MemoryBear/api/app/services/multi_agent_orchestrator.py

"""多 Agent 编排器 - Master Agent 作为决策中心"""
import uuid
import time
import asyncio
from typing import Dict, Any, List, Optional, AsyncIterator, Tuple
from sqlalchemy.orm import Session

from app.models import MultiAgentConfig, AgentConfig, ModelConfig
from app.models.multi_agent_model import AggregationStrategy, OrchestrationMode
from app.repositories.model_repository import ModelApiKeyRepository
from app.services.agent_registry import AgentRegistry
from app.services.master_agent_router import MasterAgentRouter
from app.services.conversation_state_manager import ConversationStateManager
from app.core.exceptions import BusinessException, ResourceNotFoundException
from app.core.error_codes import BizCode
from app.core.logging_config import get_business_logger

logger = get_business_logger()


class MultiAgentOrchestrator:
    """多 Agent 编排器 - 协调多个 Agent 协作完成任务"""

    def __init__(self, db: Session, config: MultiAgentConfig):
        """初始化编排器

        Args:
            db: 数据库会话
            config: 多 Agent 配置
        """
        self.db = db
        self.config = config
        self.registry = AgentRegistry(db)

        # 兼容处理：旧的 orchestration_mode 值映射到新值
        # collaboration | supervisor 是新值，其他旧值默认使用 supervisor
        self._normalized_mode = self._normalize_orchestration_mode(config.orchestration_mode)

        # 加载主 Agent
        # self.master_agent = self._load_agent(config.master_agent_id)
        # self. config.d
        self.default_model_config_id = config.default_model_config_id
        self.model_parameters = config.model_parameters

        # 加载子 Agent
        self.sub_agents = {}
        for sub_agent_info in config.sub_agents:
            agent_id = uuid.UUID(sub_agent_info["agent_id"])
            agent = self._load_agent(agent_id)
            self.sub_agents[str(agent_id)] = {
                "config": agent,
                "info": sub_agent_info
            }

        # 初始化会话状态管理器
        self.state_manager = ConversationStateManager()

        # 只有 supervisor 模式才需要 default_model_config_id 和 router
        self.master_model_config = None
        self.router = None

        if self._normalized_mode == OrchestrationMode.SUPERVISOR:
            # 获取 Master Agent 的模型配置
            if not self.default_model_config_id:
                raise BusinessException("Supervisor 模式需要配置默认模型", BizCode.AGENT_CONFIG_MISSING)

            self.master_model_config = self.db.get(ModelConfig, self.default_model_config_id)
            if not self.master_model_config:
                raise BusinessException("Master Agent 模型配置不存在", BizCode.AGENT_CONFIG_MISSING)

            # 初始化 Master Agent 路由器
            self.router = MasterAgentRouter(
                db=db,
                master_model_config=self.master_model_config,
                model_parameters=self.model_parameters,
                sub_agents=self.sub_agents,
                state_manager=self.state_manager,
                enable_rule_fast_path=config.execution_config.get("enable_rule_fast_path", True)
            )

        logger.info(
            "多 Agent 编排器初始化完成",
            extra={
                "config_id": str(config.id),
                "model": self.master_model_config.name if self.master_model_config else None,
                "sub_agent_count": len(self.sub_agents),
                "orchestration_mode": self._normalized_mode
            }
        )

    def _normalize_orchestration_mode(self, mode: str) -> str:
        """标准化 orchestration_mode，兼容旧值

        Args:
            mode: 原始的 orchestration_mode 值

        Returns:
            标准化后的模式：collaboration 或 supervisor
        """
        if mode in [OrchestrationMode.SUPERVISOR, "supervisor"]:
            return OrchestrationMode.SUPERVISOR
        # 其他所有值（包括旧的 sequential、parallel、conditional、loop 和 collaboration）都映射到 collaboration
        return OrchestrationMode.COLLABORATION

    async def execute_stream(
        self,
        message: str,
        conversation_id: Optional[uuid.UUID] = None,
        user_id: Optional[str] = None,
        variables: Optional[Dict[str, Any]] = None,
        use_llm_routing: bool = True,
        web_search: bool = True,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ):
        """执行多 Agent 任务（流式返回）

        Args:
            message: 用户消息
            conversation_id: 会话 ID
            user_id: 用户 ID
            variables: 变量参数
            use_llm_routing: 是否使用 LLM 路由

        Yields:
            SSE 格式的事件流
        """
        import json

        start_time = time.time()

        logger.info(
            "开始执行多 Agent 任务（流式）",
            extra={
                "mode": self._normalized_mode,
                "message_length": len(message)
            }
        )

        try:
            # 发送开始事件
            yield self._format_sse_event("start", {
                "mode": self._normalized_mode,
                "timestamp": time.time()
            })

            # 2. 根据模式执行（流式）
            # Collaboration 模式：Agent 之间可以相互 handoff（使用 handoffs_service）
            if self._normalized_mode == OrchestrationMode.COLLABORATION:
                async for event in self._execute_collaboration_mode_stream(
                    message,
                    conversation_id,
                    user_id,
                    web_search,
                    memory,
                    storage_type,
                    user_rag_memory_id
                ):
                    yield event
            # Supervisor 模式：由主 Agent 统一调度子 Agent
            elif self._normalized_mode == OrchestrationMode.SUPERVISOR:
                # 1. 主 Agent 分析任务
                task_analysis = await self._analyze_task(message, variables)
                task_analysis["use_llm_routing"] = use_llm_routing

                async for event in self._execute_supervisor_stream(
                    task_analysis,
                    conversation_id,
                    user_id,
                    web_search,
                    memory,
                    storage_type,
                    user_rag_memory_id
                ):
                    yield event
            else:
                raise BusinessException(
                    f"不支持的编排模式: {self._normalized_mode}",
                    BizCode.INVALID_PARAMETER
                )

            elapsed_time = time.time() - start_time

            # 发送结束事件
            yield self._format_sse_event("end", {
                "elapsed_time": elapsed_time,
                "timestamp": time.time()
            })

            logger.info(
                "多 Agent 任务完成（流式）",
                extra={
                    "mode": self._normalized_mode,
                    "elapsed_time": elapsed_time
                }
            )

        except Exception as e:
            logger.error(
                "多 Agent 任务执行失败（流式）",
                extra={"error": str(e), "mode": self._normalized_mode}
            )
            # 发送错误事件
            yield self._format_sse_event("error", {
                "error": str(e),
                "timestamp": time.time()
            })

    async def execute(
        self,
        message: str,
        conversation_id: Optional[uuid.UUID] = None,
        user_id: Optional[str] = None,
        variables: Optional[Dict[str, Any]] = None,
        use_llm_routing: bool = True,
        web_search: bool = False,
        memory: bool = True
    ) -> Dict[str, Any]:
        """执行多 Agent 任务（基于 Master Agent 决策）

        Args:
            message: 用户消息
            conversation_id: 会话 ID
            user_id: 用户 ID
            variables: 变量参数
            use_llm_routing: 是否使用 LLM 路由（保留参数，实际总是使用 Master Agent）

        Returns:
            执行结果
        """
        start_time = time.time()

        logger.info(
            "开始执行多 Agent 任务",
            extra={
                "message_length": len(message),
                "mode": self._normalized_mode
            }
        )

        try:
            # Collaboration 模式：使用 handoffs_service
            if self._normalized_mode == OrchestrationMode.COLLABORATION:
                return await self._execute_collaboration_mode(
                    message,
                    conversation_id,
                    user_id,
                    variables
                )

            # Supervisor 模式：由 Master Agent 统一调度
            # 1. Master Agent 分析任务并做出决策
            task_analysis = await self._analyze_task(message, variables)

            routing_decision = task_analysis.get("routing_decision")
            if not routing_decision:
                raise BusinessException("Master Agent 未返回路由决策", BizCode.AGENT_CONFIG_MISSING)

            logger.info(
                "Master Agent 决策",
                extra={
                    "need_collaboration": routing_decision.get("need_collaboration"),
                    "strategy": routing_decision.get("collaboration_strategy"),
                    "confidence": routing_decision.get("confidence")
                }
            )

            # 2. 根据 Master Agent 的决策执行
            results = await self._execute_conditional(
                task_analysis,
                conversation_id,
                user_id
            )

            # 3. 整合结果
            final_result = await self._aggregate_results(results)

            elapsed_time = time.time() - start_time

            # 4. 提取子 Agent 的 conversation_id（用于多轮对话）
            sub_conversation_id = None
            total_tokens = 0
            
            if isinstance(results, dict):
                sub_conversation_id = results.get("conversation_id") or results.get("result", {}).get("conversation_id")
                # 提取 token 信息
                usage = results.get("usage", {}) or results.get("result", {}).get("usage", {})
                total_tokens += usage.get("total_tokens", 0)
            elif isinstance(results, list) and results:
                for item in results:
                    if "result" in item:
                        sub_conversation_id = item["result"].get("conversation_id")
                        if sub_conversation_id:
                            break
                    # 累加每个子 Agent 的 token
                    usage = item.get("usage", {}) or item.get("result", {}).get("usage", {})
                    total_tokens += usage.get("total_tokens", 0)

            logger.info(
                "多 Agent 任务完成",
                extra={
                    "strategy": routing_decision.get("collaboration_strategy", "single"),
                    "elapsed_time": elapsed_time,
                    "sub_conversation_id": sub_conversation_id
                }
            )

            return {
                "message": final_result,
                "conversation_id": sub_conversation_id,
                "mode": OrchestrationMode.SUPERVISOR,
                "elapsed_time": elapsed_time,
                "strategy": routing_decision.get("collaboration_strategy", "single"),
                "sub_results": results,
                "usage": {
                    "prompt_tokens": 0,
                    "completion_tokens": 0,
                    "total_tokens": total_tokens
                }
            }

        except Exception as e:
            logger.error(
                "多 Agent 任务执行失败",
                extra={"error": str(e)}
            )
            raise

    async def _analyze_task(
        self,
        message: str,
        variables: Optional[Dict[str, Any]] = None
    ) -> Dict[str, Any]:
        """Master Agent 分析任务并做出路由决策

        Args:
            message: 用户消息
            variables: 变量参数

        Returns:
            任务分析结果，包含路由决策
        """
        logger.info(
            "Master Agent 开始分析任务",
            extra={"message_length": len(message)}
        )

