diff --git a/apps/api/src/services/alert_rule_engine.py b/apps/api/src/services/alert_rule_engine.py index 49f5d692..c8cda91e 100644 --- a/apps/api/src/services/alert_rule_engine.py +++ b/apps/api/src/services/alert_rule_engine.py @@ -163,15 +163,19 @@ 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，新增多層 label 查找順序: + 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 + 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", {}) @@ -184,10 +188,12 @@ def _extract_vars(alert_context: dict) -> dict[str, str]: # 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 @@ -195,26 +201,45 @@ def _extract_vars(alert_context: dict) -> dict[str, str]: 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", "service"): + 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, @@ -410,6 +435,7 @@ def match_rule(alert_context: dict) -> dict[str, Any] | None: "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": { diff --git a/apps/api/src/services/approval_execution.py b/apps/api/src/services/approval_execution.py index 2318c19c..d77d07ad 100644 --- a/apps/api/src/services/approval_execution.py +++ b/apps/api/src/services/approval_execution.py @@ -1096,18 +1096,37 @@ class ApprovalExecutionService: 在 automation_operation_log 寫一筆 'pending' 紀錄,回傳 op_id 供 _log_aol_completed 更新。 失敗時 (DB 異常) 回 None,主流程繼續 — aol 寫入永不阻塞執行。 + + 2026-04-20 P0.3: input 補 target / operation_type / namespace， + 失敗時 aol.input 就能直接看到 target 是什麼（追 awoooi-service 類誤判的 source trace）。 """ try: from sqlalchemy import text as _sql from src.db.base import get_db_context import json as _json + # 2026-04-20 P0.3: 先嘗試從 action 解析 target / op_type，失敗不阻塞 + _parsed_target: str | None = None + _parsed_op: str | None = None + _parsed_ns: str | None = None + try: + _parsed = parse_operation_from_action(approval.action or "") + _parsed_target = _parsed.resource_name + _parsed_op = _parsed.operation_type.value if _parsed.operation_type else None + _parsed_ns = _parsed.namespace + except Exception: + pass + input_payload = { "approval_id": str(approval.id), "incident_id": approval.incident_id or "", "action": (approval.action or "")[:500], "risk_level": getattr(approval, "risk_level", None) or "", "requested_by": getattr(approval, "requested_by", "") or "", + # 2026-04-20 P0.3: target source trace + "parsed_target": _parsed_target or "", + "parsed_operation": _parsed_op or "", + "parsed_namespace": _parsed_ns or "", } async with get_db_context() as db: diff --git a/apps/api/src/services/drift_narrator_service.py b/apps/api/src/services/drift_narrator_service.py index 94832f3f..0305bfc3 100644 --- a/apps/api/src/services/drift_narrator_service.py +++ b/apps/api/src/services/drift_narrator_service.py @@ -55,6 +55,10 @@ TRIGGER_MEDIUM_MIN = 3 # ============================================================ # 2026-04-18 ogt + Claude Opus 4.7: B 方案 — LLM 驅動智能摘要（取代 Python str()[:30] 截斷） # 架構鐵律: 捨棄 Python 寫死字串解析，結構化 diff 直接餵 LLM,由 LLM 產出繁中 Top 5 摘要 +# 2026-04-20 P0.2 ogt + Claude Opus 4.7: 加 recommendation 輸出，LLM 推薦該按哪顆按鈕 +# - action ∈ {adopt, revert, ignore, investigate} +# - confidence 0.0-1.0（統帥指令：先不 auto-execute，門檻 0.85 保留給未來） +# - reason 一行繁中解釋 _NARRATIVE_PROMPT = """你是 AWOOOI SRE 維運助理。以下是 K8s Config Drift 報告的原始結構化資料。 ## 漂移項目原始資料（JSON） @@ -72,9 +76,20 @@ _NARRATIVE_PROMPT = """你是 AWOOOI SRE 維運助理。