"""
OpenClaw 告警規則匹配引擎
============================================================
從 alert_rules.yaml 載入規則，取代 openclaw.py 中硬編碼的 if/elif 規則匹配。

設計原則:
- 規則在 YAML 定義，不需要改 Python 代碼
- 匹配邏輯: alertname 完全匹配 > alert_type 部分匹配 > message 關鍵字
- priority 越小越優先，999 = 通用兜底
- 變數替換: {target} {host} {container} {instance} {job} {namespace} {unit}

自動規則生成:
- 只有 generic_fallback 觸發時才生成（具體規則不觸發）
- 呼叫 Ollama (deepseek-r1:14b) 或 Gemini 生成 YAML 規則片段
- 生成後 append 到 alert_rules.yaml，清除 lru_cache 立即生效
- 同一 alertname 已有規則時跳過（去重）

2026-04-09 ogt: 初版 + 自動規則生成
"""

from __future__ import annotations

import json
import re
from datetime import datetime
from functools import lru_cache
from pathlib import Path
from typing import Any

import httpx
import structlog
import yaml

from src.constants.alert_types import ALERTNAME_TO_TYPE
from src.services.action_parser import parse_kubectl_action

logger = structlog.get_logger(__name__)

RULES_FILE = Path(__file__).parent.parent.parent / "alert_rules.yaml"

# 進程級去重（保留作為 Redis 不可用時的 fallback）
_generating: set[str] = set()

# Redis 分散式鎖 TTL (秒)，覆蓋 Ollama + Gemini 最長生成時間
_RULE_GEN_LOCK_TTL = 120

# ── action parser 注入防護 (SPF-2, 2026-04-30) ───────────────
# kubectl 走 structured token parser；非 kubectl 保留簡單 dangerous-fragment
# 掃描，避免舊式巨型 regex 誤殺安全的單一 delete pod / deployment resource forms。
_RULE_ENGINE_DANGEROUS_FRAGMENTS = (
    "rm -rf",
    "rm -f /",
    "drop table",
    "drop database",
    "truncate table",
    "$(",
    "`",
)

# ── kubectl 注入防護 公開 API ───────────────────────────────


def validate_kubectl_command(command: str) -> bool:
    """
    Action 注入安全驗證（Task 2.3, ADR-076; SPF-2 parser upgrade）。

    Returns:
        True  — 指令安全，可執行
        False — 含破壞性模式，呼叫方應清空指令並記錄 warning

    通過條件（直接 True）:
    - 空字串 — 無動作規則
    - "ssh ..." 開頭 — SSH 層指令，不走 kubectl 路徑

    阻擋條件（返回 False）:
    - kubectl parser 不支援的語法（deployment delete / drain / cordon /
      replicas=0 / shell metachar / compound command）
    - 非 kubectl 指令內含主機/SQL/command-substitution 危險片段
    """
    command = (command or "").strip()
    if not command:
        return True
    if command.startswith("ssh "):
        return True
    if command.startswith("kubectl"):
        return parse_kubectl_action(command).ok
    command_lower = command.lower()
    return not any(fragment in command_lower for fragment in _RULE_ENGINE_DANGEROUS_FRAGMENTS)


# ── 變數提取 ────────────────────────────────────────────────


_POD_SUFFIX_DEPLOYMENT_RE = __import__("re").compile(
    r"-[a-z0-9]{5,10}-[a-z0-9]{5}$"
)
_POD_SUFFIX_LEGACY_RE = __import__("re").compile(
    r"-[a-z0-9]{5}$"
)
_POD_SUFFIX_STATEFULSET_RE = __import__("re").compile(
    r"-\d+$"
)


def _strip_pod_suffix(pod_name: str) -> str:
    """
    由 Pod 名稱推斷 Deployment/StatefulSet base name。

    優先順序（由嚴格到寬鬆）:
    1. Deployment: {name}-{rs_hash 5-10 chars}-{pod_hash 5 chars}
       範例: awoooi-api-7d6b776f78-4sgjl → awoooi-api
    2. StatefulSet: {name}-{ordinal}
       範例: postgresql-0 → postgresql
    3. Legacy single-hash Pod: {name}-{hash 5 chars}
       範例: my-job-x2m4k → my-job

