feat(agent): enhance streaming with in-memory event manager and fallback polling

- Implement dual-mode streaming: prioritize in-memory EventManager for running tasks with thinking_token support
- Add fallback to database polling for completed tasks without thinking_token replay capability
- Introduce SSE event formatter utility for consistent event serialization across streaming modes
- Add 10ms micro-delay for thinking_token events to ensure proper TCP packet separation and frontend incremental rendering
- Refactor stream_agent_with_thinking endpoint to support both runtime and historical event streaming
- Update event filtering logic to handle both in-memory and database event sources
- Improve logging with debug markers for thinking_token tracking and stream mode selection
- Optimize polling intervals: 0.3s for running tasks, 2.0s for completed tasks
- Reduce idle timeout from 10 minutes to 1 minute for completed task streams
- Update frontend useAgentStream hook to handle unified event format from dual-mode streaming
- Enhance AgentAudit UI to properly display streamed events from both sources

backend/app/api/v1/endpoints/agent_tasks.py

@@ -601,25 +601,13 @@ async def stream_agent_with_thinking(
     增强版事件流 (SSE)
     
     支持:
-    - LLM 思考过程的 Token 级流式输出
+    - LLM 思考过程的 Token 级流式输出 (仅运行时)
     - 工具调用的详细输入/输出
     - 节点执行状态
     - 发现事件
     
-    事件类型:
-    - thinking_start: LLM 开始思考
-    - thinking_token: LLM 输出 Token
-    - thinking_end: LLM 思考结束
-    - tool_call_start: 工具调用开始
-    - tool_call_end: 工具调用结束
-    - node_start: 节点开始
-    - node_end: 节点结束
-    - finding_new: 新发现
-    - finding_verified: 验证通过
-    - progress: 进度更新
-    - task_complete: 任务完成
-    - task_error: 任务错误
-    - heartbeat: 心跳
+    优先使用内存中的事件队列 (支持 thinking_token)，
+    如果任务未在运行，则回退到数据库轮询 (不支持 thinking_token 复盘)。
     """
     task = await db.get(AgentTask, task_id)
     if not task:
@@ -629,119 +617,156 @@ async def stream_agent_with_thinking(
     if not project or project.owner_id != current_user.id:
         raise HTTPException(status_code=403, detail="无权访问此任务")
     
+    # 定义 SSE 格式化函数
+    def format_sse_event(event_data: Dict[str, Any]) -> str:
+        """格式化为 SSE 事件"""
+        event_type = event_data.get("event_type") or event_data.get("type")
+        
+        # 统一字段
+        if "type" not in event_data:
+            event_data["type"] = event_type
+            
+        return f"event: {event_type}\ndata: {json.dumps(event_data, ensure_ascii=False)}\n\n"
+
     async def enhanced_event_generator():
         """生成增强版 SSE 事件流"""
-        last_sequence = after_sequence
-        poll_interval = 0.3  # 更短的轮询间隔以支持流式
-        heartbeat_interval = 15  # 心跳间隔
-        max_idle = 600  # 10 分钟无事件后关闭
-        idle_time = 0
-        last_heartbeat = 0
+        # 1. 检查任务是否在运行中 (内存)
+        runner = _running_tasks.get(task_id)
         
-        # 事件类型过滤
-        skip_types = set()
-        if not include_thinking:
-            skip_types.update(["thinking_start", "thinking_token", "thinking_end"])
-        if not include_tool_calls:
-            skip_types.update(["tool_call_start", "tool_call_input", "tool_call_output", "tool_call_end"])
-        
-        while True:
+        if runner:
+            logger.info(f"Stream {task_id}: Using in-memory event manager")
             try:
-                async with async_session_factory() as session:
-                    # 查询新事件
-                    result = await session.execute(
-                        select(AgentEvent)
-                        .where(AgentEvent.task_id == task_id)
-                        .where(AgentEvent.sequence > last_sequence)
-                        .order_by(AgentEvent.sequence)
-                        .limit(100)
-                    )
-                    events = result.scalars().all()
+                # 使用 EventManager 的流式接口
+                # 过滤选项
+                skip_types = set()
+                if not include_thinking:
+                    skip_types.update(["thinking_start", "thinking_token", "thinking_end"])
+                if not include_tool_calls:
+                    skip_types.update(["tool_call_start", "tool_call_input", "tool_call_output", "tool_call_end"])
                 
-                    # 获取任务状态
-                    current_task = await session.get(AgentTask, task_id)
-                    task_status = current_task.status if current_task else None
+                async for event in runner.event_manager.stream_events(task_id, after_sequence=after_sequence):
+                    event_type = event.get("event_type")
                     
-                if events:
-                    idle_time = 0
-                    for event in events:
-                        last_sequence = event.sequence
+                    if event_type in skip_types:
+                        continue
                     
-                        # 获取事件类型字符串（event_type 已经是字符串）
-                        event_type = str(event.event_type)
+                    # 🔥 Debug: 记录 thinking_token 事件
+                    if event_type == "thinking_token":
+                        token = event.get("metadata", {}).get("token", "")[:20]
+                        logger.debug(f"Stream {task_id}: Sending thinking_token: '{token}...'")
                         
