wooo/awoooi
1
1
Fork 0
Code Issues 2 Pull Requests 1 Actions 4 Packages Projects Releases Wiki Activity
377 Commits 194 Branches 2 Tags
19b00a1ca05f5c59c9e81d19ba0deeb697e50d24
Commit Graph

188 Commits

Author SHA1 Message Date
OG T
19b00a1ca0 fix(api): 移除 Consensus Engine 假信心分數 
🔴 違反鐵律: feedback_confidence_truthfulness.md
Expert System 必須 confidence = 0.0，禁止假裝 AI 仲裁

修正:
- SREAgent: 0.85/0.80/0.75/0.60 → 0.0
- SecurityAgent: 0.70/0.85 → 0.0
- CostAgent: 0.75 → 0.0
- PerformanceAgent: 0.80/0.70 → 0.0

所有規則匹配現在正確顯示為「⚙️ 規則匹配」而非「🤖 AI 仲裁」

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:57:04 +08:00
OG T
89a2339796 feat(api): ADR-038 Circuit Breaker 整合 + Graceful Degradation 
sentry_webhook.py:
- 整合 OpenClawGuard (Circuit Breaker + Semaphore)
- 斷路狀態快速失敗，不呼叫 OpenClaw
- 並發控制: 最多 3 個同時 LLM 推理

anomaly_counter.py:
- record_anomaly() Redis 故障 Graceful Degradation
- 失敗時返回預設 AnomalyFrequency (count=0)
- 不中斷主流程

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:55:51 +08:00
OG T
39396dc57a feat(worker): Wave 1 Signal Worker XCLAIM + Graceful Shutdown 
ADR-038/039 Wave 1 強化:
- 新增 Active Sweeper: XPENDING + XCLAIM 回收閒置訊息
- PENDING_IDLE_MS: 60秒無ACK則可被回收
- SWEEP_INTERVAL_S: 每30秒掃描一次
- Graceful Shutdown: 75秒超時 (搭配 K8s 90秒)
- 超過 MAX_RETRIES 的訊息強制 ACK

K8s Worker Deployment:
- 新增 terminationGracePeriodSeconds: 90

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:53:05 +08:00
OG T
27509db212 feat(api): Wave 1 安全網 - Circuit Breaker + Global Repair Cooldown 
ADR-038: OpenClaw 雙層保護
- Layer 1: Circuit Breaker (5 failures → 60s cooldown)
- Layer 2: Concurrency Semaphore (max 3 concurrent)
- 新增 src/core/circuit_breaker.py

ADR-039: 全域修復熔斷
- Global Cooldown: 5 repairs/15min → freeze
- StatefulSet Blacklist: postgres/redis/clickhouse 禁止自動重啟
- 新增 src/services/global_repair_cooldown.py
- 整合到 auto_repair_service.py

測試:
- test_circuit_breaker.py (狀態轉換 + Semaphore)
- test_global_repair_cooldown.py (黑名單 + 計數閾值)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:48:03 +08:00
OG T
2c79cba629 fix(api): 修復最後 2 個 bare except 錯誤 
- scripts/test_nemotron_tool_calling.py: except -> except Exception

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:37:02 +08:00
OG T
d89f0520f9 fix(api): 修復 34 個 Ruff lint 錯誤 
- 自動修復 import 排序、unused imports
- 手動修復 raise from、isinstance union、unused variable
- scripts/ 暫時保留 (非 CI 阻擋)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:27:49 +08:00
OG T
5f9a6a7e55 fix(ai): 移除假信心分數 + 顯示 AI 模型來源 
問題: AI 仲裁顯示硬編碼信心分數 (0.75/0.88/0.92/0.70)

修復:
- decision_manager: 預設 confidence 0.75 → 0.0
- decision_manager: Expert System confidence=0.0 + is_rule_based
- openclaw: 所有 Mock Response confidence → 0.0
- telegram_gateway: 新增 ai_provider 欄位
- telegram_gateway: 動態來源標籤 (Ollama/Gemini/Claude/規則匹配)

