feat(governance): AI 治理事件處理鏈四軌交付(C/D/B/A)
Some checks failed
Code Review / ai-code-review (push) Successful in 48s
run-migration / migrate (push) Failing after 45s
CD Pipeline / tests (push) Successful in 3m46s
Type Sync Check / check-type-sync (push) Successful in 2m8s
CD Pipeline / build-and-deploy (push) Failing after 31m14s
CD Pipeline / post-deploy-checks (push) Has been skipped
Some checks failed
Code Review / ai-code-review (push) Successful in 48s
run-migration / migrate (push) Failing after 45s
CD Pipeline / tests (push) Successful in 3m46s
Type Sync Check / check-type-sync (push) Successful in 2m8s
CD Pipeline / build-and-deploy (push) Failing after 31m14s
CD Pipeline / post-deploy-checks (push) Has been skipped
【十二人專家團隊全景掃描 + 並行四軌實施】
統帥質疑「有讓 12-agent 一起協作嗎」後,依照團隊規則完成全鏈路交付:
onboarder + critic + db-expert + debugger + frontend-designer 並行掃描,
找到 6 大 Gap,再由 fullstack-engineer × 4、refactor-specialist 協作落地。
【Track C — trust_drift 雙寫整併】
兩條獨立寫 event_type=trust_drift 路徑互不呼叫,下游 consumer 拿到雙份資料
無法判定 source-of-truth。整併保留 governance_agent.check_trust_drift(功能
更全:auto-deprecate + Telegram + PG),TrustDriftDetector 降為純統計 lib,
W-6 watchdog 改呼叫 governance_agent。新增 TestSinglePgWritePerDriftScenario
驗證同一 drift 場景只觸發一次 PG 寫入。
變更:
- apps/api/src/services/trust_drift_detector.py(lib only,不再寫 PG)
- apps/api/tests/test_trust_drift_watchdog.py(W-6 改 mock governance_agent)
【Track D — governance_remediation_dispatch 派遣表】
ai_governance_events 是不可變 Event Sourcing,不能塞執行狀態。新建派遣表
作為投影層:1 event → 0..N dispatches,狀態可變、可重試、可審計。
- PgEnum 5 種 event_type + 7 階段狀態機(pending → dispatched → executing →
succeeded/failed/cancelled/skipped)
- 失敗重試 INSERT 新 row(不改舊 row 的 status,保留審計痕跡)
- Partial unique index ux_grd_one_active_per_event 強制「同事件唯一活躍」
- 4 個複合 index 支援 worker poll、去重查詢、觀測面板
- FK 對應 ai_governance_events / playbooks / incidents / approval_records
全部 SET NULL(avoid cascade lock,但 governance_event 用 RESTRICT)
變更:
- apps/api/src/db/models.py(GovernanceRemediationDispatch ORM class)
- apps/api/migrations/governance_remediation_dispatch_2026-05-03.sql
- apps/api/src/repositories/governance_remediation_dispatch_repo.py
(6 個 async 函式 + 3 個自訂例外:DispatchAlreadyActive /
InvalidStatusTransition / DispatchNotFound)
- apps/api/src/models/governance_dispatch.py(DecisionContextV1 等 4 schema)
- apps/api/tests/test_governance_remediation_dispatch.py(29 tests)
【Track B — /governance 頁面】
後端 PR1 三個 endpoint + 前端 PR2-5 完整三 Tab。
PR1 後端:
- GET /api/v1/ai/governance/events(events_tab,含 event_type/severity/
狀態/時間範圍篩選 + 分頁)
- GET /api/v1/ai/governance/queue(queue_tab,含 graceful fallback:
dispatch 表不存在時回 table_pending=True 不拋 500)
- GET /api/v1/ai/governance/summary(slo_tab 30d 違反時序圖)
- severity 映射規則寫死(critic 建議未來移 settings)
PR2-5 前端:
- /governance 路由 + AppLayout + Compliance Badge 橫幅 + PageTabs
- SLO Tab:3 KPI 卡片(Syne 28px + StatusOrb + 7d sparkline)+
30d 違反 stacked BarChart
- Events Tab:篩選列 + 表格 + inline 展開行(JSON / 修復建議 / 派遣記錄)
- Queue Tab:HITL 待辦卡片 + 信任度進度條 + 批准/拒絕按鈕(本 PR console.log)
- Sidebar 加入「AI 治理」入口(ShieldCheck icon)
- i18n 雙語完整(governance namespace + nav.governance)
- 7 個新元件:slo-kpi-card / slo-violation-chart / events-table /
events-filter-bar / event-detail-drawer / queue-item-card / queue-history-tabs
變更:
- apps/api/src/api/v1/ai_governance.py(router)
- apps/api/src/services/governance_query_service.py
- apps/api/src/models/governance.py(Pydantic V2 schemas)
- apps/api/tests/test_ai_governance_endpoints.py(21 tests)
- apps/web/src/app/[locale]/governance/(page + 3 tabs)
- apps/web/src/components/governance/(7 元件)
- apps/web/messages/{zh-TW,en}.json(governance namespace)
- apps/web/src/components/layout/sidebar.tsx(+1 行)
- apps/api/src/main.py(router include)
