深入Dify工作流模块：可视化业务流程引擎架构与实现

graph TB
    subgraph "工作流引擎层"
        GraphEngine[GraphEngine 图执行引擎]
        WorkflowEntry[WorkflowEntry 工作流入口]
        ThreadPool[ThreadPool 线程池]
    end
    
    subgraph "图结构层"
        Graph[Graph 图结构]
        GraphEdge[GraphEdge 图边]
        GraphParallel[GraphParallel 并行分支]
    end
    
    subgraph "节点系统层"
        BaseNode[BaseNode 基础节点]
        
        subgraph "控制节点"
            StartNode[StartNode 开始节点]
            EndNode[EndNode 结束节点]
            IfElseNode[IfElseNode 条件分支]
            LoopNode[LoopNode 循环节点]
            IterationNode[IterationNode 迭代节点]
        end
        
        subgraph "处理节点"
            LLMNode[LLMNode 语言模型节点]
            ToolNode[ToolNode 工具节点]
            CodeNode[CodeNode 代码执行节点]
            KnowledgeRetrievalNode[KnowledgeRetrievalNode 知识检索节点]
            HTTPRequestNode[HTTPRequestNode HTTP请求节点]
        end
        
        subgraph "数据节点"
            VariableAssignerNode[VariableAssignerNode 变量赋值节点]
            VariableAggregatorNode[VariableAggregatorNode 变量聚合节点]
            TemplateTransformNode[TemplateTransformNode 模板转换节点]
            ParameterExtractorNode[ParameterExtractorNode 参数提取节点]
        end
        
        subgraph "输出节点"
            AnswerNode[AnswerNode 答案节点]
            QuestionClassifierNode[QuestionClassifierNode 问题分类节点]
            DocumentExtractorNode[DocumentExtractorNode 文档提取节点]
            ListOperatorNode[ListOperatorNode 列表操作节点]
        end
    end
    
    subgraph "状态管理层"
        GraphRuntimeState[GraphRuntimeState 图运行时状态]
        VariablePool[VariablePool 变量池]
        NodeExecutionState[NodeExecutionState 节点执行状态]
    end
    
    subgraph "事件系统层"
        EventBus[EventBus 事件总线]
        GraphEngineEvent[GraphEngineEvent 图引擎事件]
        NodeEvent[NodeEvent 节点事件]
        StreamProcessor[StreamProcessor 流处理器]
    end
    
    WorkflowEntry --> GraphEngine
    GraphEngine --> Graph
    GraphEngine --> ThreadPool
    
    Graph --> GraphEdge
    Graph --> GraphParallel
    
    BaseNode --> StartNode
    BaseNode --> LLMNode
    BaseNode --> ToolNode
    BaseNode --> AnswerNode
    
    GraphEngine --> GraphRuntimeState
    GraphRuntimeState --> VariablePool
    GraphRuntimeState --> NodeExecutionState
    
    GraphEngine --> EventBus
    EventBus --> GraphEngineEvent
    EventBus --> NodeEvent
    
    style GraphEngine fill:#e3f2fd
    style BaseNode fill:#e8f5e8
    style GraphRuntimeState fill:#fff3e0
    style EventBus fill:#fce4ec

sequenceDiagram
    participant Client as 客户端
    participant WorkflowEntry as 工作流入口
    participant GraphEngine as 图执行引擎
    participant NodeFactory as 节点工厂
    participant BaseNode as 节点实例
    participant VariablePool as 变量池
    participant EventBus as 事件总线
    
    Note over Client,EventBus: 工作流执行完整流程
    
    Client->>WorkflowEntry: 启动工作流执行
    WorkflowEntry->>GraphEngine: 初始化图执行引擎
    GraphEngine->>GraphEngine: 加载工作流图结构
    
    loop 节点执行循环
        GraphEngine->>NodeFactory: 创建节点实例
        NodeFactory->>BaseNode: 实例化具体节点
        
        GraphEngine->>VariablePool: 获取节点输入变量
        VariablePool-->>GraphEngine: 返回变量值
        
        GraphEngine->>BaseNode: 执行节点逻辑
        BaseNode->>BaseNode: 处理节点业务逻辑
        
        alt 节点执行成功
            BaseNode-->>GraphEngine: 返回执行结果
            GraphEngine->>VariablePool: 更新输出变量
            GraphEngine->>EventBus: 发布成功事件
        else 节点执行失败
            BaseNode-->>GraphEngine: 返回错误信息
            GraphEngine->>EventBus: 发布失败事件
            
            alt 有错误处理策略
                GraphEngine->>GraphEngine: 执行错误分支
            else 无错误处理
                GraphEngine->>GraphEngine: 终止工作流
            end
        end
        
        GraphEngine->>GraphEngine: 计算下一个节点
        
        alt 有下一个节点
            Note over GraphEngine: 继续执行下一节点
        else 到达结束节点
            Note over GraphEngine: 工作流执行完成
        end
    end
    
    GraphEngine-->>WorkflowEntry: 返回执行结果
    WorkflowEntry-->>Client: 返回最终结果

class GraphEngine:
    """
    图执行引擎
    负责工作流图的解析、调度和执行
    """
    
    # 全局线程池映射，支持多个工作流并发执行
    workflow_thread_pool_mapping: dict[str, GraphEngineThreadPool] = {}

    def __init__(
        self,
        tenant_id: str,
        app_id: str,
        workflow_type: WorkflowType,
        workflow_id: str,
        user_id: str,
        user_from: UserFrom,
        invoke_from: InvokeFrom,
        call_depth: int,
        graph: Graph,
        graph_config: Mapping[str, Any],
        graph_runtime_state: GraphRuntimeState,
        max_execution_steps: int,
        max_execution_time: int,
        thread_pool_id: Optional[str] = None,
    ):
        """
        初始化图执行引擎
        
        Args:
            tenant_id: 租户ID
            app_id: 应用ID
            workflow_type: 工作流类型
            workflow_id: 工作流ID
            user_id: 用户ID
            user_from: 用户来源
            invoke_from: 调用来源
            call_depth: 调用深度（用于防止无限递归）
            graph: 图结构实例
            graph_config: 图配置
            graph_runtime_state: 图运行时状态
            max_execution_steps: 最大执行步数
            max_execution_time: 最大执行时间（秒）
            thread_pool_id: 可选的线程池ID，用于复用线程池
        """
        # 线程池配置
        thread_pool_max_submit_count = dify_config.MAX_SUBMIT_COUNT
        thread_pool_max_workers = 10

        # 初始化线程池
        if thread_pool_id:
            # 复用现有线程池
            if thread_pool_id not in GraphEngine.workflow_thread_pool_mapping:
                raise ValueError(f"线程池 {thread_pool_id} 不存在")
            
            self.thread_pool_id = thread_pool_id
            self.thread_pool = GraphEngine.workflow_thread_pool_mapping[thread_pool_id]
            self.is_main_thread_pool = False
        else:
            # 创建新的线程池
            self.thread_pool = GraphEngineThreadPool(
                max_workers=thread_pool_max_workers,
                max_submit_count=thread_pool_max_submit_count
            )
            self.thread_pool_id = str(uuid.uuid4())
            self.is_main_thread_pool = True
            GraphEngine.workflow_thread_pool_mapping[self.thread_pool_id] = self.thread_pool

