect(streamed).not.toContain('mermaid-block')
+
+    const final = renderMarkdown(md, { streaming: false })
+    expect(final).toContain('class="mermaid-block"')
+    expect(final).toContain('data-mermaid')
+  })
+})

mateclaw-ui/src/composables/useMarkdownRenderer.ts

@@ -268,6 +268,23 @@ function renderProductCards(rawCode: string): string {
   return `<div class="product-cards">${cards}</div>`
 }
 
+/**
+ * Lightweight loading placeholder shown in place of an echarts/mermaid block
+ * during throttled mid-stream renders. These post-mount renderers parse the
+ * fenced source (echarts: JSON.parse, mermaid: diagram grammar), which fails
+ * loudly on the half-emitted source of a still-streaming block. Emitting a
+ * neutral placeholder (no `.echarts-block` / `.mermaid-block` class, so the
+ * observers ignore it) avoids the parse churn and the mount/dispose flicker;
+ * the final non-streaming render emits the real block and mounts once.
+ */
+function chartLoadingPlaceholder(): string {
+  return '<div class="chart-loading" aria-label="Loading chart">'
+    + '<span class="chart-loading__dot"></span>'
+    + '<span class="chart-loading__dot"></span>'
+    + '<span class="chart-loading__dot"></span>'
+    + '</div>'
+}
+
 // ---------------------------------------------------------------------------
 // Custom renderer (marked v15 requires a plain object — class instances are
 // NOT dispatched).
@@ -283,6 +300,8 @@ const customRenderer = {
     // so users can copy the diagram source even before render completes; the
     // composable paints the SVG into `.mermaid-block__body`.
     if (infoStr === 'mermaid') {
+      // Mid-stream: defer to a placeholder; mermaid can't parse a partial diagram.
+      if (streamingRenderMode) return chartLoadingPlaceholder()
       const encoded = encodeURIComponent(rawCode)
       const copySvg = `<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><rect x="9" y="9" width="13" height="13" rx="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>`
       const downloadSvg = `<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M21 15v4a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2v-4"/><polyline points="7 10 12 15 17 10"/><line x1="12" y1="15" x2="12" y2="3"/></svg>`
@@ -306,6 +325,8 @@ const customRenderer = {
 
     // ECharts: same pattern, mounted by useEChartsRenderer.
     if (infoStr === 'echarts') {
+      // Mid-stream: defer to a placeholder; the option JSON is still truncated.
+      if (streamingRenderMode) return chartLoadingPlaceholder()
       return `<div class="echarts-block" data-echarts-option="${encodeURIComponent(rawCode)}"></div>`
     }
 
@@ -323,9 +344,18 @@ const customRenderer = {
 
     let highlighted: string
     try {
-      highlighted = hasLanguage
-        ? hljs.highlight(rawCode, { language: detectedLang }).value
-        : hljs.highlightAuto(rawCode).value
+      if (hasLanguage) {
+        highlighted = hljs.highlight(rawCode, { language: detectedLang }).value
+      } else if (streamingRenderMode) {
+        // Mid-stream throttled render: skip language auto-detection. hljs
+        // probes every registered grammar, which is the single most expensive
+        // step in the pipeline and would re-run on each throttled pass over a
+        // still-growing block. Show escaped plain text now; the final
+        // (non-streaming) render does the real auto-highlight once.
+        highlighted = escapeHtml(rawCode)
+      } else {
+        highlighted = hljs.highlightAuto(rawCode).value
+      }
     } catch {
       highlighted = escapeHtml(rawCode)
     }
@@ -511,19 +541,40 @@ export type WikilinkMode = 'legacy' | 'none'
 export interface RenderMarkdownOptions {
   /** How to handle `[[...]]` syntax. Defaults to `'legacy'`. */
   wikilink?: WikilinkMode
+  /**
+   * Streaming-friendly render. When `true`, the renderer skips code-block
+   * language auto-detection (the most expensive step) and bypasses the LRU
+   * cache. Use it for the throttled mid-stream renders driven by
+   * {@link useStreamingMarkdown}; the final render must run with this off so
+   * the completed message gets full-fidelity highlighting.
+   */
+  streaming?: boolean
 }
 
+// Module-level flag read by the custom code renderer. Safe because
+// `markedInstance.parse()` runs fully synchronously (JS single-threaded) — the
+// flag is set immediately before the parse and cleared in a `finally`, so it
+// can never leak across renders.
+let streamingRenderMode = false
+
 export function useMarkdownRenderer() {
   function renderMarkdown(content: string, opts?: RenderMarkdownOptions): string {
     if (!content) return ''
     const wikilink: WikilinkMode = opts?.wikilink ?? 'legacy'
+    const streaming = opts?.streaming ?? false
     const k = cacheKey(content, wikilink)
-    const cached = RENDER_CACHE.get(k)
-    if (cached !== undefined) {
-      // Refresh LRU position — re-insert at the tail.
-      RENDER_CACHE.delete(k)
-      RENDER_CACHE.set(k, cached)
-      return cached
+    // Streaming renders bypass the cache entirely: their length-based keys
+    // collide with the final full-fidelity render of the same text, and a
+    // streaming entry (no auto-highlight) must never be served as the final
+    // result.
+    if (!streaming) {
+      const cached = RENDER_CACHE.get(k)
+      if (cached !== undefined) {
+        // Refresh LRU position — re-insert at the tail.
+        RENDER_CACHE.delete(k)
+        RENDER_CACHE.set(k, cached)
+        return cached
+      }
     }
 