以下是 K8s Config Dri "field": "簡化後的欄位路徑 (40 字內)", "summary": "30 字內繁體中文口語摘要,說明從什麼變成什麼" }} - ] + ], + "recommendation": {{ + "action": "adopt 或 revert 或 ignore 或 investigate", + "confidence": 0.85, + "reason": "一行繁體中文解釋為何推薦此動作（含關鍵證據）" + }} }} +## recommendation action 語意 +- adopt: 現狀合理，應把 K8s 狀態寫回 Git (例：HPA 自動擴縮、緊急 hotfix 已驗證) +- revert: 漂移有風險，應回滾到 Git 狀態 (例：image tag 被誤改、secret 被外部改) +- ignore: 噪音，K8s controller 自動補齊 (例：空 list/dict 差異) +- investigate: 不確定，需要人工查清楚 + ## 規則 - 繁體中文 - items 最多挑 5 筆最重要的（HIGH 優先） @@ -84,6 +99,7 @@ _NARRATIVE_PROMPT = """你是 AWOOOI SRE 維運助理。以下是 K8s Config Dri - "新增 pod anti-affinity 規則" - 禁止 markdown、反引號、emoji - 只輸出純 JSON,不要包在 code block 裡 +- recommendation.confidence 要誠實（HIGH drift 且意圖不明 → 0.3-0.5；trivial noise → 0.9） """ @@ -126,8 +142,9 @@ class DriftNarratorService: return # 2026-04-18 B 方案: LLM 同時產 narrative + 結構化 items（取代 str()[:30]） - narrative, items = await self._generate_narrative_and_items(report, interpretation) - await self._send_telegram(report, narrative, items) + # 2026-04-20 P0.2: 追加 recommendation（action/confidence/reason） + narrative, items, recommendation = await self._generate_narrative_and_items(report, interpretation) + await self._send_telegram(report, narrative, items, recommendation) # 寫入 DB narrative_text (Phase 30 ADR-067) try: @@ -162,13 +179,17 @@ class DriftNarratorService: self, report: "DriftReport", interpretation: "DriftInterpretation | None", - ) -> tuple[str, list[dict]]: + ) -> tuple[str, list[dict], dict]: """ 2026-04-18 ogt + Claude Opus 4.7: B 方案 — LLM 產生 narrative + 結構化 items + 2026-04-20 P0.2 ogt + Claude Opus 4.7: 追加 recommendation（AI 推薦按鈕） - 回傳 (narrative, items): + 回傳 (narrative, items, recommendation): narrative: 繁中 4-5 行敘述 items: [{level, field, summary}, ...] 最多 5 筆 + recommendation: {action, confidence, reason} + action ∈ {adopt, revert, ignore, investigate} + confidence 0.0-1.0（統帥指令：先不 auto-execute，僅顯示供統帥參考） LLM 失敗則 fallback 到 Python 智能截斷（不是 str()[:30] 暴力砍） @@ -189,6 +210,7 @@ class DriftNarratorService: started_ms = time.time() narrative: str = "" items: list[dict] = [] + recommendation: dict = {} # 2026-04-20 P0.2 raw_response: str | None = None provider: str = "unknown" status: str = "failed" @@ -260,6 +282,33 @@ class DriftNarratorService: if not clean_items: clean_items = self._fallback_items(report) + # 2026-04-20 P0.2: 解析 recommendation（若 LLM 給了） + _rec = None + try: + if isinstance(_parsed, dict): + _rec = _parsed.get("recommendation") + # Path 2 場景：recommendation 也可能藏在 _inner + if _rec is None and _parsed.get("description", "").startswith("{"): + _inner_txt = str(_parsed["description"]).strip() + _inner = _json.loads(_inner_txt) + if isinstance(_inner, dict): + _rec = _inner.get("recommendation") + except (_json.JSONDecodeError, ValueError, KeyError): + _rec = None + if isinstance(_rec, dict) and _rec.get("action"): + _act = str(_rec.get("action", "")).strip().lower() + if _act in ("adopt", "revert", "ignore", "investigate"): + try: + _conf = float(_rec.get("confidence", 0.0)) + except (TypeError, ValueError): + _conf = 0.0 + _conf = max(0.0, min(1.0, _conf)) + recommendation = { + "action": _act, + "confidence": _conf, + "reason": str(_rec.get("reason", ""))[:200], + } + narrative = _parsed_narrative items = clean_items