Merge branch 'master' into bot_manager

routes/chat.py

@@ -155,6 +155,7 @@ async def enhanced_generate_stream_response(
 
         # 输出控制器：确保 preamble 先输出，然后是 agent stream
         preamble_output_done = False
+        last_yield_time = time.time()
 
         while True:
             try:
@@ -165,6 +166,7 @@ async def enhanced_generate_stream_response(
                     # 立即输出 preamble 内容
                     if item_data:
                         yield item_data
+                        last_yield_time = time.time()
                     preamble_output_done = True
 
                 elif item_type == "preamble_done":
@@ -175,6 +177,7 @@ async def enhanced_generate_stream_response(
                     # Agent stream 内容，需要等待 preamble 输出完成
                     if preamble_output_done:
                         yield item_data
+                        last_yield_time = time.time()
                     else:
                         # preamble 还没输出，先放回队列
                         await output_queue.put((item_type, item_data))
@@ -191,9 +194,11 @@ async def enhanced_generate_stream_response(
                 if all(task.done() for task in [preamble_task_handle, agent_task_handle]):
                     # 所有任务都完成了，退出循环
                     break
-                # 发送空内容心跳包保持连接活跃，防止 nginx/客户端超时断开
+                # 15秒无消息输出时才发送心跳包保持连接活跃
-                heartbeat_chunk = create_stream_chunk(f"chatcmpl-heartbeat", config.model_name, "")
+                if time.time() - last_yield_time >= 15:
-                yield f"data: {json.dumps(heartbeat_chunk, ensure_ascii=False)}\n\n"
+                    heartbeat_chunk = create_stream_chunk(f"chatcmpl-heartbeat", config.model_name, "")
+                    yield f"data: {json.dumps(heartbeat_chunk, ensure_ascii=False)}\n\n"
+                    last_yield_time = time.time()
                 continue
 
         # 发送结束标记