fix avatar cache TTL

apps/desktop/src/main/avatar_protocol.ts

@@ -57,9 +57,10 @@ export async function handleAvatarRequest(request: Request): Promise<Response> {
       status: 200,
       headers: {
         'Content-Type': blob.contentType,
-        // Let the renderer / Chromium memory-cache it too; the on-disk cache
-        // is authoritative, this just avoids re-asking the protocol.
-        'Cache-Control': 'public, max-age=86400',
+        // Let the renderer / Chromium memory-cache it too; the on-disk cache is
+        // authoritative (and applies the TTL), so keep this short — a long
+        // max-age would outlive the disk entry and pin a stale avatar.
+        'Cache-Control': 'public, max-age=300',
       },
     });
   } catch {

apps/desktop/src/main/context/app_context.ts

@@ -488,6 +488,12 @@ export interface AppContext {
    * no account is open / QQ is offline / harvest failed.
    */
   refreshRkeysNow(): Promise<boolean>;
+  /**
+   * Transcribe a SILK file on disk with the globally selected model. Account
+   * independent (the model is a global setting), so callers that already have a
+   * path — e.g. the Ptt cache browser — can skip the message-lookup dance.
+   */
+  transcribeSilk(silkPath: string): Promise<{ ok: boolean; text?: string; error?: string }>;
 }
 
 let cached: AppContext | undefined;
@@ -529,6 +535,9 @@ export function initAppContext(): AppContext {
       refreshRkeysNow(): Promise<boolean> {
         return Promise.resolve(false);
       },
+      transcribeSilk(): Promise<{ ok: boolean; text?: string; error?: string }> {
+        return Promise.resolve({ ok: false, error: '原生组件未就绪' });
+      },
     };
     return cached;
   }
@@ -605,7 +614,7 @@ export function initAppContext(): AppContext {
     keys: new Win32KeyService(platform, stubHooks),
     userConfig,
     globalConfig: new GlobalConfigService(platform, userConfig),
-    avatarCache: new AvatarCacheService(platform, userConfig),
+    avatarCache: new AvatarCacheService(userConfig),
     linkPreview,
     agentLabConfig: new AgentLabConfigService(userConfig),
     voiceTranscribe: new VoiceTranscribeService(platform),
@@ -663,6 +672,7 @@ export function initAppContext(): AppContext {
     account: null,
     services: null,
     scheduler: null,
+    transcribeSilk,
     async setAccount(accountCtx: AccountContext, metadata: AccountConfigMetadata = {}): Promise<void> {
       logger.info('opening account session', {
         event: 'open-account-start',

apps/desktop/src/main/file_response.ts

@@ -72,20 +72,47 @@ function parseRange(header: string | null, size: number): { start: number; end:
 /**
  * Stream `path` as a `Response`, honoring the request's `Range` header.
  * Returns 404 when the file is missing, 206 for a partial response.
+ *
+ * `revalidate` opts the file into `Last-Modified` + `Cache-Control: no-cache`,
+ * so Chromium re-asks with `If-Modified-Since` and we answer 304 when the file
+ * hasn't changed. Use it for paths whose NAME is stable but whose CONTENT is
+ * rewritten in place — QQ's `nt_data/avatar` files are keyed by a uid hash, so
+ * changing your avatar overwrites the same path and a plain `max-age` would
+ * pin the old picture in the renderer's cache until it expired.
  */
-export async function fileResponse(path: string, request?: Request): Promise<Response> {
+export async function fileResponse(
+  path: string,
+  request?: Request,
+  opts: { revalidate?: boolean } = {},
+): Promise<Response> {
   let size: number;
+  let mtimeMs: number;
   try {
     const st = await stat(path);
     if (!st.isFile()) return new Response('not a file', { status: 404 });
     size = st.size;
+    mtimeMs = st.mtimeMs;
   } catch {
     return new Response('not found', { status: 404 });
   }
 
   const type = mimeOf(path);
   const range = parseRange(request?.headers.get('range') ?? null, size);
 