-                        # 过滤事件
-                        if event_type in skip_types:
-                            continue
+                    # 格式化并 yield
+                    yield format_sse_event(event)
                     
-                        # 构建事件数据
-                        data = {
-                            "id": event.id,
-                            "type": event_type,
-                            "phase": str(event.phase) if event.phase else None,
-                            "message": event.message,
-                            "sequence": event.sequence,
-                            "timestamp": event.created_at.isoformat() if event.created_at else None,
-                        }
-                        
-                        # 添加工具调用详情
-                        if include_tool_calls and event.tool_name:
-                            data["tool"] = {
-                                "name": event.tool_name,
-                                "input": event.tool_input,
-                                "output": event.tool_output,
-                                "duration_ms": event.tool_duration_ms,
-                            }
-                        
-                        # 添加元数据
-                        if event.event_metadata:
-                            data["metadata"] = event.event_metadata
-                        
-                        # 添加 Token 使用
-                        if event.tokens_used:
-                            data["tokens_used"] = event.tokens_used
-                        
-                        # 使用标准 SSE 格式
-                        yield f"event: {event_type}\ndata: {json.dumps(data, ensure_ascii=False)}\n\n"
-                else:
-                    idle_time += poll_interval
-                
-                # 检查任务是否结束
-                if task_status:
-                    status_str = str(task_status)
-                    if status_str in ["completed", "failed", "cancelled"]:
-                        end_data = {
-                            "type": "task_end",
-                            "status": status_str,
-                            "message": f"任务{'完成' if status_str == 'completed' else '结束'}",
-                        }
-                        yield f"event: task_end\ndata: {json.dumps(end_data, ensure_ascii=False)}\n\n"
-                        break
-                
-                # 发送心跳
-                last_heartbeat += poll_interval
-                if last_heartbeat >= heartbeat_interval:
-                    last_heartbeat = 0
-                    heartbeat_data = {
-                        "type": "heartbeat",
-                        "timestamp": datetime.now(timezone.utc).isoformat(),
-                        "last_sequence": last_sequence,
-                    }
-                    yield f"event: heartbeat\ndata: {json.dumps(heartbeat_data)}\n\n"
-                
-                # 检查空闲超时
-                if idle_time >= max_idle:
-                    timeout_data = {"type": "timeout", "message": "连接超时"}
-                    yield f"event: timeout\ndata: {json.dumps(timeout_data)}\n\n"
-                    break
-                
-                await asyncio.sleep(poll_interval)
+                    # 🔥 CRITICAL: 为 thinking_token 添加微小延迟
+                    # 确保事件在不同的 TCP 包中发送，让前端能够逐个处理
+                    # 没有这个延迟，所有 token 会在一次 read() 中被接收，导致 React 批量更新
+                    if event_type == "thinking_token":
+                        await asyncio.sleep(0.01)  # 10ms 延迟
                     
