JARVIS/backend/app/agents/graph.py

"""
Jarvis LangGraph Agent 主图定义 - 优化重构版
"""

import json
import logging
import re
from typing import Literal, Union, List, Any

from langchain_core.messages import (
    BaseMessage, 
    HumanMessage, 
    AIMessage, 
    SystemMessage, 
    ToolMessage
)
from langgraph.graph import StateGraph, END

from app.agents.state import AgentState, AgentRole
from app.agents.prompts import (
    MASTER_SYSTEM_PROMPT,
    SCHEDULE_PLANNER_SYSTEM_PROMPT,
    EXECUTOR_SYSTEM_PROMPT,
    LIBRARIAN_SYSTEM_PROMPT,
    ANALYST_SYSTEM_PROMPT,
    JSON_ACTION_FALLBACK_PROMPT,
)
from app.agents.tools import ALL_TOOLS, SUB_COMMANDER_TOOLSETS
from app.agents.tools.time_reasoning import normalize_tool_time_arguments
from app.agents.skill_registry import build_skill_context
from app.services.llm_service import (
    get_llm, 
    create_llm_from_config, 
    resolve_provider_capabilities, 
    default_provider_capabilities
)
from app.logging_utils import summarize_llm_config

logger = logging.getLogger("jarvis.agent")

# ===================== 工具辅助函数 =====================

def _get_llm_for_state(state: AgentState):
    """获取配置好的 LLM 实例"""
    user_llm_config = state.get("user_llm_config")
    llm = create_llm_from_config(user_llm_config) if user_llm_config else get_llm()
    
    # 注入解析到的能力
    capabilities = getattr(llm, "_jarvis_provider_capabilities", None)
    if capabilities is None:
        capabilities = resolve_provider_capabilities(user_llm_config) if user_llm_config else default_provider_capabilities()
    
    state["provider_capabilities"] = {
        "provider": capabilities.provider,
        "supports_native_tools": capabilities.supports_native_tools,
        "preferred_tool_strategy": capabilities.preferred_tool_strategy,
    }
    return llm, capabilities


def _filter_user_messages(messages: list[BaseMessage]) -> list[BaseMessage]:
    return [m for m in messages if m.type in ("human", "user")]


def _get_role_tools(role: AgentRole) -> list:
    """获取角色对应的所有可用工具集"""
    if role == AgentRole.SCHEDULE_PLANNER:
        # 合并分析和规划工具
        return list(set(SUB_COMMANDER_TOOLSETS["schedule_analysis"] + SUB_COMMANDER_TOOLSETS["schedule_planning"]))
    if role == AgentRole.EXECUTOR:
        return list(set(SUB_COMMANDER_TOOLSETS["executor_tasks"] + SUB_COMMANDER_TOOLSETS["executor_forum"]))
    if role == AgentRole.LIBRARIAN:
        return list(set(SUB_COMMANDER_TOOLSETS["librarian_retrieval"] + SUB_COMMANDER_TOOLSETS["librarian_graph"]))
    if role == AgentRole.ANALYST:
        return list(set(SUB_COMMANDER_TOOLSETS["analyst_progress"] + SUB_COMMANDER_TOOLSETS["analyst_insights"]))
    return []


# ===================== 核心执行逻辑 (ReAct) =====================

async def call_agent_llm(state: AgentState, role: AgentRole, system_prompt: str) -> dict:
    """通用的 LLM 调用节点逻辑"""
    llm, capabilities = _get_llm_for_state(state)
    tools = _get_role_tools(role)
    
    # 构建消息序列
    messages = []
    
    # 1. 系统提示词
    messages.append(SystemMessage(content=system_prompt))
    
    # 2. 环境上下文 (时间、记忆等)
    if state.get("current_datetime_context"):
        messages.append(SystemMessage(content=f"当前时间上下文: {state['current_datetime_context']}"))
    
    if state.get("memory_context"):
        messages.append(SystemMessage(content=f"长期记忆上下文: {state['memory_context']}"))
        
    # 3. 技能增强
    role_skill_key = role.value.replace("agent_", "")
    skill_ctx = build_skill_context(role_skill_key)
    if skill_ctx:
        messages.append(SystemMessage(content=skill_ctx))
        