Telegram 卡片顯示:
- confidence > 0 + provider=ollama → 🤖 Ollama 仲裁
- confidence > 0 + provider=gemini → 🤖 Gemini 仲裁
- confidence > 0 + provider=claude → 🤖 Claude 仲裁
- confidence == 0 → ⚙️ 規則匹配

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 15:19:51 +08:00
OG T
49f21dc4e1 test(api): P1-3/P1-4 ApprovalRequestCreate + Telegram 測試 
P1-3: ApprovalRequestCreate 欄位對齊測試 (13 tests)
- 必填欄位驗證 (action, description, requested_by)
- BlastRadius Model 驗證
- SignOz/Sentry/GitHub Webhook 格式驗證
- Pydantic v2 額外欄位行為驗證

P1-4: Telegram 整合驗證測試 (19 tests)
- SignOzMetricsBlock 格式化
- TelegramMessage 結構
- 風險等級 Emoji 映射
- Webhook → Telegram 訊息流程

遵循: feedback_no_mock_testing.md (禁止 Mock)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 11:28:33 +08:00
OG T
ac2715e541 fix(api): P1-2 ApprovalRequestCreate 欄位對齊 
修正 SignOz + GitHub Webhook 的 ApprovalRequestCreate:

Before (錯誤欄位):
- action_type, target_resource, source
- blast_radius=BlastRadius.SINGLE (enum 不存在)
- dry_run_check=DryRunCheck.SKIPPED (錯誤格式)
- 缺少 action, description, requested_by

After (正確欄位):
- action, description (必填)
- blast_radius=BlastRadius(...) (Pydantic Model)
- dry_run_checks=[] (list)
- requested_by (必填)
- 其他欄位移至 metadata

遵循: ApprovalRequestBase schema (approval.py)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 11:17:27 +08:00
OG T
50c055b547 feat(api): Phase D-G P0 修正 - Learning Repository 積木化 
新增:
- ILearningRepository Protocol (interfaces.py)
- LearningRepository (Redis 持久化層)
- Learning API 端點 (/api/v1/learning/*)
- LearningService.get_recommended_fix() 方法
- LearningService.get_learning_summary() 方法

修正:
- Service 不直接依賴 Redis Client (透過 Repository)
- 符合 leWOOOgo 積木化原則
- 首席架構師審查: 74/100 → 92/100

更新:
- ADR-030: 新增 Phase D-G P0 修正章節
- Skill 02: v1.9 → v2.0
- Runner 修復: 序列建構解決 _runner_file_commands 衝突

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 11:03:51 +08:00
OG T
ae21ba2cc6 feat(ai): Phase 20 P3 優化 - Circuit Breaker + 指數退避 + Prometheus 
P3-1: Circuit Breaker 狀態機 (CLOSED/OPEN/HALF_OPEN)
- 連續 3 次失敗觸發斷路
- 60 秒後自動嘗試恢復
- 防止連鎖故障

P3-2: 指數退避重試
- 基礎延遲 1s，最大 30s
- 含 10% jitter 避免雷鳴

P3-3: Prometheus Metrics
- nvidia_tool_call_requests_total (status, tool_name)
- nvidia_tool_call_latency_seconds (histogram)
- nvidia_circuit_breaker_state_changes_total

測試: 25 → 34 PASSED

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 01:49:08 +08:00
OG T
d9a6f9d066 feat(api): Sentry Session Replay UX 自動監控 
Phase 19 UX 監控 - 善用 Sentry Session Replay:
- SentryService: 新增 list_replays, get_ux_audit_summary
- 偵測: 憤怒點擊 (Rage Clicks) + 死亡點擊 (Dead Clicks)
- 偵測: 有錯誤的 Session Replay
- 偵測: UI 相關錯誤 (TypeError/render)
- API: GET /api/v1/errors/ux-audit 端點
- 腳本: audit_ux_sentry.py CLI 工具

統帥回饋: "AI都要全自動化！" ✅

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 01:48:59 +08:00
OG T
d6dc80bcbc fix(sentry): OpenClaw URL 修正 8088→8089 
ADR-028 端口統一，Sentry webhook 漏掉更新

