"""五智能体定义 — 复用 mini-OpenClaw 的 LLM 与工具，各自带差异化 system prompt。

每个智能体 = 一个独立的 LangChain create_agent：
  - 共享底层 LLM(DeepSeek) 与读文件工具
  - 差异化的 system prompt(角色、规则、输出约定)
  - 专用工具: read_xingbang_data(读兴邦数据包)、read_output(读上游产出)
  - 属于同一个全局 AgentManager 编排, 但各 agent 独立 invoke

轨迹采集: 每个 agent 的每次调用(输入/system/prompt/工具序列/输出)都会被
evidence 回调记录, 供 evolution 模块使用。
"""

from __future__ import annotations

import json
from pathlib import Path
from typing import Any, Callable

from langchain.agents import create_agent
from langchain_core.messages import HumanMessage, AIMessage
from langchain_core.tools import tool

from config import get_rag_mode
from graph.prompt_builder import build_system_prompt
from xingbang.config import agent_specs, XINGBANG_DATA_DIR, DATA_FILES


# ── 专用工具:读取兴邦多源数据包 ────────────────────────
@tool
def read_xingbang_data(filename: str) -> str:
    """读取兴邦金租客户的多源数据文件(JSON/CSV/MD/TXT)，返回内容。

    可用文件名(需带编号前缀): 01_客户申请信息.json, 02_信用评分卡.json,
    03_风险评级报告.md, 04_授信额度测算.csv, 05_征信记录.json,
    06_司法执行记录.json, 07_舆情风险评分.json, 08_反欺诈模型输出.json,
    09_财务预警模型.csv, 10_财务分析报告.csv, 11_租赁物估值报告.md,
    12_行业调研报告.md, 13_历史审批案例库.json, 14_公司制度要点.md,
    15_银行流水扫描件.txt
    """
    if not XINGBANG_DATA_DIR.exists():
        return f"[错误] 兴邦数据目录不存在: {XINGBANG_DATA_DIR}"
    path = XINGBANG_DATA_DIR / filename
    if not path.exists():
        return f"[错误] 文件不存在: {filename}"
    content = path.read_text(encoding="utf-8", errors="ignore")
    # 限制长度避免超上下文
    if len(content) > 4000:
        content = content[:4000] + "\n...[已截断]"
    return content


@tool
def read_output(filename: str, base_dir: str = "") -> str:
    """读取当前兴邦工作区中上游智能体产出的 Markdown 文件(如
    customer_intel.md, due_diligence.md, case_reference.md)。"""
    file_path = Path(base_dir) / filename if base_dir else Path(filename)
    if not file_path.exists():
        return f"[错误] 上游产出文件不存在: {filename}"
    content = file_path.read_text(encoding="utf-8", errors="ignore")
    if len(content) > 8000:
        content = content[:8000] + "\n...[已截断]"
    return content


# ── 单个评审智能体 ─────────────────────────────────────
class XingbangAgent:
    """一个兴邦评审智能体: 差异化角色 + 复用 LLM + 工具。"""

    def __init__(
        self,
        spec,
        llm=None,
        base_dir: Path | None = None,
        output_dir: Path | None = None,
        on_evidence: Callable[[str, dict], None] | None = None,
    ):
        self.spec = spec
        self._llm = llm
        self._base_dir = base_dir
        self._output_dir = output_dir or (base_dir / "xingbang_output" if base_dir else None)
        self._on_evidence = on_evidence
        self._agent = None

    def _ensure_agent(self):
        """构建带专属 system prompt 的独立 agent(惰性)。"""
        if self._agent is not None:
            return self._agent
        tools = [read_xingbang_data]
        tools.append(read_output)
        prompt = self.spec.system_prompt
        # 追加读文件工具说明(需要 base_dir 才能读产出)
        if self._base_dir:
            prompt += (
                f"\n\n[工作目录] 上游产出文件位于 {self._base_dir}。"
                "读取上游产出时使用 read_output 工具并把 filename 设为产出文件名。"
            )
        self._agent = create_agent(model=self._llm, tools=tools, system_prompt=prompt)
        return self._agent

    def _emit(self, kind: str, data: dict):
        if self._on_evidence:
            self._on_evidence(kind, data)

    async def run(self, task: str, context: str = "") -> dict:
        """执行该智能体任务,返回 {'output': 结果文本, 'file': 保存路径}。"""
        self._emit("agent_start", {"agent": self.spec.key, "name": self.spec.name})
        agent = self._ensure_agent()
        messages = []
        if context:
            messages.append(HumanMessage(content=f"[上游上下文]\n{context}"))
        messages.append(HumanMessage(content=task))
        self._emit("prompt", {
            "agent": self.spec.key,
            "system": self.spec.system_prompt[:500],
            "task": task[:300],
        })
        import asyncio
        max_retries = 4
        backoff = 5  # 初始等待秒
        last_err = None
        for attempt in range(max_retries):
            try:
                result = await agent.ainvoke({"messages": messages})
                final = self._extract_text(result)
                break
            except Exception as e:  # noqa
                last_err = e
                msg = str(e)
                # 仅对 503/限流 重试
                if "503" in msg or "overloaded" in msg or "429" in msg:
                    if attempt < max_retries - 1:
                        await asyncio.sleep(backoff * (2 ** attempt))
                        continue
                else:
                    raise  # 非限流错误立即失败
        else:
            raise RuntimeError(f"网关过载重试后仍失败: {last_err}")
        # 保存产出到文件
        file_path = None
        if self._output_dir:
            self._output_dir.mkdir(parents=True, exist_ok=True)
            file_path = self._output_dir / self.spec.output_file
            file_path.write_text(final, encoding="utf-8")
        self._emit("agent_end", {
            "agent": self.spec.key,
            "output_file": str(file_path) if file_path else None,
            "output_preview": final[:300],
        })
        return {"output": final, "file": file_path}

    @staticmethod
    def _extract_text(result: dict) -> str:
        messages = result.get("messages", [])
        text_parts = []
        for m in messages:
            if hasattr(m, "content") and m.type == "ai" and m.content:
                text_parts.append(m.content)
        if text_parts:
            return "\n\n".join(text_parts).strip()
        content = result.get("output", "")
        return content if content else "[智能体未生成有效输出]"


def build_agents(
    llm,
    base_dir: Path,
    output_dir: Path | None = None,
    on_evidence: Callable[[str, dict], None] | None = None,
    keys: list[str] | None = None,
) -> dict[str, XingbangAgent]:
    """按流水线顺序构建五智能体实例。"""
    specs = agent_specs()
    order = keys or ["data_curator", "risk_scout", "case_finder",
                     "approval_advisor", "risk_monitor"]
    return {
        key: XingbangAgent(
            spec=specs[key], llm=llm,
            base_dir=base_dir, output_dir=output_dir, on_evidence=on_evidence,
        )
        for key in order
    }