perf(ui): tighten Shiki highlight caches and parallelize fences

Use content fingerprints instead of full source as cache keys, record
entry sizes once (no TextEncoder/JSON.stringify on get/evict), and
highlight multiple markdown fences concurrently.

Co-authored-by: Serhii Dziupin <makeittech@users.noreply.github.com>

packages/ui/src/components/chat/MarkdownRendererImpl.tsx

@@ -1117,18 +1117,14 @@ const SimpleMarkdownRendererImpl: React.FC<{
 
   const syntaxVars = React.useMemo(() => getMarkdownSyntaxVars(currentTheme), [currentTheme]);
   const ctx = useDecorateContext(currentTheme, false, undefined, mermaidControls);
-  // Content-addressed render cache (see markdownCore) ignores this string for
-  // lookup; keep a stable per-content key so effect deps stay aligned with text.
-  const cacheKey = React.useMemo(
-    () => `simple:${variant}:${renderedContent.length}:${renderedContent.slice(0, 64)}`,
-    [variant, renderedContent],
-  );
 
   useMorphdomMarkdown({
     containerRef,
     text: renderedContent,
     streaming: false,
-    cacheKey,
+    // Identity is unused for cache lookup (content-addressed in markdownCore);
+    // keep a stable per-variant key for effect deps alongside `text`.
+    cacheKey: `simple:${variant}`,
     syntaxVars,
     ctx,
   });

packages/ui/src/components/chat/markdown/highlightResultCache.ts

@@ -1,38 +1,61 @@
 // Bounded LRU for Shiki highlight results.
 //
-// Used on both the main-thread worker client and inside the markdown-shiki
-// worker so unchanged code is never re-tokenized. Keys are exact
-// (lang + source [+ theme]); eviction is by entry count and approximate byte
-// size so long sessions cannot grow unbounded.
+// Used on the main-thread worker client and inside the markdown-shiki worker so
+// unchanged code is never re-tokenized. Keys are short content fingerprints
+// (not the full source) so cache maps do not duplicate large strings. Entry
+// byte sizes are recorded once at insert time — get/evict never re-walk the
+// payload.
 
 export type HighlightResultCacheOptions = {
   maxEntries: number;
   maxBytes: number;
 };
 
-const utf8Bytes = (value: string): number => {
-  // Prefer TextEncoder when available (browser / modern Bun); fall back to
-  // UTF-16 length × 2 as a conservative upper bound for older runtimes.
-  if (typeof TextEncoder !== 'undefined') {
-    return new TextEncoder().encode(value).length;
+type CacheEntry<T> = {
+  value: T;
+  bytes: number;
+};
+
+/** UTF-16 storage estimate for a JS string (chars × 2). Avoids TextEncoder allocs. */
+export const utf16Bytes = (value: string): number => value.length * 2;
+
+/**
+ * Short stable fingerprint for cache keys: length + FNV-1a 32-bit.
+ * Collision of same length + same hash is vanishingly rare at session scale;
+ * highlight results are pure functions of (kind, lang, theme, source).
+ */
+export const contentFingerprint = (value: string): string => {
+  let h = 0x811c9dc5;
+  for (let i = 0; i < value.length; i += 1) {
+    h ^= value.charCodeAt(i);
+    h = Math.imul(h, 0x01000193);
   }
-  return value.length * 2;
+  return `${value.length.toString(36)}_${(h >>> 0).toString(36)}`;
+};
+
+/** Approximate byte cost of token-run lines without JSON.stringify. */
+export const estimateTokenRunsBytes = (
+  lines: ReadonlyArray<ReadonlyArray<readonly [number, string, number]>>,
+): number => {
+  let total = 0;
+  for (const line of lines) {
+    total += 4;
+    for (const run of line) {
+      total += 8 + utf16Bytes(run[1]);
+    }
+  }
+  return total;
 };
 
 export class HighlightResultCache<T> {
   private readonly maxEntries: number;
   private readonly maxBytes: number;
-  private readonly sizeOf: (key: string, value: T) => number;
-  private readonly map = new Map<string, T>();
+  private readonly map = new Map<string, CacheEntry<T>>();
   private totalBytes = 0;
 