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 01:46:28 +08:00
OG T
b0b91a59e5 fix(telegram): 修復簽核按鈕無作用 - 方法名稱錯誤 
根本原因:
- telegram_gateway.py 呼叫 service.add_signature() 但該方法不存在
- telegram.py 呼叫 service.reject() 但該方法不存在
- 正確方法為 sign_approval() 和 reject_approval()

修復:
- _execute_approval_action: add_signature → sign_approval
- _execute_approval_action: reject → reject_approval
- telegram webhook: 同步修復

影響範圍:
- Telegram 簽核/拒絕/稍後/靜默按鈕現在正常運作
- 前端 Y/n 按鈕本就使用正確 API (不受影響)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 01:36:38 +08:00
OG T
179e659f14 chore: 清理 Playwright 產物 + kube-state-metrics 告警擴充 
清理工作:
- .gitignore 新增 playwright-report/ 和 test-results/ 排除
- 保留 phase19/ 參考截圖目錄

kube-state-metrics 告警擴充 (P3):
- CronJobLastRunFailed: Job 執行失敗
- DaemonSetMissingPods: DaemonSet 缺少 Pod
- StatefulSetReplicasMismatch: StatefulSet 副本不足
- ContainerWaiting: ImagePullBackOff/CrashLoopBackOff 偵測
- PDBViolation: PDB 健康 Pod 數不足
- NodeUnschedulable: 節點標記為不可排程

新增:
- apps/api/scripts/test_nemotron_tool_calling.py (E2E 比較測試)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 01:28:35 +08:00
OG T
4f7282a97a fix(ai): Phase 20 P2 修復 - Protocol + 邊界測試 + model_registry 
P2-1: 定義 INvidiaProvider Protocol (@runtime_checkable)
P2-2: 補充邊界測試 15 → 25 案例
P2-3: model_registry 新增 NVIDIA + tool_calling_fallback_order

首席架構師評分: 82 → 86 → 90/100

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 01:24:17 +08:00
OG T
6de1c0ff3b fix(ai): 修復 Pydantic validation error + tuple unpacking 
1. kubectl_command 允許 None (LLM 可能返回 null)
2. 加入 field_validator 將 null 轉換為空字串
3. generate_incident_proposal 完整解包 6 值 (含 ai_tokens/ai_cost)

2026-03-29 ogt: Gemini API validation 修復

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 00:46:02 +08:00
OG T
fb643eb645 feat(ai): ADR-036 Nemotron E2E 驗證腳本 
新增 verify_nemotron_e2e.py:
- 測試 NVIDIA API 連線
- 測試 AIRouter 整合
- 測試高風險 Tool 檢測
- 測試繁體中文 Tool Calling

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 00:11:40 +08:00
OG T
7c905c4bf3 fix(ai): 修復 generate_incident_proposal tuple unpacking 錯誤 
- _call_with_cache 返回 6 值 (含 ai_tokens/ai_cost)
- generate_incident_proposal 解包只取 4 值導致 ValueError
- 修復: 完整解包 6 值並傳遞 ai_tokens/ai_cost 到 proposal_dict

2026-03-29 ogt: Token/Cost 追蹤補遺

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 00:03:22 +08:00
OG T
b77e151387 feat(ai): ADR-036 NVIDIA Nemotron Tool Calling 整合 
Phase 20 - 提升 Tool Calling 精準度 50% → 83.3%

新增:
- src/models/nvidia.py: Pydantic Schema
- src/services/nvidia_provider.py: NvidiaProvider 類別
- tests/test_nvidia_provider.py: 15 項單元測試 (全部通過)

修改:
- ai_router.py: AIProvider.NVIDIA + route_tool_calling()
- ai_rate_limiter.py: NVIDIA 限制 (5 RPM, 100/day)
- models.json: NVIDIA 配置
- cd.yaml: Secrets 注入 NVIDIA_API_KEY

