feat(Phase 3): 學習閉環重建 — 三根因修復 + 2x EWMA + Evolver Agent

ADR-083 Phase 3 學習閉環重建：

**三根因修復**
- approval_execution.py: fire-and-forget create_task → await asyncio.wait_for(timeout=30) × 2
  （成功路徑 L265 + 失敗路徑 L353，超時記錄 learning_trigger_timeout metric，主流程不 crash）
- models/approval.py: ApprovalRequestBase 新增 matched_playbook_id 欄位
- decision_manager.py: _auto_execute 建立 ApprovalRequest 時填充 matched_playbook_id
- learning_service.py: 雙路徑查找 _matched_pb_id（matched_playbook_id + metadata fallback）

**2x EWMA 負向強化**
- models/playbook.py: 新增 trust_score: float = 0.3（EWMA 動態信任度欄位）
- repositories/playbook_repository.py: update_stats 加 EWMA
  成功: trust = 0.9 × old + 0.1 × 1.0
  失敗: trust = 0.8 × old + 0.2 × 0.0（衰減速度 2x）
  trust < 0.1 → log warning，等 Evolver 封存

**Evolver Agent（新建）**
- services/playbook_evolver.py: 三功能全靜態規則
  1. 低信任封存: trust < 0.1 → DEPRECATED
  2. 休眠封存: 30d 未使用 AND trust < 0.5 → DEPRECATED
  3. 相似合併: 症狀 Jaccard > 0.9 → 保留高 trust，封存低 trust
  AIOPS_P3_EVOLVER_ENABLED=False 預設關閉

**文件**
- ADR-083 學習閉環重建
- MASTER §8 Phase 3 完工記錄

AIOPS_P3_ENABLED=False（預設），骨架就位等統帥批准開啟

Co-Authored-By: Claude Sonnet 4.6（亞太）<noreply@anthropic.com>

apps/api/src/models/approval.py

@@ -143,6 +143,9 @@ class ApprovalRequestBase(BaseModel):
     # 2026-04-14 Claude Sonnet 4.6: 上移 incident_id 到 Base，
     # 讓 ApprovalRequestCreate 也能攜帶（修 9b9ff5b 的 NoneAttr bug）
     incident_id: str | None = Field(default=None, description="關聯的 Incident ID")
+    # ADR-083 Phase 3: 命中的 Playbook ID（學習迴路必填）
+    # 2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 matched_playbook_id 傳遞修復
+    matched_playbook_id: str | None = Field(default=None, description="命中的 Playbook ID，供學習服務 EWMA 更新")
 
 
 class ApprovalRequestCreate(ApprovalRequestBase):

apps/api/src/models/playbook.py

@@ -205,6 +205,12 @@ class Playbook(BaseModel):
     success_count: int = Field(default=0, ge=0, description="成功執行次數")
     failure_count: int = Field(default=0, ge=0, description="失敗執行次數")
     last_used_at: datetime | None = Field(None, description="最後使用時間")
+    # ADR-083 Phase 3: EWMA 信任度（0.0-1.0，初值 0.3）
+    # 成功: trust_new = 0.9 * trust_old + 0.1 * 1.0
+    # 失敗: trust_new = 0.8 * trust_old + 0.2 * 0.0（2x 衰減）
+    # trust < 0.1 → 自動封存（由 Evolver Agent 處理）
+    # 2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 EWMA 負向強化
+    trust_score: float = Field(default=0.3, ge=0.0, le=1.0, description="EWMA 動態信任度（Phase 3 新增）")
 