status = "success" @@ -277,6 +326,10 @@ class DriftNarratorService: items = self._fallback_items(report) status = "failed" + # 2026-04-20 P0.2: LLM 未給 recommendation 就走 Python fallback + if not recommendation: + recommendation = self._fallback_recommendation(report, interpretation) + # ADR-090-C: 同步寫 DB 稽核（永不 propagate error,保護主流程） duration_ms = int((time.time() - started_ms) * 1000) try: @@ -294,7 +347,50 @@ class DriftNarratorService: except Exception as e: logger.warning("drift_narrator_audit_write_failed", error=str(e)) - return narrative, items + return narrative, items, recommendation + + def _fallback_recommendation( + self, + report: "DriftReport", + interpretation: "DriftInterpretation | None", + ) -> dict: + """ + 2026-04-20 P0.2 ogt + Claude Opus 4.7: LLM 沒給 recommendation 時的 Python fallback + + 規則式推薦（保守）： + - 全部 trivial/白名單 → ignore (0.8) + - 有 HIGH drift + intent=emergency_hotfix → adopt (0.5) (不確定，降信心) + - 有 HIGH drift + intent=human_error → revert (0.7) + - 其他 → investigate (0.4)（請人工介入） + """ + actionable = self._count_nontrivial_drift(report) + if actionable == 0: + return { + "action": "ignore", + "confidence": 0.8, + "reason": "全部為白名單或 K8s 預設值補齊，無實質變更。", + } + + _has_high = report.high_count > 0 + _intent = interpretation.intent.value if interpretation else "unknown" + + if _has_high and _intent == "emergency_hotfix": + return { + "action": "adopt", + "confidence": 0.5, + "reason": "HIGH drift 但意圖分析為緊急 hotfix，建議採納並補 Git（請人工複核）。", + } + if _has_high and _intent == "human_error": + return { + "action": "revert", + "confidence": 0.7, + "reason": "HIGH drift 且意圖分析為人為誤操作，建議回滾 Git 狀態。", + } + return { + "action": "investigate", + "confidence": 0.4, + "reason": f"有 {actionable} 項可操作漂移，意圖={_intent}，需人工查清楚再決定。", + } async def _log_ai_action_to_db( self, @@ -540,24 +636,29 @@ class DriftNarratorService: report: "DriftReport", narrative: str, items: list[dict], + recommendation: dict | None = None, ) -> None: """ 推送 TYPE-4D Config Drift 卡片（ADR-075）+ B 方案智能摘要 2026-04-18 ogt + Claude Opus 4.7: 改用 LLM 產的結構化 items, 取代 str()[:30] 暴力截斷產生的亂碼 + 2026-04-20 P0.2 ogt + Claude Opus 4.7: recommendation 顯示在卡片頂部 + （統帥指令：先不 auto-execute，純顯示推薦讓人一眼知道按哪顆） """ from src.services.telegram_gateway import get_telegram_gateway - diff_summary = self._render_telegram_body(report, narrative, items) + diff_summary = self._render_telegram_body(report, narrative, items, recommendation) try: tg = get_telegram_gateway() + # 2026-04-20 P0.2: 500 → 1500 字上限，讓 AI 推薦 + narrative + items 都能容納 + # （send_drift_card 已同步放寬 HTML 顯示上限至 1500） await tg.send_drift_card( incident_id=report.report_id, approval_id=report.report_id, resource_name=report.namespace, - diff_summary=diff_summary[:500], + diff_summary=diff_summary[:1500], detected_at="", ) except Exception as e: @@ -603,21 +704,38 @@ class DriftNarratorService: report: "DriftReport", narrative: str, items: list[dict], + recommendation: dict | None = None, ) -> str: """ - 組裝 Telegram 卡片 body（B 方案格式） + 組裝 Telegram 卡片 body（B 方案格式 + P0.2 AI 推薦） 範例輸出: + 🎯 AI 建議：⏪ 回滾 (85%) — image tag 被手動改到未驗證版本 + 🤖 AI 研判 volumes 與 affinity 被手動修改... 📊 漂移明細 (HIGH: 1 | MEDIUM: 29) 🔴 spec.template.spec.volumes: 新增 2 項 repair-ssh-key 掛載 🟡 spec.template.spec.serviceAccount: (未設) → awoooi-executor - 🟡 spec.template.spec.affinity.podAntiAffinity: 新增 preferred 規則 ... 