        # 设置核心属性
        self.graph = graph
        self.init_params = GraphInitParams(
            tenant_id=tenant_id,
            app_id=app_id,
            workflow_type=workflow_type,
            workflow_id=workflow_id,
            graph_config=graph_config,
            user_id=user_id,
            user_from=user_from,
            invoke_from=invoke_from,
            call_depth=call_depth,
        )
        
        self.graph_runtime_state = graph_runtime_state
        self.max_execution_steps = max_execution_steps
        self.max_execution_time = max_execution_time

    def run(self) -> Generator[GraphEngineEvent, None, None]:
        """
        运行图执行引擎
        从根节点开始执行整个工作流图
        
        Yields:
            GraphEngineEvent: 图执行过程中的各种事件
        """
        try:
            # 发布图开始执行事件
            yield GraphRunStartedEvent(
                graph_run_id=self.graph_runtime_state.graph_run_id,
                graph_id=self.graph.root_node_id,
                started_at=naive_utc_now()
            )

            # 从根节点开始执行
            yield from self._run(start_node_id=self.graph.root_node_id)

            # 发布图执行成功事件
            yield GraphRunSucceededEvent(
                graph_run_id=self.graph_runtime_state.graph_run_id,
                graph_id=self.graph.root_node_id,
                outputs=self.graph_runtime_state.variable_pool.to_dict(),
                succeeded_at=naive_utc_now()
            )

        except GraphRunFailedError as e:
            # 图执行失败
            logger.error(f"工作流图执行失败: {e}")
            yield GraphRunFailedEvent(
                graph_run_id=self.graph_runtime_state.graph_run_id,
                graph_id=self.graph.root_node_id,
                error=str(e),
                failed_at=naive_utc_now()
            )
            raise e

        except Exception as e:
            # 未预期的异常
            logger.exception(f"工作流图执行出现未预期异常: {e}")
            yield GraphRunFailedEvent(
                graph_run_id=self.graph_runtime_state.graph_run_id,
                graph_id=self.graph.root_node_id,
                error=f"未预期异常: {str(e)}",
                failed_at=naive_utc_now()
            )
            raise e

        finally:
            # 清理资源
            if self.is_main_thread_pool:
                self._release_thread()

    def _run(
        self,
        start_node_id: str,
        in_parallel_id: Optional[str] = None,
        parent_parallel_id: Optional[str] = None,
        parent_parallel_start_node_id: Optional[str] = None,
        handle_exceptions: list[str] = [],
    ) -> Generator[GraphEngineEvent, None, None]:
        """
        内部执行方法
        实现图的深度优先遍历执行
        
        Args:
            start_node_id: 起始节点ID
            in_parallel_id: 当前并行分支ID
            parent_parallel_id: 父并行分支ID
            parent_parallel_start_node_id: 父并行起始节点ID
            handle_exceptions: 处理的异常列表
            
        Yields:
            GraphEngineEvent: 执行事件
        """
        parallel_start_node_id = None
        if in_parallel_id:
            parallel_start_node_id = start_node_id

        next_node_id = start_node_id
        previous_route_node_state: Optional[RouteNodeState] = None

        # 主执行循环
        while True:
            # 检查最大执行步数限制
            if self.graph_runtime_state.node_run_steps > self.max_execution_steps:
                raise GraphRunFailedError(f"超过最大执行步数 {self.max_execution_steps}")

            # 检查最大执行时间限制
            if self._is_timed_out(
                start_at=self.graph_runtime_state.start_at,
                max_execution_time=self.max_execution_time
            ):
                raise GraphRunFailedError(f"超过最大执行时间 {self.max_execution_time}秒")

            # 创建节点路由状态
            route_node_state = self.graph_runtime_state.node_run_state.create_node_state(
                node_id=next_node_id
            )

            # 获取节点配置
            node_id = route_node_state.node_id
            node_config = self.graph.node_id_config_mapping.get(node_id)
            if not node_config:
                raise GraphRunFailedError(f"节点 {node_id} 配置未找到")

            # 确定节点类型和版本
            node_type = NodeType(node_config.get("data", {}).get("type"))
            node_version = node_config.get("data", {}).get("version", "1")

            # 从节点映射中获取节点类
            from core.workflow.nodes.node_mapping import NODE_TYPE_CLASSES_MAPPING
            node_cls = NODE_TYPE_CLASSES_MAPPING[node_type][node_version]

            previous_node_id = (
                previous_route_node_state.node_id 
                if previous_route_node_state else None
            )

            # 创建节点实例
            node_instance = node_cls(
                id=node_id,
                config=node_config,
                graph_init_params=self.init_params,
                graph=self.graph,
                graph_runtime_state=self.graph_runtime_state,
                previous_node_id=previous_node_id,
                thread_pool_id=self.thread_pool_id,
            )

            # 初始化节点数据
            node_instance.init_node_data(node_config.get("data", {}))

            # 执行节点
            yield from self._run_node(
                node_instance=node_instance,
                route_node_state=route_node_state,
                in_parallel_id=in_parallel_id,
                parent_parallel_id=parent_parallel_id,
                parent_parallel_start_node_id=parent_parallel_start_node_id,
            )

            # 获取执行结果
            node_run_result = route_node_state.node_run_result
            if not node_run_result:
                raise GraphRunFailedError(f"节点 {node_id} 执行结果为空")

            # 处理执行状态
            if node_run_result.status == WorkflowNodeExecutionStatus.FAILED:
                # 节点执行失败，检查是否有错误处理策略
                if self._should_handle_error(node_instance, handle_exceptions):
                    # 执行错误处理分支
                    yield from self._handle_node_error(
                        node_instance=node_instance,
                        error=node_run_result.error,
                        route_node_state=route_node_state
                    )
                else:
                    # 没有错误处理，抛出异常
                    raise GraphRunFailedError(
                        f"节点 {node_id} 执行失败: {node_run_result.error}"
                    )

            # 确定下一个节点
            next_node_candidates = self._get_next_node_candidates(
                node_instance=node_instance,
                route_node_state=route_node_state
            )

            if not next_node_candidates:
                # 没有下一个节点，执行结束
                break

            # 处理多个候选节点（并行执行）
            if len(next_node_candidates) > 1:
                yield from self._handle_parallel_execution(
                    candidates=next_node_candidates,
                    current_node_id=node_id,
                    in_parallel_id=in_parallel_id
                )
                break
            else:
                # 单个下一节点，继续串行执行
                next_node_id = next_node_candidates[0]
                previous_route_node_state = route_node_state

    def _run_node(
        self,
        node_instance: BaseNode,
        route_node_state: RouteNodeState,
        in_parallel_id: Optional[str] = None,
        parent_parallel_id: Optional[str] = None,
        parent_parallel_start_node_id: Optional[str] = None,
    ) -> Generator[GraphEngineEvent, None, None]:
        """
        执行单个节点
        包含节点的完整生命周期管理
        
        Args:
            node_instance: 节点实例
            route_node_state: 路由节点状态
            in_parallel_id: 并行分支ID
            parent_parallel_id: 父并行分支ID
            parent_parallel_start_node_id: 父并行起始节点ID
            
        Yields:
            GraphEngineEvent: 节点执行事件
        """
        node_id = node_instance.node_id
        node_type = node_instance._node_type
        
        # 发布节点开始执行事件
        yield NodeRunStartedEvent(
            node_execution_id=route_node_state.node_execution_id,
            node_id=node_id,
            node_type=node_type,
            parallel_id=in_parallel_id,
            parallel_start_node_id=parent_parallel_start_node_id,
            started_at=naive_utc_now()
        )

        try:
            # 准备节点输入
            node_inputs = self._prepare_node_inputs(node_instance)
            