【Track A — GovernanceDispatcher 決策融合】
把治理事件接到 remediation 執行器,走北極星方向決策融合(LLM × Playbook trust
× MCP),符合「禁寫死規則」鐵律。
- 設計鐵律:DecisionFusionAdapter 是新增 wrapper,**不修改任何 Tier 3 檔**
(decision_manager / learning_service / trust_engine),只 consume 既有 API
- 三維融合公式:confidence = 0.4×llm + 0.3×playbook_trust + 0.3×mcp_consistency
(權重加 TODO 標明未來由 AI 自學調整)
- 三分支決策路徑:
confidence ≥ 0.85 → auto_dispatch(status=dispatched)
0.65 ≤ confidence < 0.85 → pending_approval(HITL)
confidence < 0.65 → skip + log
- decision_context JSONB 完整記錄三維輸入快照(給未來 fine-tune 用)
- poll 30s 掃 unresolved 事件,仿 governance loop 模式
- 重複事件擋去重(呼叫 get_active_for_event)
變更:
- apps/api/src/services/governance_dispatcher.py
- apps/api/src/services/decision_fusion_adapter.py
- apps/api/tests/test_governance_dispatcher.py(14 tests)
- apps/api/src/main.py(lifespan task 接 run_governance_dispatcher_loop)
【驗證】
1836 個 unit test 全過(29 skipped 為既有 PG integration env 問題)
【調度教訓 — 已記入 memory】
- vuln-verifier 應在 fullstack-engineer **之前**跑(避免並行讀到已修代碼誤判)
- critic 雙輪審查不可省(第二輪抓到 NaN sentinel + Prom rule 連鎖)
- 北極星「禁寫死規則」搭配 decision-fusion 確實實施
【未動 Tier 3 — 已驗證】
git diff 確認本 commit 完全沒改 decision_manager.py / learning_service.py /
trust_engine.py,只新增 wrapper service consume 既有 API。
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -1,7 +1,16 @@
|
||||
"""
|
||||
AWOOOI AIOps Phase 6 — Trust Drift Detector(信任度漂移偵測器)
|
||||
===============================================================
|
||||
職責:偵測 Playbook trust_score 分布的兩種極端偏態:
|
||||
【LIB ONLY — NO SIDE EFFECTS】
|
||||
|
||||
2026-05-02 ogt + Claude Sonnet 4.6(亞太): 整併雙寫路徑
|
||||
背景:原本 watchdog W-6 呼叫 detector.run() 會直接寫 event_type=trust_drift 到
|
||||
ai_governance_events;governance_agent.check_trust_drift() 每 1h 也寫同一 event_type。
|
||||
造成雙寫、語義混淆,下游 consumer 無法區分 source-of-truth。
|
||||
整併決策:governance_agent.check_trust_drift() 為唯一 source-of-truth(功能更完整:
|
||||
含 auto-deprecate + Telegram 推送)。本模組降為純統計 lib,不再自行寫 PG。
|
||||
|
||||
職責(整併後):純統計 lib,偵測 Playbook trust_score 分布的兩種極端偏態:
|
||||
|
||||
極端 A「盲目樂觀」:> 70% Playbook trust_score > 0.9
|
||||
→ 可能是 PostExecutionVerifier 失效,或 RAG 資料被污染,讓所有 AI 都以為「我很棒」
|
||||
@@ -11,13 +20,16 @@ AWOOOI AIOps Phase 6 — Trust Drift Detector(信任度漂移偵測器)
|
||||
→ 可能是 EWMA 計算出錯,或所有執行都被誤判失敗,讓 AI 對自己完全沒信心
|
||||
→ 學習機制可能卡死
|
||||
|
||||
設計原則:
|
||||
設計原則(整併後):
|
||||
1. 只讀 DB,不修改任何數據
|
||||
2. 違反 → 寫 trust_drift 事件到 ai_governance_events
|
||||
3. 樣本不足(< 10 個 approved Playbook)→ 跳過偵測,不告警
|
||||
2. detect() / run() 只回傳 TrustDistribution,不寫 ai_governance_events
|
||||
3. save_drift_event() 保留供呼叫方(如需要分布事件)顯式呼叫,不在 run() 內自動觸發
|
||||
4. 樣本不足(< 10 個 approved Playbook)→ 跳過偵測,不告警
|
||||
5. AI 治理事件的唯一寫入點:governance_agent.check_trust_drift()
|
||||
|
||||
ADR-087: AI 自我治理閉環
|
||||
2026-04-15 ogt + Claude Sonnet 4.6(亞太): Phase 6 初始建立
|
||||
2026-05-02 ogt + Claude Sonnet 4.6(亞太): 降為 lib only,移除 run() 自動 PG 寫入
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -222,11 +234,14 @@ class TrustDriftDetector:
|
||||
logger.error("trust_drift_event_save_error", error=str(e))
|
||||
|
||||
async def run(self) -> TrustDistribution:
|
||||
"""完整執行:偵測 → 如有漂移則寫事件。"""
|
||||
dist = await self.detect()
|
||||
if dist.drift_detected:
|
||||
await self.save_drift_event(dist)
|
||||
return dist
|
||||
"""統計偵測(LIB ONLY):只回傳 TrustDistribution,不寫 ai_governance_events。
|
||||
|
||||
2026-05-02 ogt + Claude Sonnet 4.6(亞太): 整併雙寫路徑
|
||||
原行為:detect() 後若 drift_detected 自動呼叫 save_drift_event() 寫 PG。
|
||||
改為:只回傳結果,由呼叫方決定是否寫入。
|
||||
ai_governance_events 的唯一寫入點:governance_agent.check_trust_drift()。
|
||||
"""
|
||||
return await self.detect()
|
||||
|
||||
|
||||
# ─────────────────────────────────────────────────────────────────────────────
|
||||
|
||||
Reference in New Issue
Block a user