    GAP-A4 (2026-04-14 Claude Sonnet 4.6): Placeholder 解析缺漏修復。
    """
    # 先試 Deployment 格式（最常見）
    stripped = _POD_SUFFIX_DEPLOYMENT_RE.sub("", pod_name)
    if stripped != pod_name and stripped:
        return stripped
    # 再試 StatefulSet
    stripped = _POD_SUFFIX_STATEFULSET_RE.sub("", pod_name)
    if stripped != pod_name and stripped:
        return stripped
    # 最後試 legacy single-hash
    stripped = _POD_SUFFIX_LEGACY_RE.sub("", pod_name)
    if stripped != pod_name and stripped and "-" in stripped:
        return stripped
    return pod_name


def _is_bad_target(target: str, alertname: str) -> bool:
    """
    判斷 target 是否為「垃圾值」，不得組合成 kubectl 指令。

    垃圾值:
    - 空字串 / "unknown"
    - 包含空白、冒號（IP:port）、括號、引號
    - 等於 alertname 本身（LLM/規則填錯）
    - 純數字或 IP 格式
    """
    if not target or target in ("unknown", "none", "null", ""):
        return True
    # 2026-04-15 Claude Sonnet 4.6 (GAP-A4 Phase 3): 擴充 fallback magic string
    # 截圖實證：Telegram 卡顯示 target=unknown-service 通過 _is_bad_target
    # _resolve_target_from_k8s fallback 也會產 unknown-pod/unknown-container
    _BAD_MAGIC_PREFIXES = ("unknown-", "none-", "null-", "undefined-")
    if any(target.startswith(p) for p in _BAD_MAGIC_PREFIXES):
        return True
    if target == alertname:
        return True
    if any(c in target for c in (" ", ":", "(", ")", '"', "'", "<", ">", "{", "}")):
        return True
    # 純 IP 格式
    if target.replace(".", "").isdigit() and target.count(".") == 3:
        return True
    return False


def _extract_vars(alert_context: dict) -> dict[str, str]:
    """
    從 alert_context 提取模板變數。

    GAP-A4 (2026-04-14 Claude Sonnet 4.6): 強化 target 解析，多層 label 查找順序:
    1. labels.deployment (最權威)
    2. labels.app / labels.app.kubernetes.io/name
    3. labels.statefulset
    4. labels.pod → 去除 replicaset/pod hash 後綴
    5. labels.container / labels.name
    6. labels.service（2026-04-20 降級：K8s Service 名 != Deployment 名，
       改記 target_source=label.service 讓下游可疑點觸發 pre-flight 驗證）
    7. target_resource（但排除 IP:port 和 alertname）

    target_source 欄位回傳讓 decision/execution 層能追 trace；
    若 P0.3 本次 trace 不夠清楚，下次觸發就有 aol.context.target_source。

    若全部提取失敗 → target="unknown"，由 match_rule() 的後置驗證丟棄此規則。
    """
    labels = alert_context.get("labels", {})
    alertname = labels.get("alertname", alert_context.get("alert_type", ""))
    raw_target = alert_context.get("target_resource", "unknown")
    instance = labels.get("instance", raw_target)
    host = instance.split(":")[0] if ":" in instance else instance
    job = labels.get("job", "exporter")
    namespace = alert_context.get("namespace", "awoooi-prod")
    unit = labels.get("unit", "")

    # GAP-A4: 多層 label 查找，由最權威到最弱
    target = ""
    target_source = ""  # 2026-04-20: 追蹤 target 從哪個 label 來（供 aol 留痕）
    for key in ("deployment", "app", "app.kubernetes.io/name", "statefulset"):
        val = labels.get(key, "")
        if val and not _is_bad_target(val, alertname):
            target = val
            target_source = f"label.{key}"
            break

    # Pod label 需去除 hash 後綴還原 Deployment 名稱
    if not target:
        pod = labels.get("pod", "")
        if pod and not _is_bad_target(pod, alertname):
            target = _strip_pod_suffix(pod)
            target_source = "label.pod(stripped)"

    # container / name 次優
    if not target:
        for key in ("container", "name"):
            val = labels.get(key, "")
            if val and not _is_bad_target(val, alertname):
                target = val
                target_source = f"label.{key}"
                break

    # service label 最末位（K8s Service 名非 Deployment 名，常產生 awoooi-service 幻覺）
    # 2026-04-20 P0.3：若 service 以 '-service' 結尾，先去 suffix 視作 base name，
    # 仍留 target_source=label.service(demoted) 讓下游觸發 pre-flight 驗證
    if not target:
        svc = labels.get("service", "")
        if svc and not _is_bad_target(svc, alertname):
            if svc.endswith("-service"):
                # awoooi-service → awoooi (通常對應 awoooi-api 或類似，仍需 pre-flight 驗)
                target = svc[: -len("-service")] or svc
                target_source = "label.service(stripped -service suffix)"
            else:
                target = svc
                target_source = "label.service(demoted)"