            # 执行节点逻辑
            node_run_generator = node_instance.run(
                node_inputs=node_inputs
            )

            # 处理节点执行过程中的事件
            for node_event in node_run_generator:
                if isinstance(node_event, RunStreamChunkEvent):
                    # 流式输出事件
                    yield NodeRunStreamChunkEvent(
                        node_execution_id=route_node_state.node_execution_id,
                        node_id=node_id,
                        node_type=node_type,
                        chunk_content=node_event.chunk,
                        parallel_id=in_parallel_id,
                        parallel_start_node_id=parent_parallel_start_node_id,
                    )
                
                elif isinstance(node_event, RunRetrieverResourceEvent):
                    # 检索资源事件
                    yield NodeRunRetrieverResourceEvent(
                        node_execution_id=route_node_state.node_execution_id,
                        node_id=node_id,
                        node_type=node_type,
                        retriever_resources=node_event.retriever_resources,
                        parallel_id=in_parallel_id,
                        parallel_start_node_id=parent_parallel_start_node_id,
                    )
                
                elif isinstance(node_event, RunCompletedEvent):
                    # 节点执行完成事件
                    route_node_state.node_run_result = node_event.run_result
                    
                    # 更新变量池
                    if node_event.run_result.outputs:
                        self.graph_runtime_state.variable_pool.add(
                            variables=node_event.run_result.outputs
                        )
                    
                    # 发布节点成功事件
                    yield NodeRunSucceededEvent(
                        node_execution_id=route_node_state.node_execution_id,
                        node_id=node_id,
                        node_type=node_type,
                        inputs=node_event.run_result.inputs or {},
                        outputs=node_event.run_result.outputs or {},
                        execution_metadata=node_event.run_result.metadata or {},
                        succeeded_at=naive_utc_now(),
                        parallel_id=in_parallel_id,
                        parallel_start_node_id=parent_parallel_start_node_id,
                    )
                    break

            # 增加执行步数计数
            self.graph_runtime_state.node_run_steps += 1

        except NodeRunFailedException as e:
            # 节点执行失败
            route_node_state.node_run_result = NodeRunResult(
                status=WorkflowNodeExecutionStatus.FAILED,
                error=str(e)
            )
            
            yield NodeRunFailedEvent(
                node_execution_id=route_node_state.node_execution_id,
                node_id=node_id,
                node_type=node_type,
                error=str(e),
                failed_at=naive_utc_now(),
                parallel_id=in_parallel_id,
                parallel_start_node_id=parent_parallel_start_node_id,
            )

        except Exception as e:
            # 未预期异常
            logger.exception(f"节点 {node_id} 执行时发生未预期异常")
            
            route_node_state.node_run_result = NodeRunResult(
                status=WorkflowNodeExecutionStatus.FAILED,
                error=f"未预期异常: {str(e)}"
            )
            
            yield NodeRunExceptionEvent(
                node_execution_id=route_node_state.node_execution_id,
                node_id=node_id,
                node_type=node_type,
                exception=e,
                failed_at=naive_utc_now(),
                parallel_id=in_parallel_id,
                parallel_start_node_id=parent_parallel_start_node_id,
            )

    def _prepare_node_inputs(self, node_instance: BaseNode) -> dict[str, Any]:
        """
        准备节点输入变量
        从变量池中提取节点所需的输入变量
        
        Args:
            node_instance: 节点实例
            
        Returns:
            dict[str, Any]: 节点输入变量字典
        """
        node_inputs = {}
        
        # 获取节点的输入变量配置
        if hasattr(node_instance, '_node_data') and hasattr(node_instance._node_data, 'inputs'):
            for input_config in node_instance._node_data.inputs:
                variable_selector = input_config.variable_selector
                
                try:
                    # 从变量池获取变量值
                    variable_value = self.graph_runtime_state.variable_pool.get(
                        variable_selector
                    )
                    
                    node_inputs[input_config.variable] = variable_value
                    
                except VariableNotFoundError:
                    # 变量不存在，检查是否为必填
                    if input_config.required:
                        raise VariableNotFoundError(
                            f"节点 {node_instance.node_id} 的必填变量 "
                            f"{variable_selector} 未找到"
                        )
                    else:
                        # 非必填变量，使用默认值
                        node_inputs[input_config.variable] = input_config.default_value

        return node_inputs

    def _get_next_node_candidates(
        self,
        node_instance: BaseNode,
        route_node_state: RouteNodeState
    ) -> list[str]:
        """
        获取下一个节点候选列表
        根据节点执行结果和图结构确定后续节点
        
        Args:
            node_instance: 当前节点实例
            route_node_state: 路由节点状态
            
        Returns:
            list[str]: 下一个节点ID列表
        """
        node_id = node_instance.node_id
        node_run_result = route_node_state.node_run_result
        
        if not node_run_result:
            return []

        # 获取节点的出边
        outgoing_edges = self.graph.edge_mapping.get(node_id, [])
        
        if not outgoing_edges:
            # 没有出边，执行结束
            return []

        next_nodes = []
        
        for edge in outgoing_edges:
            # 检查边的条件（如果有）
            if self._should_follow_edge(edge, node_run_result):
                next_nodes.append(edge.target_node_id)

        return next_nodes

    def _should_follow_edge(self, edge: GraphEdge, node_run_result: NodeRunResult) -> bool:
        """
        判断是否应该遵循某条边
        评估边的条件表达式
        
        Args:
            edge: 图边对象
            node_run_result: 节点执行结果
            
        Returns:
            bool: 是否应该遵循该边
        """
        # 如果边没有条件，直接遵循
        if not edge.condition:
            return True

        try:
            # 使用条件管理器评估条件
            condition_manager = ConditionManager()
            
            return condition_manager.evaluate(
                condition=edge.condition,
                variable_pool=self.graph_runtime_state.variable_pool,
                node_outputs=node_run_result.outputs or {}
            )
            
        except Exception as e:
            logger.warning(f"边条件评估失败: {e}")
            return False

    def _handle_parallel_execution(
        self,
        candidates: list[str],
        current_node_id: str,
        in_parallel_id: Optional[str]
    ) -> Generator[GraphEngineEvent, None, None]:
        """
        处理并行执行
        当一个节点有多个后续节点时，创建并行分支执行
        
        Args:
            candidates: 候选节点ID列表
            current_node_id: 当前节点ID
            in_parallel_id: 当前并行分支ID
            
        Yields:
            GraphEngineEvent: 并行执行事件
        """
        parallel_id = str(uuid.uuid4())
        
        # 发布并行分支开始事件
        yield ParallelBranchRunStartedEvent(
            parallel_id=parallel_id,
            parallel_start_node_id=current_node_id,
            parallel_branch_nodes=candidates,
            started_at=naive_utc_now()
        )

        try:
            # 创建并行执行任务
            parallel_futures = []
            
            for candidate_node_id in candidates:
                # 为每个候选节点创建独立的执行线程
                future = self.thread_pool.submit(
                    preserve_flask_contexts(self._run_parallel_branch),
                    candidate_node_id,
                    parallel_id,
                    in_parallel_id,
                    current_node_id
                )
                
                # 添加完成回调
                future.add_done_callback(self.thread_pool.task_done_callback)
                parallel_futures.append(future)

            # 等待所有并行分支完成
            completed_results = []
            for future in parallel_futures:
                try:
                    branch_result = future.result()
                    completed_results.append(branch_result)
                except Exception as e:
                    logger.exception(f"并行分支执行失败: {e}")
                    # 发布并行分支失败事件
                    yield ParallelBranchRunFailedEvent(
                        parallel_id=parallel_id,
                        error=str(e),
                        failed_at=naive_utc_now()
                    )
                    raise e