-  constructor(
-    options: HighlightResultCacheOptions,
-    sizeOf: (key: string, value: T) => number,
-  ) {
+  constructor(options: HighlightResultCacheOptions) {
     this.maxEntries = Math.max(1, options.maxEntries);
     this.maxBytes = Math.max(1, options.maxBytes);
-    this.sizeOf = sizeOf;
   }
 
   get size(): number {
@@ -44,39 +67,37 @@ export class HighlightResultCache<T> {
   }
 
   get(key: string): T | undefined {
-    const value = this.map.get(key);
-    if (value === undefined) return undefined;
-    // Refresh LRU order.
+    const entry = this.map.get(key);
+    if (entry === undefined) return undefined;
+    // Refresh LRU order without recomputing size.
     this.map.delete(key);
-    this.map.set(key, value);
-    return value;
+    this.map.set(key, entry);
+    return entry.value;
   }
 
-  set(key: string, value: T): void {
+  set(key: string, value: T, bytes: number): void {
     const existing = this.map.get(key);
     if (existing !== undefined) {
-      this.totalBytes -= this.sizeOf(key, existing);
+      this.totalBytes -= existing.bytes;
       this.map.delete(key);
     }
 
-    const entrySize = this.sizeOf(key, value);
+    const entryBytes = Math.max(0, bytes);
     while (
       this.map.size > 0
-      && (this.map.size >= this.maxEntries || this.totalBytes + entrySize > this.maxBytes)
+      && (this.map.size >= this.maxEntries || this.totalBytes + entryBytes > this.maxBytes)
     ) {
       const oldest = this.map.keys().next().value;
       if (oldest === undefined) break;
-      const oldestValue = this.map.get(oldest);
-      if (oldestValue !== undefined) {
-        this.totalBytes -= this.sizeOf(oldest, oldestValue);
-      }
+      const oldestEntry = this.map.get(oldest);
+      if (oldestEntry !== undefined) this.totalBytes -= oldestEntry.bytes;
       this.map.delete(oldest);
       // Always allow a single oversized entry so huge files still cache once.
       if (this.map.size === 0) break;
     }
 
-    this.map.set(key, value);
-    this.totalBytes += entrySize;
+    this.map.set(key, { value, bytes: entryBytes });
+    this.totalBytes += entryBytes;
   }
 
   clear(): void {
@@ -84,14 +105,3 @@ export class HighlightResultCache<T> {
     this.totalBytes = 0;
   }
 }
-
-/** Byte size for string→string highlight HTML caches (key + value). */
-export const stringPairSize = (key: string, value: string): number =>
-  utf8Bytes(key) + utf8Bytes(value);
-
-/** Byte size for string→string[] line caches. */
-export const stringLinesSize = (key: string, value: string[]): number => {
-  let total = utf8Bytes(key);
-  for (const line of value) total += utf8Bytes(line);
-  return total;
-};

packages/ui/src/components/chat/markdown/markdown-shiki.worker.ts

@@ -2,9 +2,10 @@
 
 import { bundledLanguages, createHighlighter, type BundledLanguage, type ThemedToken } from 'shiki';
 import {
+  contentFingerprint,
+  estimateTokenRunsBytes,
   HighlightResultCache,
-  stringLinesSize,
-  stringPairSize,
+  utf16Bytes,
 } from './highlightResultCache';
 import { MARKDOWN_SHIKI_THEME, MARKDOWN_SHIKI_THEME_DEFINITION } from './markdownShikiThemeDefinition';
 import type { MarkdownTokenRun, MarkdownWorkerRequest, MarkdownWorkerResponse } from './markdown-worker-protocol';
@@ -37,24 +38,24 @@ let highlighter: ReturnType<typeof createHighlighter> | undefined;
 // Serialize work so language loading / tokenization never overlaps.
 let queue = Promise.resolve();
 