    # raw_target 末位（且必須通過 bad_target 驗證）
    if not target and not _is_bad_target(raw_target, alertname):
        target = raw_target
        target_source = "alert_context.target_resource"

    # 若全部失敗 → 保留 "unknown" 讓後置驗證層 reject
    if not target:
        target = "unknown"
        target_source = "none(fallback)"

    container = labels.get("name", labels.get("container", "")) or target
    return {
        "target": target,
        "target_source": target_source,  # 2026-04-20 P0.3 新增
        "host": host,
        "container": container,
        "instance": instance,
        "job": job,
        "namespace": namespace,
        "unit": unit,
    }


def _fill(template: str, vars: dict[str, str]) -> str:
    """填充模板變數，保留未知變數原樣"""
    try:
        return template.format_map(vars)
    except (KeyError, ValueError):
        return template


# ── 規則載入 ────────────────────────────────────────────────


@lru_cache(maxsize=1)
def _load_rules() -> list[dict]:
    """載入並快取規則（進程內不重載，重啟 Pod 才更新）"""
    if not RULES_FILE.exists():
        logger.warning("alert_rules_file_not_found", path=str(RULES_FILE))
        return []
    with RULES_FILE.open("r", encoding="utf-8") as f:
        data = yaml.safe_load(f)
    rules = sorted(data.get("rules", []), key=lambda r: r.get("priority", 999))
    logger.info("alert_rules_loaded", count=len(rules), path=str(RULES_FILE))
    return rules


# ── 匹配邏輯 ────────────────────────────────────────────────


def _matches(rule: dict, alertname: str, alert_type: str, message: str, instance: str = "") -> bool:
    """判斷規則是否匹配。通用兜底規則（alertname=["*"]）永遠回傳 False，由 match_rule 單獨處理。"""
    match = rule.get("match", {})

    # S1-3 修正: 通用兜底規則不參與 _matches，防止其 alert_type/message 關鍵字意外命中
    alertnames = match.get("alertname", [])
    if alertnames == ["*"]:
        return False

    # instance/target 前綴匹配 (最高優先，用於 E2E test 告警識別)
    # 2026-04-09 Claude Sonnet 4.6: 支援 instance_prefix 匹配，讓 E2E test 告警不走 generic_fallback
    if instance and match.get("instance_prefix"):
        for prefix in match["instance_prefix"]:
            if instance.lower().startswith(prefix.lower()):
                return True

    # alertname 完全匹配
    if alertnames and alertname in alertnames:
        return True

    # alert_type 部分匹配
    for kw in match.get("alert_type", []):
        if kw.lower() in alert_type.lower():
            return True

    # message 關鍵字匹配（不分大小寫）
    msg_lower = message.lower()
    for kw in match.get("message", []):
        if kw.lower() in msg_lower:
            return True

    return False


def _is_generic(rule: dict) -> bool:
    alertnames = rule.get("match", {}).get("alertname", [])
    return alertnames == ["*"]


# ── 公開 API ────────────────────────────────────────────────


def get_risk_for_alertname(alertname: str) -> str | None:
    """
    根據 alertname 從 alert_rules.yaml 取得 risk 等級。
    ADR-073 Phase 3-6: 供 decision_manager 做 YAML risk override。

    Returns:
        risk 字串 ("low"/"medium"/"critical") 或 None（無匹配規則）
    """
    for rule in _load_rules():
        if _matches(rule, alertname, "", "", ""):
            return rule.get("response", {}).get("risk")
    return None


def match_rule(alert_context: dict) -> dict[str, Any] | None:
    """
    根據 alert_context 匹配規則，回傳填充後的 response dict。

    Returns:
        匹配到的規則 response，未匹配則回傳 None。
        呼叫方應自行處理 None（走 AI 分析或通用兜底）。
    """
    labels = alert_context.get("labels", {})
    alertname = labels.get("alertname", alert_context.get("alert_type", "custom"))
    alert_type = alert_context.get("alert_type", "custom")
    message = alert_context.get("message", "")
    severity = alert_context.get("severity", "warning")

    rules = _load_rules()
    vars = _extract_vars(alert_context)

    instance = labels.get("instance", alert_context.get("target_resource", ""))

    matched_rule = None
    for rule in rules:
        if _is_generic(rule):
            continue  # 通用兜底最後才用
        if _matches(rule, alertname, alert_type, message, instance):
            matched_rule = rule
            break