            # 发布并行分支成功事件
            yield ParallelBranchRunSucceededEvent(
                parallel_id=parallel_id,
                results=completed_results,
                succeeded_at=naive_utc_now()
            )

        except Exception as e:
            logger.exception(f"并行执行失败: {e}")
            yield ParallelBranchRunFailedEvent(
                parallel_id=parallel_id,
                error=str(e),
                failed_at=naive_utc_now()
            )
            raise e

    def _run_parallel_branch(
        self,
        start_node_id: str,
        parallel_id: str,
        parent_parallel_id: Optional[str],
        parent_start_node_id: str
    ):
        """
        运行并行分支
        在独立线程中执行单个并行分支
        
        Args:
            start_node_id: 分支起始节点ID
            parallel_id: 并行分支ID
            parent_parallel_id: 父并行分支ID
            parent_start_node_id: 父起始节点ID
            
        Returns:
            并行分支执行结果
        """
        try:
            # 创建分支专用的状态副本
            branch_state = self._create_branch_state(parallel_id)
            
            # 在分支状态中执行
            branch_events = list(self._run(
                start_node_id=start_node_id,
                in_parallel_id=parallel_id,
                parent_parallel_id=parent_parallel_id,
                parent_parallel_start_node_id=parent_start_node_id
            ))
            
            return {
                "parallel_id": parallel_id,
                "start_node_id": start_node_id,
                "events": branch_events,
                "status": "success"
            }
            
        except Exception as e:
            logger.exception(f"并行分支 {parallel_id} 执行失败")
            return {
                "parallel_id": parallel_id,
                "start_node_id": start_node_id,
                "error": str(e),
                "status": "failed"
            }

    def _is_timed_out(self, start_at: float, max_execution_time: int) -> bool:
        """
        检查是否超时
        
        Args:
            start_at: 开始时间戳
            max_execution_time: 最大执行时间（秒）
            
        Returns:
            bool: 是否超时
        """
        if max_execution_time <= 0:
            return False
        
        current_time = time.time()
        elapsed_time = current_time - start_at
        
        return elapsed_time > max_execution_time

class GraphEngineThreadPool(ThreadPoolExecutor):
    """
    图执行引擎专用线程池
    扩展标准线程池，增加提交计数和资源控制
    """
    
    def __init__(
        self,
        max_workers=None,
        thread_name_prefix="",
        initializer=None,
        initargs=(),
        max_submit_count=dify_config.MAX_SUBMIT_COUNT,
    ):
        """
        初始化专用线程池
        
        Args:
            max_workers: 最大工作线程数
            thread_name_prefix: 线程名前缀
            initializer: 线程初始化器
            initargs: 初始化参数
            max_submit_count: 最大提交任务数
        """
        super().__init__(max_workers, thread_name_prefix, initializer, initargs)
        self.max_submit_count = max_submit_count
        self.submit_count = 0
        self._lock = threading.Lock()

    def submit(self, fn, /, *args, **kwargs):
        """
        提交任务到线程池
        增加提交计数和容量检查
        
        Args:
            fn: 要执行的函数
            *args: 位置参数
            **kwargs: 关键字参数
            
        Returns:
            Future: 任务Future对象
        """
        with self._lock:
            self.submit_count += 1
            self.check_is_full()

        future = super().submit(fn, *args, **kwargs)
        
        # 添加完成回调，用于减少计数
        future.add_done_callback(self.task_done_callback)
        
        return future

    def task_done_callback(self, future):
        """
        任务完成回调
        减少提交计数，释放资源
        
        Args:
            future: 完成的Future对象
        """
        with self._lock:
            self.submit_count -= 1

    def check_is_full(self):
        """
        检查线程池是否已满
        防止过多任务导致资源耗尽
        """
        if self.submit_count > self.max_submit_count:
            raise ValueError(
                f"线程池已达最大提交任务数 {self.max_submit_count}"
            )

    def get_current_load(self) -> float:
        """
        获取当前负载
        
        Returns:
            float: 当前负载比例 (0.0-1.0)
        """
        with self._lock:
            return self.submit_count / self.max_submit_count

class Graph(BaseModel):
    """
    工作流图结构
    表示节点和边构成的有向无环图
    """
    
    # 根节点ID
    root_node_id: str = Field(..., description="图的根节点ID")
    
    # 节点ID到配置的映射
    node_id_config_mapping: dict[str, Any] = Field(
        default_factory=dict, description="节点ID到配置的映射"
    )
    
    # 边映射：源节点ID -> 目标边列表
    edge_mapping: dict[str, list[GraphEdge]] = Field(
        default_factory=dict, description="边映射"
    )
    
    # 反向边映射：目标节点ID -> 源边列表
    reverse_edge_mapping: dict[str, list[GraphEdge]] = Field(
        default_factory=dict, description="反向边映射"
    )
    
    # 并行映射
    parallel_mapping: dict[str, GraphParallel] = Field(
        default_factory=dict, description="并行映射"
    )
    
    # 节点到并行的映射
    node_parallel_mapping: dict[str, str] = Field(
        default_factory=dict, description="节点到并行的映射"
    )

    @classmethod
    def init(
        cls, 
        graph_config: Mapping[str, Any], 
        root_node_id: Optional[str] = None
    ) -> "Graph":
        """
        从配置初始化图结构
        解析工作流配置并构建图数据结构
        
        Args:
            graph_config: 图配置字典
            root_node_id: 可选的根节点ID
            
        Returns:
            Graph: 初始化的图实例
        """
        # 提取节点配置
        nodes_config = graph_config.get("nodes", [])
        edges_config = graph_config.get("edges", [])

        # 构建节点映射
        all_node_id_config_mapping = {
            node_config.get("id"): node_config 
            for node_config in nodes_config
        }

        # 构建边映射
        edge_mapping = {}
        reverse_edge_mapping = {}
        
        for edge_config in edges_config:
            source_node_id = edge_config.get("source")
            target_node_id = edge_config.get("target")
            
            # 创建图边对象
            graph_edge = GraphEdge(
                source_node_id=source_node_id,
                target_node_id=target_node_id,
                condition=edge_config.get("condition")
            )
            
            # 添加到正向映射
            if source_node_id not in edge_mapping:
                edge_mapping[source_node_id] = []
            edge_mapping[source_node_id].append(graph_edge)
            
            # 添加到反向映射
            if target_node_id not in reverse_edge_mapping:
                reverse_edge_mapping[target_node_id] = []
            reverse_edge_mapping[target_node_id].append(graph_edge)

        # 确定根节点
        if not root_node_id:
            # 查找START类型的节点作为根节点
            root_node_configs = [
                node_config for node_config in nodes_config
                if node_config.get("data", {}).get("type") == NodeType.START.value
            ]
            
            if not root_node_configs:
                raise ValueError("未找到START类型的根节点")
            
            root_node_id = root_node_configs[0].get("id")

        # 验证根节点存在
        if root_node_id not in all_node_id_config_mapping:
            raise ValueError(f"根节点 {root_node_id} 在图中不存在")

        # 检查图的连通性
        cls._validate_graph_connectivity(
            root_node_id=root_node_id,
            edge_mapping=edge_mapping
        )

        # 获取所有可达节点
        reachable_node_ids = cls._get_reachable_nodes(
            root_node_id=root_node_id,
            edge_mapping=edge_mapping
        )

        # 构建可达节点的配置映射
        node_id_config_mapping = {
            node_id: all_node_id_config_mapping[node_id] 
            for node_id in reachable_node_ids
        }

