fix(voice): preserve pending message intent across reconnects

talkingq-url/handlers/websocket_message_handler.py

@@ -176,15 +176,13 @@ async def handle_binary_message(websocket: WebSocket, device_id: str, serial_num
             return current_active_session, first_audio_received_time
         session.start_received = True
         session.voice_target = await device_target_cache.claim_target(device_id, websocket, session_id)
-        if session.voice_target:
-            # Acknowledge START now; the first audio packet selects AI (0) or
-            # message (4). A pending message must not swallow an AI recording.
-            await websocket.send_text("START_ACK")
-            session.start_ack_sent = True
-            return session_key, None
-        current_active_session = await start_ai_recording(
-            websocket, device_id, session_id, session_key, session, current_active_session, sample_rate
-        )
+        # START does not identify AI versus message audio. Decide on the first
+        # audio packet, even without a target, so an invalid message cannot
+        # start ASR and later report an unrelated recognition failure.
+        current_active_session = await handle_start_packet(
+            device_id, session_id, session_key, session, current_active_session)
+        await websocket.send_text("START_ACK")
+        session.start_ack_sent = True
         return current_active_session, None
 
     if not session.start_received or session.recording_finished:
@@ -209,26 +207,32 @@ async def handle_binary_message(websocket: WebSocket, device_id: str, serial_num
             session_logger.info(device_id, session_id, f"收到第一个音频数据包时间: {first_audio_received_time}")
     elif packet_type == 4:
         target = session.voice_target
-        if target and not session.ai_started:
-            if await handle_target_audio_packet(device_id, audio_data, target, websocket, session_id):
-                session.message_audio_received = True
+        if not session.ai_started:
+            session.message_audio_received = True
+            if target:
+                await handle_target_audio_packet(device_id, audio_data, target, websocket, session_id)
+            else:
+                session_logger.warning(device_id, session_id, "留言录音没有对应目标，将返回留言失败提示")
         else:
-            session_logger.warning(device_id, session_id, "忽略没有对应留言目标的留言音频包")
+            session_logger.warning(device_id, session_id, "忽略普通对话中混入的留言音频")
         return current_active_session, first_audio_received_time
     elif packet_type == 3:
-        if session.voice_target:
-            await device_target_cache.remove_target(device_id, expected=session.voice_target, reason="interrupted")
+        if not session.ai_started:
+            if session.voice_target:
+                await device_target_cache.remove_target(device_id, expected=session.voice_target, reason="interrupted")
             session.voice_target = None
             session.recording_finished = True
             await websocket.send_text("INTERRUPT_ACK")
             await websocket.send_text("INTERRUPT_PROCESSED")
+            await interrupt_handler.remove_session(session_key)
             return None if current_active_session == session_key else current_active_session, None
         current_active_session = await handle_interrupt_packet(
             device_id, session_id, session_key, session, current_active_session
         )
     elif packet_type == 2:
-        if session.voice_target:
-            target = await device_target_cache.take_target(device_id, session.voice_target, websocket, session_id)
+        if session.voice_target or session.message_audio_received:
+            target = (await device_target_cache.take_target(device_id, session.voice_target, websocket, session_id)
+                      if session.voice_target else None)
             session.voice_target = None
             session.recording_finished = True
             try:
@@ -240,10 +244,19 @@ async def handle_binary_message(websocket: WebSocket, device_id: str, serial_num
                     else:
                         await process_cached_audio(device_id, target.target_device_id, serial_number,
                                                    audio_cache_key=target.audio_cache_key)
+                elif current_active_session is None or current_active_session == session_key:
+                    await websocket.send_text(
+                        f"PROMPT_SOUND_URL:http://{settings.server_host}:{settings.server_port}/assets/audio/save_audio_fail.mp3")
             finally:
                 if target:
                     await target_audio_cache.clear_audio_data(target.audio_cache_key)
+                await interrupt_handler.remove_session(session_key)
             return None if current_active_session == session_key else current_active_session, None
+        if not session.ai_started:
+            # Preserve the existing silent-recording response for START/END
+            # without audio while keeping unowned messages out of the AI path.
+            current_active_session = await start_ai_recording(
+                websocket, device_id, session_id, session_key, session, current_active_session, sample_rate)
         if current_active_session == session_key:
             current_active_session = None
 
@@ -254,9 +267,6 @@ async def handle_binary_message(websocket: WebSocket, device_id: str, serial_num
 
 
 async def start_ai_recording(websocket, device_id, session_id, session_key, session, current_active_session, sample_rate):
-    current_active_session = await handle_start_packet(
-        device_id, session_id, session_key, session, current_active_session
-    )
     await handle_websocket_data(websocket, session_key, session, 1, b"", sample_rate)
     session.ai_started = True
     return current_active_session