        # 使用 Master Agent 路由器进行决策
        routing_decision = await self.router.route(
            message=message,
            conversation_id=None,  # 会在后续传入
            variables=variables
        )

        logger.info(
            "Master Agent 分析完成",
            extra={
                "selected_agent": routing_decision.get("selected_agent_id"),
                "confidence": routing_decision.get("confidence"),
                "strategy": routing_decision.get("strategy")
            }
        )

        return {
            "message": message,
            "variables": variables or {},
            "sub_agents": self.config.sub_agents,
            "initial_context": variables or {},
            "routing_decision": routing_decision
        }

    async def _execute_sequential(
        self,
        task_analysis: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ) -> List[Dict[str, Any]]:
        """顺序执行子 Agent

        Args:
            task_analysis: 任务分析结果
            conversation_id: 会话 ID
            user_id: 用户 ID

        Returns:
            执行结果列表
        """
        results = []
        context = task_analysis.get("initial_context", {})
        message = task_analysis.get("message", "")

        # 按优先级排序
        sub_agents = sorted(
            task_analysis["sub_agents"],
            key=lambda x: x.get("priority", 0)
        )

        for sub_agent_info in sub_agents:
            agent_id = sub_agent_info["agent_id"]
            agent_data = self.sub_agents.get(agent_id)

            if not agent_data:
                logger.warning(f"子 Agent 不存在: {agent_id}")
                continue

            logger.info(
                "执行子 Agent",
                extra={
                    "agent_id": agent_id,
                    "agent_name": sub_agent_info.get("name"),
                    "priority": sub_agent_info.get("priority")
                }
            )

            # 执行子 Agent
            result = await self._execute_sub_agent(
                agent_data["config"],
                message,
                context,
                conversation_id,
                user_id,
                web_search,
                memory,
                storage_type,
                user_rag_memory_id
            )

            results.append({
                "agent_id": agent_id,
                "agent_name": sub_agent_info.get("name"),
                "result": result,
                "conversation_id": result.get("conversation_id")  # 保存会话 ID
            })

            # 更新上下文（后续 Agent 可以使用前面的结果）
            context[f"result_from_{sub_agent_info.get('name', agent_id)}"] = result.get("message")

        return results

    async def _execute_parallel(
        self,
        task_analysis: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ) -> List[Dict[str, Any]]:
        """并行执行子 Agent

        Args:
            task_analysis: 任务分析结果
            conversation_id: 会话 ID
            user_id: 用户 ID

        Returns:
            执行结果列表
        """
        context = task_analysis.get("initial_context", {})
        message = task_analysis.get("message", "")

        # 获取并发限制
        parallel_limit = self.config.execution_config.get("parallel_limit", 3)

        # 创建任务列表
        tasks = []
        for sub_agent_info in task_analysis["sub_agents"]:
            agent_id = sub_agent_info["agent_id"]
            agent_data = self.sub_agents.get(agent_id)

            if not agent_data:
                continue

            task = self._execute_sub_agent(
                agent_data["config"],
                message,
                context,
                conversation_id,
                user_id,
                web_search,
                memory,
                storage_type,
                user_rag_memory_id
            )
            tasks.append((agent_id, sub_agent_info.get("name"), task))

        # 并行执行（带限制）
        results = []
        for i in range(0, len(tasks), parallel_limit):
            batch = tasks[i:i + parallel_limit]
            batch_results = await asyncio.gather(
                *[task for _, _, task in batch],
                return_exceptions=True
            )

            for (agent_id, agent_name, _), result in zip(batch, batch_results, strict=False):
                if isinstance(result, Exception):
                    logger.error(f"子 Agent 执行失败: {agent_name}", extra={"error": str(result)})
                    results.append({
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "error": str(result)
                    })
                else:
                    results.append({
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "result": result,
                        "conversation_id": result.get("conversation_id")  # 保存会话 ID
                    })

        return results

    async def _execute_collaboration_stream(
        self,
        task_analysis: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        routing_decision: Dict[str, Any]
    ):
        """多 Agent 协作流式执行

        Args:
            task_analysis: 任务分析结果
            conversation_id: 会话 ID
            user_id: 用户 ID
            routing_decision: 路由决策

        Yields:
            SSE 格式的事件流
        """
        message = task_analysis.get("message", "")
        initial_context = task_analysis.get("initial_context", {})
        collaboration_strategy = routing_decision.get("collaboration_strategy", "sequential")

        # 获取协作信息
        if collaboration_strategy == "decomposition":
            collaboration_agents = routing_decision.get("sub_questions", [])
        else:
            collaboration_agents = routing_decision.get("collaboration_agents", [])

        logger.info(
            "开始流式协作执行",
            extra={
                "strategy": collaboration_strategy,
                "agent_count": len(collaboration_agents)
            }
        )

        # 1. 发送编排计划事件（在执行前）
        # 构建子任务信息
        sub_tasks = []
        for item in collaboration_agents:
            if collaboration_strategy == "decomposition":
                # 问题拆分模式
                agent_id = item.get("agent_id")
                agent_data = self.sub_agents.get(agent_id)
                if agent_data:
                    sub_tasks.append({
                        "agent_id": agent_id,
                        "agent_name": agent_data.get("info", {}).get("name", agent_id),
                        "sub_question": item.get("question", ""),
                        "order": item.get("order", 0)
                    })
            else:
                # 其他协作模式
                agent_id = item.get("agent_id")
                agent_data = self.sub_agents.get(agent_id)
                if agent_data:
                    sub_tasks.append({
                        "agent_id": agent_id,
                        "agent_name": agent_data.get("info", {}).get("name", agent_id),
                        "role": item.get("role", "secondary"),
                        "order": item.get("order", 0)
                    })

        yield self._format_sse_event("orchestration_plan", {
            "agent_count": len(sub_tasks),
            "strategy": collaboration_strategy,
            "sub_tasks": sub_tasks
        })

        # 2. 流式执行所有子 Agent
        results = []

        # 获取执行模式配置
        execution_mode = self.config.execution_config.get("sub_agent_execution_mode", "parallel")

        if collaboration_strategy == "decomposition":
            # 问题拆分模式
            # 检查是否有依赖关系
            has_dependencies = self._check_dependencies(collaboration_agents)

            if has_dependencies or execution_mode == "sequential":
                # 有依赖或配置为串行：串行流式执行
                logger.info("使用串行流式执行（问题拆分）")
                for sub_q in sorted(collaboration_agents, key=lambda x: x.get("order", 0)):
                    sub_question = sub_q.get("question", "")
                    agent_id = sub_q.get("agent_id")

                    agent_data = self.sub_agents.get(agent_id)
                    if not agent_data:
                        continue

                    agent_name = agent_data.get("info", {}).get("name", agent_id)

                    # 发送子问题开始事件
                    yield self._format_sse_event("sub_question_start", {
                        "question": sub_question,
                        "agent_name": agent_name
                    })

                    # 流式执行子 Agent，收集结果
                    result_content = ""
                    async for event in self._execute_sub_agent_stream(
                        agent_data["config"],
                        sub_question,
                        initial_context,
                        conversation_id,
                        user_id
                    ):
                        # 解析原始事件
                        if "data:" in event:
                            try:
                                import json
                                data_line = event.split("data: ", 1)[1].strip()
                                data = json.loads(data_line)

                                # 提取内容
                                if "content" in data:
                                    content = data["content"]
                                    result_content += content

                                    # 转换为子 Agent 专用事件，带上 agent 信息
                                    yield self._format_sse_event("sub_agent_message", {
                                        "content": content,
                                        "agent_id": agent_id,
                                        "agent_name": agent_name,
                                        "sub_question": sub_question
                                    })
                            except Exception:
                                pass
                        else:
                            # 非 data 事件直接转发
                            yield event

                    results.append({
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "sub_question": sub_question,
                        "result": {"message": result_content}
                    })

                    # 发送子问题完成事件
                    yield self._format_sse_event("sub_question_end", {
                        "agent_name": agent_name
                    })
            else:
                # 无依赖且配置为并行：并行流式执行
                logger.info(f"使用并行流式执行（问题拆分），共 {len(collaboration_agents)} 个子问题")

                # 准备并行任务
                agent_tasks = []
                agent_info_map = {}
                result_contents = {}

                for sub_q in collaboration_agents:
                    sub_question = sub_q.get("question", "")
                    agent_id = sub_q.get("agent_id")

                    agent_data = self.sub_agents.get(agent_id)
                    if not agent_data:
                        continue

                    agent_name = agent_data.get("info", {}).get("name", agent_id)
                    agent_info_map[agent_id] = {
                        "name": agent_name,
                        "sub_question": sub_question
                    }
                    result_contents[agent_id] = ""

                    agent_tasks.append((
                        agent_id,
                        agent_name,
                        agent_data["config"],
                        sub_question,
                        initial_context
                    ))

                    # 发送子问题开始事件
                    yield self._format_sse_event("sub_question_start", {
                        "question": sub_question,
                        "agent_name": agent_name
                    })

                # 并行流式执行
                async for agent_id, agent_name, event_type, content in self._parallel_stream_agents(
                    agent_tasks,
                    conversation_id,
                    user_id
                ):
                    if event_type == "content":
                        # 累积结果
                        result_contents[agent_id] += content

                        # 实时返回
                        yield self._format_sse_event("sub_agent_message", {
                            "content": content,
                            "agent_id": agent_id,
                            "agent_name": agent_name,
                            "sub_question": agent_info_map[agent_id]["sub_question"]
                        })

                    elif event_type == "done":
                        # Agent 完成
                        results.append({
                            "agent_id": agent_id,
                            "agent_name": agent_name,
                            "sub_question": agent_info_map[agent_id]["sub_question"],
                            "result": {"message": result_contents[agent_id]}
                        })

                        yield self._format_sse_event("sub_question_end", {
                            "agent_name": agent_name
                        })

                    elif event_type == "error":
                        logger.error(f"Agent {agent_name} 执行失败: {content}")
        else:
            # 其他协作模式（sequential/parallel/hierarchical）
            if collaboration_strategy == "parallel" and execution_mode == "parallel":
                # 并行协作 + 并行流式执行
                logger.info(f"使用并行流式执行（并行协作），共 {len(collaboration_agents)} 个 Agent")