        # 构建并行映射
        parallel_mapping = {}
        node_parallel_mapping = {}
        cls._build_parallel_mapping(
            edge_mapping=edge_mapping,
            reverse_edge_mapping=reverse_edge_mapping,
            start_node_id=root_node_id,
            parallel_mapping=parallel_mapping,
            node_parallel_mapping=node_parallel_mapping,
        )

        # 验证并行层次不超过限制
        cls._validate_parallel_depth(
            parallel_mapping=parallel_mapping,
            max_depth=dify_config.WORKFLOW_PARALLEL_DEPTH_LIMIT
        )

        return cls(
            root_node_id=root_node_id,
            node_id_config_mapping=node_id_config_mapping,
            edge_mapping=edge_mapping,
            reverse_edge_mapping=reverse_edge_mapping,
            parallel_mapping=parallel_mapping,
            node_parallel_mapping=node_parallel_mapping
        )

    @classmethod
    def _validate_graph_connectivity(
        cls,
        root_node_id: str,
        edge_mapping: dict[str, list[GraphEdge]]
    ):
        """
        验证图的连通性
        确保图中没有孤立节点和环路
        
        Args:
            root_node_id: 根节点ID
            edge_mapping: 边映射
        """
        # 使用深度优先搜索检查连通性
        visited = set()
        stack = [root_node_id]
        
        while stack:
            current_node = stack.pop()
            if current_node in visited:
                continue
            
            visited.add(current_node)
            
            # 添加后续节点到栈中
            for edge in edge_mapping.get(current_node, []):
                if edge.target_node_id not in visited:
                    stack.append(edge.target_node_id)

    @classmethod
    def _get_reachable_nodes(
        cls,
        root_node_id: str,
        edge_mapping: dict[str, list[GraphEdge]]
    ) -> list[str]:
        """
        获取从根节点可达的所有节点
        
        Args:
            root_node_id: 根节点ID
            edge_mapping: 边映射
            
        Returns:
            list[str]: 可达节点ID列表
        """
        reachable_nodes = [root_node_id]
        visited = {root_node_id}
        queue = [root_node_id]
        
        while queue:
            current_node = queue.pop(0)
            
            for edge in edge_mapping.get(current_node, []):
                target_node = edge.target_node_id
                if target_node not in visited:
                    visited.add(target_node)
                    reachable_nodes.append(target_node)
                    queue.append(target_node)
        
        return reachable_nodes

class GraphEdge(BaseModel):
    """
    图边表示
    连接两个节点，可包含条件表达式
    """
    
    # 源节点ID
    source_node_id: str
    
    # 目标节点ID
    target_node_id: str
    
    # 可选的条件表达式
    condition: Optional[str] = None
    
    # 边的元数据
    metadata: dict[str, Any] = Field(default_factory=dict)

class GraphParallel(BaseModel):
    """
    图并行分支表示
    定义并行执行的分支结构
    """
    
    # 并行分支ID
    parallel_id: str
    
    # 并行起始节点ID
    start_node_id: str
    
    # 并行分支节点列表
    branch_node_ids: list[str]
    
    # 父并行分支ID（支持嵌套并行）
    parent_parallel_id: Optional[str] = None
    
    # 并行策略
    strategy: ParallelStrategy = ParallelStrategy.ALL
    
    # 超时配置
    timeout: Optional[int] = None

class ParallelStrategy(Enum):
    """并行执行策略"""
    
    # 等待所有分支完成
    ALL = "all"
    
    # 等待任意一个分支完成
    ANY = "any" 
    
    # 等待指定数量的分支完成
    COUNT = "count"

class BaseNode:
    """
    工作流节点基类
    定义所有节点的通用接口和行为
    """
    
    # 节点类型（子类必须定义）
    _node_type: ClassVar[NodeType]

    def __init__(
        self,
        id: str,
        config: Mapping[str, Any],
        graph_init_params: "GraphInitParams",
        graph: "Graph",
        graph_runtime_state: "GraphRuntimeState",
        previous_node_id: Optional[str] = None,
        thread_pool_id: Optional[str] = None,
    ):
        """
        初始化节点基类
        
        Args:
            id: 节点唯一标识符
            config: 节点配置字典
            graph_init_params: 图初始化参数
            graph: 图结构实例
            graph_runtime_state: 图运行时状态
            previous_node_id: 前置节点ID
            thread_pool_id: 线程池ID
        """
        # 基础标识信息
        self.id = id
        self.node_id = config.get("id")
        
        # 租户和应用信息
        self.tenant_id = graph_init_params.tenant_id
        self.app_id = graph_init_params.app_id
        self.workflow_type = graph_init_params.workflow_type
        self.workflow_id = graph_init_params.workflow_id
        
        # 用户信息
        self.user_id = graph_init_params.user_id
        self.user_from = graph_init_params.user_from
        self.invoke_from = graph_init_params.invoke_from
        
        # 执行上下文
        self.workflow_call_depth = graph_init_params.call_depth
        self.graph = graph
        self.graph_runtime_state = graph_runtime_state
        self.previous_node_id = previous_node_id
        self.thread_pool_id = thread_pool_id
        self.graph_config = graph_init_params.graph_config

        # 验证节点ID
        if not self.node_id:
            raise ValueError("节点ID是必需的")

    @abstractmethod
    def init_node_data(self, data: Mapping[str, Any]):
        """
        初始化节点数据
        子类必须实现，用于解析和验证节点特定的配置
        
        Args:
            data: 节点数据字典
        """
        raise NotImplementedError("子类必须实现init_node_data方法")

    @abstractmethod
    def run(self, node_inputs: Mapping[str, Any]) -> Generator[NodeEvent, None, None]:
        """
        执行节点逻辑
        子类必须实现，定义节点的具体执行逻辑
        
        Args:
            node_inputs: 节点输入变量
            
        Yields:
            NodeEvent: 节点执行过程中的事件
        """
        raise NotImplementedError("子类必须实现run方法")

    def _get_error_strategy(self) -> Optional[ErrorStrategy]:
        """
        获取错误处理策略
        子类可以覆盖此方法以定义特定的错误处理策略
        
        Returns:
            Optional[ErrorStrategy]: 错误处理策略
        """
        return None

    def _get_retry_config(self) -> RetryConfig:
        """
        获取重试配置
        子类可以覆盖此方法以定义特定的重试策略
        
        Returns:
            RetryConfig: 重试配置
        """
        return RetryConfig()

    def get_variable(self, variable_selector: VariableSelector) -> Any:
        """
        从变量池获取变量值
        
        Args:
            variable_selector: 变量选择器
            
        Returns:
            Any: 变量值
        """
        return self.graph_runtime_state.variable_pool.get(variable_selector)

    def set_variable(self, variable_key: str, variable_value: Any):
        """
        设置变量到变量池
        
        Args:
            variable_key: 变量键
            variable_value: 变量值
        """
        self.graph_runtime_state.variable_pool.add_variable(
            node_id=self.node_id,
            variable_key=variable_key,
            variable_value=variable_value
        )

    def create_node_run_result(
        self,
        status: WorkflowNodeExecutionStatus,
        inputs: Optional[dict[str, Any]] = None,
        outputs: Optional[dict[str, Any]] = None,
        metadata: Optional[dict[str, Any]] = None,
        error: Optional[str] = None
    ) -> NodeRunResult:
        """
        创建节点运行结果
        
        Args:
            status: 执行状态
            inputs: 输入变量
            outputs: 输出变量
            metadata: 元数据
            error: 错误信息
            
