ewoooc/services/event_router.py

"""
EventRouter — 事件分流入口（ADR-012 Phase 1 骨幹）

所有系統事件（exception / 排程完成 / 告警 / 資訊通報）**應**統一透過
`dispatch(event)` 進入，由 EventRouter 依 severity × event_type 分流到：
  L0 Direct / L1 Hermes Observer / L2 NemoTron Investigator / L3 OpenClaw Operator

設計原則（ADR-012 §⑥）：無論 AI 狀況，**通知鏈絕不中斷**。
每一級失敗立即降級到下一級，最終保底 L0 直出模板。

Phase 1 實作範圍：
- 骨幹 + 分類邏輯
- L0 模板直出（已可用）
- L1 Hermes / L2 NemoTron / L3 OpenClaw 為 stub（附 TODO 標記）
- 完整 fallback 鏈（AI 掛必降級）
- 靜音檢查 + Audit Trail
"""

from __future__ import annotations

import os
import time
from datetime import datetime
from enum import Enum
from typing import Any

import requests

from services.logger_manager import SystemLogger
from services import telegram_templates as tpl
from services import agent_actions

sys_log = SystemLogger("EventRouter").get_logger()


class Tier(str, Enum):
    L0_DIRECT = "L0"
    L1_OBSERVER = "L1"
    L2_INVESTIGATOR = "L2"
    L3_OPERATOR = "L3"


class Severity(str, Enum):
    INFO = "info"
    SUCCESS = "success"
    WARNING = "warning"
    ALERT = "alert"  # P0/P1


# =====================================================================
# 分類規則（ADR-012 §③）
# =====================================================================
def _classify(event: dict) -> Tier:
    sev = event.get("severity", "info")
    has_trace = bool(event.get("trace"))
    event_type = event.get("event_type", "")

    # L3 OpenClaw 由週期任務主動觸發（週報、Meta-Analysis），不走 router
    # 這裡只處理 L0/L1/L2

    if sev in (Severity.INFO, Severity.SUCCESS):
        return Tier.L0_DIRECT

    if sev == Severity.WARNING:
        # 有技術 trace → L1 Hermes 翻譯
        return Tier.L1_OBSERVER if has_trace else Tier.L0_DIRECT

    if sev == Severity.ALERT:
        # 符合 L2 白名單 event_type → NemoTron 介入
        L2_EVENT_TYPES = {
            "price_threat", "db_connection_error", "crawler_timeout",
            "nim_quota_exhausted", "embedding_failure",
        }
        if event_type in L2_EVENT_TYPES:
            return Tier.L2_INVESTIGATOR
        return Tier.L1_OBSERVER

    return Tier.L0_DIRECT


# =====================================================================
# 主入口
# =====================================================================
def dispatch(event: dict, admin_chat_ids: list[int] | None = None) -> dict:
    """
    主要入口。回傳 dict: {tier, sent, insight_id, errors, latency_ms}

    event 格式見 ADR-012 §③。必要欄位：source, event_type, severity, title, summary
    可選欄位：trace, payload, time

    admin_chat_ids 若未給，從 DB telegram_users where is_admin=true 取
    """
    t0 = time.time()
    event_key = f"{event.get('source', '?')}:{event.get('event_type', '?')}"

    # 靜音檢查
    if agent_actions.is_silenced(event_key):
        sys_log.info(f"[EventRouter] 事件被靜音略過: {event_key}")
        return {"tier": "silenced", "sent": 0, "event_key": event_key}

    tier = _classify(event)
    sys_log.info(f"[EventRouter] dispatch {event_key} → {tier.value}")

    # 執行對應 Tier
    try:
        if tier == Tier.L0_DIRECT:
            text = _render_l0(event)
        elif tier == Tier.L1_OBSERVER:
            text = _render_l1_with_fallback(event)
        elif tier == Tier.L2_INVESTIGATOR:
            text = _render_l2_with_fallback(event)
        else:
            text = _render_l0(event)  # 未知 Tier 保底
    except Exception as e:
        sys_log.error(f"[EventRouter] 渲染失敗，降級 L0: {e}")
        text = _render_l0(event)

    # 發送 Telegram
    result = _send(text, admin_chat_ids)
    result["tier"] = tier.value
    result["event_key"] = event_key
    result["latency_ms"] = round((time.time() - t0) * 1000, 1)

    # 審計（每次 dispatch 都入 KM）
    _audit_dispatch(event, tier, result)
    return result


# =====================================================================
# Tier 渲染器
# =====================================================================
def _render_l0(event: dict) -> str:
    """L0 直出：根據 severity 選用對應模板"""
    sev = event.get("severity", "info")
    title = event.get("title", "未命名事件")
    module = event.get("source", "unknown")
    summary = event.get("summary", "")
    details = event.get("payload") if isinstance(event.get("payload"), dict) else None

    if sev == Severity.SUCCESS:
        return tpl.success(title=title, module=module, stats=summary)
    if sev == Severity.INFO:
        return tpl.info(title=title, module=module, content=summary)
    if sev == Severity.WARNING:
        return tpl.warning(title=title, module=module, summary=summary, details=details)
    # alert 但降級到 L0
    return tpl.alert(
        title=title, module=module,
        status=event.get("status", "未知"),
        impact=event.get("impact", "未評估"),
        summary=summary,
        actions=event.get("suggested_actions"),
        trace=event.get("trace"),
    )


def _render_l1_with_fallback(event: dict) -> str:
    """L1 Hermes 翻譯 stack trace。Phase 1 stub — 直接降 L0 + 標記"""
    # TODO Phase 2: 呼叫 Hermes 做 stack trace 翻譯與摘要
    try:
        ai_summary = _hermes_observe(event)  # stub
        if ai_summary:
            return _compose_triaged(event, tier_label="L1 · Hermes", ai_summary=ai_summary)
    except Exception as e:
        sys_log.warning(f"[EventRouter] L1 Hermes 失敗，降 L0: {e}")

