feat(flywheel): W2 三件 + KMWriter critic 修法（1635 tests 全綠）

W2 (onboarder 4 週飛輪 80→90 路徑第二週) + critic PR review 5 個 critical/major 全部修完，default flag=false 安全無爆炸風險。 ## W2 三件 PR ### PR-R2 — AOL → catalog confidence EWMA 回灌（修飛輪斷鏈 C2） - 新檔 `apps/api/src/jobs/aol_to_catalog_writeback_job.py` - 邏輯：每小時掃 AOL 計算 EWMA confidence (alpha=0.3) 回灌 alert_rule_catalog - 失敗閾值 N=5 連續低成功率 → review_status='draft' - Hermes _fetch_noisy_rules SQL 加 OR review_status='draft' - ENABLE_AOL_WRITEBACK_JOB=false (default) - 8 個測試（mock path 修正：lazy import → patch src.db.base.get_db_context） ### PR-V1 — self_healing_validator 串接 (修飛輪斷鏈 C6) - 新檔 `apps/api/src/services/self_healing_validator.py`（純函數 assess_self_healing） - post_execution_verifier.py step 5 串接（feature flag gate） - evidence_snapshot.py 加 self_healing_score / self_healing_detail 欄位 - db/models.py + base.py ALTER IF NOT EXISTS - score < 0.5 → 觸發 rollback 提案 Telegram alert（不自動執行） - ENABLE_SELF_HEALING_VALIDATOR=false (default) - 7 個測試 ### PR-L1 — KM ↔ Playbook 雙向回路 (修飛輪斷鏈 C3+C4) - learning_service.py 三條新邏輯： 1. _write_playbook_evolution_km：promote/demote 寫 KM 演化條目 2. _check_and_mark_playbook_review：N=5 累積觸發 review_required 3. _demote_alert_rule_catalog_confidence：DEPRECATED → confidence×=0.5 - PlaybookRecord 加 review_required 欄位（schema migration via base.py） - ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP=false (default) - KM_PLAYBOOK_REVIEW_THRESHOLD=5 可調 - 6 個測試 ## KMWriter Critic 5 個 Critical/Major 修復（之前 critic PR review 發現）之前 push commit c5753e1c 已修，本 commit 補回 stash 中的對應檔案： - C1 km_writer.py:194 backfill 自打臉（已修：同步 await + DLQ） - C2 km_writer.py:391 KM_WRITE_AWAIT=false 路徑收緊 - M1 decision_manager.py:2178/2203 移除 _fire_and_forget - M2 incident_service.py:1099 自製 path 加 retry+DLQ - M3 km_writer.py:166 冪等聲明對齊（UPSERT + partial unique index） ## 驗證 - 1635 unit tests 全綠（+27 from 1608） - 與 fb0c72db (推翻 A2 Ollama primary) 共存無衝突 - 所有新 Job/Service default flag=false（不爆炸） ## 期望影響飛輪斷鏈 C2 + C3 + C4 + C6 全修飛輪自主化評分：65 → 85 預估（W2 完成後）啟用順序（待 prod fb0c72db 驗證 OLLAMA primary 跑得起來後）： 1. ENABLE_AOL_WRITEBACK_JOB=true 2. ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP=true 3. ENABLE_SELF_HEALING_VALIDATOR=true Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-29 19:44:04 +08:00
parent fb0c72db42
commit 3668d49f2f
13 changed files with 2294 additions and 6 deletions
--- a/apps/api/src/services/learning_service.py
+++ b/apps/api/src/services/learning_service.py
@@ -389,13 +389,40 @@ class LearningService:
        playbook_id: str,
        success: bool,
    ) -> None:
-        """更新 Playbook 統計"""
+        """
+        更新 Playbook 統計
+
+        W2 PR-L1: 統計更新後，取 Playbook 的 symptom_pattern hash 觸發邏輯 2
+        （KM 累積門檻檢查 → review_required 標記）。
+        """
        try:
            from src.services.playbook_service import get_playbook_service

            service = get_playbook_service()
            await service.record_execution(playbook_id, success)