                # 准备并行任务
                agent_tasks = []
                agent_info_map = {}
                result_contents = {}

                for agent_info in collaboration_agents:
                    agent_id = agent_info.get("agent_id")
                    agent_data = self.sub_agents.get(agent_id)
                    if not agent_data:
                        continue

                    agent_name = agent_data.get("info", {}).get("name", agent_id)
                    agent_info_map[agent_id] = {
                        "name": agent_name,
                        "role": agent_info.get("role", "secondary"),
                        "task": agent_info.get("task", "")
                    }
                    result_contents[agent_id] = ""

                    # 构建该 Agent 的消息
                    agent_task = agent_info.get("task", "处理任务")
                    agent_message = f"""原始问题：{message}

你的任务：{agent_task}

请完成你的任务。"""

                    agent_tasks.append((
                        agent_id,
                        agent_name,
                        agent_data["config"],
                        agent_message,
                        initial_context.copy()
                    ))

                    # 发送 Agent 开始事件
                    yield self._format_sse_event("agent_start", {
                        "agent_name": agent_name
                    })

                # 并行流式执行
                async for agent_id, agent_name, event_type, content in self._parallel_stream_agents(
                    agent_tasks,
                    conversation_id,
                    user_id
                ):
                    if event_type == "content":
                        # 累积结果
                        result_contents[agent_id] += content

                        # 实时返回
                        yield self._format_sse_event("sub_agent_message", {
                            "content": content,
                            "agent_id": agent_id,
                            "agent_name": agent_name,
                            "role": agent_info_map[agent_id]["role"]
                        })

                    elif event_type == "done":
                        # Agent 完成
                        results.append({
                            "agent_id": agent_id,
                            "agent_name": agent_name,
                            "role": agent_info_map[agent_id]["role"],
                            "task": agent_info_map[agent_id]["task"],
                            "result": {"message": result_contents[agent_id]}
                        })

                        yield self._format_sse_event("agent_end", {
                            "agent_name": agent_name
                        })

                    elif event_type == "error":
                        logger.error(f"Agent {agent_name} 执行失败: {content}")
            else:
                # 顺序协作或层级协作 - 串行流式执行
                logger.info(f"使用串行流式执行（{collaboration_strategy}）")
                for agent_info in collaboration_agents:
                    agent_id = agent_info.get("agent_id")
                    agent_data = self.sub_agents.get(agent_id)
                    if not agent_data:
                        continue

                    agent_name = agent_data.get("info", {}).get("name", agent_id)

                    # 发送 Agent 开始事件
                    yield self._format_sse_event("agent_start", {
                        "agent_name": agent_name
                    })

                    # 流式执行子 Agent，收集结果
                    result_content = ""
                    async for event in self._execute_sub_agent_stream(
                        agent_data["config"],
                        message,
                        initial_context,
                        conversation_id,
                        user_id
                    ):
                        # 解析原始事件
                        if "data:" in event:
                            try:
                                import json
                                data_line = event.split("data: ", 1)[1].strip()
                                data = json.loads(data_line)

                                # 提取内容
                                if "content" in data:
                                    content = data["content"]
                                    result_content += content

                                    # 转换为子 Agent 专用事件，带上 agent 信息
                                    yield self._format_sse_event("sub_agent_message", {
                                        "content": content,
                                        "agent_id": agent_id,
                                        "agent_name": agent_name,
                                        "role": agent_info.get("role", "secondary")
                                    })
                            except:
                                pass
                        else:
                            # 非 data 事件直接转发
                            yield event

                    results.append({
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "result": {"message": result_content}
                    })

                    # 发送 Agent 完成事件
                    yield self._format_sse_event("agent_end", {
                        "agent_name": agent_name
                    })

        # 3. 智能整合结果
        merge_mode = self.config.execution_config.get("result_merge_mode", "smart")
        # merge_mode = "master"
        # 智能判断是否需要整合
        need_merge = self._should_merge_results(results, collaboration_strategy)

        if not need_merge:
            # 不需要整合：用户已经看到所有内容了
            logger.info("跳过整合阶段（用户已看到所有 Agent 输出）")
            # 输出 message 事件，将 sub_agent_message 的内容汇总输出
            if results:
                all_content = ""
                for result in results:
                    msg = result.get("result", {}).get("message", "")
                    if msg:
                        all_content += msg + "\n\n"
                if all_content:
                    yield self._format_sse_event("message", {
                        "content": all_content.strip()
                    })
        elif merge_mode == "master" and len(results) > 1:
            # Master Agent 流式整合
            logger.info("开始 Master Agent 流式整合")

            # 发送整合开始提示
            yield self._format_sse_event("merge_start", {
                "merge_mode": "master",
                "agent_count": len(results),
                "message": "正在整合多个专家的回答..."
            })

            # 流式整合
            try:
                async for event in self._master_merge_results_stream(
                    results,
                    collaboration_strategy,
                    message
                ):
                    yield event
            except Exception as e:
                logger.error(f"Master Agent 流式整合失败，降级到 smart 模式: {str(e)}")
                final_response = self._smart_merge_results(results, collaboration_strategy)
                yield self._format_sse_event("message", {
                    "content": final_response
                })
        else:
            # Smart 模式：快速整合
            logger.info("使用 Smart 模式整合")

            yield self._format_sse_event("merge_start", {
                "merge_mode": "smart",
                "agent_count": len(results)
            })

            final_response = self._smart_merge_results(results, collaboration_strategy)

            # 只有在需要时才发送整合结果
            if final_response and final_response != "":
                yield self._format_sse_event("message", {
                    "content": final_response
                })

    async def _execute_supervisor_stream(
        self,
        task_analysis: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ):
        """条件路由执行（流式，重构版 - 使用 Master Agent 决策）

        Args:
            task_analysis: 任务分析结果（包含 Master Agent 的决策）
            conversation_id: 会话 ID
            user_id: 用户 ID

        Yields:
            SSE 格式的事件流
        """
        if not task_analysis["sub_agents"]:
            raise BusinessException("没有可用的子 Agent", BizCode.AGENT_CONFIG_MISSING)

        message = task_analysis.get("message", "")
        routing_decision = task_analysis.get("routing_decision")
        yield self._format_sse_event("routing_decision", {
            "routing_decision": routing_decision
        })

        # 1. 检查是否需要协作
        if routing_decision and routing_decision.get("need_collaboration"):
            # 需要多 Agent 协作，使用流式整合
            logger.info("检测到需要多 Agent 协作，使用流式整合")

            async for event in self._execute_collaboration_stream(
                task_analysis,
                conversation_id,
                user_id,
                routing_decision
            ):
                yield event
            return

        # 2. 单 Agent 模式：如果有 Master Agent 的决策，直接使用
        if routing_decision and routing_decision.get("selected_agent_id"):
            agent_id = routing_decision["selected_agent_id"]

            logger.info(
                "使用 Master Agent 的路由决策（流式）",
                extra={
                    "agent_id": agent_id,
                    "confidence": routing_decision.get("confidence"),
                    "reasoning": routing_decision.get("reasoning")
                }
            )
        else:
            # 2. 降级：使用旧的路由逻辑
            logger.warning("未获取到 Master Agent 决策，使用旧路由逻辑（流式）")
            # use_llm = task_analysis.get("use_llm_routing", True)
            # selected_agent_info = await self._route_by_rules(
            #     message,
            #     task_analysis["sub_agents"],
            #     use_llm=use_llm,
            #     conversation_id=str(conversation_id) if conversation_id else None
            # )
            #
            # if not selected_agent_info:
            selected_agent_info = task_analysis["sub_agents"][0]
            logger.info("未匹配到路由规则，使用默认 Agent")

            agent_id = selected_agent_info["agent_id"]

        # 3. 获取 Agent 配置
        agent_data = self.sub_agents.get(agent_id)
        if not agent_data:
            raise BusinessException(f"子 Agent 不存在: {agent_id}", BizCode.AGENT_CONFIG_MISSING)

        agent_info = agent_data.get("info", {})

        # 4. 发送路由信息事件
        yield self._format_sse_event("agent_selected", {
            "agent_id": agent_id,
            "agent_name": agent_info.get("name"),
            "routing_decision": {
                "confidence": routing_decision.get("confidence") if routing_decision else None,
                "reasoning": routing_decision.get("reasoning") if routing_decision else None,
                "strategy": routing_decision.get("strategy") if routing_decision else None
            }
        })

        # 5. 流式执行子 Agent
        sub_conversation_id = None
        async for event in self._execute_sub_agent_stream(
            agent_data["config"],
            message,
            task_analysis.get("initial_context", {}),
            conversation_id,
            user_id,
            web_search,
            memory,
            storage_type,
            user_rag_memory_id
        ):
            # 解析事件以提取 conversation_id
            if "data:" in event:
                try:
                    import json
                    data_line = event.split("data: ", 1)[1].strip()
                    data = json.loads(data_line)
                    if "conversation_id" in data:
                        sub_conversation_id = data["conversation_id"]
                except:
                    pass

            yield event

        # 6. 如果有会话 ID，发送一个包含它的事件
        if sub_conversation_id:
            yield self._format_sse_event("conversation", {
                "conversation_id": sub_conversation_id
            })

    async def _execute_conditional(
        self,
        task_analysis: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ) -> Dict[str, Any]:
        """条件路由执行（重构版 - 使用 Master Agent 的决策）

        Args:
            task_analysis: 任务分析结果（包含 Master Agent 的决策）
            conversation_id: 会话 ID
            user_id: 用户 ID

        Returns:
            执行结果
        """
        if not task_analysis["sub_agents"]:
            raise BusinessException("没有可用的子 Agent", BizCode.AGENT_CONFIG_MISSING)

        message = task_analysis.get("message", "")
        routing_decision = task_analysis.get("routing_decision")

        if not routing_decision:
            raise BusinessException("缺少路由决策", BizCode.AGENT_CONFIG_MISSING)

        agent_id = routing_decision["selected_agent_id"]

        logger.info(
            "执行 Master Agent 的路由决策",
            extra={
                "agent_id": agent_id,
                "confidence": routing_decision.get("confidence"),
                "reasoning": routing_decision.get("reasoning")
            }
        )

        # 检查是否需要协作
        if routing_decision.get("need_collaboration"):
            collaboration_strategy = routing_decision.get("collaboration_strategy", "sequential")