             except Exception as e:
-                logger.error(f"Stream error: {e}")
-                error_data = {"type": "error", "message": str(e)}
-                yield f"event: error\ndata: {json.dumps(error_data)}\n\n"
-                break
+                logger.error(f"In-memory stream error: {e}")
+                err_data = {"type": "error", "message": str(e)}
+                yield format_sse_event(err_data)
+                
+        else:
+            logger.info(f"Stream {task_id}: Task not running, falling back to DB polling")
+            # 2. 回退到数据库轮询 (无法获取 thinking_token)
+            last_sequence = after_sequence
+            poll_interval = 2.0  # 完成的任务轮询可以慢一点
+            heartbeat_interval = 15
+            max_idle = 60  # 1分钟无事件关闭
+            idle_time = 0
+            last_heartbeat = 0
+            
+            skip_types = set()
+            if not include_thinking:
+                skip_types.update(["thinking_start", "thinking_token", "thinking_end"])
+            
+            while True:
+                try:
+                    async with async_session_factory() as session:
+                        # 查询新事件
+                        result = await session.execute(
+                            select(AgentEvent)
+                            .where(AgentEvent.task_id == task_id)
+                            .where(AgentEvent.sequence > last_sequence)
+                            .order_by(AgentEvent.sequence)
+                            .limit(100)
+                        )
+                        events = result.scalars().all()
+                        
+                        # 获取任务状态
+                        current_task = await session.get(AgentTask, task_id)
+                        task_status = current_task.status if current_task else None
+                    
+                    if events:
+                        idle_time = 0
+                        for event in events:
+                            last_sequence = event.sequence
+                            event_type = str(event.event_type)
+                            
+                            if event_type in skip_types:
+                                continue
+                            
+                            # 构建数据
+                            data = {
+                                "id": event.id,
+                                "type": event_type,
+                                "phase": str(event.phase) if event.phase else None,
+                                "message": event.message,
+                                "sequence": event.sequence,
+                                "timestamp": event.created_at.isoformat() if event.created_at else None,
+                            }
+                            
+                            # 添加详情
+                            if include_tool_calls and event.tool_name:
+                                data["tool"] = {
+                                    "name": event.tool_name,
+                                    "input": event.tool_input,
+                                    "output": event.tool_output,
+                                    "duration_ms": event.tool_duration_ms,
+                                }
+                                
+                            if event.event_metadata:
+                                data["metadata"] = event.event_metadata
+                                
+                            if event.tokens_used:
+                                data["tokens_used"] = event.tokens_used
+                            
+                            yield format_sse_event(data)
+                    else:
+                        idle_time += poll_interval
+                        
+                        # 检查是否应该结束
+                        if task_status:
+                            status_str = str(task_status)
+                            # 如果任务已完成且没有新事件，结束流
+                            if status_str in ["completed", "failed", "cancelled"]:
+                                end_data = {
+                                    "type": "task_end",
+                                    "status": status_str,
+                                    "message": f"任务已{status_str}"
+                                }
+                                yield format_sse_event(end_data)
+                                break
+                    
+                    # 心跳
+                    last_heartbeat += poll_interval
+                    if last_heartbeat >= heartbeat_interval:
+                        last_heartbeat = 0
+                        yield format_sse_event({"type": "heartbeat", "timestamp": datetime.now(timezone.utc).isoformat()})
+                    
+                    # 超时
+                    if idle_time >= max_idle:
+                        break
+                    
+                    await asyncio.sleep(poll_interval)
+                    
+                except Exception as e:
+                    logger.error(f"DB poll stream error: {e}")
+                    yield format_sse_event({"type": "error", "message": str(e)})
+                    break
     
     return StreamingResponse(
         enhanced_event_generator(),

backend/app/services/agent/agents/analysis.py

@@ -187,6 +187,10 @@ class AnalysisAgent(BaseAgent):
         thought_match = re.search(r'Thought:\s*(.*?)(?=Action:|Final Answer:|$)', response, re.DOTALL)
         if thought_match:
             step.thought = thought_match.group(1).strip()
+        elif not re.search(r'Action:|Final Answer:', response):
+             # 🔥 Fallback: If no markers found, treat the whole response as Thought
+             if response.strip():
+                 step.thought = response.strip()
         
         # 检查是否是最终答案
         final_match = re.search(r'Final Answer:\s*(.*?)$', response, re.DOTALL)
@@ -331,6 +335,16 @@ class AnalysisAgent(BaseAgent):
                 
                 self._total_tokens += tokens_this_round
 
+                # 🔥 Handle empty LLM response to prevent loops
+                if not llm_output or not llm_output.strip():
+                    logger.warning(f"[{self.name}] Empty LLM response in iteration {self._iteration}")
+                    await self.emit_llm_decision("收到空响应", "LLM 返回内容为空，尝试重试通过提示")
+                    self._conversation_history.append({
+                        "role": "user",
+                        "content": "Received empty response. Please output your Thought and Action.",
+                    })
+                    continue
+
                 # 解析 LLM 响应
                 step = self._parse_llm_response(llm_output)
                 self._steps.append(step)
@@ -406,6 +420,11 @@ class AnalysisAgent(BaseAgent):
             # 标准化发现
             standardized_findings = []
             for finding in all_findings:
+                # 确保 finding 是字典
+                if not isinstance(finding, dict):
+                    logger.warning(f"Skipping invalid finding (not a dict): {finding}")
+                    continue
+                    
                 standardized = {
                     "vulnerability_type": finding.get("vulnerability_type", "other"),
                     "severity": finding.get("severity", "medium"),