     // 1. LaTeX placeholders (skips fenced/inline code).
@@ -554,9 +605,20 @@ export function useMarkdownRenderer() {
             },
           )
     // 3. Marked → 4. DOMPurify.
-    const rawHtml = markedInstance.parse(withWikiLinks) as string
+    let rawHtml: string
+    streamingRenderMode = streaming
+    try {
+      rawHtml = markedInstance.parse(withWikiLinks) as string
+    } finally {
+      streamingRenderMode = false
+    }
     const result = DOMPurify.sanitize(rawHtml, purifyConfig)
 
+    if (streaming) {
+      // Throwaway render — don't pollute the LRU with low-fidelity entries.
+      return result
+    }
+
     // Evict oldest entry when at capacity (Map preserves insertion order).
     if (RENDER_CACHE.size >= RENDER_CACHE_CAP) {
       const oldestKey = RENDER_CACHE.keys().next().value

mateclaw-ui/src/composables/useStreamingMarkdown.ts

@@ -0,0 +1,87 @@
+import { ref, watch, onUnmounted, type Ref } from 'vue'
+import { useMarkdownRenderer, type RenderMarkdownOptions } from '@/composables/useMarkdownRenderer'
+
+/**
+ * Throttle interval (ms) for mid-stream markdown re-renders.
+ *
+ * Streaming emits one `content_delta` per token; rendering the full accumulated
+ * markdown on every delta is O(n²) over the message and re-runs marked +
+ * highlight.js + DOMPurify each time — the dominant source of streaming jank.
+ * Throttling to ~7 renders/sec keeps the text feeling live while cutting the
+ * render count ~10× on a fast model. The final render (when streaming stops)
+ * always runs immediately at full fidelity.
+ */
+const STREAM_RENDER_INTERVAL_MS = 140
+
+/**
+ * Reactive, throttled markdown rendering for streaming message content.
+ *
+ * While `streaming()` is true the rendered HTML is refreshed at most once per
+ * {@link STREAM_RENDER_INTERVAL_MS} and uses the renderer's `streaming` mode
+ * (skips code auto-highlight, bypasses the cache). The moment `streaming()`
+ * flips to false — or the source changes while already idle — it renders once
+ * immediately with full fidelity (auto-highlight + cache).
+ *
+ * @param source     getter for the raw markdown text
+ * @param streaming  getter that is true while the text is still being streamed
+ * @param opts       passed through to `renderMarkdown` (e.g. `wikilink`)
+ */
+export function useStreamingMarkdown(
+  source: () => string,
+  streaming: () => boolean,
+  opts?: RenderMarkdownOptions,
+): { html: Ref<string> } {
+  const { renderMarkdown } = useMarkdownRenderer()
+  const html = ref('')
+
+  let throttleTimer: ReturnType<typeof setTimeout> | null = null
+  let pendingText: string | null = null
+  let lastRenderMs = 0
+
+  const clearTimer = () => {
+    if (throttleTimer) {
+      clearTimeout(throttleTimer)
+      throttleTimer = null
+    }
+  }
+
+  const renderNow = (text: string, stream: boolean) => {
+    html.value = text ? renderMarkdown(text, { ...opts, streaming: stream }) : ''
+    lastRenderMs = Date.now()
+  }
+
+  watch(
+    [source, streaming],
+    ([text, isStreaming]) => {
+      if (!isStreaming) {
+        // Idle / completed: full-fidelity render now, drop any pending throttle.
+        clearTimer()
+        pendingText = null
+        renderNow(text, false)
+        return
+      }
+      // Streaming: leading-edge render if we're past the interval, otherwise
+      // coalesce into a single trailing render.
+      pendingText = text
+      const elapsed = Date.now() - lastRenderMs
+      if (elapsed >= STREAM_RENDER_INTERVAL_MS) {
+        clearTimer()
+        renderNow(text, true)
+        pendingText = null
+      } else if (!throttleTimer) {
+        throttleTimer = setTimeout(() => {
+          throttleTimer = null
+          if (pendingText !== null) {
+            renderNow(pendingText, true)
+            pendingText = null
+          }
+        }, STREAM_RENDER_INTERVAL_MS - elapsed)
+      }
+    },
+    { immediate: true },
+  )
+
+  onUnmounted(clearTimer)
+
+  return { html }
+}