-// Memoize tokenization by exact source. Without this, every remount / cache
-// miss on the main thread re-runs codeToHtml for unchanged content and pins a
-// core (openchamber/openchamber#2769).
+// Memoize tokenization by content fingerprint. Defense in depth for the
+// main-thread client cache (openchamber/openchamber#2769): keys stay short so
+// large sources are not duplicated in the Map.
 const WORKER_CACHE_MAX_ENTRIES = 1500;
 const WORKER_CACHE_MAX_BYTES = 16 * 1024 * 1024;
 
-const htmlCache = new HighlightResultCache<string>(
-  { maxEntries: WORKER_CACHE_MAX_ENTRIES, maxBytes: WORKER_CACHE_MAX_BYTES },
-  stringPairSize,
-);
-const linesCache = new HighlightResultCache<string[]>(
-  { maxEntries: WORKER_CACHE_MAX_ENTRIES, maxBytes: WORKER_CACHE_MAX_BYTES },
-  stringLinesSize,
-);
-const tokensCache = new HighlightResultCache<MarkdownTokenRun[][]>(
-  { maxEntries: WORKER_CACHE_MAX_ENTRIES, maxBytes: WORKER_CACHE_MAX_BYTES },
-  (key, value) => stringPairSize(key, JSON.stringify(value)),
-);
+const htmlCache = new HighlightResultCache<string>({
+  maxEntries: WORKER_CACHE_MAX_ENTRIES,
+  maxBytes: WORKER_CACHE_MAX_BYTES,
+});
+const linesCache = new HighlightResultCache<string[]>({
+  maxEntries: WORKER_CACHE_MAX_ENTRIES,
+  maxBytes: WORKER_CACHE_MAX_BYTES,
+});
+const tokensCache = new HighlightResultCache<MarkdownTokenRun[][]>({
+  maxEntries: WORKER_CACHE_MAX_ENTRIES,
+  maxBytes: WORKER_CACHE_MAX_BYTES,
+});
 
 const ensureHighlighter = (): ReturnType<typeof createHighlighter> => {
   highlighter ??= createHighlighter({
@@ -96,9 +97,14 @@ const resolveLanguage = async (instance: Instance, requested: string): Promise<s
   return lang;
 };
 
+const htmlKey = (lang: string, code: string): string => `${lang}:${contentFingerprint(code)}`;
+const linesKey = (lang: string, code: string): string => `${lang}:${contentFingerprint(code)}`;
+const tokensKey = (themeName: string, lang: string, code: string): string =>
+  `${themeName}:${lang}:${contentFingerprint(code)}`;
+
 async function highlight(request: Extract<MarkdownWorkerRequest, { type: 'highlight' }>): Promise<void> {
   try {
-    const cacheKey = `${request.lang}\0${request.code}`;
+    const cacheKey = htmlKey(request.lang, request.code);
     const cached = htmlCache.get(cacheKey);
     if (cached !== undefined) {
       post({ type: 'highlight', id: request.id, html: cached });
@@ -111,7 +117,7 @@ async function highlight(request: Extract<MarkdownWorkerRequest, { type: 'highli
       theme: MARKDOWN_SHIKI_THEME,
       tabindex: false,
     });
-    htmlCache.set(cacheKey, html);
+    htmlCache.set(cacheKey, html, utf16Bytes(cacheKey) + utf16Bytes(html));
     post({ type: 'highlight', id: request.id, html });
   } catch (error) {
     post({ type: 'error', id: request.id, message: error instanceof Error ? error.message : String(error) });
@@ -120,7 +126,7 @@ async function highlight(request: Extract<MarkdownWorkerRequest, { type: 'highli
 