backend/app/services/agent/agents/base.py

@@ -409,17 +409,38 @@ class BaseAgent(ABC):
         """发射事件"""
         if self.event_emitter:
             from ..event_manager import AgentEventData
+            
+            # 准备 metadata
+            metadata = kwargs.get("metadata", {}) or {}
+            if "agent_name" not in metadata:
+                metadata["agent_name"] = self.name
+            
+            # 分离已知字段和未知字段
+            known_fields = {
+                "phase", "tool_name", "tool_input", "tool_output", 
+                "tool_duration_ms", "finding_id", "tokens_used"
+            }
+            
+            event_kwargs = {}
+            for k, v in kwargs.items():
+                if k in known_fields:
+                    event_kwargs[k] = v
+                elif k != "metadata":
+                    # 将未知字段放入 metadata
+                    metadata[k] = v
+            
             await self.event_emitter.emit(AgentEventData(
                 event_type=event_type,
                 message=message,
-                **kwargs
+                metadata=metadata,
+                **event_kwargs
             ))
     
     # ============ LLM 思考相关事件 ============
     
     async def emit_thinking(self, message: str):
         """发射 LLM 思考事件"""
-        await self.emit_event("thinking", f"[{self.name}] {message}")
+        await self.emit_event("thinking", message)
     
     async def emit_llm_start(self, iteration: int):
         """发射 LLM 开始思考事件"""
@@ -444,7 +465,7 @@ class BaseAgent(ABC):
     
     async def emit_thinking_start(self):
         """发射开始思考事件（流式输出用）"""
-        await self.emit_event("thinking_start", f"[{self.name}] 开始思考...")
+        await self.emit_event("thinking_start", "开始思考...")
     
     async def emit_thinking_token(self, token: str, accumulated: str):
         """发射思考 token 事件（流式输出用）"""
@@ -461,7 +482,7 @@ class BaseAgent(ABC):
         """发射思考结束事件（流式输出用）"""
         await self.emit_event(
             "thinking_end",
-            f"[{self.name}] 思考完成",
+            "思考完成",
             metadata={"accumulated": full_response}
         )
     
@@ -690,6 +711,9 @@ class BaseAgent(ABC):
                     token = chunk["content"]
                     accumulated = chunk["accumulated"]
                     await self.emit_thinking_token(token, accumulated)
+                    # 🔥 CRITICAL: 让出控制权给事件循环，让 SSE 有机会发送事件
+                    # 如果不这样做，所有 token 会在循环结束后一起发送
+                    await asyncio.sleep(0)
                     
                 elif chunk["type"] == "done":
                     accumulated = chunk["content"]

backend/app/services/agent/agents/orchestrator.py

@@ -421,6 +421,9 @@ class OrchestratorAgent(BaseAgent):
 ### 发现摘要
 """
                 for i, f in enumerate(findings[:10]):  # 最多显示 10 个
+                    if not isinstance(f, dict):
+                         continue
+                         
                     observation += f"""
 {i+1}. [{f.get('severity', 'unknown')}] {f.get('title', 'Unknown')}
    - 类型: {f.get('vulnerability_type', 'unknown')}
@@ -452,6 +455,9 @@ class OrchestratorAgent(BaseAgent):
         type_counts = {}
         
         for f in self._all_findings:
+            if not isinstance(f, dict):
+                continue
+                
             sev = f.get("severity", "low")
             severity_counts[sev] = severity_counts.get(sev, 0) + 1
             
@@ -475,7 +481,8 @@ class OrchestratorAgent(BaseAgent):
         
         summary += "\n### 详细列表\n"
         for i, f in enumerate(self._all_findings):
-            summary += f"{i+1}. [{f.get('severity')}] {f.get('title')} ({f.get('file_path')})\n"
+            if isinstance(f, dict):
+                summary += f"{i+1}. [{f.get('severity')}] {f.get('title')} ({f.get('file_path')})\n"
         
         return summary
     
@@ -484,8 +491,9 @@ class OrchestratorAgent(BaseAgent):
         severity_counts = {"critical": 0, "high": 0, "medium": 0, "low": 0}
         
         for f in self._all_findings:
-            sev = f.get("severity", "low")
-            severity_counts[sev] = severity_counts.get(sev, 0) + 1
+            if isinstance(f, dict):
+                sev = f.get("severity", "low")
+                severity_counts[sev] = severity_counts.get(sev, 0) + 1
         
         return {
             "total_findings": len(self._all_findings),

backend/app/services/agent/agents/recon.py

@@ -157,6 +157,11 @@ class ReconAgent(BaseAgent):
         thought_match = re.search(r'Thought:\s*(.*?)(?=Action:|Final Answer:|$)', response, re.DOTALL)
         if thought_match:
             step.thought = thought_match.group(1).strip()
+        elif not re.search(r'Action:|Final Answer:', response):
+             # 🔥 Fallback: If no markers found, treat the whole response as Thought
+             # This prevents empty steps loops "Decision: Continue Thinking"
+             if response.strip():
+                 step.thought = response.strip()
         
