feat(telegram): transcribe replied-to voice messages as context

gateway/platforms/base.py

@@ -901,6 +901,7 @@ class MessageEvent:
     # Reply context
     reply_to_message_id: Optional[str] = None
     reply_to_text: Optional[str] = None  # Text of the replied-to message (for context injection)
+    reply_to_audio_path: Optional[str] = None  # Local path to cached audio of replied-to voice message
 
     # Auto-loaded skill(s) for topic/channel bindings (e.g., Telegram DM Topics,
     # Discord channel_skill_bindings).  A single name or ordered list.

gateway/platforms/telegram.py

@@ -2887,6 +2887,7 @@ async def _handle_text_message(self, update: Update, context: ContextTypes.DEFAU
 
         event = self._build_message_event(update.message, MessageType.TEXT, update_id=update.update_id)
         event.text = self._clean_bot_trigger_text(event.text)
+        await self._enrich_event_with_replied_audio(event, update.message)
         self._enqueue_text_event(event)
 
     async def _handle_command(self, update: Update, context: ContextTypes.DEFAULT_TYPE) -> None:
@@ -3089,6 +3090,9 @@ async def _handle_media_message(self, update: Update, context: ContextTypes.DEFA
         if msg.caption:
             event.text = self._clean_bot_trigger_text(msg.caption)
 
+        # Enrich with replied-to voice/audio if applicable
+        await self._enrich_event_with_replied_audio(event, msg)
+
         # Handle stickers: describe via vision tool with caching
         if msg.sticker:
             await self._handle_sticker(msg, event)
@@ -3524,6 +3528,7 @@ def _build_message_event(
         # Extract reply context if this message is a reply
         reply_to_id = None
         reply_to_text = None
+        reply_to_audio_path = None
         if message.reply_to_message:
             reply_to_id = str(message.reply_to_message.message_id)
             reply_to_text = message.reply_to_message.text or message.reply_to_message.caption or None
@@ -3551,6 +3556,29 @@ def _build_message_event(
             timestamp=message.date,
         )
 
+    async def _enrich_event_with_replied_audio(self, event: "MessageEvent", message: Message) -> None:
+        """If the triggering message is a reply to a voice/audio note, download and cache
+        that audio so the STT pipeline can transcribe it as context."""
+        if not message.reply_to_message:
+            return
+        replied = message.reply_to_message
+        if replied.voice:
+            try:
+                file_obj = await replied.voice.get_file()
+                audio_bytes = await file_obj.download_as_bytearray()
+                event.reply_to_audio_path = cache_audio_from_bytes(bytes(audio_bytes), ext=".ogg")
+                logger.info("[Telegram] Cached replied-to voice at %s", event.reply_to_audio_path)
+            except Exception as e:
+                logger.warning("[Telegram] Failed to cache replied-to voice: %s", e, exc_info=True)
+        elif replied.audio:
+            try:
+                file_obj = await replied.audio.get_file()
+                audio_bytes = await file_obj.download_as_bytearray()
+                event.reply_to_audio_path = cache_audio_from_bytes(bytes(audio_bytes), ext=".mp3")
+                logger.info("[Telegram] Cached replied-to audio at %s", event.reply_to_audio_path)
+            except Exception as e:
+                logger.warning("[Telegram] Failed to cache replied-to audio: %s", e, exc_info=True)
+
     # ── Message reactions (processing lifecycle) ──────────────────────────
 
     def _reactions_enabled(self) -> bool:

gateway/run.py

@@ -5542,6 +5542,16 @@ async def _prepare_inbound_message_text(
             # guess (or answer for both subjects). Token overhead is minimal.
             reply_snippet = event.reply_to_text[:500]
             message_text = f'[Replying to: "{reply_snippet}"]\n\n{message_text}'
+        elif getattr(event, "reply_to_audio_path", None) and event.reply_to_message_id:
+            # Replied-to message was a voice/audio note — transcribe it and inject as context
+            try:
+                transcribed = await self._enrich_message_with_transcription("", [event.reply_to_audio_path])
+                transcribed = transcribed.strip()
+                if transcribed and not any(m in transcribed for m in ("No STT provider", "STT is disabled", "can't listen")):
+                    message_text = f'[Replying to voice message: "{transcribed[:500]}"]\n\n{message_text}'
+                    logger.info("[Gateway] Injected replied-to voice transcription as context")
+            except Exception as _e:
+                logger.warning("[Gateway] Failed to transcribe replied-to audio: %s", _e)
 
         if "@" in message_text:
             try:

hermes_cli/models.py

@@ -203,6 +203,7 @@ def _xai_curated_models() -> list[str]:
         "copilot-acp",
     ],
     "copilot": [
+        "gpt-5.5",
         "gpt-5.4",
         "gpt-5.4-mini",
         "gpt-5-mini",
@@ -211,6 +212,8 @@ def _xai_curated_models() -> list[str]:
         "gpt-4.1",
         "gpt-4o",
         "gpt-4o-mini",
+        "claude-opus-4.7",
+        "claude-opus-4.6",
         "claude-sonnet-4.6",
         "claude-sonnet-4",
         "claude-sonnet-4.5",