+  // `Last-Modified` has one-second granularity, so compare at that resolution —
+  // otherwise a sub-second mtime always reads as "newer" and we never 304.
+  const lastModified = new Date(Math.floor(mtimeMs / 1000) * 1000);
+  const revalidateHeaders = opts.revalidate
+    ? { 'Last-Modified': lastModified.toUTCString(), 'Cache-Control': 'no-cache' }
+    : undefined;
+
+  if (revalidateHeaders && !range) {
+    const since = Date.parse(request?.headers.get('if-modified-since') ?? '');
+    if (Number.isFinite(since) && lastModified.getTime() <= since) {
+      return new Response(null, { status: 304, headers: revalidateHeaders });
+    }
+  }
+
   if (range) {
     const stream = createReadStream(path, { start: range.start, end: range.end });
     return new Response(Readable.toWeb(stream) as ReadableStream, {
@@ -95,6 +122,7 @@ export async function fileResponse(path: string, request?: Request): Promise<Res
         'Content-Length': String(range.end - range.start + 1),
         'Content-Range': `bytes ${range.start}-${range.end}/${size}`,
         'Accept-Ranges': 'bytes',
+        ...revalidateHeaders,
       },
     });
   }
@@ -106,6 +134,7 @@ export async function fileResponse(path: string, request?: Request): Promise<Res
       'Content-Type': type,
       'Content-Length': String(size),
       'Accept-Ranges': 'bytes',
+      ...revalidateHeaders,
     },
   });
 }

apps/desktop/src/main/ipc/routers/media_resource.ts

@@ -5,7 +5,8 @@
  * listings. Thin tRPC skin over `MediaResourceService` (see `@weq/service`); all
  * scanning / merging lives there. Bytes are NOT returned here — the renderer
  * points `<img>`/`<video>` at `weq-media://localmedia`, resolved via the same
- * service. Read-only.
+ * service. Read-only apart from `transcribeVoice`, which runs the recognition
+ * engine on a cached clip (it writes nothing).
  */
 
 import { z } from 'zod';
@@ -66,4 +67,23 @@ export const mediaResourceRouter = router({
   analyzeTree: procedure.input(z.object({ key: treeKey })).query(({ input }) => {
     return requireServices().mediaResource.analyzeTree(input.key);
   }),
+
+  /**
+   * Transcribe one cached voice clip (语音 → 转文字). `rel` is the same Ptt-tree
+   * path the browser streams through `weq-media://localvoice`; the service
+   * re-validates it stays inside the tree. There's no message behind a cache
+   * entry, so unlike `account.transcribeVoice` nothing is written back — the
+   * text only lives in the browser's UI.
+   *
+   * Returns `{ success:false, error }` for every failure mode so the card can
+   * show a friendly message instead of throwing.
+   */
+  transcribeVoice: procedure
+    .input(z.object({ rel: z.string() }))
+    .mutation(async ({ input }): Promise<{ success: boolean; text?: string; error?: string }> => {
+      const silk = await requireServices().mediaResource.resolveFile('ptt', input.rel);
+      if (!silk) return { success: false, error: '语音文件不存在' };
+      const r = await getAppContext().transcribeSilk(silk);
+      return r.ok ? { success: true, text: r.text ?? '' } : { success: false, error: r.error };
+    }),
 });

apps/desktop/src/main/media_protocol.ts

@@ -110,11 +110,24 @@ function fileResponse(path: string): Promise<Response> {
   return streamFile(path, currentRequest.getStore());
 }
 
+/**
+ * Like {@link fileResponse} but asking the renderer to revalidate (304 on an
+ * unchanged mtime). For the local avatar cache: the file NAME is a uid hash, so
+ * QQ overwrites it in place when someone changes their picture — caching by
+ * max-age would keep showing the old one.
+ */
+function revalidatingFileResponse(path: string): Promise<Response> {
+  return streamFile(path, currentRequest.getStore(), { revalidate: true });
+}
+
 /**
  * CDN fallback for an avatar the local cache didn't have. Routes through the
  * shared {@link AvatarCacheService} disk cache (so the fetched bytes warm the
  * same cache the rest of the app reads), returning 404 on any failure so the
  * renderer shows its glyph.
+ *
+ * Short `max-age`: the disk cache is the real cache (and applies its own TTL),
+ * so a long browser cache would only delay a changed avatar from showing.
  */
 async function avatarFallbackResponse(src: string): Promise<Response> {
   if (!/^https?:\/\//i.test(src)) return notFound('avatar fallback needs http url');
@@ -124,7 +137,7 @@ async function avatarFallbackResponse(src: string): Promise<Response> {
     const blob = await cache.get(src);
     return new Response(new Uint8Array(blob.data), {
       status: 200,
-      headers: { 'Content-Type': blob.contentType, 'Cache-Control': 'public, max-age=86400' },
+      headers: { 'Content-Type': blob.contentType, 'Cache-Control': 'public, max-age=300' },
     });
   } catch {
     return notFound('avatar fallback failed');
@@ -394,7 +407,7 @@ export function handleMediaRequest(request: Request): Promise<Response> {
           if (hash) path = await services.avatarResource.resolveFile(scope, hash, variant);
           else if (uid) path = await services.avatarResource.resolveByUid(scope, uid, variant);
           else if (uin) path = await services.avatarResource.resolveByUin(scope, uin, variant);
-          if (path) return fileResponse(path);
+          if (path) return revalidatingFileResponse(path);
           // Local miss → CDN fallback (disk-cached by AvatarCacheService).
           if (fb) return avatarFallbackResponse(fb);
           return notFound('avatar not found');

apps/desktop/src/renderer/src/styles/cache.css

@@ -3946,6 +3946,58 @@ html[data-theme="dark"] .weq-filerow-name {
   font-variant-numeric: tabular-nums;
 }
 