還有 27 項 """ - lines = [f"🤖 AI 研判\n{narrative}\n"] + lines = [] + + # 2026-04-20 P0.2 AI 推薦（頂部，純推薦不自動執行） + if recommendation and recommendation.get("action"): + _act = recommendation["action"] + _conf = float(recommendation.get("confidence", 0.0)) + _reason = recommendation.get("reason", "") + _emoji_action = { + "adopt": "✅ 採納", + "revert": "⏪ 回滾", + "ignore": "🔕 忽略", + "investigate": "🔍 人工調查", + }.get(_act, _act) + lines.append(f"🎯 AI 建議：{_emoji_action} ({int(_conf * 100)}%) — {_reason}\n") + + lines.append(f"🤖 AI 研判\n{narrative}\n") # 用非 trivial + 非白名單 的實際可操作數顯示 actionable = self._count_nontrivial_drift(report) diff --git a/apps/api/src/services/security_interceptor.py b/apps/api/src/services/security_interceptor.py index 4780cd6f..6946807d 100644 --- a/apps/api/src/services/security_interceptor.py +++ b/apps/api/src/services/security_interceptor.py @@ -435,7 +435,9 @@ class TelegramSecurityInterceptor: - 格式二: {action, incident_id, is_info_action: True} """ # 2026-04-01 Claude Code (ADR-050): 支援 read-only info actions (2-part format) - INFO_ACTIONS = {"detail", "reanalyze", "history"} + # 2026-04-20 P0.1 ogt + Claude Opus 4.7: drift_view_page 納入 INFO_ACTIONS + # payload 格式: drift_view_page:{report_id}_{page}（底線分隔，不跟冒號衝突） + INFO_ACTIONS = {"detail", "reanalyze", "history", "drift_view_page"} parts = callback_data.split(":") if len(parts) == 2 and parts[0] in INFO_ACTIONS: return { diff --git a/apps/api/src/services/telegram_gateway.py b/apps/api/src/services/telegram_gateway.py index ea8b0902..ffc7184a 100644 --- a/apps/api/src/services/telegram_gateway.py +++ b/apps/api/src/services/telegram_gateway.py @@ -1961,7 +1961,8 @@ class TelegramGateway: # 根因：
 在 Telegram HTML mode 渲染為 code block，但 diff_summary 是 AI
         #      研判敘述 + emoji 清單（非 code），應以純文字顯示
         # Diff 長度處理 (ADR-071, Section 14.9.6)
-        if len(diff_summary) <= 500:
+        # 2026-04-20 P0.2 ogt + Claude Opus 4.7: 500 → 1500 讓 AI 建議 + narrative + items 完整顯示
+        if len(diff_summary) <= 1500:
             diff_block = f"\n━━━━━━━━━━━━━━━━━━━\n{html.escape(diff_summary)}"
         else:
             web_url = f"https://aiops.wooo.work/incidents/{incident_id}/drift-diff"
@@ -2084,10 +2085,37 @@ class TelegramGateway:
 
         return {"action": action, "approval_id": approval_id, "user": user, "success": False}
 
-    async def _send_drift_diff_detail(self, report_id: str) -> None:
+    # 2026-04-20 P0.1 ogt + Claude Opus 4.7: drift_view 分頁 + 分類桶
+    # 原邏輯: _send_drift_diff_detail 一次列 3800 字元 → 30 項洗版
+    # 新邏輯: 分頁 10 項/頁、header 顯示 3 桶分類計數、⬅️/➡️ 按鈕切頁
+    _DRIFT_PAGE_SIZE = 10
+
+    def _classify_drift_item(self, item) -> str:
         """
-        送完整 Drift Diff 到 Telegram (drift_view 按鈕回應)
-        展示全部 items (含 HIGH + MEDIUM + 可操作+trivial 分群)
+        分類 drift item 到 3 桶（規則式，不走 LLM 省 token）:
+          - k8s_default: K8s controller 自動補齊（白名單或空↔空）
+          - human_high: HIGH level 且非 trivial（像是 image/env/ports 被人工改）
+          - routine_medium: MEDIUM 非 trivial（一般設定調整）
+        """
+        level = getattr(item.drift_level, "value", str(item.drift_level))
+        # 白名單或 trivial → K8s 自動補齊
+        if item.is_allowlisted:
+            return "k8s_default"
+        _g, _a = item.git_value, item.actual_value