    # Fallback：L0 模板 + 降級標記
    text = _render_l0(event)
    return text + "\n\n🟡 _AI 分析暫不可用，以原始資料呈現_"


def _render_l2_with_fallback(event: dict) -> str:
    """L2 NemoTron 介入（含 tool call）。Phase 1 stub — 降 L1"""
    # TODO Phase 3: 呼叫 NemoTron dispatcher，允許執行 SAFE_ACTIONS 中的 tool
    try:
        ai_result = _nemoton_investigate(event)  # stub
        if ai_result:
            return _compose_triaged(
                event, tier_label="L2 · NemoTron",
                ai_summary=ai_result.get("summary", ""),
                ai_actions=ai_result.get("actions_taken", []),
            )
    except Exception as e:
        sys_log.warning(f"[EventRouter] L2 NemoTron 失敗，降 L1: {e}")

    return _render_l1_with_fallback(event)


def _compose_triaged(event: dict, tier_label: str, ai_summary: str,
                     ai_actions: list | None = None) -> str:
    """三層式訊息：AI 摘要 + 原始事實 + 建議行動（ADR-012 §④）"""
    base = _render_l0(event)
    parts = [f"🤖 *AI 摘要（{tier_label}）：*", ai_summary, ""]
    if ai_actions:
        parts.append("🛠️ *AI 已執行動作：*")
        for a in ai_actions:
            parts.append(f"• {a}")
        parts.append("")
    return base + "\n\n" + "\n".join(parts)


# =====================================================================
# AI 介入 stub（Phase 2/3 填實作）
# =====================================================================
def _hermes_observe(event: dict) -> str | None:
    """Phase 1 stub：Phase 2 會呼叫 hermes_analyst_service"""
    # 不呼叫 AI，回傳 None 讓上層降級
    return None


def _nemoton_investigate(event: dict) -> dict | None:
    """Phase 1 stub：Phase 3 會呼叫 nemoton_dispatcher_service"""
    return None


# =====================================================================
# Telegram 發送 + Audit
# =====================================================================
def _send(text: str, admin_chat_ids: list[int] | None) -> dict:
    token = os.getenv("TELEGRAM_BOT_TOKEN")
    if not token:
        return {"sent": 0, "errors": ["TELEGRAM_BOT_TOKEN 未設定"]}

    if admin_chat_ids is None:
        admin_chat_ids = _load_admin_chat_ids()
    if not admin_chat_ids:
        return {"sent": 0, "errors": ["無管理員 chat_id"]}

    url = f"https://api.telegram.org/bot{token}/sendMessage"
    sent = 0
    errors = []
    for cid in admin_chat_ids:
        try:
            r = requests.post(url, json={
                "chat_id": int(cid), "text": text, "parse_mode": "Markdown",
            }, timeout=10)
            if r.ok:
                sent += 1
            else:
                errors.append(f"chat={cid} status={r.status_code}")
        except Exception as e:
            errors.append(f"chat={cid} exc={e}")
    return {"sent": sent, "errors": errors}


def _load_admin_chat_ids() -> list[int]:
    """從 DB 撈 is_admin=true 的 chat_id；fallback 到 .env TELEGRAM_CHAT_IDS"""
    try:
        from database.manager import get_session
        from sqlalchemy import text as sa_text
        session = get_session()
        try:
            rows = session.execute(sa_text(
                "SELECT telegram_id FROM telegram_users WHERE is_active=true AND is_admin=true"
            )).fetchall()
            if rows:
                return [int(r[0]) for r in rows]
        finally:
            session.close()
    except Exception as e:
        sys_log.warning(f"[EventRouter] 查 telegram_users 失敗，fallback .env: {e}")

    # Fallback 到 env
    import re
    raw = os.getenv("TELEGRAM_CHAT_IDS", "")
    return [int(x.strip()) for x in re.sub(r'[\[\]"\' ]', "", raw).split(",") if x.strip()]


def _audit_dispatch(event: dict, tier: Tier, result: dict) -> None:
    """每次 dispatch 都寫入 ai_insights 作為審計軌跡"""
    try:
        from services.openclaw_learning_service import store_insight
        store_insight(
            insight_type="agent_action",
            content=f"dispatch tier={tier.value} event={event.get('event_type')}",
            period=datetime.now().strftime("%Y-%m-%d"),
            metadata={
                "tier": tier.value,
                "event": {k: event.get(k) for k in ("source", "event_type", "severity", "title")},
                "result": {k: result.get(k) for k in ("sent", "latency_ms", "errors")},
                "ts": datetime.now().isoformat(),
            },
        )
    except Exception as e:
        sys_log.error(f"[EventRouter] audit 失敗: {e}")