    # 4. 历史对话 (add_messages 已经处理好了)
    messages.extend(state["messages"])
    
    # 绑定工具
    if tools and capabilities.supports_native_tools:
        llm_with_tools = llm.bind_tools(tools)
    else:
        llm_with_tools = llm
        if tools: # 如果有工具但不支持原生，注入 JSON Fallback 提示
             messages.append(SystemMessage(content=JSON_ACTION_FALLBACK_PROMPT))
             tool_names = [t.name for t in tools]
             messages.append(SystemMessage(content=f"本次可用工具列表: {', '.join(tool_names)}"))

    logger.info(
        f"agent_node_started",
        extra={
            "details": {
                "role": role.value,
                "message_count": len(messages),
                "tool_count": len(tools),
                "provider": capabilities.provider
            }
        }
    )
    
    # 执行调用
    response = await llm_with_tools.ainvoke(messages)
    
    logger.info(
        f"agent_node_finished",
        extra={
            "details": {
                "role": role.value,
                "has_tool_calls": bool(getattr(response, "tool_calls", None)),
                "content_length": len(response.content) if response.content else 0
            }
        }
    )
    
    return {"messages": [response]}


async def execute_tools_node(state: AgentState) -> dict:
    """执行工具调用并返回 ToolMessage 的通用节点"""
    last_message = state["messages"][-1]
    if not hasattr(last_message, "tool_calls") or not last_message.tool_calls:
        return {"messages": []}
    
    tool_map = {t.name: t for t in ALL_TOOLS}
    tool_messages = []
    created_entities = []
    
    for tool_call in last_message.tool_calls:
        tool_name = tool_call["name"]
        tool_args = tool_call["args"]
        tool_id = tool_call.get("id")
        
        logger.info(
            f"tool_execution_started",
            extra={
                "details": {
                    "tool_name": tool_name,
                    "tool_args": tool_args,
                    "tool_id": tool_id
                }
            }
        )
        
        try:
            # 时间参数归一化
            normalized_args = normalize_tool_time_arguments(
                tool_name, 
                tool_args, 
                state.get("current_datetime_context")
            )
            
            tool = tool_map.get(tool_name)
            if not tool:
                result = f"Error: Tool {tool_name} not found."
            else:
                result = await tool.ainvoke(normalized_args) if hasattr(tool, "ainvoke") else tool.invoke(normalized_args)
            
            # 实体识别（用于业务追踪）
            if any(k in tool_name for k in ["create", "add", "new"]):
                 created_entities.append({"tool": tool_name, "result": str(result)})
            
            status = "success"
        except Exception as e:
            logger.exception(f"tool_execution_failed: {tool_name}")
            result = f"Error executing tool {tool_name}: {str(e)}"
            status = "failed"
            
        tool_messages.append(ToolMessage(
            tool_call_id=tool_id,
            content=str(result),
            name=tool_name
        ))
        
        logger.info(
            f"tool_execution_finished",
            extra={
                "details": {
                    "tool_name": tool_name,
                    "status": status,
                    "result_preview": str(result)[:200]
                }
            }
        )

    return {
        "messages": tool_messages,
        "created_entities": state.get("created_entities", []) + created_entities
    }


# ===================== 各角色节点定义 =====================

async def master_node(state: AgentState) -> dict:
    """主控节点：负责意图识别与初步分发"""
    user_messages = _filter_user_messages(state["messages"])
    if not user_messages:
        return {"final_response": "未收到有效输入。"}
    
    query = user_messages[-1].content.strip()
    
    # 快捷回复逻辑 (保留原有的人性化设计)
    if re.match(r"^(你好|早|在吗|嗨|hi|hello)", query.lower()):
        return {"final_response": "您好。我在。\n\n您把问题给我，我先帮您收束重点，再往下推。", "messages": [AIMessage(content="您好。我在。")]}

    llm, capabilities = _get_llm_for_state(state)
    
    # 路由判断：让 LLM 决定跳转到哪个角色，或者直接回答
    # 这里我们使用一个简洁的提示词让 LLM 输出角色名称或直接回答
    system_msg = SystemMessage(content=MASTER_SYSTEM_PROMPT + "\n\n请直接输出接下来该由哪个 Agent 接手(role_name)，如果直接回答，请正常输出。")
    
    response = await llm.ainvoke([system_msg] + state["messages"])
    content = response.content.strip().lower()
    