    # 未匹配到具體規則 → 用通用兜底
    if matched_rule is None:
        for rule in rules:
            if _is_generic(rule):
                matched_rule = rule
                break

    if matched_rule is None:
        return None

    resp = matched_rule["response"]
    risk = resp.get("risk", "medium")
    # severity=critical 強制升級風險等級
    if severity == "critical" and risk == "medium":
        risk = "critical"

    optimization = [
        {
            "type": o["type"],
            "description": _fill(o.get("description", ""), vars),
            "kubectl_or_config": _fill(o.get("command", ""), vars),
        }
        for o in resp.get("optimization", [])
    ]

    # Task 2.3: kubectl 注入防護 — 變數替換後驗證，阻擋破壞性模式
    kubectl_command = _fill(resp.get("kubectl_command", ""), vars)
    if not validate_kubectl_command(kubectl_command):
        logger.warning(
            "rule_kubectl_command_blocked",
            rule_id=matched_rule["id"],
            reason="destructive_pattern_detected",
            command_snippet=kubectl_command[:80],
        )
        kubectl_command = ""

    # GAP-A4 (2026-04-14 Claude Sonnet 4.6): 後置驗證 — 垃圾 target 丟棄 command
    # 避免 `kubectl rollout restart deployment unknown/HostHighCpuLoad/...` 這類無效指令
    # 清空 kubectl_command 讓 decision_manager 降級給 LLM 處理
    # 2026-05-04 ogt + Claude Sonnet 4.6: SSH 指令豁免 bad_target 驗證
    # 根因：host_resource_alert 規則的 kubectl_command 以 "ssh {host} '...'" 組成，
    #       {host} 由 instance label 派生（_extract_vars 第 184-185 行），與 {target} 無關。
    #       但 host 告警缺少 K8s deployment label → target="unknown" → _is_bad_target=True
    #       → kubectl_command 被清空 → auto_approve 以 no_executable_action 拒絕 → 人工攔截。
    # 修復：SSH 指令不依賴 target，跳過 bad_target 驗證，保留指令讓自動診斷路徑通行。
    _invalid_target = False
    _is_ssh_command = kubectl_command.startswith("ssh ")
    if kubectl_command and not _is_ssh_command and _is_bad_target(vars["target"], alertname):
        logger.warning(
            "rule_kubectl_command_discarded_bad_target",
            rule_id=matched_rule["id"],
            target=vars["target"],
            alertname=alertname,
            reason="target 未解析為真實 deployment/app，拒絕組裝指令 → fallback LLM",
            original_command=kubectl_command[:120],
        )
        kubectl_command = ""
        _invalid_target = True

    # 還有 {var} 殘留 → 模板變數未被 _fill 填滿（可能 vars 缺少對應 key）
    if kubectl_command and ("{" in kubectl_command or "}" in kubectl_command):
        logger.warning(
            "rule_kubectl_command_discarded_unfilled_placeholder",
            rule_id=matched_rule["id"],
            command=kubectl_command[:120],
        )
        kubectl_command = ""
        _invalid_target = True

    # 2026-04-16 ogt + Claude Sonnet 4.6: blocked_reason 讓 decision_manager 感知 target 失敗
    # 根因：rule engine 清空 kubectl_command 但不告訴下游原因 → decision_manager 繼續嘗試執行
    # 修復：設定 blocked_reason="INVALID_TARGET - ..." 讓 auto_execute 路徑提早返回 + TYPE-4
    _blocked_reason_out = (
        f"INVALID_TARGET - target='{vars['target']}' 未解析為有效 Deployment，需 SRE 人工確認"
        if _invalid_target else ""
    )