         # 检查是否是最终答案
         final_match = re.search(r'Final Answer:\s*(.*?)$', response, re.DOTALL)
@@ -170,6 +175,12 @@ class ReconAgent(BaseAgent):
                 answer_text, 
                 default={"raw_answer": answer_text}
             )
+            # 确保 findings 格式正确
+            if "initial_findings" in step.final_answer:
+                step.final_answer["initial_findings"] = [
+                    f for f in step.final_answer["initial_findings"] 
+                    if isinstance(f, dict)
+                ]
             return step
         
         # 提取 Action
@@ -256,6 +267,16 @@ class ReconAgent(BaseAgent):
                 
                 self._total_tokens += tokens_this_round
                 
+                # 🔥 Handle empty LLM response to prevent loops
+                if not llm_output or not llm_output.strip():
+                    logger.warning(f"[{self.name}] Empty LLM response in iteration {self._iteration}")
+                    await self.emit_llm_decision("收到空响应", "LLM 返回内容为空，尝试重试通过提示")
+                    self._conversation_history.append({
+                        "role": "user",
+                        "content": "Received empty response. Please output your Thought and Action.",
+                    })
+                    continue
+
                 # 解析 LLM 响应
                 step = self._parse_llm_response(llm_output)
                 self._steps.append(step)

backend/app/services/agent/agents/verification.py

@@ -169,6 +169,10 @@ class VerificationAgent(BaseAgent):
         thought_match = re.search(r'Thought:\s*(.*?)(?=Action:|Final Answer:|$)', response, re.DOTALL)
         if thought_match:
             step.thought = thought_match.group(1).strip()
+        elif not re.search(r'Action:|Final Answer:', response):
+             # 🔥 Fallback: If no markers found, treat the whole response as Thought
+             if response.strip():
+                 step.thought = response.strip()
         
         # 检查是否是最终答案
         final_match = re.search(r'Final Answer:\s*(.*?)$', response, re.DOTALL)
@@ -338,6 +342,16 @@ class VerificationAgent(BaseAgent):
                 
                 self._total_tokens += tokens_this_round
 
+                # 🔥 Handle empty LLM response to prevent loops
+                if not llm_output or not llm_output.strip():
+                    logger.warning(f"[{self.name}] Empty LLM response in iteration {self._iteration}")
+                    await self.emit_llm_decision("收到空响应", "LLM 返回内容为空，尝试重试通过提示")
+                    self._conversation_history.append({
+                        "role": "user",
+                        "content": "Received empty response. Please output your Thought and Action.",
+                    })
+                    continue
+
                 # 解析 LLM 响应
                 step = self._parse_llm_response(llm_output)
                 self._steps.append(step)

backend/app/services/agent/event_manager.py

@@ -300,16 +300,19 @@ class EventManager:
         }
         
         # 保存到数据库（跳过高频事件如 thinking_token）
-        skip_db_events = {"thinking_token", "thinking_start", "thinking_end"}
+        skip_db_events = {"thinking_token"}
         if self.db_session_factory and event_type not in skip_db_events:
             try:
                 await self._save_event_to_db(event_data)
             except Exception as e:
                 logger.error(f"Failed to save event to database: {e}")
         
-        # 推送到队列
+        # 推送到队列（非阻塞）
         if task_id in self._event_queues:
-            await self._event_queues[task_id].put(event_data)
+            try:
+                self._event_queues[task_id].put_nowait(event_data)
+            except asyncio.QueueFull:
+                logger.warning(f"Event queue full for task {task_id}, dropping event: {event_type}")
         
         # 调用回调
         if task_id in self._event_callbacks:
@@ -348,9 +351,10 @@ class EventManager:
             await db.commit()
     
     def create_queue(self, task_id: str) -> asyncio.Queue:
-        """创建事件队列"""
+        """创建或获取事件队列"""
         if task_id not in self._event_queues:
-            self._event_queues[task_id] = asyncio.Queue()
+            # 🔥 使用较大的队列容量，缓存更多 token 事件
+            self._event_queues[task_id] = asyncio.Queue(maxsize=1000)
         return self._event_queues[task_id]
     
     def remove_queue(self, task_id: str):
@@ -398,13 +402,43 @@ class EventManager:
         task_id: str,
         after_sequence: int = 0,
     ) -> AsyncGenerator[Dict, None]:
-        """流式获取事件"""
-        queue = self.create_queue(task_id)
+        """流式获取事件
         