            # 根据策略获取协作信息
            if collaboration_strategy == "decomposition":
                # 问题拆分模式：使用 sub_questions
                collaboration_agents = routing_decision.get("sub_questions", [])
                logger.info(
                    "Master Agent 建议问题拆分",
                    extra={
                        "sub_question_count": len(collaboration_agents),
                        "strategy": collaboration_strategy
                    }
                )
            else:
                # 其他协作模式：使用 collaboration_agents
                collaboration_agents = routing_decision.get("collaboration_agents", [])
                logger.info(
                    "Master Agent 建议多 Agent 协作",
                    extra={
                        "collaboration_agent_count": len(collaboration_agents),
                        "strategy": collaboration_strategy
                    }
                )

            # 执行多 Agent 协作
            return await self._execute_collaboration(
                message=message,
                collaboration_agents=collaboration_agents,
                strategy=collaboration_strategy,
                initial_context=task_analysis.get("initial_context", {}),
                conversation_id=conversation_id,
                user_id=user_id,
                routing_decision=routing_decision
            )

        # 3. 获取 Agent 配置
        agent_data = self.sub_agents.get(agent_id)
        if not agent_data:
            raise BusinessException(f"子 Agent 不存在: {agent_id}", BizCode.AGENT_CONFIG_MISSING)

        agent_info = agent_data.get("info", {})

        logger.info(
            "执行选中的 Agent",
            extra={
                "agent_id": agent_id,
                "agent_name": agent_info.get("name"),
                "message_preview": message[:50]
            }
        )

        # 4. 执行 Agent
        result = await self._execute_sub_agent(
            agent_data["config"],
            message,
            task_analysis.get("initial_context", {}),
            conversation_id,
            user_id,
            web_search,
            memory,
            storage_type,
            user_rag_memory_id
        )

        # 5. 返回结果
        return {
            "agent_id": agent_id,
            "agent_name": agent_info.get("name"),
            "result": result,
            "conversation_id": result.get("conversation_id"),
            "routing_decision": routing_decision  # 包含 Master Agent 的决策信息
        }

    async def _execute_loop(
        self,
        task_analysis: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ) -> Dict[str, Any]:
        """循环执行（迭代优化）

        Args:
            task_analysis: 任务分析结果
            conversation_id: 会话 ID
            user_id: 用户 ID

        Returns:
            执行结果
        """
        max_iterations = self.config.execution_config.get("max_iterations", 5)

        if not task_analysis["sub_agents"]:
            raise BusinessException("没有可用的子 Agent", BizCode.AGENT_CONFIG_MISSING)

        agent_info = task_analysis["sub_agents"][0]
        agent_id = agent_info["agent_id"]
        agent_data = self.sub_agents.get(agent_id)

        if not agent_data:
            raise BusinessException(f"子 Agent 不存在: {agent_id}", BizCode.AGENT_CONFIG_MISSING)

        context = task_analysis.get("initial_context", {})
        message = task_analysis.get("message", "")

        result = None
        for i in range(max_iterations):
            logger.info(
                "循环执行 Agent",
                extra={
                    "iteration": i + 1,
                    "max_iterations": max_iterations,
                    "agent_name": agent_info.get("name")
                }
            )

            result = await self._execute_sub_agent(
                agent_data["config"],
                message,
                context,
                conversation_id,
                user_id,
                web_search,
                memory,
                storage_type,
                user_rag_memory_id
            )

            # 简化版本：执行一次就返回
            # 在实际应用中，应该验证结果是否满足条件
            break

        return {
            "agent_id": agent_id,
            "agent_name": agent_info.get("name"),
            "iterations": i + 1,
            "result": result,
            "conversation_id": result.get("conversation_id") if result else None  # 保存会话 ID
        }

    async def _execute_sub_agent_stream(
        self,
        agent_config: AgentConfig,
        message: str,
        context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ):
        """执行单个子 Agent（流式）

        Args:
            agent_config: Agent 配置
            message: 消息
            context: 上下文
            conversation_id: 会话 ID
            user_id: 用户 ID

        Yields:
            SSE 格式的事件流
        """
        from app.services.draft_run_service import DraftRunService

        # 获取模型配置
        model_config = self.db.get(ModelConfig, agent_config.default_model_config_id)
        if not model_config:
            raise BusinessException(
                "Agent 模型配置不存在",
                BizCode.AGENT_CONFIG_MISSING
            )

        # 流式执行 Agent
        draft_service = DraftRunService(self.db)
        async for event in draft_service.run_stream(
            agent_config=agent_config,
            model_config=model_config,
            message=message,
            workspace_id=agent_config.app.workspace_id,
            conversation_id=str(conversation_id) if conversation_id else None,
            user_id=user_id,
            variables=context,
            storage_type=storage_type,
            user_rag_memory_id=user_rag_memory_id,
            web_search=web_search,
            memory=memory,
            sub_agent=True
        ):
            yield event

    async def _execute_sub_agent(
        self,
        agent_config: AgentConfig,
        message: str,
        context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ) -> Dict[str, Any]:
        """执行单个子 Agent

        Args:
            agent_config: Agent 配置
            message: 消息
            context: 上下文
            conversation_id: 会话 ID
            user_id: 用户 ID


        Returns:
            执行结果
        """
        from app.services.draft_run_service import DraftRunService

        # 获取模型配置
        model_config = self.db.get(ModelConfig, agent_config.default_model_config_id)
        if not model_config:
            raise BusinessException(
                "Agent 模型配置不存在",
                BizCode.AGENT_CONFIG_MISSING
            )

        # 执行 Agent
        draft_service = DraftRunService(self.db)
        result = await draft_service.run(
            agent_config=agent_config,
            model_config=model_config,
            message=message,
            workspace_id=agent_config.app.workspace_id,
            conversation_id=str(conversation_id) if conversation_id else None,
            user_id=user_id,
            variables=context,
            web_search=web_search,
            memory=memory,
            storage_type=storage_type,
            user_rag_memory_id=user_rag_memory_id,
            sub_agent=True
        )

        return result

    async def _aggregate_results(
        self,
        results: Any
    ) -> str:
        """整合子 Agent 的结果

        Args:
            results: 子 Agent 执行结果

        Returns:
            整合后的结果
        """
        strategy = self.config.aggregation_strategy

        if strategy == AggregationStrategy.MERGE:
            return self._merge_results(results)
        elif strategy == AggregationStrategy.VOTE:
            return self._vote_results(results)
        elif strategy == AggregationStrategy.PRIORITY:
            return self._priority_results(results)
        else:
            return self._merge_results(results)

    def _merge_results(self, results: Any) -> str:
        """合并所有结果

        Args:
            results: 执行结果

        Returns:
            合并后的结果
        """
        if isinstance(results, list):
            # 顺序或并行执行的结果
            merged = []
            for item in results:
                if "result" in item:
                    agent_name = item.get("agent_name", "Agent")
                    message = item["result"].get("message", "")
                    merged.append(f"【{agent_name}】\n{message}")
                elif "error" in item:
                    agent_name = item.get("agent_name", "Agent")
                    merged.append(f"【{agent_name}】\n错误: {item['error']}")

            return "\n\n".join(merged)
        elif isinstance(results, dict):
            # 条件或循环执行的结果
            if "result" in results:
                return results["result"].get("message", "")
            return str(results)

        return str(results)

    def _vote_results(self, results: Any) -> str:
        """投票选择最佳结果（简化版本）

        Args:
            results: 执行结果

        Returns:
            最佳结果
        """
        # 简化版本：返回第一个成功的结果
        if isinstance(results, list):
            for item in results:
                if "result" in item:
                    return item["result"].get("message", "")

        return self._merge_results(results)

    def _priority_results(self, results: Any) -> str:
        """按优先级选择结果（简化版本）

        Args:
            results: 执行结果

        Returns:
            优先级最高的结果
        """
        # 简化版本：返回第一个结果
        if isinstance(results, list) and results:
            if "result" in results[0]:
                return results[0]["result"].get("message", "")

        return self._merge_results(results)

    async def _execute_collaboration_mode_stream(
        self,
        message: str,
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        web_search: bool = False,
        memory: bool = True,
        storage_type: str = '',
        user_rag_memory_id: str = ''
    ):
        """Collaboration 模式流式执行 - Agent 之间可以相互 handoff

        使用 handoffs_service 实现 Agent 之间的动态切换

        Args:
            message: 用户消息
            conversation_id: 会话 ID
            user_id: 用户 ID
            web_search: 是否启用网络搜索
            memory: 是否启用记忆
            storage_type: 存储类型
            user_rag_memory_id: RAG 记忆 ID

        Yields:
            SSE 格式的事件流
        """
        from app.services.handoffs_service import (
            convert_multi_agent_config_to_handoffs,
            HandoffsService
        )

        try:
            # 1. 构建 multi_agent_config 字典
            multi_agent_config = {
                "sub_agents": self.config.sub_agents,
                "orchestration_mode": self.config.orchestration_mode
            }

            # 2. 转换配置（每个 Agent 包含自己的 model_config）
            agent_configs = convert_multi_agent_config_to_handoffs(
                multi_agent_config,
                self.db
            )

            if not agent_configs:
                raise BusinessException("没有可用的子 Agent", BizCode.AGENT_CONFIG_MISSING)

            # 3. 创建 HandoffsService
            handoffs_service = HandoffsService(
                agent_configs=agent_configs,
                streaming=True
            )

            # 4. 使用 handoffs_service 的流式聊天
            conv_id = str(conversation_id) if conversation_id else None

            async for event in handoffs_service.chat_stream(
                message=message,
                conversation_id=conv_id
            ):
                # handoffs_service 返回的已经是 SSE 格式，直接 yield
                yield event

        except Exception as e:
            logger.error(f"Collaboration 模式执行失败: {str(e)}", exc_info=True)
            yield self._format_sse_event("error", {
                "error": str(e),
                "timestamp": time.time()
            })

    async def _execute_collaboration_mode(
        self,
        message: str,
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        variables: Optional[Dict[str, Any]] = None
    ) -> Dict[str, Any]:
        """Collaboration 模式非流式执行 - Agent 之间可以相互 handoff

        使用 handoffs_service 实现 Agent 之间的动态切换

        Args:
            message: 用户消息
            conversation_id: 会话 ID
            user_id: 用户 ID
            variables: 变量参数