    # 简单的角色映射识别
    roles = {r.value: r for r in AgentRole}
    target_role = None
    for r_val, r_enum in roles.items():
        if r_val in content and len(content) < 50: # 如果内容很短且包含角色名，视为路由
            target_role = r_enum
            break
    
    if target_role and target_role != AgentRole.MASTER:
        logger.info(f"master_routing_decided: {target_role.value}")
        return {
            "current_agent": target_role.value,
            "agent_trace": state.get("agent_trace", []) + [target_role.value],
            "messages": [AIMessage(content=f"已分发至 {target_role.value} 处理。")]
        }
    
    return {"final_response": response.content, "messages": [response]}


async def planner_node(state: AgentState) -> dict:
    return await call_agent_llm(state, AgentRole.SCHEDULE_PLANNER, SCHEDULE_PLANNER_SYSTEM_PROMPT)

async def executor_node(state: AgentState) -> dict:
    return await call_agent_llm(state, AgentRole.EXECUTOR, EXECUTOR_SYSTEM_PROMPT)

async def librarian_node(state: AgentState) -> dict:
    return await call_agent_llm(state, AgentRole.LIBRARIAN, LIBRARIAN_SYSTEM_PROMPT)

async def analyst_node(state: AgentState) -> dict:
    return await call_agent_llm(state, AgentRole.ANALYST, ANALYST_SYSTEM_PROMPT)


# ===================== 路由逻辑 =====================

def route_after_agent(state: AgentState) -> Literal["tools", "__end__"]:
    """判断 Agent 执行后是该走工具节点还是结束"""
    last_message = state["messages"][-1]
    if hasattr(last_message, "tool_calls") and last_message.tool_calls:
        return "tools"
    return END

def route_master(state: AgentState) -> str:
    """主控路由逻辑"""
    if state.get("final_response"):
        return END
    return state.get("current_agent", END)


# ===================== 图构建 =====================

def create_agent_graph(callbacks: list | None = None):
    workflow = StateGraph(AgentState)

    # 添加节点
    workflow.add_node(AgentRole.MASTER.value, master_node)
    workflow.add_node(AgentRole.SCHEDULE_PLANNER.value, planner_node)
    workflow.add_node(AgentRole.EXECUTOR.value, executor_node)
    workflow.add_node(AgentRole.LIBRARIAN.value, librarian_node)
    workflow.add_node(AgentRole.ANALYST.value, analyst_node)
    workflow.add_node("tools", execute_tools_node)

    # 设置入口
    workflow.set_entry_point(AgentRole.MASTER.value)

    # 主控分发逻辑
    workflow.add_conditional_edges(
        AgentRole.MASTER.value,
        route_master,
        {
            AgentRole.SCHEDULE_PLANNER.value: AgentRole.SCHEDULE_PLANNER.value,
            AgentRole.EXECUTOR.value: AgentRole.EXECUTOR.value,
            AgentRole.LIBRARIAN.value: AgentRole.LIBRARIAN.value,
            AgentRole.ANALYST.value: AgentRole.ANALYST.value,
            END: END
        }
    )

    # 各角色节点的 ReAct 循环
    for role in [AgentRole.SCHEDULE_PLANNER, AgentRole.EXECUTOR, AgentRole.LIBRARIAN, AgentRole.ANALYST]:
        workflow.add_conditional_edges(
            role.value,
            route_after_agent,
            {
                "tools": "tools",
                END: END
            }
        )
        
    # 工具执行完后回到当前 Agent 角色继续处理
    workflow.add_conditional_edges(
        "tools",
        lambda s: s.get("current_agent", AgentRole.MASTER.value),
        {
            AgentRole.SCHEDULE_PLANNER.value: AgentRole.SCHEDULE_PLANNER.value,
            AgentRole.EXECUTOR.value: AgentRole.EXECUTOR.value,
            AgentRole.LIBRARIAN.value: AgentRole.LIBRARIAN.value,
            AgentRole.ANALYST.value: AgentRole.ANALYST.value,
        }
    )

    # 编译
    if callbacks:
        return workflow.compile(callbacks=callbacks)
    return workflow.compile()


_agent_graph = None

def get_agent_graph(callbacks: list | None = None):
    global _agent_graph
    if _agent_graph is None:
        from app.config_tracing import get_langsmith_callbacks
        langsmith_callbacks = get_langsmith_callbacks()
        all_callbacks = (callbacks or []) + langsmith_callbacks
        _agent_graph = create_agent_graph(callbacks=all_callbacks or None)
    return _agent_graph