+            # W2 PR-L1 邏輯 2: 取得 Playbook symptom_pattern hash，觸發 KM 累積檢查
+            from src.core.config import settings
+            if settings.ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP:
+                try:
+                    from src.repositories.playbook_repository import get_playbook_repository
+                    from src.models.playbook import SymptomPattern
+                    repo = get_playbook_repository()
+                    playbook = await repo.get_by_id(playbook_id)
+                    if playbook and playbook.symptom_pattern:
+                        sp = playbook.symptom_pattern
+                        # symptom_pattern 可能是 Pydantic model 或 dict（ORM 載入）
+                        if isinstance(sp, dict):
+                            sp = SymptomPattern.model_validate(sp)
+                        symptoms_hash = sp.compute_hash()
+                        await self._check_and_mark_playbook_review(symptoms_hash)
+                except Exception as inner_e:
+                    logger.warning(
+                        "playbook_review_check_failed",
+                        playbook_id=playbook_id,
+                        error=str(inner_e),
+                    )
+
        except Exception as e:
            logger.warning(
                "playbook_stats_update_error",
@@ -459,6 +486,7 @@ class LearningService:
        - 尋找 source_incident_ids 包含此 incident_id 的 Playbooks
        - 提升 ai_confidence +0.1 (上限 1.0)
        - 若信心度 >= 0.9 且 status == DRAFT → 自動升級為 APPROVED
+        - W2 PR-L1: 寫 KM 演化條目（ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP 開啟時）
        """
        try:
            from src.repositories.playbook_repository import get_playbook_repository
@@ -478,6 +506,7 @@ class LearningService:

            updated_count = 0
            for playbook in playbooks:
+                previous_trust = playbook.trust_score
                result = await repo.adjust_confidence(
                    playbook_id=playbook.playbook_id,
                    delta=CONFIDENCE_BOOST,
@@ -485,6 +514,13 @@ class LearningService:
                )
                if result:
                    updated_count += 1
+                    # W2 PR-L1: promote 觸發 → 寫 KM 演化條目
+                    await self._write_playbook_evolution_km(
+                        playbook=playbook,
+                        previous_trust=previous_trust,
+                        evolution_type="promote",
+                        incident_id=incident_id,
+                    )

            logger.info(
                "playbook_promoted",
@@ -513,6 +549,7 @@ class LearningService:
        - 尋找 source_incident_ids 包含此 incident_id 的 Playbooks
        - 降低 ai_confidence -0.15 (下限 0.0)
        - 若信心度 < 0.3 且 failure_rate > 50% → 自動降級為 DEPRECATED
+        - W2 PR-L1: 寫 KM 演化條目；DEPRECATED 時回灌 alert_rule_catalog（飛輪 C4 修復）
        """
        try:
            from src.repositories.playbook_repository import get_playbook_repository
@@ -532,6 +569,7 @@ class LearningService:

            updated_count = 0
            for playbook in playbooks:
+                previous_trust = playbook.trust_score
                result = await repo.adjust_confidence(
                    playbook_id=playbook.playbook_id,
                    delta=CONFIDENCE_PENALTY,
@@ -539,6 +577,17 @@ class LearningService:
                )
                if result:
                    updated_count += 1
+                    # W2 PR-L1: demote 觸發 → 寫 KM 演化條目
+                    await self._write_playbook_evolution_km(
+                        playbook=playbook,
+                        previous_trust=previous_trust,
+                        evolution_type="demote",
+                        incident_id=incident_id,
+                    )
+                    # W2 PR-L1 邏輯 3: DEPRECATED 時回灌 alert_rule_catalog（飛輪 C4 修復）
+                    from src.models.playbook import PlaybookStatus
+                    if playbook.status == PlaybookStatus.DEPRECATED:
+                        await self._demote_alert_rule_catalog_confidence(playbook)

            logger.info(
                "playbook_demoted",
@@ -557,6 +606,241 @@ class LearningService:
            )
            return False