 async function highlightTokens(request: Extract<MarkdownWorkerRequest, { type: 'highlightTokens' }>): Promise<void> {
   try {
-    const cacheKey = `${request.themeName}\0${request.lang}\0${request.code}`;
+    const cacheKey = tokensKey(request.themeName, request.lang, request.code);
     const cached = tokensCache.get(cacheKey);
     if (cached !== undefined) {
       post({ type: 'highlightTokens', id: request.id, lines: cached });
@@ -139,7 +145,7 @@ async function highlightTokens(request: Extract<MarkdownWorkerRequest, { type: '
     const lines = tokens.map((line) =>
       line.map((token) => [token.content.length, token.color ?? '', token.fontStyle ?? 0] as [number, string, number]),
     );
-    tokensCache.set(cacheKey, lines);
+    tokensCache.set(cacheKey, lines, utf16Bytes(cacheKey) + estimateTokenRunsBytes(lines));
     post({ type: 'highlightTokens', id: request.id, lines });
   } catch (error) {
     post({ type: 'error', id: request.id, message: error instanceof Error ? error.message : String(error) });
@@ -148,7 +154,7 @@ async function highlightTokens(request: Extract<MarkdownWorkerRequest, { type: '
 
 async function highlightLines(request: Extract<MarkdownWorkerRequest, { type: 'highlightLines' }>): Promise<void> {
   try {
-    const cacheKey = `${request.lang}\0${request.code}`;
+    const cacheKey = linesKey(request.lang, request.code);
     const cached = linesCache.get(cacheKey);
     if (cached !== undefined) {
       post({ type: 'highlightLines', id: request.id, lines: cached });
@@ -161,7 +167,9 @@ async function highlightLines(request: Extract<MarkdownWorkerRequest, { type: 'h
       theme: MARKDOWN_SHIKI_THEME,
     });
     const lines = tokens.map((line) => line.map(tokenSpan).join(''));
-    linesCache.set(cacheKey, lines);
+    let bytes = utf16Bytes(cacheKey);
+    for (const line of lines) bytes += utf16Bytes(line);
+    linesCache.set(cacheKey, lines, bytes);
     post({ type: 'highlightLines', id: request.id, lines });
   } catch (error) {
     post({ type: 'error', id: request.id, message: error instanceof Error ? error.message : String(error) });

packages/ui/src/components/chat/markdown/markdown-worker.ts

@@ -1,8 +1,9 @@
 import MarkdownShikiWorkerUrl from './markdown-shiki.worker.ts?worker&url';
 import {
+  contentFingerprint,
+  estimateTokenRunsBytes,
   HighlightResultCache,
-  stringLinesSize,
-  stringPairSize,
+  utf16Bytes,
 } from './highlightResultCache';
 import type { MarkdownTokenRun, MarkdownWorkerRequest, MarkdownWorkerResponse } from './markdown-worker-protocol';
 
@@ -12,10 +13,11 @@ import type { MarkdownTokenRun, MarkdownWorkerRequest, MarkdownWorkerResponse }
 // tokenization error) the promise resolves to `null` and the caller keeps the
 // escaped plain-text code — highlighting never falls back onto the main thread.
 //
-// Results are memoized by exact source (+ lang / theme). Unchanged content must
-// not re-enter the worker — that was the sustained ~40 msg/s re-highlight load
-// in openchamber/openchamber#2769. In-flight requests with the same key coalesce
-// so remount storms share one round-trip.
+// Results are memoized by content fingerprint (+ lang / theme). Unchanged
+// content must not re-enter the worker — that was the sustained ~40 msg/s
+// re-highlight load in openchamber/openchamber#2769. In-flight requests with
+// the same key coalesce so remount storms share one round-trip. Cache keys are
+// fingerprints (not full source) so large files are not duplicated in the Map.
 
 type PendingResolver = (response: MarkdownWorkerResponse | null) => void;
 
@@ -27,17 +29,10 @@ type CachedHighlight =
 const CLIENT_CACHE_MAX_ENTRIES = 2000;
 const CLIENT_CACHE_MAX_BYTES = 24 * 1024 * 1024;
 