-        # 先发送历史事件
-        history = await self.get_events(task_id, after_sequence)
-        for event in history:
-            yield event
+        🔥 重要: 此方法会先排空队列中已缓存的事件（在 SSE 连接前产生的），
+        然后继续实时推送新事件。
+        """
+        # 获取现有队列（由 AgentRunner 在初始化时创建）
+        queue = self._event_queues.get(task_id)
+        
+        if not queue:
+            # 如果队列不存在，创建一个新的（回退逻辑）
+            queue = self.create_queue(task_id)
+            logger.warning(f"Queue not found for task {task_id}, created new one")
+        
+        # 🔥 先排空队列中已缓存的事件（这些是在 SSE 连接前产生的）
+        buffered_count = 0
+        while not queue.empty():
+            try:
+                buffered_event = queue.get_nowait()
+                buffered_count += 1
+                yield buffered_event
+                
+                # 🔥 为所有缓存事件添加延迟，确保不会一起输出
+                event_type = buffered_event.get("event_type")
+                if event_type == "thinking_token":
+                    await asyncio.sleep(0.015)  # 15ms for tokens
+                else:
+                    await asyncio.sleep(0.005)  # 5ms for other events
+                
+                # 检查是否是结束事件
+                if event_type in ["task_complete", "task_error", "task_cancel"]:
+                    logger.info(f"Task {task_id} already completed, sent {buffered_count} buffered events")
+                    return
+            except asyncio.QueueEmpty:
+                break
+        
+        if buffered_count > 0:
+            logger.info(f"Drained {buffered_count} buffered events for task {task_id}")
         
         # 然后实时推送新事件
         try:
@@ -413,6 +447,10 @@ class EventManager:
                     event = await asyncio.wait_for(queue.get(), timeout=30)
                     yield event
                     
+                    # 🔥 为 thinking_token 添加微延迟确保流式效果
+                    if event.get("event_type") == "thinking_token":
+                        await asyncio.sleep(0.01)  # 10ms
+                    
                     # 检查是否是结束事件
                     if event.get("event_type") in ["task_complete", "task_error", "task_cancel"]:
                         break
@@ -421,8 +459,10 @@ class EventManager:
                     # 发送心跳
                     yield {"event_type": "heartbeat", "timestamp": datetime.now(timezone.utc).isoformat()}
                     
-        finally:
-            self.remove_queue(task_id)
+        except GeneratorExit:
+            # SSE 连接断开
+            logger.debug(f"SSE stream closed for task {task_id}")
+        # 🔥 不要移除队列，让 AgentRunner 管理队列的生命周期
     
     def create_emitter(self, task_id: str) -> AgentEventEmitter:
         """创建事件发射器"""

backend/app/services/agent/graph/runner.py

@@ -72,6 +72,10 @@ class AgentRunner:
         self.event_manager = EventManager(db_session_factory=async_session_factory)
         self.event_emitter = AgentEventEmitter(task.id, self.event_manager)
         
+        # 🔥 CRITICAL: 立即创建事件队列，确保在 Agent 开始执行前队列就存在
+        # 这样即使前端 SSE 连接稍晚，token 事件也不会丢失
+        self.event_manager.create_queue(task.id)
+        
         # 🔥 LLM 服务 - 使用用户配置（从系统配置页面获取）
         self.llm_service = LLMService(user_config=self.user_config)
         
@@ -708,6 +712,11 @@ class AgentRunner:
         
         for finding in findings:
             try:
+                # 确保 finding 是字典
+                if not isinstance(finding, dict):
+                    logger.warning(f"Skipping invalid finding (not a dict): {finding}")
+                    continue
+                    
                 db_finding = AgentFinding(
                     id=str(uuid.uuid4()),
                     task_id=self.task.id,

frontend/src/hooks/useAgentStream.ts

@@ -12,7 +12,7 @@ import {
   AgentStreamState,
 } from '../shared/api/agentStream';
 
-export interface UseAgentStreamOptions extends Omit<StreamOptions, 'onEvent'> {
+export interface UseAgentStreamOptions extends StreamOptions {
   autoConnect?: boolean;
   maxEvents?: number;
 }
@@ -73,6 +73,10 @@ export function useAgentStream(
     ...callbackOptions
   } = options;
 