     # === 人工標記 ===
     approved_by: str | None = Field(None, description="核准者")

apps/api/src/repositories/playbook_repository.py

@@ -292,7 +292,15 @@ class PlaybookRepository:
         playbook_id: str,
         success: bool,
     ) -> bool:
-        """更新執行統計"""
+        """
+        更新執行統計 + EWMA 信任度
+
+        ADR-083 Phase 3: 負向 2x 強化 EWMA 公式
+        成功: trust_new = 0.9 * trust_old + 0.1 * 1.0
+        失敗: trust_new = 0.8 * trust_old + 0.2 * 0.0（衰減速度 2x）
+        trust < 0.1 → 記錄警告，由 Evolver Agent 封存
+        2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 EWMA 實裝
+        """
         try:
             playbook = await self.get_by_id(playbook_id)
             if not playbook:
@@ -300,17 +308,32 @@ class PlaybookRepository:
 
             if success:
                 playbook.success_count += 1
+                # 正向 EWMA：alpha=0.1，正向結果權重較小（保守更新）
+                playbook.trust_score = 0.9 * playbook.trust_score + 0.1 * 1.0
             else:
                 playbook.failure_count += 1
+                # 負向 EWMA：alpha=0.2，失敗懲罰 2x（快速衰退）
+                playbook.trust_score = 0.8 * playbook.trust_score + 0.2 * 0.0
 
+            # 邊界保護
+            playbook.trust_score = max(0.0, min(1.0, playbook.trust_score))
             playbook.last_used_at = now_taipei()
             await self.update(playbook)
 
+            if playbook.trust_score < 0.1:
+                logger.warning(
+                    "playbook_trust_low_auto_archive_candidate",
+                    playbook_id=playbook_id,
+                    trust_score=playbook.trust_score,
+                    hint="Evolver Agent 應將此 Playbook 封存",
+                )
+
             logger.info(
                 "playbook_stats_updated",
                 playbook_id=playbook_id,
                 success=success,
                 success_rate=playbook.success_rate,
+                trust_score=playbook.trust_score,
             )
             return True
 

apps/api/src/services/approval_execution.py

@@ -261,14 +261,25 @@ class ApprovalExecutionService:
                 self._trigger_playbook_extraction(approval)
             )
 
-            # ADR-030 Phase 5: 觸發學習服務 (fire-and-forget)
-            asyncio.create_task(
-                self._trigger_learning(
-                    approval=approval,
-                    success=True,
-                    duration_seconds=result.duration_ms / 1000 if result.duration_ms else 0,
+            # ADR-030 Phase 5 / ADR-083 Phase 3: 觸發學習服務
+            # Phase 3 修復：移除 fire-and-forget，改用 await + 30s 熔斷
+            # 超時 → 記錄 metric，主流程繼續（不 crash）
+            # 2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 fire-and-forget 修復
+            try:
+                await asyncio.wait_for(
+                    self._trigger_learning(
+                        approval=approval,
+                        success=True,
+                        duration_seconds=result.duration_ms / 1000 if result.duration_ms else 0,
+                    ),
+                    timeout=30.0,
+                )
+            except asyncio.TimeoutError:
+                logger.warning(
+                    "learning_trigger_timeout",
+                    approval_id=str(approval.id),
+                    timeout_sec=30.0,
                 )
-            )
 
             # ADR-081 Phase 1: 執行後驗證 (fire-and-forget)
             # PostExecutionVerifier 等待 K8s 收斂後抓取後狀態，補填 EvidenceSnapshot
@@ -349,15 +360,25 @@ class ApprovalExecutionService:
                 )
             )
 
-            # ADR-030 Phase 5: 觸發學習服務 (失敗案例)
-            asyncio.create_task(
-                self._trigger_learning(
-                    approval=approval,
-                    success=False,
-                    error_message=result.error,
-                    duration_seconds=result.duration_ms / 1000 if result.duration_ms else 0,
+            # ADR-030 Phase 5 / ADR-083 Phase 3: 觸發學習服務（失敗案例）
+            # Phase 3 修復：fire-and-forget → await + 30s 熔斷
+            # 2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 fire-and-forget 修復
+            try:
+                await asyncio.wait_for(
+                    self._trigger_learning(
+                        approval=approval,
+                        success=False,
+                        error_message=result.error,
+                        duration_seconds=result.duration_ms / 1000 if result.duration_ms else 0,
+                    ),
+                    timeout=30.0,
+                )
+            except asyncio.TimeoutError:
+                logger.warning(
+                    "learning_trigger_timeout",
+                    approval_id=str(approval.id),
+                    timeout_sec=30.0,
                 )
-            )
 
     async def _push_execution_result_to_alert(
         self,

apps/api/src/services/decision_manager.py

@@ -1478,6 +1478,10 @@ class DecisionManager:
 
