feat: 实现真正的 LLM 流式 token 发送

backend/app/agent/agent_service.py

@@ -15,6 +15,7 @@ from ..model_services import get_cached_chat_services
 from ..main_graph.main_graph_builder import build_agent_graph
 from backend.app.logger import debug, info, warning, error
 from ..main_graph.state import AgentState
+from .stream_context import token_queue_var
 
 
 class AIAgentService:
@@ -251,71 +252,107 @@ class AIAgentService:
         chunk_count = 0
         full_message_content = ""
 
-        try:
-            info(f"📡 开始调用 graph.astream()...")
-            
-            event_count = 0
-            
-            async for chunk in self.graph.astream(
-                input_state,
-                config=config,
-                stream_mode=["messages", "updates"],
-                version="v2",
-                subgraphs=True
-            ):
-                chunk_count += 1
-                chunk_type = chunk["type"]
+        # 创建 token 队列
+        token_queue = asyncio.Queue()
+
+        # 设置上下文变量
+        token_queue_var.set(token_queue)
+
+        # 事件：graph 执行完成
+        graph_done = asyncio.Event()
+        graph_error = None
+
+        async def run_graph():
+            """在后台运行 graph，并把 chunk 放进队列，同时也处理 events"""
+            nonlocal chunk_count, full_message_content, graph_error
+            try:
+                info(f"📡 开始调用 graph.astream()...")
                 
-                # 记录原始 chunk 信息（前 10 个和后 10 个）
-                if chunk_count <= 10 or chunk_count % 50 == 0:
-                    info(f"  [{chunk_count}] chunk_type={chunk_type}, data={type(chunk.get('data'))}")
+                event_count = 0
+                
+                async for chunk in self.graph.astream(
+                    input_state,
+                    config=config,
+                    stream_mode=["messages", "updates"],
+                    version="v2",
+                    subgraphs=True
+                ):
+                    chunk_count += 1
+                    chunk_type = chunk["type"]
+                    
+                    # 记录原始 chunk 信息（前 10 个和后 10 个）
+                    if chunk_count <= 10 or chunk_count % 50 == 0:
+                        info(f"  [{chunk_count}] chunk_type={chunk_type}, data={type(chunk.get('data'))}")
 
-                if chunk_type == "messages":
-                    async for event in self._handle_message_chunk(
-                        chunk, current_node, tool_calls_in_progress
-                    ):
-                        if event.get("type") == "_update_state":
-                            current_node = event.get("current_node", current_node)
-                        else:
-                            event_count += 1
-                            # 记录前 10 个事件
-                            if event_count <= 10:
-                                info(f"    → yield event #{event_count}: {event.get('type')}")
-                            
-                            # 如果是 agent 节点的 token，收集完整消息
-                            if (
-                                event.get("type") == "llm_token"
-                                and event.get("node") == "agent"
-                                and "token" in event
-                            ):
-                                full_message_content += event["token"]
-                            yield event
+                    if chunk_type == "messages":
+                        async for event in self._handle_message_chunk(
+                            chunk, current_node, tool_calls_in_progress
+                        ):
+                            if event.get("type") == "_update_state":
+                                nonlocal current_node
+                                current_node = event.get("current_node", current_node)
+                            else:
+                                event_count += 1
+                                # 记录前 10 个事件
+                                if event_count <= 10:
+                                    info(f"    → yield event #{event_count}: {event.get('type')}")
+                                
+                                # 如果是 agent 节点的 token，收集完整消息
+                                if (
+                                    event.get("type") == "llm_token"
+                                    and event.get("node") == "agent"
+                                    and "token" in event
+                                ):
+                                    full_message_content += event["token"]
+                                await token_queue.put(event)
 
-                elif chunk_type == "updates":
-                    async for event in self._handle_updates_chunk(
-                        chunk, tool_calls_in_progress, actual_model_used
-                    ):
-                        if event.get("type") == "_update_state":
-                            actual_model_used = event.get("actual_model_used", actual_model_used)
-                        else:
-                            event_count += 1
-                            if event_count <= 10:
-                                info(f"    → yield event #{event_count}: {event.get('type')}")
-                            yield event
+                    elif chunk_type == "updates":
+                        async for event in self._handle_updates_chunk(
+                            chunk, tool_calls_in_progress, actual_model_used
+                        ):
+                            if event.get("type") == "_update_state":
+                                nonlocal actual_model_used
+                                actual_model_used = event.get("actual_model_used", actual_model_used)
+                            else:
+                                event_count += 1
+                                if event_count <= 10:
+                                    info(f"    → yield event #{event_count}: {event.get('type')}")
+                                await token_queue.put(event)
 
-            # 完整消息集合完成后，一次性打印
-            info(f"✅ graph.astream() 完成，共 {chunk_count} 个 chunks, {event_count} 个 events")
-            if full_message_content:
-                info(f"📄 完整消息内容: {repr(full_message_content)}")
+                # 完整消息集合完成后，一次性打印
+                info(f"✅ graph.astream() 完成，共 {chunk_count} 个 chunks, {event_count} 个 events")
+                if full_message_content:
+                    info(f"📄 完整消息内容: {repr(full_message_content)}")
+
+            except Exception as e:
+                error(f"❌ 执行图时出错: {e}")
+                import traceback
+                error(f"📋 堆栈: {traceback.format_exc()}")
+                graph_error = e
+                await token_queue.put({
+                    "type": "error",
+                    "message": str(e)
+                })
+            finally:
+                graph_done.set()
+
+        # 启动后台任务运行 graph
+        graph_task = asyncio.create_task(run_graph())
+
+        try:
+            # 从队列里取事件并 yield
+            while True:
+                # 尝试从队列取事件，超时检查 graph 是否完成
+                try:
+                    event = await asyncio.wait_for(token_queue.get(), timeout=0.1)
+                    yield event
+                except asyncio.TimeoutError:
+                    # 检查 graph 是否完成
+                    if graph_done.is_set():
+                        break
+
+            # 如果 graph 有错误，已经在 run_graph 里 yield error 了
 