    return {
        "rule_id": matched_rule["id"],
        "action_title": _fill(resp["action_title"], vars),
        "description": _fill(resp["description"], vars),
        "suggested_action": resp["suggested_action"],
        "kubectl_command": kubectl_command,
        "target_resource": vars["target"],
        "target_source": vars.get("target_source", ""),  # 2026-04-20 P0.3 留痕
        "namespace": vars["namespace"],
        "risk_level": risk,
        "blast_radius": {
            "affected_pods": 1,
            "estimated_downtime": resp.get("estimated_downtime", "unknown"),
            "related_services": [vars["target"]],
            "data_impact": "NONE",
        },
        "primary_responsibility": resp.get("responsibility", "COLLAB"),
        "responsibility_reasoning": resp.get("responsibility_reasoning", ""),
        "secondary_teams": resp.get("secondary_teams", []),
        "optimization_suggestions": optimization,
        "reasoning": _fill(resp.get("reasoning", ""), vars),
        "deviation_analysis": "規則引擎觸發，監控指標偏離正常基準",
        "confidence": 0.0,  # 🔴 規則匹配固定 0.0，禁止偽造
        "affected_services": [vars["target"]],
        "signoz_correlation": "",
        "blocked_reason": _blocked_reason_out,  # "" 正常 | "INVALID_TARGET - ..." 目標解析失敗
    }


# ── 自動規則生成 ─────────────────────────────────────────────

_AUTO_RULE_PROMPT = """\
你是 SRE 專家。根據以下 Prometheus 告警資訊，生成一條 YAML 規則供 OpenClaw 規則引擎使用。

告警資訊:
- alertname: {alertname}
- alert_type: {alert_type}
- message: {message}
- labels: {labels}

請嚴格輸出以下 YAML 格式，不要加任何說明文字，只輸出 YAML 區塊:

- id: {rule_id}
  priority: {priority}
  description: <一行中文描述>
  match:
    alertname:
      - {alertname}
  response:
    action_title: <含 {{target}} 變數的動作標題>
    description: "⚙️ 規則匹配: <含 {{target}} 的描述，說明根因>"
    suggested_action: <RESTART_DEPLOYMENT|DELETE_POD|SCALE_DEPLOYMENT|NO_ACTION>
    kubectl_command: "<含 {{target}} {{namespace}} 的具體指令>"
    estimated_downtime: "<停機估計>"
    risk: <low|medium|critical>
    responsibility: <FE|BE|INFRA|DB|COLLAB>
    responsibility_reasoning: "<責任歸屬說明>"
    secondary_teams: []
    optimization:
      - type: <優化類型>
        description: "<優化說明>"
        command: "<含 {{target}} {{namespace}} 的指令>"
    reasoning: "[規則匹配] <處置邏輯說明>"

規則 id 使用 snake_case，不含特殊字符。\
"""


def get_incident_type(alertname: str) -> str:
    """
    I1 整合 ADR-064 Rule Engine (2026-04-11):
    從 YAML 規則動態推斷 incident_type，取代 webhooks.py 靜態 dict。

    優先順序:
    1. alert_rules.yaml 中 match.alertname 完全匹配 且 rule 有 incident_type 欄位 → 使用之
    2. ALERTNAME_TO_TYPE 靜態 dict fallback（constants/alert_types.py）
    3. "custom"（兜底）

    注意：YAML rule.id 語意為「規則識別符」，不等於 incident_type；
    兩者命名空間不同，不可混用。YAML 未設 incident_type 時一律走 fallback。
    """
    try:
        rules = _load_rules()
        for rule in rules:
            if _is_generic(rule):
                continue
            alertnames = rule.get("match", {}).get("alertname", [])
            if alertname in alertnames:
                # 只有 YAML 明確設定 incident_type 才採用，否則 fall through 到靜態 dict
                incident_type = rule.get("incident_type")
                if incident_type:
                    return incident_type
                break
    except Exception:
        pass

    return ALERTNAME_TO_TYPE.get(alertname, "custom")


def _rule_id_exists(alertname: str) -> bool:
    """檢查 alertname 是否已有規則（排除通用兜底）"""
    try:
        rules = _load_rules()
        for rule in rules:
            if _is_generic(rule):
                continue
            if alertname in rule.get("match", {}).get("alertname", []):
                return True
    except Exception:
        pass
    return False


def _append_rule_to_yaml(rule_yaml: str, alertname: str) -> bool:
    """驗證並 append 規則到 alert_rules.yaml，清除 lru_cache"""
    try:
        parsed = yaml.safe_load(rule_yaml)
        if not isinstance(parsed, list) or not parsed:
            logger.warning("auto_rule_invalid_yaml", alertname=alertname, raw=rule_yaml[:200])
            return False

        rule = parsed[0]
        required = {"id", "match", "response"}
        if not required.issubset(rule.keys()):
            logger.warning("auto_rule_missing_fields", alertname=alertname, keys=list(rule.keys()))
            return False

        resp = rule.get("response", {})
        if not resp.get("kubectl_command") or not resp.get("suggested_action"):
            logger.warning("auto_rule_empty_response", alertname=alertname)
            return False