        Returns:
            NodeRunResult: 节点运行结果
        """
        return NodeRunResult(
            status=status,
            inputs=inputs or {},
            outputs=outputs or {},
            metadata=metadata or {},
            error=error
        )

class BaseNodeData(BaseModel):
    """
    节点数据基类
    定义所有节点配置的通用字段
    """
    
    # 节点标题
    title: str
    
    # 节点描述
    desc: Optional[str] = None
    
    # 错误处理策略
    error_strategy: Optional[ErrorStrategy] = None
    
    # 重试配置
    retry_config: RetryConfig = Field(default_factory=RetryConfig)
    
    # 自定义配置
    custom_config: dict[str, Any] = Field(default_factory=dict)

class RetryConfig(BaseModel):
    """重试配置"""
    
    # 是否启用重试
    enabled: bool = False
    
    # 最大重试次数
    max_attempts: int = 3
    
    # 重试间隔（秒）
    retry_interval: int = 1
    
    # 重试策略
    strategy: RetryStrategy = RetryStrategy.FIXED_INTERVAL

class RetryStrategy(Enum):
    """重试策略枚举"""
    
    # 固定间隔
    FIXED_INTERVAL = "fixed_interval"
    
    # 指数退避
    EXPONENTIAL_BACKOFF = "exponential_backoff"
    
    # 线性退避
    LINEAR_BACKOFF = "linear_backoff"

# 基于实际使用场景的节点分类
NODE_CATEGORY_MATRIX = {
    "输入输出控制": {
        "start": "工作流入口，初始化变量和上下文",
        "end": "工作流出口，收集最终结果",
        "answer": "用户可见的响应输出节点"
    },
    
    "AI推理决策": {
        "llm": "大语言模型推理，支持结构化输出",
        "agent": "智能体节点，支持工具调用和多轮推理", 
        "knowledge-retrieval": "知识库检索，RAG核心节点",
        "question-classifier": "问题分类，智能路由决策"
    },
    
    "数据处理转换": {
        "code": "代码执行，支持Python/JS/TS多语言",
        "template-transform": "模板转换，变量替换和格式化",
        "parameter-extractor": "参数提取，从文本中提取结构化数据",
        "document-extractor": "文档解析，提取关键信息",
        "list-operator": "列表操作，支持过滤、排序、聚合"
    },
    
    "外部系统集成": {
        "tool": "外部工具调用，API集成",
        "http-request": "HTTP请求，RESTful API调用"
    },
    
    "流程控制": {
        "if-else": "条件分支，基于条件表达式的路由",
        "loop": "循环控制，支持条件循环和计数循环",
        "iteration": "迭代处理，对列表数据逐项处理"
    },
    
    "变量管理": {
        "variable-assigner": "变量赋值，支持复杂表达式",
        "variable-aggregator": "变量聚合，合并多个数据源"
    }
}

# 节点性能特征分析
NODE_PERFORMANCE_CHARACTERISTICS = {
    "高计算密集型": ["llm", "agent", "code"],           # CPU/GPU密集
    "I/O密集型": ["knowledge-retrieval", "http-request", "tool"], # 网络I/O
    "内存密集型": ["document-extractor", "list-operator"],        # 大数据处理
    "实时性要求高": ["if-else", "variable-assigner"],            # 毫秒级响应
    "可并行化": ["iteration", "loop", "template-transform"]       # 支持并行优化
}

# 基于实际部署的节点配置最佳实践
PRODUCTION_NODE_CONFIG = {
    # LLM节点生产优化
    "llm_node_optimization": {
        "timeout_config": {
            "default_timeout": 120,        # 2分钟默认超时
            "streaming_timeout": 300,      # 流式输出5分钟超时
            "batch_timeout": 600          # 批处理10分钟超时
        },
        "retry_config": {
            "max_retries": 3,
            "retry_delay": [1, 2, 4],     # 递增延迟重试
            "retry_on_errors": ["timeout", "rate_limit", "server_error"]
        },
        "resource_limits": {
            "max_prompt_tokens": 100000,   # 最大提示令牌数
            "max_completion_tokens": 4096, # 最大完成令牌数
            "memory_limit": "1Gi"          # 内存限制
        }
    },
    
    # Code节点安全优化
    "code_node_security": {
        "execution_environment": "docker_sandbox",
        "allowed_libraries": {
            "python": ["json", "math", "datetime", "re", "pandas", "numpy"],
            "javascript": ["lodash", "moment", "crypto-js"],
            "typescript": ["lodash", "@types/node"]
        },
        "resource_limits": {
            "cpu_limit": "500m",           # CPU限制
            "memory_limit": "512Mi",       # 内存限制
            "execution_timeout": 30,       # 执行超时
            "network_access": False        # 禁用网络访问
        },
        "security_policies": {
            "file_system_access": "read_only",
            "process_spawning": False,
            "system_calls": "restricted"
        }
    },
    
    # 知识检索节点性能优化
    "knowledge_retrieval_optimization": {
        "search_config": {
            "default_top_k": 5,
            "max_top_k": 20,
            "score_threshold": 0.7,
            "rerank_enabled": True
        },
        "cache_config": {
            "query_cache_ttl": 300,        # 查询缓存5分钟
            "embedding_cache_ttl": 3600,   # 向量缓存1小时
            "result_cache_enabled": True
        },
        "performance_config": {
            "concurrent_searches": 3,      # 并发搜索数
            "batch_embedding_size": 50,    # 批量嵌入大小
            "connection_pool_size": 10     # 连接池大小
        }
    }
}

class LLMNode(BaseNode):
    """
    LLM语言模型节点
    负责调用大语言模型进行文本生成和理解
    """
    
    _node_type = NodeType.LLM
    _node_data: LLMNodeData
    
    # 编译的正则表达式，用于提取<think>块
    _THINK_PATTERN = re.compile(r"<think[^>]*>(.*?)</think>", re.IGNORECASE | re.DOTALL)
    
    # 多模态文件输出
    _file_outputs: list["File"]
    
    # 文件保存器
    _llm_file_saver: LLMFileSaver

    def __init__(
        self,
        id: str,
        config: Mapping[str, Any],
        graph_init_params: "GraphInitParams",
        graph: "Graph",
        graph_runtime_state: "GraphRuntimeState",
        previous_node_id: Optional[str] = None,
        thread_pool_id: Optional[str] = None,
        *,
        llm_file_saver: LLMFileSaver | None = None,
    ):
        """
        初始化LLM节点
        
        Args:
            id: 节点ID
            config: 节点配置
            graph_init_params: 图初始化参数
            graph: 图实例
            graph_runtime_state: 图运行时状态
            previous_node_id: 前置节点ID
            thread_pool_id: 线程池ID
            llm_file_saver: 可选的文件保存器
        """
        super().__init__(
            id=id,
            config=config,
            graph_init_params=graph_init_params,
            graph=graph,
            graph_runtime_state=graph_runtime_state,
            previous_node_id=previous_node_id,
            thread_pool_id=thread_pool_id,
        )
        
        # 初始化文件输出列表
        self._file_outputs = []

        # 设置文件保存器
        if llm_file_saver is None:
            llm_file_saver = FileSaverImpl(
                user_id=graph_init_params.user_id,
                tenant_id=graph_init_params.tenant_id,
            )
        self._llm_file_saver = llm_file_saver

    def init_node_data(self, data: Mapping[str, Any]):
        """
        初始化LLM节点数据
        解析和验证LLM节点的配置参数
        