+/* 转文字入口：常态低调，hover 卡片时点亮。 */
+.weq-voice-t9n-btn {
+  flex: none;
+  display: inline-flex;
+  align-items: center;
+  gap: 3px;
+  padding: 2px 7px;
+  border: 1px solid color-mix(in srgb, var(--weq-accent-effective) 30%, transparent);
+  border-radius: 999px;
+  color: var(--weq-accent-effective);
+  font-size: 10.5px;
+  line-height: 1.4;
+  white-space: nowrap;
+  cursor: pointer;
+  opacity: 0;
+  transition: opacity 120ms ease, background-color 120ms ease;
+}
+
+.weq-voice-card:hover .weq-voice-t9n-btn,
+.weq-voice-t9n-btn:focus-visible,
+.weq-voice-t9n-btn:disabled {
+  opacity: 1;
+}
+
+.weq-voice-t9n-btn:hover:not(:disabled) {
+  background: color-mix(in srgb, var(--weq-accent-effective) 12%, transparent);
+}
+
+.weq-voice-t9n-btn:disabled {
+  cursor: progress;
+  color: var(--weq-fg-muted);
+  border-color: color-mix(in srgb, var(--weq-fg-primary) 14%, transparent);
+}
+
+.weq-voice-t9n {
+  padding: 7px 10px;
+  border: 1px solid var(--weq-border-subtle);
+  border-radius: 10px;
+  background: color-mix(in srgb, var(--weq-accent-effective) 6%, transparent);
+  color: var(--weq-fg-secondary);
+  font-size: 12px;
+  line-height: 1.55;
+  white-space: pre-wrap;
+  word-break: break-word;
+}
+
+.weq-voice-t9n.is-error {
+  border-color: color-mix(in srgb, #d94c4c 30%, transparent);
+  background: color-mix(in srgb, #d94c4c 8%, transparent);
+  color: #c0392b;
+}
+
 /* ══════════════════════════════════════════════════════════════════════════
  * 整体资源分析弹窗
  * ════════════════════════════════════════════════════════════════════════ */

apps/desktop/src/renderer/src/views/cache/MediaResourceExplorers.tsx

@@ -10,7 +10,8 @@
  *   - {@link VoiceExplorer}     — 语音 (Ptt): month-bucketed SILK clips, rendered
  *     as cards with a (simulated) waveform + duration + play/pause. Clicking
  *     decodes the SILK to WAV via `weq-media://localvoice` and plays it; only one
- *     clip plays at a time.
+ *     clip plays at a time. With a transcription model configured, each card also
+ *     offers 转文字.
  *
  * All share {@link useCursorPaged}, a cursor-based infinite-scroll loader (the
  * backend pages by bucket, so a cursor — not an offset — resumes the walk). All
@@ -19,9 +20,9 @@
  */
 
 import { useCallback, useEffect, useMemo, useRef, useState, type ReactElement } from 'react';
-import { RefreshCw, Play, Pause } from 'lucide-react';
+import { RefreshCw, Play, Pause, FileText, Loader2 } from 'lucide-react';
 import type { FlatMediaEntry, MonthMediaEntry, VoiceMediaEntry } from '@weq/service';
-import { client } from '../../trpc/client';
+import { client, trpc } from '../../trpc/client';
 import { localMediaUrl, localVoiceUrl } from '../../lib/resourceUrl';
 import { openLightbox } from '../../components/ImageLightbox';
 import { openVideoLightbox } from '../../components/VideoLightbox';
@@ -326,6 +327,13 @@ export function VoiceExplorer(): ReactElement {
   );
   const { entries, loading, error, done, sentinelRef } = useCursorPaged<VoiceMediaEntry>(fetchPage);
 