             # 建立虛擬 ApprovalRequest (auto_execute — 不需人工審核)
             _risk = token.proposal_data.get("risk_level", "low")
+            # ADR-083 Phase 3: 傳遞 matched_playbook_id，學習迴路 EWMA 才能觸發
+            # Playbook RAG 命中時 proposal_data["playbook_id"] 會有值（見 _try_playbook_match）
+            # 2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 matched_playbook_id 修復
+            _matched_playbook_id = token.proposal_data.get("playbook_id")
             approval = ApprovalRequest(
                 incident_id=incident.incident_id,
                 action=action,
@@ -1486,6 +1490,7 @@ class DecisionManager:
                 required_signatures=0,
                 status=ApprovalStatus.APPROVED,
                 risk_level=_risk,
+                matched_playbook_id=_matched_playbook_id,
             )
 
             # ADR-071-I: 執行前抓 metrics_before 快照 (2026-04-11 Claude Sonnet 4.6)

apps/api/src/services/learning_service.py

@@ -205,24 +205,33 @@ class LearningService:
         trust_after = trust_record.score if trust_record else 0
 
         # 2. 更新 Playbook 統計 (如果有匹配)
+        # ADR-083 Phase 3: 雙路徑查找 matched_playbook_id
+        # 路徑 A: ApprovalRequest.matched_playbook_id（auto_execute 路徑，Phase 3 修復）
+        # 路徑 B: approval.metadata["playbook_id"]（人工審核路徑，透過 proposal_service 存入 metadata）
+        # 2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 Playbook EWMA 修復
+        _matched_pb_id: str | None = (
+            getattr(approval, "matched_playbook_id", None)
+            or (approval.metadata or {}).get("matched_playbook_id")
+            or (approval.metadata or {}).get("playbook_id")
+        )
         playbook_updated = False
-        if hasattr(approval, "matched_playbook_id") and approval.matched_playbook_id:
+        if _matched_pb_id:
             try:
                 await self._update_playbook_stats(
-                    playbook_id=approval.matched_playbook_id,
+                    playbook_id=_matched_pb_id,
                     success=result.success,
                 )
                 playbook_updated = True
             except Exception as e:
                 logger.warning(
                     "playbook_stats_update_failed",
-                    playbook_id=approval.matched_playbook_id,
+                    playbook_id=_matched_pb_id,
                     error=str(e),
                 )
 
         # 3. 嘗試萃取新 Playbook (成功且無匹配 Playbook)
         new_playbook_id = None
-        if result.success and not getattr(approval, "matched_playbook_id", None):
+        if result.success and not _matched_pb_id:
             try:
                 new_playbook_id = await self._try_extract_playbook(
                     incident_id=result.incident_id,