        Args:
            data: 节点数据字典
        """
        self._node_data = LLMNodeData.model_validate(data)
        
        # 验证模型配置
        if not self._node_data.model_config:
            raise LLMNodeError("LLM节点必须配置模型")
        
        # 验证提示配置
        if not self._node_data.prompt_template:
            raise LLMNodeError("LLM节点必须配置提示模板")

    def run(self, node_inputs: Mapping[str, Any]) -> Generator[NodeEvent, None, None]:
        """
        执行LLM节点
        包含提示构建、模型调用、结果处理的完整流程
        
        Args:
            node_inputs: 节点输入变量
            
        Yields:
            NodeEvent: 节点执行事件
        """
        try:
            # 1. 构建提示消息
            prompt_messages = self._build_prompt_messages(node_inputs)
            
            # 2. 获取模型实例
            model_instance = self._get_model_instance()
            
            # 3. 检查并调整模型参数
            self._adjust_model_parameters(model_instance, prompt_messages)
            
            # 4. 调用模型推理
            yield from self._invoke_llm_model(
                model_instance=model_instance,
                prompt_messages=prompt_messages,
                node_inputs=node_inputs
            )

        except Exception as e:
            logger.exception(f"LLM节点 {self.node_id} 执行失败")
            yield RunCompletedEvent(
                run_result=self.create_node_run_result(
                    status=WorkflowNodeExecutionStatus.FAILED,
                    inputs=dict(node_inputs),
                    error=str(e)
                )
            )

    def _build_prompt_messages(self, node_inputs: Mapping[str, Any]) -> list[PromptMessage]:
        """
        构建提示消息
        将模板和变量组合成完整的提示消息列表
        
        Args:
            node_inputs: 节点输入变量
            
        Returns:
            list[PromptMessage]: 提示消息列表
        """
        prompt_template = self._node_data.prompt_template
        variable_pool = self.graph_runtime_state.variable_pool
        
        messages = []
        
        # 处理不同类型的提示模板
        if prompt_template.prompt_type == PromptTemplateType.SIMPLE:
            # 简单提示模板
            prompt_text = prompt_template.simple_prompt_template
            
            # 替换模板变量
            rendered_prompt = VariableTemplateParser.render(
                template=prompt_text,
                variable_pool=variable_pool,
                node_inputs=node_inputs
            )
            
            messages.append(UserPromptMessage(content=rendered_prompt))
            
        elif prompt_template.prompt_type == PromptTemplateType.ADVANCED:
            # 高级提示模板（支持多轮对话）
            for message_template in prompt_template.advanced_chat_prompt_template.messages:
                # 渲染消息内容
                rendered_content = VariableTemplateParser.render(
                    template=message_template.text,
                    variable_pool=variable_pool,
                    node_inputs=node_inputs
                )
                
                # 根据角色创建相应的消息
                if message_template.role == PromptMessageRole.SYSTEM:
                    messages.append(SystemPromptMessage(content=rendered_content))
                elif message_template.role == PromptMessageRole.USER:
                    messages.append(UserPromptMessage(content=rendered_content))
                elif message_template.role == PromptMessageRole.ASSISTANT:
                    messages.append(AssistantPromptMessage(content=rendered_content))

        # 处理文件输入（多模态支持）
        if self._should_include_files(node_inputs):
            messages = self._add_file_content_to_messages(messages, node_inputs)

        return messages

    def _invoke_llm_model(
        self,
        model_instance: ModelInstance,
        prompt_messages: list[PromptMessage],
        node_inputs: Mapping[str, Any]
    ) -> Generator[NodeEvent, None, None]:
        """
        调用LLM模型
        执行实际的模型推理并处理结果
        
        Args:
            model_instance: 模型实例
            prompt_messages: 提示消息列表
            node_inputs: 节点输入
            
        Yields:
            NodeEvent: 模型调用事件
        """
        model_config = self._node_data.model_config
        
        try:
            # 判断是否为结构化输出
            if self._node_data.output_schema:
                # 结构化输出模式
                yield from self._invoke_structured_output(
                    model_instance=model_instance,
                    prompt_messages=prompt_messages,
                    output_schema=self._node_data.output_schema,
                    node_inputs=node_inputs
                )
            else:
                # 普通文本输出模式
                yield from self._invoke_text_generation(
                    model_instance=model_instance,
                    prompt_messages=prompt_messages,
                    node_inputs=node_inputs
                )

        except Exception as e:
            logger.exception(f"模型调用失败: {e}")
            yield RunCompletedEvent(
                run_result=self.create_node_run_result(
                    status=WorkflowNodeExecutionStatus.FAILED,
                    inputs=dict(node_inputs),
                    error=f"模型调用失败: {str(e)}"
                )
            )

    def _invoke_text_generation(
        self,
        model_instance: ModelInstance,
        prompt_messages: list[PromptMessage],
        node_inputs: Mapping[str, Any]
    ) -> Generator[NodeEvent, None, None]:
        """
        执行文本生成
        标准的LLM文本生成流程
        
        Args:
            model_instance: 模型实例
            prompt_messages: 提示消息
            node_inputs: 节点输入
            
        Yields:
            NodeEvent: 生成过程事件
        """
        model_config = self._node_data.model_config
        
        # 调用模型
        result_generator = model_instance.invoke_llm(
            prompt_messages=prompt_messages,
            model_parameters=model_config.completion_params,
            stop=model_config.stop,
            stream=True,  # 启用流式输出
            user=self.user_id,
        )

        # 处理流式结果
        full_text = ""
        usage = None
        
        for chunk in result_generator:
            if isinstance(chunk, LLMResultChunk):
                # 处理流式输出块
                if chunk.delta.message.content:
                    content = chunk.delta.message.content
                    full_text += content
                    
                    # 发布流式输出事件
                    yield RunStreamChunkEvent(
                        chunk=content,
                        chunk_type="text"
                    )
                
                # 记录使用情况
                if chunk.delta.usage:
                    usage = chunk.delta.usage
            
            elif isinstance(chunk, LLMResult):
                # 非流式结果
                full_text = chunk.message.content or ""
                usage = chunk.usage
                break

        # 后处理生成的内容
        processed_text = self._post_process_llm_output(full_text)
        
        # 处理多模态输出（如果有）
        file_outputs = self._extract_file_outputs(processed_text)
        
        # 构建输出变量
        outputs = {
            "text": processed_text,
            "usage": {
                "prompt_tokens": usage.prompt_tokens if usage else 0,
                "completion_tokens": usage.completion_tokens if usage else 0,
                "total_tokens": usage.total_tokens if usage else 0,
                "total_price": usage.total_price if usage else 0.0,
            }
        }
        
        # 添加文件输出
        if file_outputs:
            outputs["files"] = file_outputs

        # 发布完成事件
        yield RunCompletedEvent(
            run_result=self.create_node_run_result(
                status=WorkflowNodeExecutionStatus.SUCCEEDED,
                inputs=dict(node_inputs),
                outputs=outputs,
                metadata={
                    "model_provider": model_config.provider,
                    "model_name": model_config.model,
                    "prompt_messages_count": len(prompt_messages),
                }
            )
        )

    def _invoke_structured_output(
        self,
        model_instance: ModelInstance,
        prompt_messages: list[PromptMessage],
        output_schema: dict,
        node_inputs: Mapping[str, Any]
    ) -> Generator[NodeEvent, None, None]:
        """
        执行结构化输出
        使用schema约束模型输出格式
        
        Args:
            model_instance: 模型实例
            prompt_messages: 提示消息
            output_schema: 输出schema定义
            node_inputs: 节点输入
            