+  // 转文字 only shows when a transcription model is selected in the settings.
+  const settings = trpc.bootstrap.getSettings.useQuery(undefined, {
+    refetchOnWindowFocus: false,
+    staleTime: 60_000,
+  });
+  const canTranscribe = Boolean(settings.data?.voiceTranscribe.modelId);
+
   if (error && entries.length === 0) {
     return <div className="weq-cache-grid-state is-error">{error}</div>;
   }
@@ -343,7 +351,7 @@ export function VoiceExplorer(): ReactElement {
         </div>,
       );
     }
-    nodes.push(<VoiceCard key={entry.rel} entry={entry} />);
+    nodes.push(<VoiceCard key={entry.rel} entry={entry} canTranscribe={canTranscribe} />);
   }
 
   return (
@@ -360,14 +368,24 @@ export function VoiceExplorer(): ReactElement {
  * One voice clip: a simulated waveform + duration + play/pause. The SILK bytes
  * are decoded to WAV on demand (first play), and the real duration replaces the
  * byte-estimated one once the audio's metadata loads. Playback progress lights
- * up the waveform left-to-right.
+ * up the waveform left-to-right. With a transcription model configured, a
+ * 转文字 button runs the recognizer over the clip and shows the text below.
  */
-function VoiceCard({ entry }: { entry: VoiceMediaEntry }): ReactElement {
+function VoiceCard({
+  entry,
+  canTranscribe,
+}: {
+  entry: VoiceMediaEntry;
+  canTranscribe: boolean;
+}): ReactElement {
   const bars = useMemo(() => fakeWaveform(entry.hash), [entry.hash]);
   const audioRef = useRef<HTMLAudioElement | null>(null);
   const [playing, setPlaying] = useState(false);
   const [progress, setProgress] = useState(0);
   const [realDur, setRealDur] = useState<number | null>(null);
+  const transcribe = trpc.account.mediaResource.transcribeVoice.useMutation();
+  const [transcript, setTranscript] = useState<string | null>(null);
+  const [transcribeError, setTranscribeError] = useState<string | null>(null);
 
   // Stop + drop the audio if the card is recycled to a different clip.
   useEffect(() => {
@@ -413,8 +431,21 @@ function VoiceCard({ entry }: { entry: VoiceMediaEntry }): ReactElement {
     }
   };
 
+  const runTranscribe = (): void => {
+    if (transcribe.isLoading) return;
+    setTranscribeError(null);
+    transcribe
+      .mutateAsync({ rel: entry.rel })
+      .then((res) => {
+        if (res.success) setTranscript(res.text ?? '');
+        else setTranscribeError(res.error ?? '识别失败');
+      })
+      .catch((err) => setTranscribeError(err instanceof Error ? err.message : String(err)));
+  };
+
   const seconds = realDur ?? Math.min(60, Math.max(1, Math.round(entry.bytes / SILK_BYTES_PER_SEC)));
   const filled = Math.round(progress * bars.length);
+  const hasResult = transcript !== null || transcribeError !== null;
 
   return (
     <figure className="weq-voice-card" title={entry.name}>
@@ -443,8 +474,29 @@ function VoiceCard({ entry }: { entry: VoiceMediaEntry }): ReactElement {
       </button>
       <figcaption className="weq-voice-meta">
         <span className="weq-voice-hash">{entry.hash.slice(0, 8)}…</span>
+        {canTranscribe && !hasResult ? (
+          <button
+            type="button"
+            className="weq-voice-t9n-btn"
+            title="转文字"
+            onClick={runTranscribe}
+            disabled={transcribe.isLoading}
+          >
+            {transcribe.isLoading ? (
+              <Loader2 size={11} strokeWidth={2} className="weq-spin" aria-hidden />
+            ) : (
+              <FileText size={11} strokeWidth={2} aria-hidden />
+            )}
+            <span>{transcribe.isLoading ? '转写中' : '转文字'}</span>
+          </button>
+        ) : null}
         <span className="weq-voice-size">{fmtBytes(entry.bytes)}</span>
       </figcaption>
+      {hasResult ? (
+        <div className={`weq-voice-t9n${transcribeError ? ' is-error' : ''}`}>
+          {transcribeError ?? (transcript || '（未识别到语音内容）')}
+        </div>
+      ) : null}
     </figure>
   );
 }

apps/desktop/src/renderer/src/views/cache/SysEmojiExplorer.tsx

@@ -134,6 +134,7 @@ export function SysEmojiExplorer(): ReactElement {
         <span className="weq-cache-data-meta">
           {total ?? entries.length} 个 · 默认动图(APNG)预览，点击查看全部格式
           {status && !status.qqRoot ? ' · QQ 资源目录缺失，已改用下载补全' : ''}
+          {status?.usingFallback ? ' · emoji.db 无表情数据，用内置地址表兜底' : ''}
         </span>
         {missing > 0 ? (
           <button