apps/api/src/services/playbook_evolver.py

@@ -0,0 +1,356 @@
+"""
+AWOOOI AIOps Phase 3 — Playbook Evolver Agent（知識演化官）
+==========================================================
+職責：Playbook 自動合併、低信任封存、知識庫精化
+
+觸發方式：定時（每日凌晨）或手動呼叫 `run_evolver()`
+
+三大功能：
+1. 低信任封存 — trust_score < 0.1 → DEPRECATED（自動退場）
+2. 休眠封存 — 30 天未使用 AND trust_score < 0.5 → DEPRECATED
+3. 相似合併 — Jaccard 症狀相似度 > 0.9 → 合併為一（保留 trust 較高者）
+
+設計原則：
+- 純靜態規則（不依賴 LLM）— 保證確定性
+- 熔斷保護：單筆操作失敗不影響其他 Playbook
+- best-effort 審計：合併/封存寫 logger.info
+- feature flag 保護：AIOPS_P3_EVOLVER_ENABLED=False 時靜默跳過
+
+ADR-083: Phase 3 學習閉環重建
+2026-04-15 ogt + Claude Sonnet 4.6（亞太）: Phase 3 初始建立
+"""
+
+from __future__ import annotations
+
+import asyncio
+from dataclasses import dataclass, field
+from datetime import timedelta
+from typing import Any
+
+import structlog
+
+from src.models.playbook import Playbook, PlaybookStatus
+from src.utils.timezone import now_taipei
+
+logger = structlog.get_logger(__name__)
+
+# ── 閾值常數 ────────────────────────────────────────────────────────────────
+TRUST_ARCHIVE_THRESHOLD = 0.1       # trust_score < 此值 → 封存
+DORMANT_TRUST_THRESHOLD = 0.5       # 休眠封存：trust < 此值 AND 30d 未用
+DORMANT_DAYS = 30                   # 休眠天數
+MERGE_SIMILARITY_THRESHOLD = 0.9    # 症狀相似度 > 此值 → 合併候選
+MAX_MERGE_PER_RUN = 10              # 單次合併上限（防止操作過多）
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Result Types
+# ─────────────────────────────────────────────────────────────────────────────
+
+@dataclass
+class EvolverReport:
+    """Evolver 執行報告"""
+    archived_count: int = 0
+    merged_count: int = 0
+    skipped_count: int = 0
+    archived_ids: list[str] = field(default_factory=list)
+    merged_pairs: list[tuple[str, str]] = field(default_factory=list)  # (deprecated_id, kept_id)
+    errors: list[str] = field(default_factory=list)
+
+    @property
+    def total_affected(self) -> int:
+        return self.archived_count + self.merged_count
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Main Entry Point
+# ─────────────────────────────────────────────────────────────────────────────
+
+async def run_evolver() -> EvolverReport:
+    """
+    執行 Evolver Agent 全流程。
+
+    Returns:
+        EvolverReport（包含所有操作結果）
+
+    Raises:
+        不拋出 — 所有錯誤內部吸收，最壞情況返回空報告
+    """
+    from src.core.feature_flags import aiops_flags
+
+    if not aiops_flags.AIOPS_P3_EVOLVER_ENABLED:
+        logger.debug("evolver_skipped", reason="AIOPS_P3_EVOLVER_ENABLED=False")
+        return EvolverReport()
+
+    report = EvolverReport()
+
+    try:
+        playbooks = await _fetch_all_active_playbooks()
+        if not playbooks:
+            logger.info("evolver_no_playbooks")
+            return report
+
+        # Step 1: 低信任封存
+        await _archive_low_trust(playbooks, report)
+
+        # Step 2: 休眠封存（剩餘 APPROVED/DRAFT 中挑）
+        remaining = [p for p in playbooks if p.status not in (PlaybookStatus.DEPRECATED,)]
+        await _archive_dormant(remaining, report)
+
+        # Step 3: 相似合併
+        still_active = [p for p in remaining if p.status not in (PlaybookStatus.DEPRECATED,)]
+        await _merge_similar(still_active, report)
+
+        logger.info(
+            "evolver_done",
+            archived=report.archived_count,
+            merged=report.merged_count,
+            skipped=report.skipped_count,
+            errors=len(report.errors),
+        )
+
+    except Exception:
+        logger.exception("evolver_fatal")
+
+    return report
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Step Implementations
+# ─────────────────────────────────────────────────────────────────────────────
+
+async def _archive_low_trust(playbooks: list[Playbook], report: EvolverReport) -> None:
+    """Step 1: trust_score < 0.1 → DEPRECATED（自動退場）"""
+    from src.services.playbook_service import get_playbook_service
+
+    service = get_playbook_service()
+
+    for pb in playbooks:
+        if pb.status == PlaybookStatus.DEPRECATED:
+            continue
+        if pb.trust_score < TRUST_ARCHIVE_THRESHOLD:
+            try:
+                await service.update_with_validation(
+                    pb.playbook_id,
+                    {"status": PlaybookStatus.DEPRECATED.value},
+                )
+                logger.info(
+                    "evolver_archived_low_trust",
+                    playbook_id=pb.playbook_id,
+                    playbook_name=pb.name,
+                    trust_score=pb.trust_score,
+                )
+                report.archived_count += 1
+                report.archived_ids.append(pb.playbook_id)
+                # 原地更新 status 避免後續步驟重複處理
+                pb.status = PlaybookStatus.DEPRECATED
+            except Exception as e:
+                report.errors.append(f"archive_low_trust:{pb.playbook_id}:{e}")
+                logger.warning(
+                    "evolver_archive_failed",
+                    playbook_id=pb.playbook_id,
+                    error=str(e),
+                )
+
+
+async def _archive_dormant(playbooks: list[Playbook], report: EvolverReport) -> None:
+    """Step 2: 30d 未使用 AND trust < 0.5 → DEPRECATED（休眠退場）"""
+    from src.services.playbook_service import get_playbook_service
+
+    service = get_playbook_service()
+    cutoff = now_taipei() - timedelta(days=DORMANT_DAYS)
+
+    for pb in playbooks:
+        if pb.status == PlaybookStatus.DEPRECATED:
+            continue
+        if pb.last_used_at is None:
+            # 從未使用過 — 只在 trust 低於閾值時封存
+            if pb.trust_score >= DORMANT_TRUST_THRESHOLD:
+                report.skipped_count += 1
+                continue
+        elif pb.last_used_at > cutoff:
+            # 30 天內有使用 — 不封存
+            report.skipped_count += 1
+            continue
+
+        if pb.trust_score >= DORMANT_TRUST_THRESHOLD:
+            # trust 夠高 — 保留休眠 Playbook
+            report.skipped_count += 1
+            continue
+
+        try:
+            await service.update_with_validation(
+                pb.playbook_id,
+                {"status": PlaybookStatus.DEPRECATED.value},
+            )
+            logger.info(
+                "evolver_archived_dormant",
+                playbook_id=pb.playbook_id,
+                playbook_name=pb.name,
+                trust_score=pb.trust_score,
+                last_used_at=str(pb.last_used_at),
+            )
+            report.archived_count += 1
+            report.archived_ids.append(pb.playbook_id)
+            pb.status = PlaybookStatus.DEPRECATED
+        except Exception as e:
+            report.errors.append(f"archive_dormant:{pb.playbook_id}:{e}")
+            logger.warning(
+                "evolver_dormant_archive_failed",
+                playbook_id=pb.playbook_id,
+                error=str(e),
+            )
+
+
+async def _merge_similar(playbooks: list[Playbook], report: EvolverReport) -> None:
+    """
+    Step 3: 症狀 Jaccard 相似度 > 0.9 → 合併為一
+
+    策略：保留 trust_score 較高的那筆，將較差的標記 DEPRECATED
+    合併次數上限 MAX_MERGE_PER_RUN，避免單次操作影響太多。
+    """
+    from src.services.playbook_service import get_playbook_service
+    from src.utils.similarity import calculate_jaccard_similarity
+
+    service = get_playbook_service()
+    merged_set: set[str] = set()  # 已合併（或被合併）的 playbook_id
+
+    active = [p for p in playbooks if p.status not in (PlaybookStatus.DEPRECATED,)]
+    merge_count = 0
+
+    for i, pb_a in enumerate(active):
+        if pb_a.playbook_id in merged_set:
+            continue
+        if merge_count >= MAX_MERGE_PER_RUN:
+            break
+
+        for pb_b in active[i + 1:]:
+            if pb_b.playbook_id in merged_set:
+                continue
+            if merge_count >= MAX_MERGE_PER_RUN:
+                break
+
+            sim = _compute_symptom_similarity(pb_a, pb_b)
+            if sim < MERGE_SIMILARITY_THRESHOLD:
+                continue
+
+            # 相似度 >= 0.9 → 合併
+            # 保留 trust 較高的，封存較差的
+            keep, drop = (
+                (pb_a, pb_b) if pb_a.trust_score >= pb_b.trust_score else (pb_b, pb_a)
+            )
+
+            try:
+                # 把 drop 的來源 incident 合入 keep