        # S1-2 修正: dedent 正規化 LLM 可能輸出的前置空格，再加 2 spaces 縮排到 rules: 下
        import textwrap
        normalized = textwrap.dedent(rule_yaml.strip())

        # append 到 YAML 檔
        with RULES_FILE.open("a", encoding="utf-8") as f:
            now = datetime.now().strftime("%Y-%m-%d %H:%M")
            f.write(f"\n  # AUTO-GENERATED {now} — alertname={alertname}\n")
            for line in normalized.splitlines():
                f.write(f"  {line}\n")

        # 清除 lru_cache 讓新規則立即生效
        _load_rules.cache_clear()
        logger.info("auto_rule_written", rule_id=rule.get("id"), alertname=alertname)
        return True

    except Exception as e:
        logger.error("auto_rule_append_failed", alertname=alertname, error=str(e))
        return False


def _parse_for_to_seconds(for_str: str) -> int | None:
    """將 Prometheus 'for' 字串（如 '5m', '30s', '1h'）轉換為整數秒數。
    無法解析時回傳 None。
    """
    if not for_str:
        return None
    for_str = for_str.strip()
    mapping = {"s": 1, "m": 60, "h": 3600, "d": 86400}
    m = re.fullmatch(r"(\d+)([smhd]?)", for_str)
    if not m:
        return None
    value = int(m.group(1))
    unit = m.group(2) or "s"
    return value * mapping.get(unit, 1)


async def _insert_catalog_ai_generated(
    rule_dict: dict,
    llm_source: str,
    rule_id: str,
    alertname_safe: str,
) -> None:
    """寫入 alert_rule_catalog source='ai_generated'

    冪等：rule_name 唯一索引（alert_rule_catalog_rule_name_key）已存在
    使用 INSERT ON CONFLICT (rule_name) DO NOTHING

    transaction 策略：YAML + DB 不在同一 transaction
    YAML 已成功 → DB 失敗只 log warning，不回滾 YAML

    review_status='draft'：對應 DB CHECK ('draft','approved','deprecated','retired')
    confidence=0.50：新規則未驗證，保守初始值

    ADR-091 Task T1 — 2026-04-28 ogt + Claude Sonnet 4.6 (Asia/Taipei)
    """
    from src.core.config import settings as _settings

    if not _settings.ENABLE_AI_RULE_CATALOG_WRITE:
        logger.debug(
            "ai_rule_catalog_write_skipped_flag_disabled",
            alertname=alertname_safe,
        )
        return

    from sqlalchemy import text as _sql
    import src.db.base as _db_base
    import json as _json

    # 從 rule_dict 提取欄位
    # 'expr' 在 OpenClaw YAML 規則中不存在（非 PromQL），
    # 使用 alertname 作為語意佔位（與 yaml_hardcoded 同等策略）
    expr = rule_dict.get("expr") or f'alertname="{alertname_safe}"'
    for_str = rule_dict.get("for", "0s")
    duration_seconds = _parse_for_to_seconds(str(for_str))
    labels = rule_dict.get("labels", {})
    annotations = rule_dict.get("annotations", {})

    # 若 LLM 有產出 incident_type，注入 annotations 方便後續 T3 查詢
    incident_type = rule_dict.get("incident_type")
    if incident_type and "incident_type" not in annotations:
        annotations = {**annotations, "incident_type": str(incident_type)}

    # severity 從 labels 取（Prometheus 慣例），兜底空字串
    severity = (labels.get("severity", "") or "").strip()[:50] or None

    try:
        async with _db_base.get_db_context() as db:
            await db.execute(
                _sql("""
                    INSERT INTO alert_rule_catalog (
                        rule_name, source, expr, duration_seconds,
                        severity, labels, annotations,
                        created_by_agent, confidence, review_status,
                        created_at, updated_at
                    ) VALUES (
                        :rule_name, 'ai_generated', :expr, :duration_seconds,
                        :severity,
                        CAST(:labels AS jsonb),
                        CAST(:annotations AS jsonb),
                        :created_by_agent, 0.50, 'draft',
                        NOW(), NOW()
                    )
                    ON CONFLICT (rule_name) DO NOTHING
                """),
                {
                    "rule_name": alertname_safe[:200],
                    "expr": expr[:4000],
                    "duration_seconds": duration_seconds,
                    "severity": severity,
                    "labels": _json.dumps(labels, ensure_ascii=False),
                    "annotations": _json.dumps(annotations, ensure_ascii=False),
                    "created_by_agent": llm_source,
                },
            )
        logger.info(
            "ai_rule_catalog_insert_success",
            alertname=alertname_safe,
            rule_id=rule_id,
            llm_source=llm_source,
            outcome="success",
        )
    except Exception as db_err:
        # DB 失敗只 warning，不影響已成功的 YAML 寫入
        logger.warning(
            "ai_rule_catalog_insert_failed",
            alertname=alertname_safe,
            rule_id=rule_id,
            error=str(db_err),
            outcome="failed",
        )