-        except Exception as e:
-            error(f"❌ 执行图时出错: {e}")
-            import traceback
-            error(f"📋 堆栈: {traceback.format_exc()}")
-            yield {
-                "type": "error",
-                "message": str(e)
-            }
         finally:
             # 无论成功或失败，都发送结束事件，保证前端平稳关闭
             if current_node:
@@ -327,3 +364,5 @@ class AIAgentService:
                 "type": "done",
                 "model_used": actual_model_used
             }
+            # 取消任务
+            graph_task.cancel()

backend/app/agent/stream_context.py

@@ -0,0 +1,9 @@
+"""流式上下文，用于在 LangGraph 节点和 agent_service 之间传递 token 回调"""
+import contextvars
+import asyncio
+from typing import Optional, Any
+
+# 上下文变量：存储当前的 token 队列
+token_queue_var: contextvars.ContextVar[Optional[asyncio.Queue]] = contextvars.ContextVar(
+    "token_queue", default=None
+)

backend/app/main_graph/nodes/agent.py

@@ -1,10 +1,11 @@
 """Agent 节点：核心推理与工具调用"""
 
 from typing import Dict, Any, Optional
-from langchain_core.messages import SystemMessage
+from langchain_core.messages import SystemMessage, AIMessage, AIMessageChunk
 from langchain_core.runnables.config import RunnableConfig
 from ..state import AgentState
-from backend.app.logger import info, warning
+from backend.app.logger import info, warning, error
+from .stream_context import token_queue_var
 
 
 # 系统提示词（从 main_graph_builder.py 搬过来）
@@ -77,23 +78,81 @@ def create_agent_node(llm_with_tools, llm):
             # 判断是否达到步数上限
             if state.current_step >= state.max_steps:
                 info(f"[Agent] 达到步数上限 {state.max_steps}，强制结束，不绑定工具")
-                llm_no_tools = llm.bind_tools([])
-                response = await llm_no_tools.ainvoke(full_messages)
+                current_llm = llm.bind_tools([])
             else:
-                info(f"[Agent] 调用带工具的 LLM...")
-                response = await llm_with_tools.ainvoke(full_messages)
-            
+                current_llm = llm_with_tools
+
+            info(f"[Agent] 调用带工具的 LLM...")
+
+            # 获取 token 队列
+            token_queue = token_queue_var.get()
+
+            # 完整消息
+            full_content = ""
+            full_reasoning_content = ""
+            full_tool_calls = []
+
+            # 流式调用 LLM
+            async for chunk in current_llm.astream(full_messages):
+                if isinstance(chunk, AIMessageChunk):
+                    # 处理 content
+                    if chunk.content:
+                        full_content += chunk.content
+                        if token_queue:
+                            await token_queue.put({
+                                "type": "llm_token",
+                                "node": "agent",
+                                "token": chunk.content,
+                                "reasoning_token": ""
+                            })
+
+                    # 处理 reasoning_content
+                    if hasattr(chunk, 'additional_kwargs') and chunk.additional_kwargs:
+                        reasoning_content = chunk.additional_kwargs.get("reasoning_content", "")
+                        if reasoning_content:
+                            full_reasoning_content += reasoning_content
+                            if token_queue:
+                                await token_queue.put({
+                                    "type": "llm_token",
+                                    "node": "agent",
+                                    "token": "",
+                                    "reasoning_token": reasoning_content
+                                })
+
+                    # 处理 tool_calls
+                    if hasattr(chunk, 'tool_calls') and chunk.tool_calls:
+                        # 合并 tool_calls
+                        for tc in chunk.tool_calls:
+                            # 查找是否已经有这个 id 的 tool_call
+                            found = False
+                            for existing_tc in full_tool_calls:
+                                if existing_tc.get("id") == tc.get("id"):
+                                    # 合并 args
+                                    existing_tc["args"] = {**existing_tc.get("args", {}), **tc.get("args", {})}
+                                    found = True
+                                    break
+                            if not found:
+                                full_tool_calls.append(tc)
+
+            # 构建完整的 AIMessage
+            response = AIMessage(
+                content=full_content,
+                tool_calls=full_tool_calls if full_tool_calls else None
+            )
+            if full_reasoning_content:
+                response.additional_kwargs["reasoning_content"] = full_reasoning_content
+
             info(f"[Agent] LLM 调用成功！响应类型: {type(response).__name__}")
             if hasattr(response, 'tool_calls') and response.tool_calls:
                 info(f"[Agent] 检测到工具调用: {[tc['name'] for tc in response.tool_calls]}")
 
-            # 返回状态更新（注意：不原地修改 state，返回字典让 LangGraph 处理
+            # 返回状态更新
             return {
                 "messages": [response],
                 "current_step": state.current_step + 1,
                 "llm_calls": state.llm_calls + 1
             }
-        
+
         except Exception as e:
             error(f"[Agent] ❌ 第 {state.current_step} 步推理出错: {e}")
             import traceback