fix: 用户开始录音5s后ai反馈“我在认真听”

api/routers/websocket.py

@@ -32,6 +32,7 @@ LEGACY_VOICE_SESSION_ID = "legacy"
 class MessageType(str, Enum):
     """WebSocket 消息类型"""
     CONNECT = "connect"
+    RECORDING_STARTED = "recording_started"  # 客户端开始录音，用于服务端 5s 后发「我在认真听」
     AUDIO_CHUNK = "audio_chunk"
     AUDIO_SEGMENT = "audio_segment"  # 分段语音消息（长语音持续上传）
     AUDIO_MESSAGE = "audio_message"  # 完整音频消息（类似微信语音）
@@ -146,6 +147,9 @@ class SegmentStreamState:
     buffered_transcripts: Dict[int, Tuple[str, Segment]] = field(default_factory=dict)
     consumed_index: int = -1
     active_tasks: Set[asyncio.Task] = field(default_factory=set)
+    # 录音开始约 5s 后只发一次「我在认真听」；若用户提前结束录音则取消待发
+    listening_feedback_sent: bool = False
+    listening_feedback_task: Optional[asyncio.Task] = None
 
 
 def _utc_now() -> datetime:
@@ -257,17 +261,21 @@ async def _get_persisted_contiguous_segment_index(
     return contiguous_index
 
 
+LISTENING_FEEDBACK_DELAY_SEC = 5.0
+LISTENING_FEEDBACK_TEXT = "我在认真听，你继续说，我会边听边整理重点。"
+
+
 async def _send_segment_transition_feedback(
     conversation_id: str,
     segment_index: int,
     manager: ConnectionManager,
 ) -> None:
-    """ASR 处理中先给陪伴式过渡反馈，避免用户感知卡住。"""
+    """发送一次「我在认真听」陪伴式过渡反馈（由延迟任务调用）。"""
     await manager.send_message(conversation_id, {
         "type": MessageType.AGENT_RESPONSE,
         "conversation_id": conversation_id,
         "data": {
-            "text": "我在认真听，你继续说，我会边听边整理重点。",
+            "text": LISTENING_FEEDBACK_TEXT,
             "transition": True,
             "segment_index": segment_index,
         },
@@ -275,6 +283,22 @@ async def _send_segment_transition_feedback(
     })
 
 
+async def _delayed_listening_feedback(
+    conversation_id: str,
+    voice_session_id: str,
+    manager: ConnectionManager,
+) -> None:
+    """录音开始后延迟 5 秒发送一次「我在认真听」，本会话内只发一次；若用户已结束录音则不再发送。"""
+    await asyncio.sleep(LISTENING_FEEDBACK_DELAY_SEC)
+    state = manager.get_or_create_segment_state(conversation_id, voice_session_id)
+    async with state.lock:
+        if state.listening_feedback_sent:
+            return
+        state.listening_feedback_sent = True
+        state.listening_feedback_task = None
+    await _send_segment_transition_feedback(conversation_id, 0, manager)
+
+
 async def _process_audio_segment_async(
     conversation_id: str,
     user_id: str,
@@ -600,6 +624,28 @@ async def websocket_endpoint(
                                 user_message_timestamp=segment.created_at or user_message_timestamp,
                             )
 
+                    elif msg_type == MessageType.RECORDING_STARTED:
+                        # 用户点击开始录音：启动 5s 定时器，到时发一次「我在认真听」
+                        data = message.get("data", {})
+                        voice_session_id = _normalize_voice_session_id(data.get("voice_session_id"))
+                        segment_state = manager.get_or_create_segment_state(
+                            conversation_id,
+                            voice_session_id,
+                        )
+                        async with segment_state.lock:
+                            if segment_state.listening_feedback_task is not None and not segment_state.listening_feedback_task.done():
+                                continue  # 本会话已有待发任务，不重复
+                            if segment_state.listening_feedback_sent:
+                                continue
+                            delayed_task = asyncio.create_task(
+                                _delayed_listening_feedback(
+                                    conversation_id=conversation_id,
+                                    voice_session_id=voice_session_id,
+                                    manager=manager,
+                                )
+                            )
+                            segment_state.listening_feedback_task = delayed_task
+
                     elif msg_type == MessageType.AUDIO_SEGMENT:
                         # 处理分段语音消息（长语音持续上传）
                         data = message.get("data", {})
@@ -680,12 +726,13 @@ async def websocket_endpoint(
                             )
                             continue
 
-                        # 先发过渡反馈，减少“等待空白”体感
-                        await _send_segment_transition_feedback(
-                            conversation_id=conversation_id,
-                            segment_index=segment_index,
-                            manager=manager,
-                        )
+                        # 若本段是用户结束录音的最后一段，取消尚未发出的「我在认真听」，避免结束后再说
+                        if is_last:
+                            async with segment_state.lock:
+                                t = segment_state.listening_feedback_task
+                                segment_state.listening_feedback_task = None
+                            if t is not None and not t.done():
+                                t.cancel()
 
                         task = asyncio.create_task(
                             _process_audio_segment_async(