-const cachedHighlightSize = (key: string, value: CachedHighlight): number => {
-  if (value.type === 'highlight') return stringPairSize(key, value.html);
-  if (value.type === 'highlightLines') return stringLinesSize(key, value.lines);
-  // Token runs: approximate as JSON length of the lines payload.
-  return stringPairSize(key, JSON.stringify(value.lines));
-};
-
-const resultCache = new HighlightResultCache<CachedHighlight>(
-  { maxEntries: CLIENT_CACHE_MAX_ENTRIES, maxBytes: CLIENT_CACHE_MAX_BYTES },
-  cachedHighlightSize,
-);
+const resultCache = new HighlightResultCache<CachedHighlight>({
+  maxEntries: CLIENT_CACHE_MAX_ENTRIES,
+  maxBytes: CLIENT_CACHE_MAX_BYTES,
+});
 
 const inflight = new Map<string, Promise<CachedHighlight | null>>();
 
@@ -48,6 +43,17 @@ const pending = new Map<number, PendingResolver>();
 // repeat tokenization sends only the name (not the whole theme object) again.
 const sentThemes = new Set<string>();
 
+const entryBytes = (key: string, value: CachedHighlight): number => {
+  const keyBytes = utf16Bytes(key);
+  if (value.type === 'highlight') return keyBytes + utf16Bytes(value.html);
+  if (value.type === 'highlightLines') {
+    let total = keyBytes;
+    for (const line of value.lines) total += utf16Bytes(line);
+    return total;
+  }
+  return keyBytes + estimateTokenRunsBytes(value.lines);
+};
+
 const failAll = (): void => {
   pending.forEach((resolve) => resolve(null));
   pending.clear();
@@ -102,6 +108,11 @@ const coalesce = (
   return pendingRequest;
 };
 
+const cacheKeyFor = (kind: string, lang: string, code: string, themeName?: string): string => {
+  const fp = contentFingerprint(code);
+  return themeName === undefined ? `${kind}:${lang}:${fp}` : `${kind}:${themeName}:${lang}:${fp}`;
+};
+
 /** Test-only: clear client-side highlight memoization. */
 export const resetMarkdownWorkerClientCacheForTests = (): void => {
   resultCache.clear();
@@ -113,7 +124,7 @@ export const resetMarkdownWorkerClientCacheForTests = (): void => {
  * or `null` if highlighting is unavailable or failed (caller keeps plain code).
  */
 export const highlightCodeInWorker = async (code: string, lang: string): Promise<string | null> => {
-  const key = `highlight:${lang}:${code}`;
+  const key = cacheKeyFor('highlight', lang, code);
   const cached = resultCache.get(key);
   if (cached?.type === 'highlight') return cached.html;
 
@@ -121,7 +132,7 @@ export const highlightCodeInWorker = async (code: string, lang: string): Promise
     const response = await request((id) => ({ type: 'highlight', id, code, lang }));
     if (response?.type !== 'highlight') return null;
     const entry: CachedHighlight = { type: 'highlight', html: response.html };
-    resultCache.set(key, entry);
+    resultCache.set(key, entry, entryBytes(key, entry));
     return entry;
   });
   return result?.type === 'highlight' ? result.html : null;
@@ -133,7 +144,7 @@ export const highlightCodeInWorker = async (code: string, lang: string): Promise
  * round-trip instead of one per line. Resolves to `null` on failure.
  */
 export const highlightLinesInWorker = async (code: string, lang: string): Promise<string[] | null> => {
-  const key = `highlightLines:${lang}:${code}`;
+  const key = cacheKeyFor('highlightLines', lang, code);
   const cached = resultCache.get(key);
   if (cached?.type === 'highlightLines') return cached.lines;
 
