用户输入（“我要注册冲突的CS471”）
  ↓
Planner（LLM） → 输出子任务列表：
  [
    {tool: "check_course_capacity", params: {course_id: "CS471"}},
    {tool: "check_student_credits", params: {student_id: "..."}},
    {tool: "check_max_credit_limit", params: {student_id: "...", course_id: "CS471"}},
    {tool: "submit_approval_request", params: {course: "CS471", reason: "conflict override"}}
  ]
  ↓
Executor 逐条执行，每一步结果写入短期记忆
  ↓
全部成功 → 输出最终结果 → 更新数据库 → 回复用户
  ↓
某步失败（如审批超时） → Re-planner 重新规划（是否换审批人？是否转人工？）

{
  "tasks": [
    {"tool": "check_course_capacity", "params": {"course_id": "CS471"}, "depends_on": []},
    {"tool": "check_student_credits", "params": {"student_id": "12345"}, "depends_on": []},
    {"tool": "submit_approval_request", "params": {"course": "CS471"}, "depends_on": [0,1]}
  ]
}

from langgraph.graph import StateGraph, END
from typing import TypedDict, List, Dict, Any
import json

# 1. 定义状态
class AgentState(TypedDict):
    user_input: str
    plan: List[Dict]
    current_step: int
    results: Dict[str, Any]
    error: str

# 2. 定义工具函数（模拟）
def check_course_capacity(course_id: str) -> Dict:
    # 假设调用真实API
    return {"course_id": course_id, "available": True}

def check_student_credits(student_id: str) -> Dict:
    return {"student_id": student_id, "credits": 120}

def submit_approval_request(course: str, student_id: str) -> Dict:
    return {"status": "submitted"}

# 3. Planner 节点（用LLM包装，这里简化直接返回固定规划）
def planner_node(state: AgentState) -> AgentState:
    # 实际应调用LLM，这里硬编码演示
    state['plan'] = [
        {"tool": "check_course_capacity", "params": {"course_id": "CS471"}},
        {"tool": "check_student_credits", "params": {"student_id": "12345"}},
        {"tool": "submit_approval_request", "params": {"course": "CS471", "student_id": "12345"}}
    ]
    state['current_step'] = 0
    state['results'] = {}
    return state

# 4. 执行节点（每次调用一个工具）
def executor_node(state: AgentState) -> AgentState:
    plan = state['plan']
    step = state['current_step']
    if step >= len(plan):
        state['error'] = "finished"
        return state
    task = plan[step]
    tool_name = task['tool']
    params = task['params']
    # 根据工具名调用
    if tool_name == "check_course_capacity":
        result = check_course_capacity(**params)
    elif tool_name == "check_student_credits":
        result = check_student_credits(**params)
    elif tool_name == "submit_approval_request":
        result = submit_approval_request(**params)
    else:
        state['error'] = f"Unknown tool: {tool_name}"
        return state
    state['results'][tool_name] = result
    state['current_step'] += 1
    return state

# 5. 决定是否继续

def should_continue(state: AgentState) -> str:
    if state.get('error'):
        return 'error'
    if state['current_step'] >= len(state['plan']):
        return 'done'
    return 'continue'

# 6. 构建图
graph = StateGraph(AgentState)
graph.add_node("planner", planner_node)
graph.add_node("executor", executor_node)
graph.set_entry_point("planner")
graph.add_edge("planner", "executor")
graph.add_conditional_edges("executor", should_continue, {
    "continue": "executor",
    "done": END,
    "error": END
})

app = graph.compile()

# 运行
initial_state = {
    "user_input": "I want to register CS471 but it conflicts.",
    "plan": [],
    "current_step": 0,
    "results": {},
    "error": ""
}
result = app.invoke(initial_state)
print(result['results'])