+        _empty_g = _g is None or str(_g).strip() in ("", "{}", "[]", "null", "None")
+        _empty_a = _a is None or str(_a).strip() in ("", "{}", "[]", "null", "None")
+        if _empty_g and _empty_a:
+            return "k8s_default"
+        if level == "high":
+            return "human_high"
+        return "routine_medium"
+
+    async def _send_drift_diff_detail(self, report_id: str, page: int = 0) -> None:
+        """
+        送分頁 Drift Diff 到 Telegram (drift_view / drift_view_page 按鈕回應)
+
+        每頁 _DRIFT_PAGE_SIZE 項，header 顯示 3 桶分類計數 + 分頁位置，
+        底部含「⬅️ 上頁 / 下頁 ➡️」按鈕 (callback: drift_view_page:{report_id}_{page})。
         """
         try:
             from src.repositories.drift_repository import get_drift_repository
@@ -2100,25 +2128,44 @@ class TelegramGateway:
                 })
                 return
 
-            # 2026-04-19 ogt + Claude Opus 4.7: 修 HTTP 400 真因
-            # 原邏輯: _full[:3950] 切在 HTML tag/entity 中間 → Telegram parse_mode HTML 拒絕
-            # 修法: item-by-item 累計長度,超過 3800 就停,確保完整 HTML 結構
-            #       (3800 留 250 buffer 給 header + 截斷提示)
-            _MAX_LEN = 3800
+            # 1. 分類 & 排序（HIGH 優先 → routine → trivial）
+            _classified: list[tuple[str, object]] = [
+                (self._classify_drift_item(_it), _it) for _it in _rpt.items
+            ]
+            _bucket_order = {"human_high": 0, "routine_medium": 1, "k8s_default": 2}
+            _classified.sort(key=lambda x: _bucket_order[x[0]])
 
+            _bucket_counts = {"human_high": 0, "routine_medium": 0, "k8s_default": 0}
+            for _bk, _ in _classified:
+                _bucket_counts[_bk] += 1
+
+            _total = len(_classified)
+            _total_pages = max(1, (_total + self._DRIFT_PAGE_SIZE - 1) // self._DRIFT_PAGE_SIZE)
+            _page = max(0, min(page, _total_pages - 1))
+            _start = _page * self._DRIFT_PAGE_SIZE
+            _end = min(_start + self._DRIFT_PAGE_SIZE, _total)
+            _slice = _classified[_start:_end]
+
+            # 2. Header（AI 分類桶）
             _header = [
-                f"📊 完整 Drift Diff — {html.escape(report_id)}",
+                f"📊 Drift Diff (頁 {_page + 1}/{_total_pages}) — {html.escape(report_id)[:24]}",
                 f"Namespace: {html.escape(_rpt.namespace)}",
-                f"HIGH×{_rpt.high_count}  MEDIUM×{_rpt.medium_count}  INFO×{_rpt.info_count}",
+                (
+                    f"🔴 人工高風險 {_bucket_counts['human_high']} | "
+                    f"🟡 一般修改 {_bucket_counts['routine_medium']} | "
+                    f"🔧 K8s 自動 {_bucket_counts['k8s_default']}"
+                ),
                 "━" * 20,
             ]
             _lines = list(_header)
+            _MAX_LEN = 3800
             _used_len = sum(len(s) + 1 for s in _header)
-            _shown = 0
 
-            for _item in _rpt.items:
-                _level = getattr(_item.drift_level, "value", str(_item.drift_level))
-                _emoji = "🔴" if _level == "high" else ("🟡" if _level == "medium" else "⚪")
+            # 3. 本頁項目（每項仍守 _MAX_LEN 上限，極端長值時寧可提早中斷也不洗版）
+            _rendered = 0
+            _bucket_emoji = {"human_high": "🔴", "routine_medium": "🟡", "k8s_default": "🔧"}
+            for _bk, _item in _slice:
+                _emoji = _bucket_emoji[_bk]
                 _field = (_item.field_path or "")[:80]
                 _git = str(_item.git_value)[:40] if _item.git_value is not None else "(未設)"
                 _k8s = str(_item.actual_value)[:40] if _item.actual_value is not None else "(未設)"
@@ -2131,22 +2178,42 @@ class TelegramGateway:
                     break
                 _lines.append(_block)
                 _used_len += len(_block) + 1
-                _shown += 1
+                _rendered += 1
 
-            _remaining = len(_rpt.items) - _shown
-            if _remaining > 0:
-                _lines.append(f"… 還有 {_remaining} 項未顯示")
+            _skipped_in_page = len(_slice) - _rendered
+            if _skipped_in_page > 0:
+                _lines.append(f"… 本頁還有 {_skipped_in_page} 項過長未顯示，請縮小 field 範圍")
 
             _full = "\n".join(_lines)
 
-            await self._send_request("sendMessage", {
+            # 4. 分頁按鈕（INFO_ACTIONS 2-part 格式，payload 用底線分隔 report_id 與 page）
+            _rows = []
+            _nav = []
+            if _page > 0:
+                _nav.append({
+                    "text": "⬅️ 上頁",
+                    "callback_data": f"drift_view_page:{report_id}_{_page - 1}",
+                })
+            if _page < _total_pages - 1:
+                _nav.append({
+                    "text": "下頁 ➡️",
+                    "callback_data": f"drift_view_page:{report_id}_{_page + 1}",
+                })
+            if _nav:
+                _rows.append(_nav)
+            _keyboard = {"inline_keyboard": _rows} if _rows else None
+
+            _payload = {
                 "chat_id": settings.OPENCLAW_TG_CHAT_ID,
                 "text": _full,
                 "parse_mode": "HTML",
                 "disable_web_page_preview": True,
-            })
+            }
+            if _keyboard:
+                _payload["reply_markup"] = _keyboard
+            await self._send_request("sendMessage", _payload)
         except Exception as _e:
-            logger.warning("drift_diff_detail_send_failed", report_id=report_id, error=str(_e))
+            logger.warning("drift_diff_detail_send_failed", report_id=report_id, page=page, error=str(_e))
             await self._send_request("sendMessage", {
                 "chat_id": settings.OPENCLAW_TG_CHAT_ID,
                 "text": f"⚠️ Drift Diff 查詢失敗: {html.escape(str(_e)[:150])}",
@@ -2986,6 +3053,18 @@ class TelegramGateway:
                     # 2026-04-01 Claude Code (ADR-050 P2): reanalyze button handler
                     await self._answer_callback(callback_query_id, action, text="🔄 重診排程中...")
                     await self._send_reanalyze_result(incident_id)
+                elif action == "drift_view_page":
+                    # 2026-04-20 P0.1 ogt + Claude Opus 4.7: drift_view 分頁切頁
+                    # incident_id 格式: {report_id}_{page}（底線分隔）
+                    _rid, _, _page_str = incident_id.rpartition("_")
+                    try:
+                        _page_num = int(_page_str)
+                    except ValueError:
+                        _rid, _page_num = incident_id, 0
+                    await self._answer_callback(
+                        callback_query_id, action, text=f"📄 切換至第 {_page_num + 1} 頁..."
