think增加流式输出模式

agent/guideline_middleware.py

@@ -96,14 +96,11 @@ Action: Provide concise, friendly, and personified natural responses.
         # 使用回调处理器调用模型
         response = self.model.invoke(
             guideline_prompt,
-            config={"callbacks": [BaseCallbackHandler()]}
+            config={"metadata": {"message_tag": "THINK"}}
         )
 
-        # 提取<think>与</think>之间的内容作为thinking
+        response.additional_kwargs["message_tag"] = "THINK"
-        
+        response.content = f"<think>{response.content}</think>"
-        match = re.search(r'<think>(.*?)</think>', response.content, re.DOTALL)
-        response.additional_kwargs["thinking"] = match.group(1).strip() if match else response.content
-        
         messages = state['messages']+[response]
         return {
             "messages": messages
@@ -118,12 +115,10 @@ Action: Provide concise, friendly, and personified natural responses.
         # 使用回调处理器调用模型
         response = await self.model.ainvoke(
             guideline_prompt, 
-            config={"callbacks": [BaseCallbackHandler()]}
+            config={"metadata": {"message_tag": "THINK"}}
         )
-        
+        response.additional_kwargs["message_tag"] = "THINK"
-        # 提取<think>与</think>之间的内容作为thinking
+        response.content = f"<think>{response.content}</think>"
-        match = re.search(r'<think>(.*?)</think>', response.content, re.DOTALL)
-        response.additional_kwargs["thinking"] = match.group(1).strip() if match else response.content
 
         messages = state['messages']+[response]
         return {

prompt/guideline_prompt.md

@@ -56,9 +56,8 @@
 **语言要求**：所有用户交互和结果输出，必须使用[{language}]
 
 ## 输出格式
-按顺序输出执行步骤, 用<think>和</think>标签包裹，输出格式如下：
+按顺序输出执行步骤, 输出格式如下：
 
-<think>
 ### 🎯 核心目标
 [简洁描述用户的主要目标]
 
@@ -70,5 +69,3 @@
 ### ⚡ 计划阶段
 **执行步骤**:
 [按1,2,3...n顺序列出执行步骤]
-
-</think>

routes/chat.py

@@ -148,6 +148,7 @@ async def enhanced_generate_stream_response(
                 
                 async for msg, metadata in agent.astream({"messages": messages}, stream_mode="messages"):
                     new_content = ""
+                    print(metadata)
                     if isinstance(msg, AIMessageChunk):
                         # 判断是否有工具调用
                         if msg.tool_call_chunks:  # 检查工具调用块
@@ -160,23 +161,20 @@ async def enhanced_generate_stream_response(
                         elif len(msg.content) > 0:
                             preamble_completed.set()
                             await output_queue.put(("preamble_done", None))
-                            if message_tag != "ANSWER":
+                            meta_message_tag = metadata.get("message_tag", "ANSWER")
-                                message_tag = "ANSWER"
+                            if meta_message_tag != message_tag:
-                                new_content = f"[{message_tag}]\n{msg.text}"
+                                message_tag = meta_message_tag
-                            elif message_tag == "ANSWER":
+                                new_content = f"[{meta_message_tag}]\n"
-                                new_content = msg.text
+                            new_content += msg.text
                         elif message_tag == "TOOL_CALL" and \
                             (
                                 ("finish_reason" in msg.response_metadata and msg.response_metadata["finish_reason"] == "tool_calls") or \
                                 ("stop_reason" in msg.response_metadata and msg.response_metadata["stop_reason"] == "tool_use")
                             ):
                             new_content = f"[{message_tag}] {function_name}\n{tool_args}"
-                            message_tag = "TOOL_CALL"
                     elif isinstance(msg, ToolMessage) and len(msg.content) > 0:
                         message_tag = "TOOL_RESPONSE"
                         new_content = f"[{message_tag}] {msg.name}\n{msg.text}"
-                    elif isinstance(msg, AIMessage) and msg.additional_kwargs and "thinking" in msg.additional_kwargs:
-                        new_content = "[THINK]\n" + msg.additional_kwargs["thinking"] + "\n"
 
                     # 只有当有新内容时才发送chunk
                     if new_content:
@@ -326,10 +324,10 @@ async def create_agent_and_generate_response(
     response_text = ""
     for msg in append_messages:
         if isinstance(msg,AIMessage):
-            if msg.additional_kwargs and "thinking" in msg.additional_kwargs:
+            if len(msg.text)>0:
-                response_text += "[THINK]\n"+msg.additional_kwargs["thinking"]+ "\n"
+                meta_message_tag = msg.additional_kwargs.get("message_tag", "ANSWER")
-            elif len(msg.text)>0:
+                output_text = msg.text.replace("<think>","").replace("</think>","") if meta_message_tag == "THINK" else msg.text
-                response_text += "[ANSWER]\n"+msg.text+ "\n"
+                response_text += f"[{meta_message_tag}]\n"+output_text+ "\n"
             if len(msg.tool_calls)>0:
                 response_text += "".join([f"[TOOL_CALL] {tool['name']}\n{json.dumps(tool["args"]) if isinstance(tool["args"],dict) else tool["args"]}\n" for tool in msg.tool_calls])
         elif isinstance(msg,ToolMessage) and tool_response: