Harden audio transcription arg hydration and add rewrite audit event

src/daemon/routing.ts

@@ -164,6 +164,57 @@ function shouldForceNativeForCapabilityQuery(text: string): boolean {
   );
 }
 
+function providerAcceptsNativeAudioContentParts(provider: string): boolean {
+  return (
+    provider === 'openai'
+    || provider === 'github'
+    || provider === 'gemini'
+    || provider === 'openrouter'
+    || provider === 'zhipuai'
+    || provider === 'xai'
+    || provider === 'minimax'
+    || provider === 'moonshot'
+    || provider === 'vercel'
+  );
+}
+
+const LAST_AUDIO_ATTACHMENT_CONFIG_KEY = 'lastAudioAttachment';
+
+function persistLatestAudioAttachment(
+  session: { setConfig(key: string, value: string): void },
+  audioAttachments: Attachment[],
+): void {
+  const latest = [...audioAttachments].reverse().find((att) => (
+    (typeof att.data === 'string' && att.data.length > 0)
+      || (typeof att.url === 'string' && att.url.length > 0)
+  ));
+  if (!latest) {
+    return;
+  }
+
+  const payload: { data?: string; url?: string; mimeType?: string } = {
+    mimeType: latest.mimeType,
+  };
+  if (typeof latest.data === 'string' && latest.data.length > 0) {
+    payload.data = latest.data;
+  } else if (typeof latest.url === 'string' && latest.url.length > 0) {
+    payload.url = latest.url;
+  }
+
+  if (!payload.data && !payload.url) {
+    return;
+  }
+
+  try {
+    session.setConfig(LAST_AUDIO_ATTACHMENT_CONFIG_KEY, JSON.stringify(payload));
+  } catch (error) {
+    console.warn(
+      'Failed to persist latest audio attachment for tool hydration:',
+      error instanceof Error ? error.message : String(error),
+    );
+  }
+}
+
 function isTtsEnabledForChannel(config: Config, channel: string): boolean {
   if (!config.tts?.enabled) {
     return false;
@@ -1266,6 +1317,9 @@ export function createMessageRouter(deps: {
       let messageText = incomingText;
       let attachments = msg.attachments;
       const audioAttachments = (msg.attachments ?? []).filter((a: Attachment) => isSupportedAudio(a));
+      if (audioAttachments.length > 0) {
+        persistLatestAudioAttachment(session, audioAttachments);
+      }
 
       if (audioAttachments.length > 0 && !nativeAudioSupported) {
         // Model doesn't support native audio — transcribe via Whisper and strip audio attachments
@@ -1300,9 +1354,15 @@ export function createMessageRouter(deps: {
           const transcript = await transcribeAudio(att, audioConfig);
           messageText = `[Voice message]: ${transcript}\n\n${messageText}`;
         }
-        // Remove audio attachments so buildUserMessage doesn't create audio content parts
-        attachments = (msg.attachments ?? []).filter((a: Attachment) => !isSupportedAudio(a));
-        if (attachments.length === 0) { attachments = undefined; }
+        // For providers that cannot ingest native audio content parts (e.g. Anthropic),
+        // keep the original audio attachment available in the tool loop so
+        // audio.transcribe can still be hydrated from bytes if the model requests it.
+        // For providers that do accept native audio parts (OpenAI-compatible/Gemini),
+        // strip audio to avoid sending raw audio to a model tier that was marked as non-audio.
+        if (providerAcceptsNativeAudioContentParts(modelProvider)) {
+          attachments = (msg.attachments ?? []).filter((a: Attachment) => !isSupportedAudio(a));
+          if (attachments.length === 0) { attachments = undefined; }
+        }
       }
       // If native audio IS supported, we pass attachments through unchanged —
       // buildUserMessage() in the agent will create native audio content parts