# 伪代码示例：预授权规划
user_request = "为患者张三（ID: 12345）申请膝关节MRI预授权"
plan = planner.plan(user_request, available_tools)
# 返回：[{"step": 1, "tool": "extract_medical_codes", "input": {"diagnosis": "膝盖痛"}},
#         {"step": 2, "tool": "check_insurance_benefits", "input": {"patient_id": 12345}},
#         {"step": 3, "tool": "submit_prior_auth", "input": {...}},
#         {"step": 4, "tool": "poll_auth_status", "input": {"request_id": "..."}}]

class PriorAuthSubmissionTool:
    """向保险公司提交预授权请求"""
    name = "submit_prior_auth"
    description = "Submit a prior authorization request to the payer portal"
    input_schema = {
        "type": "object",
        "properties": {
            "patient_id": {"type": "string"},
            "cpt_code": {"type": "string"},
            "diagnosis_code": {"type": "string"},
            "provider_npi": {"type": "string"}
        },
        "required": ["patient_id", "cpt_code", "diagnosis_code"]
    }

# LangGraph简化版执行循环
while not done:
    agent_state = graph.invoke({"messages": messages, "memory": memory})
    tool_call = agent_state["tool_call"]
    if tool_call is None:
        # LLM判断任务完成
        done = True
    else:
        result = execute_tool(tool_call)
        messages.append(result)

graph TD
    A[收到预授权请求] --> B{提取代码}
    B -- 成功 --> C[查询保险福利]
    B -- 失败 --> D[请求人工录入代码]
    C -- 覆盖且无需授权 --> E[直接写入EMR:无需授权]
    C -- 覆盖且需要授权 --> F[提交预授权到保险门户]
    C -- 不覆盖 --> G[通知医生，提供替代CPT选项]
    F --> H[轮询状态（每30分钟，最多3天）]
    H -- 批准 --> I[写入EMR并通知用户]
    H -- 拒绝 --> J[通知医生并提供申诉选项]
    H -- 超时 --> K[升级到人工处理]
    H -- 网络错误 --> L[重试（最多3次）]

def validate_tool_output(tool_name, raw_output):
    if tool_name == "submit_prior_auth":
        if "request_id" in raw_output:
            return {"status": "success", "request_id": raw_output["request_id"]}
        elif "error" in raw_output:
            return {"status": "error", "message": raw_output["error"], "retryable": True}
        else:
            return {"status": "unknown", "raw": raw_output, "retryable": False}

-- 表结构示例
CREATE TABLE agent_sessions (
    session_id TEXT PRIMARY KEY,
    patient_id TEXT,
    plan TEXT,  -- JSON格式的任务计划
    current_step INTEGER,
    memory TEXT, -- 关键记忆摘要
    status TEXT, -- pending / in_progress / completed / failed
    created_at TIMESTAMP
);

pip install langgraph langchain-openai

from langgraph.graph import StateGraph, END
from typing import TypedDict, Optional
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool

# 定义Agent状态
class AgentState(TypedDict):
    user_input: str
    patient_id: Optional[str]
    memory: dict
    plan: list
    current_step: int
    result: str

# 模拟的工具函数
@tool
def get_patient_info(patient_id: str) -> dict:
    """获取患者基本信息"""
    # 实际中调用医院EMR系统
    return {"name": "张三", "insurance": "UnitedHealth", "status": "active"}

@tool
def get_prior_auth_status(request_id: str) -> str:
    """查询预授权状态"""
    # 模拟返回
    return "approved"

@tool
def submit_prior_auth(patient_id: str, cpt_code: str) -> dict:
    """提交预授权申请"""
    # 模拟成功
    return {"request_id": "PA-20260610-001", "status": "submitted"}

# 创建Agent节点
def plan_node(state: AgentState) -> AgentState:
    # 使用LLM制定计划（简化：根据关键词硬编码）
    if "状态" in state["user_input"]:
        state["plan"] = ["get_prior_auth_status"]
    else:
        state["plan"] = ["get_patient_info", "submit_prior_auth"]
    state["current_step"] = 0
    return state

def execute_node(state: AgentState) -> AgentState:
    if state["current_step"] >= len(state["plan"]):
        state["result"] = "任务完成"
        return state
    
    step = state["plan"][state["current_step"]]
    if step == "get_patient_info":
        info = get_patient_info(state["patient_id"])
        state["memory"] = info
    elif step == "submit_prior_auth":
        result = submit_prior_auth(state["patient_id"], "73721")
        state["memory"]["request_id"] = result["request_id"]
        state["result"] = f"已提交预授权申请，ID: {result['request_id']}"
    elif step == "get_prior_auth_status":
        status = get_prior_auth_status(state["memory"].get("request_id", "dummy"))
        state["result"] = f"预授权状态：{status}"
    
    state["current_step"] += 1
    return state

def should_continue(state: AgentState):
    if state["current_step"] < len(state["plan"]):
        return "execute"
    else:
        return "end"

# 构建图
builder = StateGraph(AgentState)
builder.add_node("plan", plan_node)
builder.add_node("execute", execute_node)
builder.set_entry_point("plan")
builder.add_conditional_edges(
    "plan",
    lambda s: "execute"
)
builder.add_conditional_edges(
    "execute",
    should_continue,
    {"execute": "execute", "end": END}
)

graph = builder.compile()

# 运行
initial_state = {
    "user_input": "我要为张三提交一个膝关节MRI的预授权申请",
    "patient_id": "12345",
    "memory": {},
    "plan": [],
    "current_step": 0,
    "result": ""
}
final = graph.invoke(initial_state)
print(final["result"])