        Returns:
            执行结果
        """
        from app.services.handoffs_service import (
            convert_multi_agent_config_to_handoffs,
            HandoffsService
        )

        start_time = time.time()

        try:
            # 1. 构建 multi_agent_config 字典
            multi_agent_config = {
                "sub_agents": self.config.sub_agents,
                "orchestration_mode": self.config.orchestration_mode
            }

            # 2. 转换配置（每个 Agent 包含自己的 model_config）
            agent_configs = convert_multi_agent_config_to_handoffs(
                multi_agent_config,
                self.db
            )

            if not agent_configs:
                raise BusinessException("没有可用的子 Agent", BizCode.AGENT_CONFIG_MISSING)

            # 3. 创建 HandoffsService
            handoffs_service = HandoffsService(
                agent_configs=agent_configs,
                streaming=False
            )

            # 4. 使用 handoffs_service 的非流式聊天
            conv_id = str(conversation_id) if conversation_id else None

            result = await handoffs_service.chat(
                message=message,
                conversation_id=conv_id
            )

            elapsed_time = time.time() - start_time

            return {
                "message": result.get("response", ""),
                "conversation_id": result.get("conversation_id"),
                "mode": OrchestrationMode.COLLABORATION,
                "elapsed_time": elapsed_time,
                "strategy": "collaboration",
                "active_agent": result.get("active_agent"),
                "sub_results": result,
                "usage": result.get("usage")
            }

        except Exception as e:
            logger.error(f"Collaboration 模式执行失败: {str(e)}", exc_info=True)
            raise

    def _format_sse_event(self, event: str, data: Dict[str, Any]) -> str:
        """格式化 SSE 事件

        Args:
            event: 事件类型
            data: 事件数据

        Returns:
            SSE 格式的字符串
        """
        import json
        return f"event: {event}\ndata: {json.dumps(data, ensure_ascii=False)}\n\n"

    def _load_agent(self, release_id: uuid.UUID):
        """从发布版本加载 Agent 配置

        Args:
            release_id: 发布版本 ID

        Returns:
            Agent 配置对象（包含发布版本的配置数据）
        """
        from app.models import AppRelease, App

        # 获取发布版本
        release = self.db.get(AppRelease, release_id)
        if not release:
            raise ResourceNotFoundException("发布版本", str(release_id))

        # 从发布版本的 config 中获取 Agent 配置
        config_data = release.config
        if not config_data:
            raise BusinessException(f"发布版本 {release_id} 缺少配置数据", BizCode.AGENT_CONFIG_MISSING)

        # 获取应用信息（用于 workspace_id）
        app = self.db.get(App, release.app_id)
        if not app:
            raise ResourceNotFoundException("应用", str(release.app_id))

        # 创建一个类似 AgentConfig 的对象，包含所有需要的属性
        class AgentConfigProxy:
            """Agent 配置代理对象，模拟 AgentConfig 的接口"""
            def __init__(self, release, app, config_data):
                self.id = release.id
                self.app_id = release.app_id
                self.app = app
                self.name = release.name
                self.description = release.description
                self.system_prompt = config_data.get("system_prompt")
                self.model_parameters = config_data.get("model_parameters")
                self.knowledge_retrieval = config_data.get("knowledge_retrieval")
                self.memory = config_data.get("memory")
                self.variables = config_data.get("variables", [])
                self.tools = config_data.get("tools", {})
                self.default_model_config_id = release.default_model_config_id

        return AgentConfigProxy(release, app, config_data)

    async def _execute_collaboration(
        self,
        message: str,
        collaboration_agents: List[Dict[str, Any]],
        strategy: str,
        initial_context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        routing_decision: Dict[str, Any]
    ) -> Dict[str, Any]:
        """执行多 Agent 协作

        Args:
            message: 用户消息
            collaboration_agents: 协作 Agent 列表
            strategy: 协作策略（sequential/parallel/hierarchical）
            initial_context: 初始上下文
            conversation_id: 会话 ID
            user_id: 用户 ID
            routing_decision: 路由决策

        Returns:
            协作执行结果
        """
        logger.info(
            "开始多 Agent 协作",
            extra={
                "agent_count": len(collaboration_agents),
                "strategy": strategy
            }
        )

        if strategy == "decomposition":
            # 问题拆分：每个 Agent 处理一个子问题
            return await self._execute_decomposition_collaboration(
                message, collaboration_agents, initial_context,
                conversation_id, user_id, routing_decision
            )
        elif strategy == "sequential":
            # 顺序协作：按顺序执行，后续 Agent 可以使用前面的结果
            return await self._execute_sequential_collaboration(
                message, collaboration_agents, initial_context,
                conversation_id, user_id, routing_decision
            )
        elif strategy == "parallel":
            # 并行协作：同时执行所有 Agent
            return await self._execute_parallel_collaboration(
                message, collaboration_agents, initial_context,
                conversation_id, user_id, routing_decision
            )
        elif strategy == "hierarchical":
            # 层级协作：主 Agent 协调，其他 Agent 辅助
            return await self._execute_hierarchical_collaboration(
                message, collaboration_agents, initial_context,
                conversation_id, user_id, routing_decision
            )
        else:
            # 默认使用顺序协作
            return await self._execute_sequential_collaboration(
                message, collaboration_agents, initial_context,
                conversation_id, user_id, routing_decision
            )

    def _check_dependencies(self, sub_questions: List[Dict[str, Any]]) -> bool:
        """检测子问题是否有依赖关系

        Args:
            sub_questions: 子问题列表

        Returns:
            True 如果有依赖关系，False 如果完全独立
        """
        for sub_q in sub_questions:
            depends_on = sub_q.get("depends_on", [])
            if depends_on and len(depends_on) > 0:
                logger.info(
                    "检测到依赖关系",
                    extra={
                        "question": sub_q.get("question", "")[:50],
                        "depends_on": depends_on
                    }
                )
                return True
        return False

    async def _execute_decomposition_collaboration(
        self,
        message: str,
        collaboration_agents: List[Dict[str, Any]],
        initial_context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        routing_decision: Dict[str, Any]
    ) -> Dict[str, Any]:
        """问题拆分执行

        每个 Agent 处理一个独立的子问题，避免重复

        示例：
        原问题："写一首关于雪的古诗，并计算3+8"
        拆分后：
        - 子问题1："写一首关于雪的古诗" → 文科导师
        - 子问题2："计算3+8" → 理科导师

        Args:
            collaboration_agents: 在 decomposition 模式下，这就是 sub_questions 列表
        """
        results = []

        # collaboration_agents 在 decomposition 模式下就是 sub_questions
        sub_questions = collaboration_agents

        if not sub_questions:
            # 如果没有子问题，降级到普通协作
            logger.warning(
                "问题拆分模式但没有子问题，降级到顺序协作",
                extra={
                    "collaboration_agents": collaboration_agents,
                    "routing_decision": routing_decision
                }
            )
            return await self._execute_sequential_collaboration(
                message, collaboration_agents, initial_context,
                conversation_id, user_id, routing_decision
            )

        logger.info(
            "开始问题拆分执行",
            extra={
                "sub_question_count": len(sub_questions),
                "original_message": message[:50]
            }
        )

        # 检测是否有依赖关系
        has_dependencies = self._check_dependencies(sub_questions)

        # 获取执行模式配置
        execution_mode = self.config.execution_config.get("sub_agent_execution_mode", "parallel")

        # 如果有依赖关系，强制使用串行模式
        if has_dependencies:
            logger.info("检测到子问题有依赖关系，强制使用串行执行")
            execution_mode = "sequential"

        if execution_mode == "sequential":
            # 串行执行模式
            logger.info(f"串行执行 {len(sub_questions)} 个子问题")

            # 用于存储已完成的子问题结果（按 order 索引）
            completed_results = {}

            for sub_q in sorted(sub_questions, key=lambda x: x.get("order", 0)):
                sub_question = sub_q.get("question", "")
                agent_id = sub_q.get("agent_id")
                order = sub_q.get("order", 0)
                depends_on = sub_q.get("depends_on", [])

                agent_data = self.sub_agents.get(agent_id)
                if not agent_data:
                    logger.warning(
                        f"子问题对应的 Agent 不存在: {agent_id}",
                        extra={
                            "sub_question": sub_question,
                            "available_agents": list(self.sub_agents.keys())
                        }
                    )
                    continue

                agent_name = agent_data.get("info", {}).get("name", agent_id)

                # 如果有依赖，构建包含依赖结果的上下文
                context_with_deps = initial_context.copy()
                if depends_on:
                    dependency_results = []
                    for dep_order in depends_on:
                        if dep_order in completed_results:
                            dep_result = completed_results[dep_order]
                            dependency_results.append({
                                "question": dep_result.get("sub_question"),
                                "answer": dep_result.get("result", {}).get("message", "")
                            })

                    if dependency_results:
                        context_with_deps["previous_results"] = dependency_results
                        logger.info(
                            "子问题依赖前置结果",
                            extra={
                                "current_order": order,
                                "depends_on": depends_on,
                                "dependency_count": len(dependency_results)
                            }
                        )

                logger.info(
                    "处理子问题（串行）",
                    extra={
                        "sub_question": sub_question,
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "has_dependencies": bool(depends_on)
                    }
                )

                # 串行执行
                try:
                    result = await self._execute_sub_agent(
                        agent_data["config"],
                        sub_question,
                        context_with_deps,  # 使用包含依赖结果的上下文
                        conversation_id,
                        user_id
                    )
                    result_entry = {
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "sub_question": sub_question,
                        "result": result,
                        "conversation_id": result.get("conversation_id"),
                        "order": order
                    }
                    results.append(result_entry)
                    completed_results[order] = result_entry  # 保存结果供后续依赖使用
                except Exception as e:
                    logger.error(f"子问题执行失败: {str(e)}")
                    results.append({
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "sub_question": sub_question,
                        "error": str(e),
                        "order": order
                    })
        else:
            # 并行执行模式（默认）
            tasks = []
            agent_infos = []

            for sub_q in sorted(sub_questions, key=lambda x: x.get("order", 0)):
                sub_question = sub_q.get("question", "")
                agent_id = sub_q.get("agent_id")

                agent_data = self.sub_agents.get(agent_id)
                if not agent_data:
                    logger.warning(f"子问题对应的 Agent 不存在: {agent_id}")
                    continue

                agent_name = agent_data.get("info", {}).get("name", agent_id)

                logger.info(
                    "准备处理子问题（并行）",
                    extra={
                        "sub_question": sub_question,
                        "agent_id": agent_id,
                        "agent_name": agent_name
                    }
                )