+  // 🔥 使用 ref 存储 callback options，避免 connect 函数依赖变化导致重连
+  const callbackOptionsRef = useRef(callbackOptions);
+  callbackOptionsRef.current = callbackOptions;
+
   // 状态
   const [events, setEvents] = useState<StreamEventData[]>([]);
   const [thinking, setThinking] = useState('');
@@ -117,6 +121,15 @@ export function useAgentStream(
       afterSequence,
 
       onEvent: (event) => {
+        // Pass to custom callback first (important for capturing metadata like agent_name)
+        callbackOptionsRef.current.onEvent?.(event);
+
+        // 忽略 thinking 事件，防止污染日志列表 (它们会通过 onThinking* 回调单独处理)
+        if (
+          event.type === 'thinking_token' ||
+          event.type === 'thinking_start' ||
+          event.type === 'thinking_end'
+        ) return;
         setEvents((prev) => [...prev.slice(-maxEvents + 1), event]);
       },
 
@@ -124,20 +137,20 @@ export function useAgentStream(
         thinkingBufferRef.current = [];
         setIsThinking(true);
         setThinking('');
-        callbackOptions.onThinkingStart?.();
+        callbackOptionsRef.current.onThinkingStart?.();
       },
 
       onThinkingToken: (token, accumulated) => {
         thinkingBufferRef.current.push(token);
         setThinking(accumulated);
-        callbackOptions.onThinkingToken?.(token, accumulated);
+        callbackOptionsRef.current.onThinkingToken?.(token, accumulated);
       },
 
       onThinkingEnd: (response) => {
         setIsThinking(false);
         setThinking(response);
         thinkingBufferRef.current = [];
-        callbackOptions.onThinkingEnd?.(response);
+        callbackOptionsRef.current.onThinkingEnd?.(response);
       },
 
       onToolStart: (name, input) => {
@@ -145,7 +158,7 @@ export function useAgentStream(
           ...prev,
           { name, input, status: 'running' as const },
         ]);
-        callbackOptions.onToolStart?.(name, input);
+        callbackOptionsRef.current.onToolStart?.(name, input);
       },
 
       onToolEnd: (name, output, durationMs) => {
@@ -156,16 +169,16 @@ export function useAgentStream(
               : tc
           )
         );
-        callbackOptions.onToolEnd?.(name, output, durationMs);
+        callbackOptionsRef.current.onToolEnd?.(name, output, durationMs);
       },
 
       onNodeStart: (nodeName, phase) => {
         setCurrentPhase(phase);
-        callbackOptions.onNodeStart?.(nodeName, phase);
+        callbackOptionsRef.current.onNodeStart?.(nodeName, phase);
       },
 
       onNodeEnd: (nodeName, summary) => {
-        callbackOptions.onNodeEnd?.(nodeName, summary);
+        callbackOptionsRef.current.onNodeEnd?.(nodeName, summary);
       },
 
       onProgress: (current, total, message) => {
@@ -174,35 +187,35 @@ export function useAgentStream(
           total,
           percentage: total > 0 ? Math.round((current / total) * 100) : 0,
         });
-        callbackOptions.onProgress?.(current, total, message);
+        callbackOptionsRef.current.onProgress?.(current, total, message);
       },
 
       onFinding: (finding, isVerified) => {
         setFindings((prev) => [...prev, finding]);
-        callbackOptions.onFinding?.(finding, isVerified);
+        callbackOptionsRef.current.onFinding?.(finding, isVerified);
       },
 
       onComplete: (data) => {
         setIsComplete(true);
         setIsConnected(false);
-        callbackOptions.onComplete?.(data);
+        callbackOptionsRef.current.onComplete?.(data);
       },
 
       onError: (err) => {
         setError(err);
         setIsComplete(true);
         setIsConnected(false);
-        callbackOptions.onError?.(err);
+        callbackOptionsRef.current.onError?.(err);
       },
 
       onHeartbeat: () => {
-        callbackOptions.onHeartbeat?.();
+        callbackOptionsRef.current.onHeartbeat?.();
       },
     });
 
     handlerRef.current.connect();
     setIsConnected(true);
-  }, [taskId, includeThinking, includeToolCalls, afterSequence, maxEvents, callbackOptions]);
+  }, [taskId, includeThinking, includeToolCalls, afterSequence, maxEvents]); // 🔥 移除 callbackOptions 依赖
 