@@ -141,7 +152,7 @@ export const highlightLinesInWorker = async (code: string, lang: string): Promis
     const response = await request((id) => ({ type: 'highlightLines', id, code, lang }));
     if (response?.type !== 'highlightLines') return null;
     const entry: CachedHighlight = { type: 'highlightLines', lines: response.lines };
-    resultCache.set(key, entry);
+    resultCache.set(key, entry, entryBytes(key, entry));
     return entry;
   });
   return result?.type === 'highlightLines' ? result.lines : null;
@@ -159,7 +170,7 @@ export const highlightTokensInWorker = async (
   themeName: string,
   theme: unknown,
 ): Promise<MarkdownTokenRun[][] | null> => {
-  const key = `highlightTokens:${themeName}:${lang}:${code}`;
+  const key = cacheKeyFor('highlightTokens', lang, code, themeName);
   const cached = resultCache.get(key);
   if (cached?.type === 'highlightTokens') return cached.lines;
 
@@ -176,7 +187,7 @@ export const highlightTokensInWorker = async (
     if (response?.type !== 'highlightTokens') return null;
     sentThemes.add(themeName);
     const entry: CachedHighlight = { type: 'highlightTokens', lines: response.lines };
-    resultCache.set(key, entry);
+    resultCache.set(key, entry, entryBytes(key, entry));
     return entry;
   });
   return result?.type === 'highlightTokens' ? result.lines : null;

packages/ui/src/components/chat/markdown/markdownCore.ts

@@ -4,7 +4,7 @@ import katex from 'katex';
 import DOMPurify from 'dompurify';
 import { buildAgentMentionUrl, parseAgentHref, parseSkillHref } from '@/lib/messages/inlineMessageLinks';
 import { isVSCodeRuntime } from '@/lib/desktop';
-import { HighlightResultCache, stringPairSize } from './highlightResultCache';
+import { contentFingerprint, HighlightResultCache, utf16Bytes } from './highlightResultCache';
 import { highlightCodeInWorker } from './markdown-worker';
 import { escapeRawMarkdownHtml, MARKDOWN_FORBIDDEN_TAGS } from './markdownSecurity';
 
@@ -254,32 +254,37 @@ const highlightCodeBlocks = async (html: string): Promise<string> => {
 
   const lineLimit = isVSCodeRuntime() ? VSCODE_CODE_HIGHLIGHT_LINE_LIMIT : CODE_HIGHLIGHT_LINE_LIMIT;
 
-  let result = html;
-  for (const match of matches) {
-    const [full, rawLang, escapedCode] = match;
-    const requested = (rawLang || 'text').toLowerCase();
-    // Leave mermaid fences untouched so the decorate pass can render them as
-    // diagrams (highlighting would strip the `language-mermaid` class).
-    if (requested === 'mermaid') continue;
+  // Highlight all eligible fences concurrently — sequential await was O(n)
+  // worker round-trips for messages with multiple code blocks.
+  const replacements = await Promise.all(
+    matches.map(async (match) => {
+      const [full, rawLang, escapedCode] = match;
+      const requested = (rawLang || 'text').toLowerCase();
+      // Leave mermaid fences untouched so the decorate pass can render them as
+      // diagrams (highlighting would strip the `language-mermaid` class).
+      if (requested === 'mermaid') return null;
 
-    const code = unescapeHtml(escapedCode ?? '');
+      const code = unescapeHtml(escapedCode ?? '');
 
-    // Oversized block: skip highlight, keep plain code but stamp the language.
-    if (exceedsLineLimit(code, lineLimit)) {
-      result = result.replace(full, () => full.replace('<pre', `<pre data-md-lang="${requested}"`));
-      continue;
-    }
+      // Oversized block: skip highlight, keep plain code but stamp the language.
+      if (exceedsLineLimit(code, lineLimit)) {
+        return { full, next: full.replace('<pre', `<pre data-md-lang="${requested}"`) };
+      }
 