        Yields:
            NodeEvent: 结构化输出事件
        """
        try:
            # 调用结构化输出接口
            result_generator = invoke_llm_with_structured_output(
                model_instance=model_instance,
                prompt_messages=prompt_messages,
                output_schema=output_schema,
                stream=True,
                user=self.user_id
            )

            structured_output = None
            full_text = ""
            usage = None

            for chunk in result_generator:
                if isinstance(chunk, LLMResultChunkWithStructuredOutput):
                    # 处理结构化输出块
                    if chunk.delta.message.content:
                        content = chunk.delta.message.content
                        full_text += content
                        
                        # 发布流式输出
                        yield RunStreamChunkEvent(
                            chunk=content,
                            chunk_type="structured"
                        )
                    
                    # 获取结构化输出
                    if chunk.delta.structured_output:
                        structured_output = chunk.delta.structured_output
                    
                    if chunk.delta.usage:
                        usage = chunk.delta.usage

            # 验证结构化输出
            if not structured_output:
                raise OutputParserError("未能生成有效的结构化输出")

            # 构建输出
            outputs = {
                "text": full_text,
                "structured_output": structured_output.data,
                "schema_valid": True,
                "usage": {
                    "prompt_tokens": usage.prompt_tokens if usage else 0,
                    "completion_tokens": usage.completion_tokens if usage else 0,
                    "total_tokens": usage.total_tokens if usage else 0,
                    "total_price": usage.total_price if usage else 0.0,
                }
            }

            yield RunCompletedEvent(
                run_result=self.create_node_run_result(
                    status=WorkflowNodeExecutionStatus.SUCCEEDED,
                    inputs=dict(node_inputs),
                    outputs=outputs,
                    metadata={
                        "output_type": "structured",
                        "schema": output_schema,
                    }
                )
            )

        except OutputParserError as e:
            # 结构化输出解析失败
            logger.error(f"结构化输出解析失败: {e}")
            yield RunCompletedEvent(
                run_result=self.create_node_run_result(
                    status=WorkflowNodeExecutionStatus.FAILED,
                    inputs=dict(node_inputs),
                    error=f"结构化输出解析失败: {str(e)}"
                )
            )

    def _post_process_llm_output(self, text: str) -> str:
        """
        后处理LLM输出
        清理和优化模型生成的文本
        
        Args:
            text: 原始输出文本
            
        Returns:
            str: 后处理的文本
        """
        if not text:
            return ""
        
        # 移除<think>标签块（如果存在）
        text = self._THINK_PATTERN.sub("", text)
        
        # 清理多余的空白字符
        text = re.sub(r'\n\s*\n\s*\n', '\n\n', text)  # 多个空行合并为两个
        text = text.strip()
        
        return text

    def _extract_file_outputs(self, text: str) -> list[dict]:
        """
        从文本中提取文件输出
        解析文本中的文件引用并生成文件对象
        
        Args:
            text: 生成的文本
            
        Returns:
            list[dict]: 文件输出列表
        """
        file_outputs = []
        
        # 查找base64编码的图片
        base64_pattern = r'data:image/([^;]+);base64,([^"]+)'
        matches = re.finditer(base64_pattern, text)
        
        for match in matches:
            image_format = match.group(1)
            base64_data = match.group(2)
            
            try:
                # 解码base64数据
                image_data = base64.b64decode(base64_data)
                
                # 保存文件
                file_obj = self._llm_file_saver.save_file(
                    content=image_data,
                    file_type=FileType.IMAGE,
                    file_extension=image_format,
                    metadata={
                        "source": "llm_generation",
                        "node_id": self.node_id,
                        "generated_at": time.time()
                    }
                )
                
                file_outputs.append({
                    "id": file_obj.id,
                    "type": "image",
                    "url": file_obj.url,
                    "filename": file_obj.filename,
                    "size": len(image_data)
                })
                
            except Exception as e:
                logger.warning(f"提取生成的图片失败: {e}")
                continue
        
        return file_outputs

    def _get_model_instance(self) -> ModelInstance:
        """
        获取模型实例
        根据节点配置创建模型实例
        
        Returns:
            ModelInstance: 模型实例
        """
        model_config = self._node_data.model_config
        
        # 创建模型管理器
        model_manager = ModelManager()
        
        # 获取模型实例
        model_instance = model_manager.get_model_instance(
            tenant_id=self.tenant_id,
            model_type=ModelType.LLM,
            provider=model_config.provider,
            model=model_config.model
        )
        
        return model_instance

    def _adjust_model_parameters(
        self,
        model_instance: ModelInstance,
        prompt_messages: list[PromptMessage]
    ):
        """
        调整模型参数
        根据提示长度和模型限制自动调整参数
        
        Args:
            model_instance: 模型实例
            prompt_messages: 提示消息列表
        """
        model_config = self._node_data.model_config
        
        # 获取模型上下文长度限制
        model_context_tokens = model_config.model_schema.model_properties.get(
            ModelPropertyKey.CONTEXT_SIZE
        )
        
        if not model_context_tokens:
            return  # 无法获取上下文限制，跳过调整

        # 计算提示令牌数
        prompt_tokens = model_instance.get_llm_num_tokens(prompt_messages)
        
        # 获取当前max_tokens设置
        max_tokens = model_config.completion_params.get("max_tokens", 2048)
        
        # 检查是否超过模型限制
        if prompt_tokens + max_tokens > model_context_tokens:
            # 调整max_tokens以适应上下文限制
            adjusted_max_tokens = max(model_context_tokens - prompt_tokens, 16)
            
            logger.info(
                f"调整max_tokens: {max_tokens} -> {adjusted_max_tokens} "
                f"(提示令牌: {prompt_tokens}, 上下文限制: {model_context_tokens})"
            )
            
            model_config.completion_params["max_tokens"] = adjusted_max_tokens

class LLMNodeData(BaseNodeData):
    """
    LLM节点数据配置
    定义LLM节点的所有配置选项
    """
    
    # 模型配置
    model_config: ModelConfig
    
    # 提示模板配置
    prompt_template: PromptTemplateConfig
    
    # 输出schema（用于结构化输出）
    output_schema: Optional[dict] = None
    
    # 记忆配置
    memory_config: Optional[MemoryConfig] = None
    
    # 上下文配置
    context_config: Optional[ContextConfig] = None
    
    # 视觉配置（多模态）
    vision_config: Optional[VisionConfig] = None

class ModelConfig(BaseModel):
    """模型配置"""
    
    # 模型提供者
    provider: str
    
    # 模型名称
    model: str
    
    # 模型模式
    mode: LLMMode
    
    # 完成参数
    completion_params: dict[str, Any] = Field(default_factory=dict)
    
    # 停止词
    stop: list[str] = Field(default_factory=list)
    
    # 模型schema
    model_schema: Optional[dict] = None

class PromptTemplateConfig(BaseModel):
    """提示模板配置"""
    
    # 提示类型
    prompt_type: PromptTemplateType
    
    # 简单提示模板
    simple_prompt_template: Optional[str] = None
    
    # 高级聊天提示模板
    advanced_chat_prompt_template: Optional[AdvancedChatPromptTemplate] = None
    
    # 高级完成提示模板
    advanced_completion_prompt_template: Optional[AdvancedCompletionPromptTemplate] = None

class PromptTemplateType(Enum):
    """提示模板类型"""
    
    SIMPLE = "simple"      # 简单文本提示
    ADVANCED = "advanced"  # 高级多轮对话提示