   // 断开连接
   const disconnect = useCallback(() => {
@@ -277,4 +290,3 @@ export function useAgentToolCalls(taskId: string | null) {
 }
 
 export default useAgentStream;
-

frontend/src/shared/api/agentStream.ts

@@ -10,6 +10,7 @@
 // 事件类型定义
 export type StreamEventType =
   // LLM 相关
+  | 'thinking' // General thinking event
   | 'thinking_start'
   | 'thinking_token'
   | 'thinking_end'
@@ -19,6 +20,8 @@ export type StreamEventType =
   | 'tool_call_output'
   | 'tool_call_end'
   | 'tool_call_error'
+  | 'tool_call'      // Backend sends this
+  | 'tool_result'    // Backend sends this
   // 节点相关
   | 'node_start'
   | 'node_end'
@@ -69,6 +72,12 @@ export interface StreamEventData {
   error?: string;           // task_error
   findings_count?: number;  // task_complete
   security_score?: number;  // task_complete
+  // Backend tool event fields
+  tool_name?: string;       // tool_call, tool_result
+  tool_input?: Record<string, unknown>;  // tool_call
+  tool_output?: unknown;    // tool_result
+  tool_duration_ms?: number; // tool_result
+  agent_name?: string;      // Extracted from metadata
 }
 
 // 事件回调类型
@@ -202,8 +211,13 @@ export class AgentStreamHandler {
         const events = this.parseSSE(buffer);
         buffer = events.remaining;
 
+        // 🔥 逐个处理事件，添加微延迟确保 React 能逐个渲染
         for (const event of events.parsed) {
           this.handleEvent(event);
+          // 为 thinking_token 添加微延迟确保打字效果
+          if (event.type === 'thinking_token') {
+            await new Promise(resolve => setTimeout(resolve, 5));
+          }
         }
       }
 
@@ -294,6 +308,11 @@ export class AgentStreamHandler {
    * 处理事件
    */
   private handleEvent(event: StreamEventData): void {
+    // Extract agent_name from metadata if present
+    if (event.metadata?.agent_name && !event.agent_name) {
+      event.agent_name = event.metadata.agent_name as string;
+    }
+
     // 通用回调
     this.options.onEvent?.(event);
 
@@ -306,17 +325,21 @@ export class AgentStreamHandler {
         break;
 
       case 'thinking_token':
-        if (event.token) {
-          this.thinkingBuffer.push(event.token);
+        // 兼容处理：token 可能在顶层，也可能在 metadata 中
+        const token = event.token || (event.metadata?.token as string);
+        const accumulated = event.accumulated || (event.metadata?.accumulated as string);
+
+        if (token) {
+          this.thinkingBuffer.push(token);
           this.options.onThinkingToken?.(
-            event.token,
-            event.accumulated || this.thinkingBuffer.join('')
+            token,
+            accumulated || this.thinkingBuffer.join('')
           );
         }
         break;
 
       case 'thinking_end':
-        const fullResponse = event.accumulated || this.thinkingBuffer.join('');
+        const fullResponse = event.accumulated || (event.metadata?.accumulated as string) || this.thinkingBuffer.join('');
         this.thinkingBuffer = [];
         this.options.onThinkingEnd?.(fullResponse);
         break;
@@ -341,6 +364,22 @@ export class AgentStreamHandler {
         }
         break;
 
+      // Alternative event names (backend sends these)
+      case 'tool_call':
+        this.options.onToolStart?.(
+          event.tool_name || 'unknown',
+          event.tool_input || {}
+        );
+        break;
+
+      case 'tool_result':
+        this.options.onToolEnd?.(
+          event.tool_name || 'unknown',
+          event.tool_output,
+          event.tool_duration_ms || 0
+        );
+        break;
+
       // 节点
       case 'node_start':
         this.options.onNodeStart?.(
@@ -469,6 +508,7 @@ export interface AgentStreamState {
   events: StreamEventData[];
   thinking: string;
   isThinking: boolean;
+  thinkingAgent?: string; // Who is thinking
   toolCalls: Array<{
     name: string;
     input: Record<string, unknown>;
@@ -494,6 +534,7 @@ export function createAgentStreamWithState(
     events: [],
     thinking: '',
     isThinking: false,
+    thinkingAgent: undefined,
     toolCalls: [],
     currentPhase: '',
     progress: { current: 0, total: 100, percentage: 0 },
@@ -509,9 +550,16 @@ export function createAgentStreamWithState(
 
   return new AgentStreamHandler(taskId, {
     onEvent: (event) => {
-      updateState({
-        events: [...state.events, event].slice(-500), // 保留最近 500 条
-      });
+      const updates: Partial<AgentStreamState> = {
+        events: [...state.events, event].slice(-500),
+      };
+
+      // Update thinking agent if available
+      if (event.agent_name && (event.type === 'thinking' || event.type === 'thinking_start' || event.type === 'thinking_token')) {
+        updates.thinkingAgent = event.agent_name;
+      }
+
+      updateState(updates);
     },
     onThinkingStart: () => {
       updateState({ isThinking: true, thinking: '' });