async def _call_ollama(prompt: str, ollama_url: str, model: str) -> str | None:
    """呼叫 Ollama 生成規則 YAML"""
    try:
        async with httpx.AsyncClient(timeout=60) as client:
            resp = await client.post(
                f"{ollama_url}/api/generate",
                json={"model": model, "prompt": prompt, "stream": False, "options": {"temperature": 0.1}},
            )
            resp.raise_for_status()
            return resp.json().get("response", "")
    except Exception as e:
        logger.warning("auto_rule_ollama_failed", error=str(e))
        return None


async def _call_gemini(prompt: str, api_key: str) -> str | None:
    """呼叫 Gemini 生成規則 YAML（Ollama 失敗時備援）"""
    try:
        url = "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent"
        async with httpx.AsyncClient(timeout=30) as client:
            resp = await client.post(
                url,
                headers={"x-goog-api-key": api_key},
                json={"contents": [{"parts": [{"text": prompt}]}],
                      "generationConfig": {"temperature": 0.1}},
            )
            resp.raise_for_status()
            data = resp.json()
            return data["candidates"][0]["content"]["parts"][0]["text"]
    except Exception as e:
        logger.warning("auto_rule_gemini_failed", error=str(e))
        return None


def _extract_yaml_block(text: str) -> str:
    """從 LLM 回應中提取 YAML 區塊"""
    # 去掉 markdown code fence
    text = re.sub(r"```(?:yaml)?\n?", "", text).strip()
    # 確保以 `- id:` 開頭
    match = re.search(r"(- id:.+)", text, re.DOTALL)
    return match.group(1).strip() if match else text


async def auto_generate_rule(
    alert_context: dict,
    ollama_url: str,
    model: str,
    gemini_api_key: str = "",
) -> None:
    """
    非同步背景任務：呼叫 AI 為未知告警自動生成規則並寫入 alert_rules.yaml。

    觸發條件: match_rule() 命中 generic_fallback
    流程: Ollama → 失敗則 Gemini → 驗證格式 → append YAML → 清除 lru_cache 立即生效

    Args:
        alert_context: 告警上下文
        ollama_url: Ollama endpoint（由呼叫方從 settings 注入，S2-1 DI 修正）
        model: Ollama 模型名稱
        gemini_api_key: Gemini API Key（空字串則跳過 Gemini 備援）

    限制:
        - 進程級去重 (_generating set)，多 Pod 環境可能重複生成（ADR-064 已記錄）
        - 寫入後清除 lru_cache，同 Pod 立即生效；其他 Pod 需重啟
    """
    labels = alert_context.get("labels", {})
    alertname = labels.get("alertname", alert_context.get("alert_type", "custom"))

    # S3-3 修正: sanitize alertname，防止含 {/} 的 alertname 在 format() 中拋出 KeyError
    alertname_safe = re.sub(r"[{}]", "", alertname)

    # 去重：同一 alertname 同時只跑一次
    # ADR-064 L1: 優先用 Redis 分散式鎖，Redis 不可用時 fallback 進程級 set
    if alertname_safe in _generating:
        return
    if _rule_id_exists(alertname_safe):
        logger.debug("auto_rule_skip_exists", alertname=alertname_safe)
        return

    # 嘗試取得 Redis 分散式鎖
    lock_key = f"rule_generating:{alertname_safe}"
    redis_lock_acquired = False
    try:
        from src.core.redis_client import get_redis
        redis = get_redis()
        # SET NX EX — 只有第一個 Pod 能 SET 成功
        redis_lock_acquired = bool(await redis.set(lock_key, "1", nx=True, ex=_RULE_GEN_LOCK_TTL))
        if not redis_lock_acquired:
            logger.info("auto_rule_skip_redis_lock", alertname=alertname_safe)
            return
    except Exception as redis_err:
        # Redis 不可用時 fallback 進程級去重（降級不中斷）
        logger.warning("auto_rule_redis_lock_unavailable", error=str(redis_err))
        if alertname_safe in _generating:
            return