+    # =========================================================================
+    # W2 PR-L1: KM → Playbook 互饋回路私有方法
+    # 飛輪斷鏈 C3 + C4 修復
+    # 2026-04-28 ogt + Claude Sonnet 4.6
+    # =========================================================================
+
+    async def _write_playbook_evolution_km(
+        self,
+        playbook: Any,
+        previous_trust: float,
+        evolution_type: str,
+        incident_id: str,
+    ) -> None:
+        """
+        邏輯 1: promote/demote 觸發 → 寫 KM 演化條目（飛輪 C3）
+
+        KM 條目 metadata 含：playbook_id, previous_trust, new_trust,
+                              success_count, failure_count, decision_chain
+        path_type='playbook_evolution'，供冪等 key 使用
+        (incident_id, path_type) = (incident_id, 'playbook_evolution') 可能重複，
+        但 playbook_id 不同的演化各自獨立，所以 path_type 加 playbook_id 作為識別。
+        """
+        from src.core.config import settings
+        if not settings.ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP:
+            return
+
+        try:
+            import json
+            from src.services.km_writer import KMWritePayload, km_write_with_flag
+            from src.utils.timezone import now_taipei
+
+            new_trust = getattr(playbook, "trust_score", previous_trust)
+            success_count = getattr(playbook, "success_count", 0)
+            failure_count = getattr(playbook, "failure_count", 0)
+
+            path_type = f"playbook_evolution:{playbook.playbook_id}"
+
+            payload = KMWritePayload(
+                path_type=path_type,
+                incident_id=incident_id,
+                entry_create_kwargs={
+                    "title": f"Playbook {evolution_type}: {playbook.name} [{playbook.playbook_id}]",
+                    "content": (
+                        f"Playbook {evolution_type} 事件記錄\n"
+                        f"Playbook ID: {playbook.playbook_id}\n"
+                        f"名稱: {playbook.name}\n"
+                        f"trust_score 變化: {previous_trust:.3f} → {new_trust:.3f}\n"
+                        f"成功次數: {success_count} / 失敗次數: {failure_count}\n"
+                        f"觸發來源: incident {incident_id}\n"
+                        f"記錄時間: {now_taipei().isoformat()}"
+                    ),
+                    "entry_type": "best_practice",
+                    "category": "AI系統",
+                    "tags": ["playbook_evolution", evolution_type, playbook.playbook_id],
+                    "source": "ai_extracted",
+                    "related_playbook_id": playbook.playbook_id,
+                    "related_incident_id": incident_id,
+                    "path_type": path_type,
+                },
+                metadata={
+                    "playbook_id": playbook.playbook_id,
+                    "previous_trust": previous_trust,
+                    "new_trust": new_trust,
+                    "success_count": success_count,
+                    "failure_count": failure_count,
+                    "evolution_type": evolution_type,
+                },
+            )
+            await km_write_with_flag(payload)
+            logger.info(
+                "playbook_evolution_km_written",
+                playbook_id=playbook.playbook_id,
+                evolution_type=evolution_type,
+                trust_change=f"{previous_trust:.3f} → {new_trust:.3f}",
+            )
+        except Exception as e:
+            logger.warning(
+                "playbook_evolution_km_write_failed",
+                playbook_id=getattr(playbook, "playbook_id", "unknown"),
+                evolution_type=evolution_type,
+                error=str(e),
+            )
+
+    async def _check_and_mark_playbook_review(self, symptoms_hash: str) -> None:
+        """
+        邏輯 2: KM 累積 N=5 條同 symptom_pattern_hash → 觸發 Playbook review_required 標記（飛輪 C3）
+
+        每次 KM 寫入後由 _update_playbook_stats 呼叫端觸發此檢查。
+        若同 symptoms_hash 在 knowledge_entries 已有 >= threshold 條，
+        則 UPDATE playbooks SET review_required=true WHERE 症狀 hash 相符。
+
+        比對策略：從 KnowledgeEntry 讀 symptoms_hash 計數，
+                  再透過 playbook.symptom_pattern 的 hash 比對 Playbook。
+        """
+        from src.core.config import settings
+        if not settings.ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP:
+            return
+        if not symptoms_hash:
+            return
+
+        try:
+            from sqlalchemy import text as sa_text
+            from src.db.base import get_db_context
+
+            async with get_db_context() as db:
+                # 計算同 symptoms_hash 的 KM 條目數
+                count_result = await db.execute(
+                    sa_text(
+                        "SELECT COUNT(*) FROM knowledge_entries "
+                        "WHERE symptoms_hash = :hash"
+                    ),
+                    {"hash": symptoms_hash},
+                )
+                count = count_result.scalar() or 0
+
+                if count < settings.KM_PLAYBOOK_REVIEW_THRESHOLD:
+                    return
+
+                # 累積達到門檻 → 標記相關 Playbook 需要 review