+                merged_source_ids = list(
+                    set(keep.source_incident_ids) | set(drop.source_incident_ids)
+                )
+                await service.update_with_validation(
+                    keep.playbook_id,
+                    {"source_incident_ids": merged_source_ids},
+                )
+                # 封存被合併的
+                await service.update_with_validation(
+                    drop.playbook_id,
+                    {"status": PlaybookStatus.DEPRECATED.value},
+                )
+                logger.info(
+                    "evolver_merged",
+                    kept_id=keep.playbook_id,
+                    dropped_id=drop.playbook_id,
+                    similarity=f"{sim:.2f}",
+                    kept_trust=keep.trust_score,
+                    dropped_trust=drop.trust_score,
+                )
+                merged_set.add(drop.playbook_id)
+                report.merged_count += 1
+                report.merged_pairs.append((drop.playbook_id, keep.playbook_id))
+                drop.status = PlaybookStatus.DEPRECATED
+                merge_count += 1
+            except Exception as e:
+                report.errors.append(f"merge:{keep.playbook_id}+{drop.playbook_id}:{e}")
+                logger.warning(
+                    "evolver_merge_failed",
+                    keep_id=keep.playbook_id,
+                    drop_id=drop.playbook_id,
+                    error=str(e),
+                )
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Helpers
+# ─────────────────────────────────────────────────────────────────────────────
+
+async def _fetch_all_active_playbooks() -> list[Playbook]:
+    """抓取所有非 DEPRECATED 的 Playbook（用於 Evolver 掃描）。"""
+    try:
+        from src.services.playbook_service import get_playbook_service
+
+        service = get_playbook_service()
+        # 不過濾 status — Evolver 需要看到 DRAFT + APPROVED
+        playbooks_page, total = await service.list_playbooks(limit=500)
+        return playbooks_page
+    except Exception as e:
+        logger.warning("evolver_fetch_playbooks_failed", error=str(e))
+        return []
+
+
+def _compute_symptom_similarity(pb_a: Playbook, pb_b: Playbook) -> float:
+    """
+    計算兩個 Playbook 症狀模式的 Jaccard 相似度。
+
+    組合三維度：
+    - alert_names Jaccard（權重 0.5）
+    - keywords Jaccard（權重 0.3）
+    - affected_services Jaccard（權重 0.2）
+    """
+    from src.utils.similarity import calculate_jaccard_similarity
+
+    sp_a = pb_a.symptom_pattern
+    sp_b = pb_b.symptom_pattern
+
+    alert_sim = calculate_jaccard_similarity(
+        set(sp_a.alert_names), set(sp_b.alert_names)
+    )
+    keyword_sim = calculate_jaccard_similarity(
+        set(sp_a.keywords), set(sp_b.keywords)
+    )
+
+    # affected_services 為空 = 通用型 Playbook → 視為完全相符
+    if not sp_a.affected_services and not sp_b.affected_services:
+        service_sim = 1.0
+    elif not sp_a.affected_services or not sp_b.affected_services:
+        service_sim = 0.5  # 一個通用一個特定 → 中等
+    else:
+        service_sim = calculate_jaccard_similarity(
+            set(sp_a.affected_services), set(sp_b.affected_services)
+        )
+
+    return 0.5 * alert_sim + 0.3 * keyword_sim + 0.2 * service_sim
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Singleton / Scheduling Hook
+# ─────────────────────────────────────────────────────────────────────────────
+
+async def schedule_daily_evolver() -> None:
+    """
+    供 startup 或 APScheduler 呼叫的每日 Evolver 觸發點。
+
+    呼叫方式（main.py lifespan 或 scheduler）：
+        asyncio.create_task(schedule_daily_evolver())
+    """
+    from src.core.feature_flags import aiops_flags
+
+    if not aiops_flags.AIOPS_P3_EVOLVER_ENABLED:
+        return
+
+    logger.info("evolver_daily_scheduled")
+    try:
+        report = await run_evolver()
+        logger.info(
+            "evolver_daily_done",
+            archived=report.archived_count,
+            merged=report.merged_count,
+        )
+    except Exception:
+        logger.exception("evolver_daily_failed")