                # 创建异步任务
                task = self._execute_sub_agent(
                    agent_data["config"],
                    sub_question,
                    initial_context,
                    conversation_id,
                    user_id
                )
                tasks.append(task)
                agent_infos.append({
                    "agent_id": agent_id,
                    "agent_name": agent_name,
                    "sub_question": sub_question
                })

            # 并行执行所有任务
            logger.info(f"并行执行 {len(tasks)} 个子问题")
            task_results = await asyncio.gather(*tasks, return_exceptions=True)

            # 处理结果
            for i, result in enumerate(task_results):
                if isinstance(result, Exception):
                    logger.error(f"子问题执行失败: {str(result)}")
                    results.append({
                        "agent_id": agent_infos[i]["agent_id"],
                        "agent_name": agent_infos[i]["agent_name"],
                        "sub_question": agent_infos[i]["sub_question"],
                        "error": str(result)
                    })
                else:
                    results.append({
                        "agent_id": agent_infos[i]["agent_id"],
                        "agent_name": agent_infos[i]["agent_name"],
                        "sub_question": agent_infos[i]["sub_question"],
                        "result": result,
                        "conversation_id": result.get("conversation_id")
                    })

        # 整合结果（问题拆分模式）
        final_response = await self._merge_decomposition_results(results, message)

        return {
            "agent_id": "decomposition",
            "agent_name": "问题拆分协作",
            "result": {
                "message": final_response,
                "conversation_id": results[0].get("conversation_id") if results else None
            },
            "conversation_id": results[0].get("conversation_id") if results else None,
            "routing_decision": routing_decision,
            "collaboration_results": results
        }

    async def _execute_sequential_collaboration(
        self,
        message: str,
        collaboration_agents: List[Dict[str, Any]],
        initial_context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        routing_decision: Dict[str, Any]
    ) -> Dict[str, Any]:
        """顺序协作执行

        每个 Agent 按顺序执行，后续 Agent 可以看到前面 Agent 的结果
        """
        results = []
        context = initial_context.copy()
        accumulated_response = []

        # 按 order 排序
        sorted_agents = sorted(collaboration_agents, key=lambda x: x.get("order", 0))

        for agent_info in sorted_agents:
            agent_id = agent_info["agent_id"]
            agent_data = self.sub_agents.get(agent_id)

            if not agent_data:
                logger.warning(f"协作 Agent 不存在: {agent_id}")
                continue

            agent_name = agent_data.get("info", {}).get("name", agent_id)
            agent_task = agent_info.get("task", "处理任务")

            logger.info(
                "执行协作 Agent",
                extra={
                    "agent_id": agent_id,
                    "agent_name": agent_name,
                    "role": agent_info.get("role"),
                    "task": agent_task,
                    "order": agent_info.get("order")
                }
            )

            # 构建该 Agent 的消息（包含任务说明和前面的结果）
            agent_message = message
            if context.get("previous_results"):
                agent_message = f"""原始问题：{message}

你的任务：{agent_task}

前面专家的分析结果：
{context['previous_results']}

请基于以上信息，完成你的任务。"""

            # 执行 Agent
            result = await self._execute_sub_agent(
                agent_data["config"],
                agent_message,
                context,
                conversation_id,
                user_id
            )

            agent_response = result.get("message", "")

            results.append({
                "agent_id": agent_id,
                "agent_name": agent_name,
                "role": agent_info.get("role"),
                "task": agent_task,
                "result": result,
                "conversation_id": result.get("conversation_id")
            })

            # 更新上下文
            context[f"result_from_{agent_name}"] = agent_response

            # 累积响应
            accumulated_response.append(f"【{agent_name}】\n{agent_response}")

            # 更新 previous_results 供下一个 Agent 使用
            context["previous_results"] = "\n\n".join(accumulated_response)

        # 整合最终结果
        final_response = await self._merge_collaboration_results(
            results,
            strategy="sequential",
            original_question=message
        )

        return {
            "agent_id": "collaboration",
            "agent_name": "多Agent协作",
            "result": {
                "message": final_response,
                "conversation_id": results[0].get("conversation_id") if results else None
            },
            "conversation_id": results[0].get("conversation_id") if results else None,
            "routing_decision": routing_decision,
            "collaboration_results": results
        }

    async def _execute_parallel_collaboration(
        self,
        message: str,
        collaboration_agents: List[Dict[str, Any]],
        initial_context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        routing_decision: Dict[str, Any]
    ) -> Dict[str, Any]:
        """并行协作执行

        所有 Agent 同时执行，互不依赖
        """
        tasks = []
        agent_infos = []

        for agent_info in collaboration_agents:
            agent_id = agent_info["agent_id"]
            agent_data = self.sub_agents.get(agent_id)

            if not agent_data:
                continue

            agent_task = agent_info.get("task", "处理任务")

            # 构建该 Agent 的消息
            agent_message = f"""原始问题：{message}

你的任务：{agent_task}

请完成你的任务。"""

            # 创建任务
            task = self._execute_sub_agent(
                agent_data["config"],
                agent_message,
                initial_context.copy(),
                conversation_id,
                user_id
            )
            tasks.append(task)
            agent_infos.append((agent_id, agent_data, agent_info))

        # 并行执行
        task_results = await asyncio.gather(*tasks, return_exceptions=True)

        # 处理结果
        results = []
        for (agent_id, agent_data, agent_info), result in zip(agent_infos, task_results, strict=False):
            agent_name = agent_data.get("info", {}).get("name", agent_id)

            if isinstance(result, Exception):
                logger.error(f"协作 Agent 执行失败: {agent_name}", extra={"error": str(result)})
                results.append({
                    "agent_id": agent_id,
                    "agent_name": agent_name,
                    "error": str(result)
                })
            else:
                results.append({
                    "agent_id": agent_id,
                    "agent_name": agent_name,
                    "role": agent_info.get("role"),
                    "task": agent_info.get("task"),
                    "result": result,
                    "conversation_id": result.get("conversation_id")
                })

        # 整合结果
        final_response = await self._merge_collaboration_results(
            results,
            strategy="parallel",
            original_question=message
        )

        return {
            "agent_id": "collaboration",
            "agent_name": "多Agent协作",
            "result": {
                "message": final_response,
                "conversation_id": results[0].get("conversation_id") if results else None
            },
            "conversation_id": results[0].get("conversation_id") if results else None,
            "routing_decision": routing_decision,
            "collaboration_results": results
        }

    async def _execute_hierarchical_collaboration(
        self,
        message: str,
        collaboration_agents: List[Dict[str, Any]],
        initial_context: Dict[str, Any],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str],
        routing_decision: Dict[str, Any]
    ) -> Dict[str, Any]:
        """层级协作执行

        主 Agent（primary）负责协调，其他 Agent 提供辅助信息
        """
        # 找到主 Agent 和辅助 Agents
        primary_agent = None
        secondary_agents = []

        for agent_info in collaboration_agents:
            if agent_info.get("role") == "primary":
                primary_agent = agent_info
            else:
                secondary_agents.append(agent_info)

        if not primary_agent:
            # 如果没有指定主 Agent，使用第一个
            primary_agent = collaboration_agents[0]
            secondary_agents = collaboration_agents[1:]

        # 1. 先执行辅助 Agents（并行）
        secondary_results = []
        if secondary_agents:
            tasks = []
            agent_infos = []

            for agent_info in secondary_agents:
                agent_id = agent_info["agent_id"]
                agent_data = self.sub_agents.get(agent_id)

                if not agent_data:
                    continue

                agent_task = agent_info.get("task", "提供专业意见")
                agent_message = f"""问题：{message}

请从你的专业角度提供意见：{agent_task}"""

                task = self._execute_sub_agent(
                    agent_data["config"],
                    agent_message,
                    initial_context.copy(),
                    conversation_id,
                    user_id
                )
                tasks.append(task)
                agent_infos.append((agent_id, agent_data, agent_info))

            # 并行执行辅助 Agents
            task_results = await asyncio.gather(*tasks, return_exceptions=True)

            for (agent_id, agent_data, agent_info), result in zip(agent_infos, task_results, strict=False):
                agent_name = agent_data.get("info", {}).get("name", agent_id)

                if not isinstance(result, Exception):
                    secondary_results.append({
                        "agent_id": agent_id,
                        "agent_name": agent_name,
                        "role": "secondary",
                        "result": result
                    })

        # 2. 执行主 Agent（整合辅助 Agents 的结果）
        primary_agent_id = primary_agent["agent_id"]
        primary_agent_data = self.sub_agents.get(primary_agent_id)

        if not primary_agent_data:
            raise BusinessException(f"主协作 Agent 不存在: {primary_agent_id}", BizCode.AGENT_CONFIG_MISSING)

        # 构建主 Agent 的消息（包含辅助 Agents 的结果）
        primary_message = f"""问题：{message}

你的任务：{primary_agent.get('task', '综合分析并给出最终答案')}
"""

        if secondary_results:
            expert_opinions = []
            for sec_result in secondary_results:
                expert_opinions.append(
                    f"【{sec_result['agent_name']}的意见】\n{sec_result['result'].get('message', '')}"
                )

            primary_message += f"""

其他专家的意见：
{chr(10).join(expert_opinions)}

请综合以上专家意见，给出你的最终答案。"""

        # 执行主 Agent
        primary_result = await self._execute_sub_agent(
            primary_agent_data["config"],
            primary_message,
            initial_context,
            conversation_id,
            user_id
        )

        primary_agent_name = primary_agent_data.get("info", {}).get("name", primary_agent_id)

        # 整合所有结果
        all_results = [*secondary_results, {"agent_id": primary_agent_id, "agent_name": primary_agent_name, "role": "primary", "result": primary_result, "conversation_id": primary_result.get("conversation_id")}]

        return {
            "agent_id": primary_agent_id,
            "agent_name": primary_agent_name,
            "result": primary_result,
            "conversation_id": primary_result.get("conversation_id"),
            "routing_decision": routing_decision,
            "collaboration_results": all_results
        }

    async def _merge_decomposition_results(
        self,
        results: List[Dict[str, Any]],
        original_question: str = None
    ) -> str:
        """整合问题拆分的结果