-    // Tokenize off the main thread. On failure the worker resolves to null and
-    // we keep the original escaped <pre><code> (no main-thread highlight).
-    const highlighted = await highlightCodeInWorker(code, requested);
-    if (highlighted) {
+      // Tokenize off the main thread. On failure the worker resolves to null and
+      // we keep the original escaped <pre><code> (no main-thread highlight).
+      const highlighted = await highlightCodeInWorker(code, requested);
+      if (!highlighted) return null;
       // Stamp the language so the decorate pass can show a header label.
-      const stamped = highlighted.replace(/^<pre/, `<pre data-md-lang="${requested}"`);
-      result = result.replace(full, () => stamped);
-    }
-  }
+      return { full, next: highlighted.replace(/^<pre/, `<pre data-md-lang="${requested}"`) };
+    }),
+  );
 
+  let result = html;
+  for (const replacement of replacements) {
+    if (!replacement) continue;
+    result = result.replace(replacement.full, () => replacement.next);
+  }
   return result;
 };
 
@@ -331,20 +336,10 @@ const sanitize = (html: string): string => {
 const HTML_CACHE_MAX_ENTRIES = 2000;
 const HTML_CACHE_MAX_BYTES = 24 * 1024 * 1024;
 
-const htmlCache = new HighlightResultCache<string>(
-  { maxEntries: HTML_CACHE_MAX_ENTRIES, maxBytes: HTML_CACHE_MAX_BYTES },
-  stringPairSize,
-);
-
-// FNV-1a 32-bit hash of the block content.
-const hash = (value: string): string => {
-  let h = 0x811c9dc5;
-  for (let i = 0; i < value.length; i += 1) {
-    h ^= value.charCodeAt(i);
-    h = Math.imul(h, 0x01000193);
-  }
-  return (h >>> 0).toString(36);
-};
+const htmlCache = new HighlightResultCache<string>({
+  maxEntries: HTML_CACHE_MAX_ENTRIES,
+  maxBytes: HTML_CACHE_MAX_BYTES,
+});
 
 /** Content-addressed cache key for a markdown block (exported for tests). */
 export const markdownBlockCacheKey = (
@@ -411,14 +406,14 @@ export const renderMarkdownBlocks = async (
   const blocks = streamBlocks(text, streaming);
   return Promise.all(
     blocks.map(async (block) => {
-      const contentHash = hash(block.raw);
+      const contentHash = contentFingerprint(block.raw);
       const id = markdownBlockCacheKey(contentHash, block.mode, block.highlight);
       const cached = htmlCache.get(id);
       if (cached !== undefined) {
         return { id, html: cached };
       }
       const html = await parseBlock(block);
-      htmlCache.set(id, html);
+      htmlCache.set(id, html, utf16Bytes(id) + utf16Bytes(html));
       return { id, html };
     }),
   );

packages/ui/src/components/chat/markdown/markdownHighlightingRepro.test.ts

@@ -8,16 +8,27 @@
  *  3. Worker/client had no result memoization.
  *
  * These tests assert the fixed contracts: content-addressed caching, room for
- * long sessions, and bounded LRU behavior.
+ * long sessions, bounded LRU behavior, and fingerprint-key helpers.
  */
 import { beforeEach, describe, expect, mock, test } from 'bun:test';
 
-import { HighlightResultCache, stringPairSize } from './highlightResultCache';
+import {
+  contentFingerprint,
+  estimateTokenRunsBytes,
+  HighlightResultCache,
+  utf16Bytes,
+} from './highlightResultCache';
 
 let highlightCalls = 0;
+let highlightInflight = 0;
+let highlightMaxInflight = 0;
 
 const highlightCodeInWorkerMock = mock(async (code: string, lang: string) => {
   highlightCalls += 1;
+  highlightInflight += 1;
+  highlightMaxInflight = Math.max(highlightMaxInflight, highlightInflight);
+  await Promise.resolve();
+  highlightInflight -= 1;
   return `<pre data-lang="${lang}"><code>${code}</code></pre>`;
 });
 