    _generating.add(alertname_safe)
    try:
        rule_id = re.sub(r"[^a-z0-9_]", "_", alertname_safe.lower()).strip("_")

        # S3-2 修正: priority 上界 890，防止超出 AI 生成範圍
        existing = [r.get("priority", 0) for r in _load_rules() if not _is_generic(r)]
        next_priority = max((p for p in existing if 500 <= p < 900), default=499) + 10
        priority = min(next_priority, 890)

        prompt = _AUTO_RULE_PROMPT.format(
            alertname=alertname_safe,
            alert_type=alert_context.get("alert_type", "custom"),
            message=alert_context.get("message", "")[:200],
            labels=json.dumps(
                {k: v for k, v in labels.items()
                 if k in ("job", "instance", "severity", "namespace", "container", "name")},
                ensure_ascii=False,
            ),
            rule_id=rule_id,
            priority=priority,
        )

        logger.info("auto_rule_generating", alertname=alertname_safe, rule_id=rule_id)

        # 1. 先試 Ollama
        raw = await _call_ollama(prompt, ollama_url, model)
        llm_source = "ollama" if raw else None

        # 2. Ollama 失敗 → Gemini
        if not raw and gemini_api_key:
            raw = await _call_gemini(prompt, gemini_api_key)
            llm_source = "gemini" if raw else None

        if not raw:
            logger.warning(
                "auto_rule_auto_generate_failed",
                alertname=alertname_safe,
                reason="llm_no_response",
            )
            return

        yaml_block = _extract_yaml_block(raw)
        success = _append_rule_to_yaml(yaml_block, alertname_safe)
        if success:
            logger.info(
                "alert_rule_auto_generated",
                alertname=alertname_safe,
                rule_id=rule_id,
                source=llm_source,
            )
            # 成功後記錄今日新增規則數（供系統報告顯示）
            import asyncio as _asyncio
            try:
                from src.core.redis_client import get_redis as _get_redis
                from src.utils.timezone import now_taipei as _now_taipei
                _today = _now_taipei().strftime("%Y%m%d")
                _key = f"stats:auto_rule_generated:{_today}"
                _redis = _get_redis()

                async def _incr_rule_stat() -> None:
                    # pipeline 原子化 incr+expire，避免 race condition
                    async with _redis.pipeline() as _p:
                        _p.incr(_key)
                        _p.expire(_key, 86400 * 7)
                        await _p.execute()

                _asyncio.create_task(_incr_rule_stat())
            except Exception as _redis_err:
                logger.debug("auto_rule_stats_redis_failed", error=str(_redis_err))
            # 立即為新規則建立 APPROVED Playbook（不等下次重啟）
            from src.services.playbook_seed_service import seed_playbooks_from_rules
            _asyncio.create_task(seed_playbooks_from_rules())
            # ADR-091 T1: 雙寫 alert_rule_catalog source='ai_generated'
            # 獨立 try/except — DB 失敗不回滾已成功的 YAML 寫入
            try:
                parsed_rules = yaml.safe_load(yaml_block)
                rule_dict = parsed_rules[0] if isinstance(parsed_rules, list) and parsed_rules else {}
                _asyncio.create_task(
                    _insert_catalog_ai_generated(
                        rule_dict=rule_dict,
                        llm_source=llm_source or "unknown",
                        rule_id=rule_id,
                        alertname_safe=alertname_safe,
                    )
                )
            except Exception as _catalog_err:
                logger.warning(
                    "ai_rule_catalog_task_create_failed",
                    alertname=alertname_safe,
                    error=str(_catalog_err),
                )
        else:
            logger.warning(
                "auto_rule_auto_generate_failed",
                alertname=alertname_safe,
                reason="yaml_validation_failed",
            )

    except Exception as e:
        logger.error("auto_rule_exception", alertname=alertname_safe, error=str(e))
    finally:
        _generating.discard(alertname_safe)
        # 釋放 Redis 分散式鎖（生成完成，讓其他 Pod 可以讀到新規則）
        if redis_lock_acquired:
            try:
                from src.core.redis_client import get_redis
                await get_redis().delete(lock_key)
            except Exception:
                pass  # TTL 到期自動釋放，不影響正確性