        每个 Agent 处理了不同的子问题，需要按顺序组合

        Args:
            results: 结果列表，每个包含 sub_question 和 result
            original_question: 原始用户问题

        Returns:
            整合后的响应
        """
        if not results:
            return "未获取到有效结果"

        # 获取整合模式
        merge_mode = self.config.execution_config.get("result_merge_mode", "smart")

        if merge_mode == "master":
            # 使用 Master Agent 整合
            return await self._master_merge_results(results, "decomposition", original_question)
        else:
            # smart 模式：直接组合答案
            parts = []
            for result in results:
                message = result.get("result", {}).get("message", "")
                if message:
                    parts.append(message)

            return "\n\n".join(parts)

    async def _merge_collaboration_results(
        self,
        results: List[Dict[str, Any]],
        strategy: str,
        original_question: str = None
    ) -> str:
        """整合协作结果（智能去重和合并）

        Args:
            results: 协作结果列表
            strategy: 协作策略
            original_question: 原始用户问题

        Returns:
            整合后的响应
        """
        if not results:
            logger.error(
                "协作结果为空",
                extra={
                    "strategy": strategy,
                    "has_original_question": bool(original_question)
                }
            )
            return "协作执行失败，没有可用结果"

        # 获取整合策略配置
        merge_mode = self.config.execution_config.get("result_merge_mode", "smart")

        if merge_mode == "master":
            # Master Agent 整合：让 Master Agent 结合原始问题和子 Agent 答案生成最终回复
            return await self._master_merge_results(results, strategy, original_question)
        else:
            # 默认使用智能整合
            return self._smart_merge_results(results, strategy)

    def _smart_merge_results(
        self,
        results: List[Dict[str, Any]],
        strategy: str
    ) -> str:
        """智能整合结果（去重、提取关键信息）

        适用场景：多个 Agent 回答相似问题，需要去重和优化

        注意：在流式场景下，用户已经看到了所有 Agent 的输出，
        这个方法主要用于生成一个"整合后的版本"供后续使用（如保存到数据库）
        """
        if not results:
            return ""

        # 提取所有消息
        messages = []
        for result in results:
            if "error" in result:
                continue
            message = result.get("result", {}).get("message", "")
            if message:
                messages.append(message)

        if not messages:
            return ""

        if len(messages) == 1:
            # 只有一个结果，直接返回
            return messages[0]

        # 多个结果：根据策略智能整合
        if strategy == "decomposition":
            # 问题拆分：将所有子问题的答案合并
            # 按顺序组合各个 Agent 的回答
            merged_parts = []
            for result in results:
                if "error" in result:
                    continue
                agent_name = result.get("agent_name", "")
                sub_question = result.get("sub_question", "")
                message = result.get("result", {}).get("message", "")
                if message:
                    if sub_question:
                        merged_parts.append(f"**{sub_question}**\n{message}")
                    else:
                        merged_parts.append(message)

            if merged_parts:
                return "\n\n".join(merged_parts)
            return ""

        elif strategy == "sequential":
            # 顺序协作：返回最后一个 Agent 的结果（它包含了前面的信息）
            return self._merge_sequential_smart(results)

        elif strategy == "parallel":
            # 并行协作：检查是否需要去重
            return self._merge_parallel_smart(results)

        elif strategy == "hierarchical":
            # 层级协作：只返回主 Agent 的结果
            return self._merge_hierarchical_smart(results)

        else:
            # 默认：返回最完整的一个
            return max(messages, key=len)

    def _merge_sequential_smart(self, results: List[Dict[str, Any]]) -> str:
        """智能整合顺序协作结果

        顺序协作的特点：后续 Agent 会引用前面的结果
        策略：只保留最后一个 Agent 的完整回答（它已经包含了前面的信息）
        """
        if not results:
            return ""

        # 获取最后一个成功的结果
        for result in reversed(results):
            if "error" not in result:
                message = result.get("result", {}).get("message", "")
                if message:
                    return message

        return "未获取到有效结果"

    def _merge_parallel_smart(self, results: List[Dict[str, Any]]) -> str:
        """智能整合并行协作结果

        并行协作的特点：多个独立观点
        策略：
        1. 如果回答高度相似（重复），只保留一个
        2. 如果回答不同，合并所有观点（但不显示 Agent 名称）
        """
        messages = []
        for result in results:
            if "error" in result:
                continue
            message = result.get("result", {}).get("message", "")
            if message:
                messages.append(message)

        if not messages:
            return "未获取到有效结果"

        if len(messages) == 1:
            return messages[0]

        # 检查相似度
        similarity = self._calculate_similarity(messages)

        if similarity > 0.7:
            # 高度相似，只返回最长的一个
            return max(messages, key=len)
        else:
            # 不同观点，合并（不显示 Agent 名称）
            # 使用分隔符区分不同部分
            return "\n\n---\n\n".join(messages)

    def _merge_hierarchical_smart(self, results: List[Dict[str, Any]]) -> str:
        """智能整合层级协作结果

        层级协作的特点：主 Agent 已经综合了辅助 Agent 的意见
        策略：只返回主 Agent 的结果
        """
        # 找到主 Agent 的结果
        for result in results:
            if result.get("role") == "primary":
                message = result.get("result", {}).get("message", "")
                if message:
                    return message

        # 如果没有找到主 Agent，返回最后一个
        if results:
            last_result = results[-1]
            return last_result.get("result", {}).get("message", "")

        return "未获取到有效结果"

    async def _master_merge_results(
        self,
        results: List[Dict[str, Any]],
        strategy: str,
        original_question: str = None
    ) -> str:
        """使用 Master Agent 整合多个子 Agent 的结果

        Args:
            results: 子 Agent 的响应结果列表
            strategy: 协作策略
            original_question: 原始用户问题

        Returns:
            Master Agent 整合后的最终回复
        """
        if not results:
            return "没有收到任何 Agent 的响应"

        if len(results) == 1:
            # 只有一个结果，直接返回
            return results[0].get('result', {}).get('message', '')

        # 构建子 Agent 回答的汇总
        agent_responses = []
        for i, result in enumerate(results, 1):
            if "error" in result:
                continue

            agent_name = result.get('agent_name', f'Agent {i}')
            task = result.get('task', '')
            message = result.get('result', {}).get('message', '')

            if message:
                response_info = {
                    'agent_name': agent_name,
                    'task': task,
                    'response': message
                }
                agent_responses.append(response_info)

        if not agent_responses:
            return "未获取到有效结果"

        # 构建 Master Agent 的整合 prompt
        responses_text = ""
        for resp in agent_responses:
            agent_name = resp['agent_name']
            task = resp['task']
            response = resp['response']

            if task:
                responses_text += f"\n### {agent_name}（任务：{task}）的回答：\n{response}\n"
            else:
                responses_text += f"\n### {agent_name} 的回答：\n{response}\n"

        # 根据策略调整整合指令
        strategy_instructions = {
            "decomposition": "这些是针对不同子问题的回答，请将它们整合成一个完整、连贯的答案。",
            "sequential": "这些是按顺序协作的结果，后面的 Agent 可能依赖前面的结果，请整合成最终答案。",
            "parallel": "这些是从不同角度并行分析的结果，请综合这些观点给出全面的答案。",
            "hierarchical": "这些是层级协作的结果，请综合各方意见给出最终答案。"
        }

        strategy_instruction = strategy_instructions.get(strategy, "请整合这些回答，生成统一的最终答案。")

        question_context = f"\n**原始问题**：{original_question}\n" if original_question else ""

        merge_prompt = f"""你是一个智能助手，现在需要整合多个专业 Agent 的回答，生成一个统一、连贯、完整的最终答案。
{question_context}
**各个专业 Agent 的回答**：
{responses_text}

**整合要求**：
{strategy_instruction}

请注意：
1. 结合原始问题和各个 Agent 的专业回答
2. 去除重复内容，保留所有有价值的信息
3. 确保答案逻辑清晰、表达流畅
4. 如果不同 Agent 的观点有冲突，请合理说明
5. 直接给出整合后的答案，不要添加"根据以上回答"等元信息

请生成最终的整合答案："""

        try:
            # 调用 Master Agent 的 LLM 进行整合
            from app.core.models import RedBearLLM
            from app.core.models.base import RedBearModelConfig
            from app.models import ModelApiKey, ModelType

            # 获取 Master Agent 的模型配置
            default_model_config_id = self.config.default_model_config_id
            if not default_model_config_id:
                logger.warning("没有配置 Master Agent，使用简单整合")
                return self._smart_merge_results(results, strategy)

            # 获取 API Key 配置
            # api_key_config = self.db.query(ModelApiKey).join(
            #     ModelConfig, ModelApiKey.model_configs
            # ).filter(
            #     ModelConfig.id == default_model_config_id,
            #     ModelApiKey.is_active.is_(True)
            # ).first()
            api_keys = ModelApiKeyRepository.get_by_model_config(self.db, default_model_config_id)
            api_key_config = api_keys[0] if api_keys else None

            if not api_key_config:
                logger.warning("Master Agent 没有可用的 API Key，使用简单整合")
                return self._smart_merge_results(results, strategy)

            logger.info(
                "使用 Master Agent 整合结果",
                extra={
                    "agent_count": len(agent_responses),
                    "strategy": strategy,
                    "has_original_question": bool(original_question)
                }
            )

            # 创建 RedBearModelConfig
            model_config = RedBearModelConfig(
                model_name=api_key_config.model_name,
                provider=api_key_config.provider,
                api_key=api_key_config.api_key,
                base_url=api_key_config.api_base,
                temperature=0.7,  # 整合任务使用中等温度
                max_tokens=2000
            )

            # 创建 LLM 实例
            llm = RedBearLLM(model_config, type=ModelType.CHAT)

            # 调用模型进行整合
            response = await llm.ainvoke(merge_prompt)

            # 提取响应内容
            if hasattr(response, 'content'):
                merged_response = response.content
            else:
                merged_response = str(response)

            logger.info(
                "Master Agent 整合完成",
                extra={
                    "merged_length": len(merged_response)
                }
            )

            return merged_response

        except Exception as e:
            logger.error(f"Master Agent 整合失败: {str(e)}")
            # 降级到智能整合
            return self._smart_merge_results(results, strategy)

    async def _master_merge_results_stream(
        self,
        results: List[Dict[str, Any]],
        strategy: str,
        original_question: str = None
    ):
        """使用 Master Agent 流式整合多个子 Agent 的结果