+                    )
+                    await self._send_drift_diff_detail(_rid or incident_id, page=_page_num)
                 else:
                     # 2026-04-14 Claude Sonnet 4.6 (Phase 5 Sprint 5.1):
                     # 未知 action → fallback dispatcher (查看 callback_action_spec.yaml 是否有註冊)
diff --git a/docs/LOGBOOK.md b/docs/LOGBOOK.md
index 09d1dbe3..ebf19fc5 100644
--- a/docs/LOGBOOK.md
+++ b/docs/LOGBOOK.md
@@ -6,6 +6,44 @@
 
 ---
 
+## 📍 2026-04-20 上午 — P0.1 + P0.2 + P0.3 三項 Drift/Target 修復
+
+### 統帥三問 RCA 後決議
+1. 全做 P0.1 + P0.2 + P0.3
+2. AI 推薦門檻 0.85 OK，**但先不 auto-execute**（純推薦）
+3. 先查 aol 找 awoooi-service 來源 trace 再修
+
+### RCA 結論（awoooi-service 失敗）
+- 透過 `/api/v1/aiops/kpi` 看到過去 24h 有 1 筆 `playbook_executed actor=approval_execution status=failed`
+- grep 全 codebase：**無任何程式碼寫死 `awoooi-service`**（只有歷史 comment）
+- 最可能來源：`alert_rule_engine._extract_vars` 從 `labels.service` 取值當 Deployment 名（K8s Service 名 ≠ Deployment 名）
+- cf59050 / 4f2e122（2026-04-18）已修 NEMOTRON 幻覺雙路徑；本次修第三條路徑（rule engine label fallback）
+
+### 修復內容（5 檔 / 281 行）
+| # | 檔案 | 內容 |
+|---|------|------|
+| P0.3a | `alert_rule_engine.py` | `_extract_vars` service label 降級：`-service` 結尾先剝 suffix，同時回傳 `target_source` 追蹤來源 |
+| P0.3c | `approval_execution.py` | `_log_aol_started` input 補 `parsed_target/operation/namespace`，下次失敗可直接從 aol 查 trace |
+| P0.3b | `approval_execution.py` | 既有 `_log_aol_completed` 本就寫 `resource_name/error/stderr`，追 trace 夠用 |
+| P0.1 | `telegram_gateway.py` | `_send_drift_diff_detail` 加分頁（10 項/頁）+ 3 桶分類 header（人工高風險/一般修改/K8s 自動）+ ⬅️/➡️ 按鈕 |
+| P0.1 | `security_interceptor.py` | INFO_ACTIONS 加 `drift_view_page` 白名單 |
+| P0.2 | `drift_narrator_service.py` | LLM prompt 加 recommendation 欄位（adopt/revert/ignore/investigate + confidence + reason）|
+| P0.2 | `drift_narrator_service.py` | `_render_telegram_body` 頂部顯示「🎯 AI 建議：⏪ 回滾 (85%) — 原因」 |
+| P0.2 | `drift_narrator_service.py` + `telegram_gateway.py` | 卡片 diff_summary 上限 500 → 1500 字，容納推薦 + narrative + items |
+
+### 驗證
+- 90 個 pytest test 全過（drift / rule_engine / approval_execution）
+- 5 檔 AST syntax check 過
+- AI 推薦**純顯示不自動執行**（依統帥指令）
+
+### 下一步
+1. 等下次 real drift 觸發，驗卡片頂部有「🎯 AI 建議」
+2. 等下次 drift_view 按下，驗分頁 + 分類 header + ⬅️/➡️ 按鈕
+3. 若 awoooi-service 再復發，查 `automation_operation_log` 的 `input.parsed_target` 直接追來源
+4. P1 留：drift 分類器 (noise/controller/human) 進 DB、auto-adopt 門檻 ≥0.85 + low risk
+
+---
+
 ## 📍 2026-04-19 晚 21:30 — Gap Review + 3 Gap 修 + AI 自主化 1/9→4/9 LLM 🎖️🎖️🎖️🎖️
 
 ### 統帥核心指示