docs/adr/ADR-083-learning-loop-reconstruction.md

@@ -0,0 +1,92 @@
+# ADR-083: 學習閉環重建與 Playbook 演化
+
+**狀態**: Completed  
+**日期**: 2026-04-15  
+**作者**: ogt + Claude Sonnet 4.6（亞太）  
+**Phase**: Phase 3  
+
+---
+
+## 背景
+
+Phase 2（ADR-082）完成後，AWOOOI AIOps 系統進入 Phase 3。
+根據 MASTER v2 架構診斷，L7 學習層有三個核彈級根因導致 Playbook 信任度**永遠停在初始值 0.3**：
+
+1. **fire-and-forget**：`approval_execution._trigger_learning()` 以 `asyncio.create_task()` 呼叫，主流程不等待，GC 可能在學習完成前回收協程。
+2. **matched_playbook_id 永 null**：`decision_manager._auto_execute` 建立 `ApprovalRequest` 時從未填充 `matched_playbook_id`，導致 `learning_service._update_playbook_stats` 的 `if _matched_pb_id:` 條件永遠為 False。
+3. **無負向強化**：`playbook_repository.update_stats` 只遞增 `success_count / failure_count`，無 EWMA 動態信任度計算。
+
+## 決策
+
+### D1: fire-and-forget → await asyncio.wait_for
+
+```python
+# 舊（火烤）
+asyncio.create_task(self._trigger_learning(...))
+
+# 新（Phase 3）
+try:
+    await asyncio.wait_for(self._trigger_learning(...), timeout=30.0)
+except asyncio.TimeoutError:
+    logger.warning("learning_trigger_timeout", ...)
+```
+
+- 超時 30s → 記錄 metric，主流程繼續（不 crash）
+- 成功路徑 + 失敗路徑各修一處
+
+### D2: matched_playbook_id 傳遞
+
+```python
+# ApprovalRequestBase 新增欄位
+matched_playbook_id: str | None = Field(default=None)
+
+# decision_manager._auto_execute 填充
+_matched_playbook_id = token.proposal_data.get("playbook_id")
+approval = ApprovalRequest(..., matched_playbook_id=_matched_playbook_id)
+
+# learning_service 雙路徑查找
+_matched_pb_id = (
+    getattr(approval, "matched_playbook_id", None)
+    or (approval.metadata or {}).get("matched_playbook_id")
+    or (approval.metadata or {}).get("playbook_id")
+)
+```
+
+### D3: 2x EWMA 負向強化
+
+```
+Playbook.trust_score 初值 = 0.3（新增欄位）
+
+成功: trust_new = 0.9 × trust_old + 0.1 × 1.0
+失敗: trust_new = 0.8 × trust_old + 0.2 × 0.0  ← 衰減係數 2x 快
+```
+
+trust < 0.1 → 記錄警告，由 Evolver Agent 封存
+
+### D4: Evolver Agent（playbook_evolver.py）
+
+三功能全靜態規則（不依賴 LLM）：
+1. **低信任封存**：trust_score < 0.1 → DEPRECATED
+2. **休眠封存**：30d 未使用 AND trust < 0.5 → DEPRECATED
+3. **相似合併**：症狀 Jaccard > 0.9 → 保留高 trust，封存低 trust
+
+Feature flag：`AIOPS_P3_EVOLVER_ENABLED=False`（預設關閉）
+
+## 後果
+
+**正面**：
+- 學習閉環首次真正打通（第 9 節點觸發率 0% → 接近 100%）
+- Playbook 信任度動態更新，失敗者加速退場
+- Evolver 防止 Playbook 重複萃取膨脹
+
+**限制**：
+- 人工審核路徑的 `matched_playbook_id` 傳遞仍依賴 `metadata` fallback（完整修復需改 approval DB schema，留 Phase 3.5）
+- Evolver 使用 Jaccard（非 cosine）相似度，向量化合併留 Phase 4
+- `AIOPS_P3_ENABLED = False` 預設值 — 學習呼叫仍執行，但 EWMA 計算不依賴此開關（直接在 repository 層）
+
+## 驗收條件
+
+- `matched_playbook_id` null 率 = 0（auto_execute 路徑）
+- Playbook trust_score 隨每次執行更新（可由 `playbook_stats_updated` log 驗證）
+- 學習超時 → `learning_trigger_timeout` log 出現（不 crash）
+- Evolver `run_evolver()` 空跑無 exception