+                # Playbook 的 symptom_pattern 存為 JSONB，無直接 hash 欄位
+                # 透過 knowledge_entries.related_playbook_id 關聯找到要標記的 Playbook
+                updated = await db.execute(
+                    sa_text(
+                        "UPDATE playbooks pb "
+                        "SET review_required = true, updated_at = NOW() "
+                        "FROM knowledge_entries ke "
+                        "WHERE ke.symptoms_hash = :hash "
+                        "  AND ke.related_playbook_id = pb.playbook_id "
+                        "  AND pb.review_required = false "
+                        "RETURNING pb.playbook_id"
+                    ),
+                    {"hash": symptoms_hash},
+                )
+                marked_ids = [row[0] for row in updated.fetchall()]
+                await db.commit()
+
+            if marked_ids:
+                logger.info(
+                    "playbook_review_required_marked",
+                    symptoms_hash=symptoms_hash,
+                    km_count=count,
+                    threshold=settings.KM_PLAYBOOK_REVIEW_THRESHOLD,
+                    playbook_ids=marked_ids,
+                )
+        except Exception as e:
+            logger.warning(
+                "playbook_review_mark_failed",
+                symptoms_hash=symptoms_hash,
+                error=str(e),
+            )
+
+    async def _demote_alert_rule_catalog_confidence(self, playbook: Any) -> None:
+        """
+        邏輯 3: Playbook DEPRECATED 時回灌 alert_rule_catalog（飛輪 C4 修復）
+
+        UPDATE alert_rule_catalog
+        SET confidence = confidence * 0.5,
+            review_status = 'draft'   -- CHECK constraint 允許 draft/approved/deprecated/retired
+        WHERE rule_name LIKE pattern(symptom_pattern.alert_names)
+
+        注意：alert_rule_catalog.review_status CHECK 限制只允許：
+              draft | approved | deprecated | retired
+              任務描述的 'needs_review' 不合法，改用 'draft'（語意等效：需要人工審核）
+
+        失敗容忍：不影響 demote 主流程。
+        """
+        from src.core.config import settings
+        if not settings.ENABLE_KM_PLAYBOOK_FEEDBACK_LOOP:
+            return
+
+        try:
+            import json
+            from sqlalchemy import text as sa_text
+            from src.db.base import get_db_context
+
+            # 從 playbook symptom_pattern 取出 alert_names 作為比對鍵
+            symptom = getattr(playbook, "symptom_pattern", None)
+            if symptom is None:
+                return
+
+            # symptom_pattern 可能是 Pydantic model 或 dict（從 ORM 載入為 dict）
+            if hasattr(symptom, "alert_names"):
+                alert_names: list[str] = symptom.alert_names or []
+            elif isinstance(symptom, dict):
+                alert_names = symptom.get("alert_names") or []
+            else:
+                return
+
+            if not alert_names:
+                logger.debug(
+                    "playbook_demote_no_alert_names",
+                    playbook_id=playbook.playbook_id,
+                )
+                return
+
+            async with get_db_context() as db:
+                updated_count = 0
+                for alert_name in alert_names:
+                    # rule_name 完全匹配或前綴匹配（去掉 * suffix）
+                    match_name = alert_name.rstrip("*")
+                    result = await db.execute(
+                        sa_text(
+                            "UPDATE alert_rule_catalog "
+                            "SET confidence = CASE "
+                            "      WHEN confidence IS NOT NULL "
+                            "      THEN GREATEST(0.01, confidence * 0.5) "
+                            "      ELSE 0.5 "
+                            "    END, "
+                            "    review_status = 'draft', "
+                            "    updated_at = NOW() "
+                            "WHERE rule_name LIKE :pattern "
+                            "  AND (review_status IS NULL OR review_status NOT IN "
+                            "       ('deprecated', 'retired')) "
+                            "RETURNING rule_id"
+                        ),
+                        {"pattern": f"{match_name}%"},
+                    )
+                    affected = result.rowcount or 0
+                    updated_count += affected
+                await db.commit()
+
+            if updated_count > 0:
+                logger.info(
+                    "alert_rule_catalog_confidence_demoted",
+                    playbook_id=playbook.playbook_id,
+                    alert_names=alert_names,
+                    rules_updated=updated_count,
+                )
+        except Exception as e:
+            logger.warning(
+                "alert_rule_catalog_demote_failed",
+                playbook_id=getattr(playbook, "playbook_id", "unknown"),
+                error=str(e),
+            )
+
    # =========================================================================
    # 🆕 Phase D-G P0 修正: 新增方法
    # =========================================================================