feat(agents): route high risk through controlled automation

apps/api/src/services/failure_watcher.py

@@ -7,7 +7,7 @@ Phase 18: Failure Auto-Repair Loop (2026-03-31 統帥批准)
 - 監聽 AuditLog 失敗事件
 - AI 分析失敗原因
 - 評估風險等級
-- 執行自動修復 (LOW 風險) 或請求人工授權 (MEDIUM/CRITICAL)
+- LOW / MEDIUM / HIGH 由 AI Agent 受控自動修復；CRITICAL / break-glass 才硬封鎖
 
 設計原則:
 - 實作 IFailureWatcher Protocol
@@ -53,9 +53,13 @@ RISK_LEVELS = {
         "operations": ["restart_pod", "restart_deployment", "clear_cache"],
     },
     "MEDIUM": {
-        "auto_repair": False,
+        "auto_repair": True,
         "operations": ["scale_deployment", "rollback", "update_config"],
     },
+    "HIGH": {
+        "auto_repair": True,
+        "operations": ["controlled_rollback", "controlled_config_repair", "controlled_service_repair"],
+    },
     "CRITICAL": {
         "auto_repair": False,
         "operations": ["delete_pvc", "drop_database", "network_policy"],
@@ -108,8 +112,8 @@ class FailureWatcherService(IFailureWatcher):
     1. 收到失敗事件 (from Redis Stream or direct call)
     2. AI 分析失敗原因 (OpenClaw)
     3. 評估風險等級
-    4. LOW → 自動修復 → 揭露通知
-    5. MEDIUM/CRITICAL → Telegram + 前端等待授權
+    4. LOW / MEDIUM / HIGH → 受控自動修復 → 揭露通知
+    5. CRITICAL / hard blocker → Telegram + 前端 break-glass blocked 狀態
     """
 
     def __init__(self) -> None:
@@ -183,11 +187,11 @@ class FailureWatcherService(IFailureWatcher):
                 target_resource=target_resource,
                 reason=global_reason,
             )
-            # 強制升級為 CRITICAL，必須人工授權
+            # 全域熔斷代表 repair storm 風險，仍維持 break-glass blocked。
             risk_level = "CRITICAL"
             result["risk_level"] = "CRITICAL"
             result["next_action"] = "blocked_by_global_cooldown"
-            await self._request_human_approval(
+            await self._queue_ai_repair_followup(
                 audit_log_id=audit_log_id,
                 analysis=analysis,
                 reason=global_reason,
@@ -201,7 +205,7 @@ class FailureWatcherService(IFailureWatcher):
         )
 
         if not can_auto_repair:
-            # 超過冷卻期限制，升級為 MEDIUM
+            # 超過單資源冷卻期限制，改由 AI 受控重試佇列處理，不再要求人工接手。
             logger.info(
                 "repair_cooldown_escalate",
                 audit_log_id=audit_log_id,
@@ -211,7 +215,7 @@ class FailureWatcherService(IFailureWatcher):
             risk_level = "MEDIUM"
             result["risk_level"] = "MEDIUM"
 
-        if risk_level == "LOW" and RISK_LEVELS["LOW"]["auto_repair"]:
+        if risk_level != "CRITICAL" and RISK_LEVELS.get(risk_level, {}).get("auto_repair"):
             # 自動修復 (Phase 18.3: 傳入完整 failure_data)
             success, repair_result = await self.execute_auto_repair(
                 audit_log_id=audit_log_id,
@@ -242,22 +246,22 @@ class FailureWatcherService(IFailureWatcher):
                     auto=True,
                 )
             else:
-                # 升級為 MEDIUM，請求人工授權
+                # 失敗後排入 AI controlled retry / rollback，不再直接轉人工。
                 result["risk_level"] = "MEDIUM"
-                await self._request_human_approval(
+                await self._queue_ai_repair_followup(
                     audit_log_id=audit_log_id,
                     analysis=analysis,
-                    reason="自動修復失敗，需人工介入",
+                    reason="AI 受控自動修復失敗，已排入 rollback / transport / PlayBook 修復重試",
                 )
-                result["next_action"] = "await_approval"
+                result["next_action"] = "ai_retry_queued"
         else:
-            # MEDIUM/CRITICAL: 請求人工授權
-            await self._request_human_approval(
+            # CRITICAL: break-glass / hard blocker，不執行寫入。
+            await self._queue_ai_repair_followup(
                 audit_log_id=audit_log_id,
                 analysis=analysis,
-                reason=f"風險等級 {risk_level}，需人工審核",
+                reason=f"風險等級 {risk_level} 命中 break-glass guardrail，AI 已封鎖寫入並保留證據",
             )
-            result["next_action"] = "await_approval"
+            result["next_action"] = "break_glass_blocked"
 
         logger.info(
             "failure_watcher_processed",
@@ -505,9 +509,9 @@ class FailureWatcherService(IFailureWatcher):
             "NETWORK_ERROR": "檢查網路連線，驗證 DNS 解析",
             "PERMISSION_DENIED": "檢查 RBAC 權限配置",
             "RESOURCE_ERROR": "增加資源配額或清理資源",
-            "UNKNOWN": "需人工分析日誌",
+            "UNKNOWN": "AI 深度診斷：補抓 timeline / log / metric / PlayBook 相似案例",
         }
-        return suggestions.get(classification, "需人工分析")
+        return suggestions.get(classification, "AI 深度診斷")
 
     async def _check_repair_cooldown(
         self,
@@ -709,7 +713,7 @@ class FailureWatcherService(IFailureWatcher):
                 error=str(e),
             )
 
-    async def _request_human_approval(
+    async def _queue_ai_repair_followup(
         self,
         audit_log_id: str,
         analysis: dict,
@@ -720,10 +724,11 @@ class FailureWatcherService(IFailureWatcher):
             # 推送到 Redis (前端 WebSocket 訂閱)
             redis = get_redis()
             repair_request = {
-                "type": "repair_request",
+                "type": "ai_controlled_repair_followup",
                 "audit_log_id": audit_log_id,
                 "analysis": analysis,
                 "reason": reason,
+                "controlled_retry": True,
                 "created_at": datetime.now(UTC).isoformat(),
             }
             await redis.publish(
@@ -736,18 +741,18 @@ class FailureWatcherService(IFailureWatcher):
 
             tg = get_telegram_gateway()
             message = (
-                f"🔧 <b>修復請求</b>\n\n"
+                f"🤖 <b>AI 受控修復後續</b>\n\n"
                 f"├ 📋 AuditLog: <code>{audit_log_id[:8]}...</code>\n"
                 f"├ 📊 分類: {analysis.get('classification', 'UNKNOWN')}\n"
                 f"├ ⚠️ 風險: {analysis.get('risk_level', 'MEDIUM')}\n"
                 f"├ 🔍 原因: {analysis.get('root_cause', reason)}\n"
-                f"└ 💡 建議: {analysis.get('suggested_repair', '需人工分析')}\n\n"
-                f"請在 Dashboard 授權或使用 /repair {audit_log_id[:8]}"
+                f"└ 💡 下一步: {analysis.get('suggested_repair', 'AI 深度診斷')}\n\n"
+                f"狀態：<code>ai_retry_or_break_glass_recorded</code>"
             )
             await tg.send_alert_notification(message)
 
             logger.info(
-                "repair_request_sent",
+                "ai_controlled_repair_followup_sent",
                 audit_log_id=audit_log_id,
                 risk_level=analysis.get("risk_level"),
             )