@@ -40,35 +51,45 @@ beforeEach(() => {
   resetMarkdownHtmlCacheForTests();
   resetMarkdownWorkerClientCacheForTests();
   highlightCalls = 0;
+  highlightInflight = 0;
+  highlightMaxInflight = 0;
 });
 
 describe('HighlightResultCache', () => {
   test('returns cached values for identical keys and refreshes LRU order', () => {
-    const cache = new HighlightResultCache<string>(
-      { maxEntries: 2, maxBytes: 10_000 },
-      stringPairSize,
-    );
-    cache.set('a', 'one');
-    cache.set('b', 'two');
+    const cache = new HighlightResultCache<string>({ maxEntries: 2, maxBytes: 10_000 });
+    cache.set('a', 'one', utf16Bytes('a') + utf16Bytes('one'));
+    cache.set('b', 'two', utf16Bytes('b') + utf16Bytes('two'));
     expect(cache.get('a')).toBe('one');
     // Touch `a` so `b` is oldest; inserting `c` should evict `b`.
-    cache.set('c', 'three');
+    cache.set('c', 'three', utf16Bytes('c') + utf16Bytes('three'));
     expect(cache.get('b')).toEqual(undefined);
     expect(cache.get('a')).toBe('one');
     expect(cache.get('c')).toBe('three');
   });
 
   test('evicts by byte budget while still caching a single oversized entry', () => {
-    const cache = new HighlightResultCache<string>(
-      { maxEntries: 10, maxBytes: 64 },
-      stringPairSize,
-    );
-    cache.set('small', 'x');
-    cache.set('huge', 'y'.repeat(200));
+    const cache = new HighlightResultCache<string>({ maxEntries: 10, maxBytes: 64 });
+    cache.set('small', 'x', utf16Bytes('small') + utf16Bytes('x'));
+    cache.set('huge', 'y'.repeat(200), utf16Bytes('huge') + utf16Bytes('y'.repeat(200)));
     expect(cache.get('huge')).toBe('y'.repeat(200));
     // Oversized insert cleared prior entries to make room.
     expect(cache.size).toBe(1);
   });
+
+  test('contentFingerprint is stable and length-qualified', () => {
+    expect(contentFingerprint('const x = 1')).toBe(contentFingerprint('const x = 1'));
+    expect(contentFingerprint('const x = 1')).not.toBe(contentFingerprint('const x = 2'));
+    expect(contentFingerprint('ab')).not.toBe(contentFingerprint('abc'));
+  });
+
+  test('estimateTokenRunsBytes avoids JSON and stays positive', () => {
+    const lines: Array<Array<[number, string, number]>> = [
+      [[3, '#fff', 0], [1, '', 1]],
+      [[8, 'var(--md-syntax-keyword)', 0]],
+    ];
+    expect(estimateTokenRunsBytes(lines)).toBeGreaterThan(0);
+  });
 });
 
 describe('markdownCore content-addressed htmlCache (#2769)', () => {
@@ -135,4 +156,19 @@ describe('markdownCore content-addressed htmlCache (#2769)', () => {
     expect(markdownBlockCacheKey('abc', 'live', false)).toBe('abc:live:0');
     expect(markdownBlockCacheKey('abc', 'full', true)).not.toBe(markdownBlockCacheKey('abc', 'full', false));
   });
+
+  test('multiple code fences in one document highlight concurrently', async () => {
+    const multi = [
+      '```ts\nconst a = 1;\n```',
+      '',
+      '```ts\nconst b = 2;\n```',
+      '',
+      '```ts\nconst c = 3;\n```',
+    ].join('\n');
+
+    await renderMarkdownBlocks(multi, false, 'multi');
+    expect(highlightCalls).toBe(3);
+    // Sequential awaits would keep max inflight at 1.
+    expect(highlightMaxInflight).toBeGreaterThan(1);
+  });
 });