docs/superpowers/specs/2026-04-15-MASTER-ai-autonomous-flywheel-v2.md

@@ -1454,3 +1454,33 @@ Phase 6 完成後
 - 狀態標頭更新：🟢 §0-§7 全填完
 
 **MASTER v2 第一版完成。** 下一步：統帥 review → 批准 ADR-080 → Phase 0 開工。
+
+---
+
+### 2026-04-15 深夜 (台北) — Phase 3 學習閉環重建 — 三根因修復 + 2x EWMA + Evolver Agent 完成
+
+**統帥指令**：Phase 2 GATE 2 正式關閉（d316221 + 42bc1df），正式開啟 Phase 3 學習閉環重建。
+
+**變更摘要（5 檔 / 1 新建）：**
+
+| 檔案 | 修改內容 | 根因 |
+|------|---------|------|
+| `services/approval_execution.py` | fire-and-forget `create_task` → `await asyncio.wait_for(..., timeout=30)` × 2 處（成功路徑 + 失敗路徑） | 學習觸發 0% |
+| `models/approval.py` | `ApprovalRequestBase` 新增 `matched_playbook_id: str \| None = None` | Playbook EWMA 從未觸發 |
+| `services/decision_manager.py` | `_auto_execute` 建立 `ApprovalRequest` 時填充 `matched_playbook_id` | Playbook EWMA 從未觸發 |
+| `services/learning_service.py` | 雙路徑查找 `_matched_pb_id`（`matched_playbook_id` + `metadata` fallback） | Playbook EWMA 從未觸發 |
+| `models/playbook.py` | 新增 `trust_score: float = 0.3`（EWMA 動態信任度） | 無 trust_score 欄位 |
+| `repositories/playbook_repository.py` | `update_stats` 加 EWMA：成功 α=0.1、失敗 α=0.2（2x 衰減）+ 邊界保護 + 低信任警告 | 無 EWMA 更新 |
+| `services/playbook_evolver.py` | **新建** Evolver Agent：低信任封存（<0.1）+ 休眠封存（30d + <0.5）+ Jaccard 合併（>0.9） | 無 Playbook 生命週期管理 |
+
+**退出驗收條件（Phase 3 §7.1）：**
+- [ ] `matched_playbook_id` null 率 = 0（所有 RAG 命中的決策都填充）
+- [ ] Playbook trust_score 每次執行後更新
+- [ ] 學習呼叫成功率 ≥ 99%（不再 fire-and-forget）
+- [ ] Evolver 合併演練 1 次成功
+
+**Feature Flags（預設全 False，等統帥開啟）：**
+- `AIOPS_P3_ENABLED`
+- `AIOPS_P3_EVOLVER_ENABLED`
+
+**下一步：** ADR-083 草稿 → Gate 3 架構審查 → Phase 3 commit push Gitea