        Args:
            results: 子 Agent 的响应结果列表
            strategy: 协作策略
            original_question: 原始用户问题

        Yields:
            SSE 格式的事件流
        """
        if not results:
            yield self._format_sse_event("message", {"content": "没有收到任何 Agent 的响应"})
            return

        if len(results) == 1:
            # 只有一个结果，直接返回
            yield self._format_sse_event("message", {
                "content": results[0].get('result', {}).get('message', '')
            })
            return

        # 构建子 Agent 回答的汇总（与非流式版本相同）
        agent_responses = []
        for i, result in enumerate(results, 1):
            if "error" in result:
                continue

            agent_name = result.get('agent_name', f'Agent {i}')
            task = result.get('task', '')
            sub_question = result.get('sub_question', '')
            message = result.get('result', {}).get('message', '')

            if message:
                response_info = {
                    'agent_name': agent_name,
                    'task': task or sub_question,
                    'response': message
                }
                agent_responses.append(response_info)

        if not agent_responses:
            yield self._format_sse_event("message", {"content": "未获取到有效结果"})
            return

        # 构建整合 prompt
        responses_text = ""
        for resp in agent_responses:
            agent_name = resp['agent_name']
            task = resp['task']
            response = resp['response']

            if task:
                responses_text += f"\n### {agent_name}（任务：{task}）的回答：\n{response}\n"
            else:
                responses_text += f"\n### {agent_name} 的回答：\n{response}\n"

        strategy_instructions = {
            "decomposition": "这些是针对不同子问题的回答，请将它们整合成一个完整、连贯的答案。",
            "sequential": "这些是按顺序协作的结果，后面的 Agent 可能依赖前面的结果，请整合成最终答案。",
            "parallel": "这些是从不同角度并行分析的结果，请综合这些观点给出全面的答案。",
            "hierarchical": "这些是层级协作的结果，请综合各方意见给出最终答案。"
        }

        strategy_instruction = strategy_instructions.get(strategy, "请整合这些回答，生成统一的最终答案。")
        question_context = f"\n**原始问题**：{original_question}\n" if original_question else ""

        merge_prompt = f"""你是一个智能助手，现在需要整合多个专业 Agent 的回答，生成一个统一、连贯、完整的最终答案。
{question_context}
**各个专业 Agent 的回答**：
{responses_text}

**整合要求**：
{strategy_instruction}

请注意：
1. 结合原始问题和各个 Agent 的专业回答
2. 去除重复内容，保留所有有价值的信息
3. 确保答案逻辑清晰、表达流畅
4. 如果不同 Agent 的观点有冲突，请合理说明
5. 直接给出整合后的答案，不要添加"根据以上回答"等元信息

请生成最终的整合答案："""

        try:
            from app.core.models import RedBearLLM
            from app.core.models.base import RedBearModelConfig
            from app.models import ModelApiKey, ModelType

            # 获取 Master Agent 的模型配置
            default_model_config_id = self.config.default_model_config_id
            if not default_model_config_id:
                logger.warning("没有配置 Master Agent，使用简单整合")
                final_response = self._smart_merge_results(results, strategy)
                yield self._format_sse_event("message", {"content": final_response})
                return

            # 获取 API Key 配置
            # api_key_config = self.db.query(ModelApiKey).join(
            #     ModelConfig, ModelApiKey.model_configs
            # ).filter(
            #     ModelConfig.id == default_model_config_id,
            #     ModelApiKey.is_active.is_(True)
            # ).first()
            api_keys = ModelApiKeyRepository.get_by_model_config(self.db, default_model_config_id)
            api_key_config = api_keys[0] if api_keys else None

            if not api_key_config:
                logger.warning("Master Agent 没有可用的 API Key，使用简单整合")
                final_response = self._smart_merge_results(results, strategy)
                yield self._format_sse_event("message", {"content": final_response})
                return

            logger.info(
                "开始 Master Agent 流式整合",
                extra={
                    "agent_count": len(agent_responses),
                    "strategy": strategy
                }
            )

            # 创建 RedBearModelConfig（启用流式）
            model_config = RedBearModelConfig(
                model_name=api_key_config.model_name,
                provider=api_key_config.provider,
                api_key=api_key_config.api_key,
                base_url=api_key_config.api_base,
                temperature=0.7,
                max_tokens=2000,
                extra_params={"streaming": True}  # 启用流式输出
            )

            # 创建 LLM 实例
            llm = RedBearLLM(model_config, type=ModelType.CHAT)

            logger.info("开始流式调用 Master Agent LLM")

            # 流式调用模型进行整合
            try:
                chunk_count = 0
                logger.debug(f"开始流式调用，provider={api_key_config.provider}")

                # 获取底层模型
                underlying_model = llm._model if hasattr(llm, '_model') else llm
                logger.debug(f"底层模型类型: {type(underlying_model).__name__}")

                # 使用底层模型的 astream 方法直接流式输出
                # 这样可以绕过可能的包装器累积问题
                async for chunk in underlying_model.astream(merge_prompt):
                    chunk_count += 1

                    # 提取内容
                    if hasattr(chunk, 'content'):
                        content = chunk.content
                    elif isinstance(chunk, str):
                        content = chunk
                    else:
                        content = str(chunk)

                    if content:
                        if chunk_count <= 5:
                            logger.debug(f"收到流式 chunk #{chunk_count}: {content[:30]}...")
                        yield self._format_sse_event("message", {"content": content})

                logger.info(f"Master Agent 流式整合完成，共 {chunk_count} 个 chunks")

            except AttributeError as e:
                # 如果底层模型不支持流式，降级到非流式
                logger.warning(f"底层模型不支持流式，降级到非流式: {str(e)}")
                response = await llm.ainvoke(merge_prompt)
                if hasattr(response, 'content'):
                    content = response.content
                else:
                    content = str(response)
                yield self._format_sse_event("message", {"content": content})

        except Exception as e:
            logger.error(f"Master Agent 流式整合失败: {str(e)}")
            # 降级到智能整合
            final_response = self._smart_merge_results(results, strategy)
            yield self._format_sse_event("message", {"content": final_response})

    def _should_merge_results(
        self,
        results: List[Dict[str, Any]],
        strategy: str
    ) -> bool:
        """判断是否需要整合结果

        Args:
            results: Agent 执行结果
            strategy: 协作策略

        Returns:
            True 如果需要整合，False 如果不需要
        """
        if not results or len(results) == 1:
            # 没有结果或只有一个结果，不需要整合
            return False

        if strategy == "decomposition":
            # 问题拆分：每个子问题独立，用户已经看到所有答案
            # 通常不需要整合（除非配置要求）
            return self.config.execution_config.get("force_merge_decomposition", True)

        if strategy == "hierarchical":
            # 层级协作：主 Agent 已经整合了，不需要再整合
            return False

        # sequential 和 parallel 模式：可能需要整合去重
        return True

    async def _parallel_stream_agents(
        self,
        agent_tasks: List[Tuple[str, str, Any, str, Dict[str, Any]]],
        conversation_id: Optional[uuid.UUID],
        user_id: Optional[str]
    ) -> AsyncIterator[Tuple[str, str, str, str]]:
        """并行流式执行多个 Agent，实时返回结果

        Args:
            agent_tasks: [(agent_id, agent_name, agent_config, message, context), ...]
            conversation_id: 会话 ID
            user_id: 用户 ID

        Yields:
            (agent_id, agent_name, event_type, content) 元组
        """
        # 为每个 Agent 创建异步生成器
        async def stream_single_agent(agent_id, agent_name, agent_config, message, context):
            """单个 Agent 的流式执行包装器"""
            try:
                async for event in self._execute_sub_agent_stream(
                    agent_config,
                    message,
                    context,
                    conversation_id,
                    user_id
                ):
                    # 解析事件
                    if "data:" in event:
                        try:
                            import json
                            data_line = event.split("data: ", 1)[1].strip()
                            data = json.loads(data_line)

                            if "content" in data:
                                yield (agent_id, agent_name, "content", data["content"])
                        except:
                            pass

                # 发送完成信号
                yield (agent_id, agent_name, "done", "")

            except Exception as e:
                logger.error(f"Agent {agent_name} 流式执行失败: {str(e)}")
                yield (agent_id, agent_name, "error", str(e))

        # 创建所有 Agent 的流式任务
        streams = []
        for agent_id, agent_name, agent_config, message, context in agent_tasks:
            stream = stream_single_agent(agent_id, agent_name, agent_config, message, context)
            streams.append(stream)

        # 使用队列来合并多个异步流
        queue = asyncio.Queue()
        active_streams = len(streams)

        async def consume_stream(stream, stream_id):
            """消费单个流并放入队列"""
            nonlocal active_streams
            try:
                async for item in stream:
                    await queue.put(item)
            finally:
                active_streams -= 1
                if active_streams == 0:
                    await queue.put(None)  # 所有流都完成了

        # 启动所有流的消费任务
        tasks = [
            asyncio.create_task(consume_stream(stream, i))
            for i, stream in enumerate(streams)
        ]

        # 从队列中读取并 yield
        while True:
            item = await queue.get()
            if item is None:  # 所有流都完成
                break
            yield item

        # 等待所有任务完成
        await asyncio.gather(*tasks, return_exceptions=True)

    def _calculate_similarity(self, messages: List[str]) -> float:
        """计算消息相似度（简化版）

        Args:
            messages: 消息列表

        Returns:
            相似度 (0-1)
        """
        if len(messages) < 2:
            return 0.0

        # 简化版：比较长度和关键词
        # 实际应用中可以使用更复杂的算法（如编辑距离、余弦相似度等）

        # 计算平均长度
        avg_length = sum(len(m) for m in messages) / len(messages)

        # 如果长度差异很大，认为不相似
        length_variance = sum(abs(len(m) - avg_length) for m in messages) / len(messages)
        if length_variance > avg_length * 0.5:
            return 0.3

        # 提取关键词（简化：取前50个字符）
        keywords = [m[:50] for m in messages]

        # 计算重复度
        unique_keywords = len(set(keywords))
        total_keywords = len(keywords)

        similarity = 1.0 - (unique_keywords / total_keywords)

        return similarity