路由策略:
- Tool Calling: Nemotron → Gemini → Claude
- 一般對話: Ollama → Gemini → Claude (不變)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-29 00:00:08 +08:00
OG T
6ac0f8c0e5 chore: force API rebuild (runner temp file fix) 
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 23:47:18 +08:00
OG T
ba521fa531 fix(ai): 更新 Gemini 模型名稱 1.5-flash → 2.0-flash (2026-03-28 ogt) 
根本原因: gemini-1.5-flash 已停用，API 返回 404
解決方案: 更新到 gemini-2.0-flash

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 23:23:52 +08:00
OG T
c76a10ad6e feat(ai): $5 USD 成本上限 + 自動切換 Ollama (2026-03-29 ogt) 
統帥要求:
1. 累積成本超過 $5 USD → 自動停用 Gemini，切換回 Ollama
2. 發送 Telegram 告警通知統帥
3. $4 USD 時發送警告

實作:
- ai_rate_limiter.py: 新增 COST_LIMITS, record_cost(), reset_cost()
- openclaw.py: 每次成功呼叫後記錄成本
- 成本存入 Redis (不過期，手動重置)
- 重置指令: redis-cli DEL ai_rate:total_cost:gemini

API 端點: GET /api/v1/health/ai-usage
- 顯示 total_cost_usd.current/limit/remaining
- 顯示 cost_exceeded: true/false

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 22:34:51 +08:00
OG T
d469a239af fix(ai): 移除 confidence 預設值，強制 LLM 真實計算 
變更:
1. models/ai.py: confidence 改為 REQUIRED (移除 default=0.8)
2. openclaw.py: 如果 LLM 沒輸出 confidence，設為 0.5 + COLLAB

根本原因:
- 原本 Pydantic default=0.8 導致信心分數永遠是 80%
- 現在強制 LLM 必須計算真實信心分數

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 22:21:29 +08:00
OG T
984d31de0c feat(ai): Gemini 優先 + Token/Cost 追蹤 (2026-03-29 ogt) 
變更:
1. ConfigMap: Gemini 優先 ["gemini","ollama","claude"]
2. openclaw.py: 捕獲 Gemini usageMetadata (tokens/cost)
3. webhooks.py: 傳遞 ai_tokens/ai_cost 到 Telegram
4. telegram_gateway.py: 顯示 💰 Tokens: X / $Y.YYYY

Gemini 1.5 Flash 定價:
- Input: $0.075/1M tokens
- Output: $0.30/1M tokens

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 22:18:24 +08:00
OG T
59c9eff83a fix(api): 修復 10 個 Lint 錯誤 (imports 排序 + unused imports + set comprehension) 
- F401: 移除未使用的 imports (TerminalSessionStatus, AutoApproveDecision, TerminalSession)
- I001: 修正 import blocks 排序
- C401: set(generator) → {set comprehension}

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 18:51:52 +08:00
OG T
d206460751 feat(security): Phase 20 CSRF 防護實作 
Phase 19 首席架構師審查指出: 核鑰 UX 安全性缺 CSRF 防護

後端:
- 新增 src/core/csrf.py (Double Submit Cookie 模式)
- 新增 src/api/v1/csrf.py (GET /api/v1/csrf/token)
- 新增 src/models/csrf.py (CSRFTokenResponse)
- 修改 approvals.py sign/reject/bulk 端點加入 CSRFToken 驗證

前端:
- 新增 hooks/useCSRF.ts (React Hook)
- 修改 approval.store.ts 整合 CSRF Token 參數

安全特性:
- 256-bit Token (secrets.token_hex)
- 時序安全比較 (secrets.compare_digest)
- SameSite=Strict Cookie
- 1 小時 Token 有效期

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 18:31:58 +08:00
OG T
7b9b0c490b feat(phase19): Omni-Terminal 100% 完成 + 首席架構師審查 47/50 
## Phase 19 Omni-Terminal (Wave 0-6 全部完成)

### 核心功能
- SSE 狀態機 (7-State 設計，10/10 分)
- GenUI 動態渲染 (6 張卡片 + Zod Schema 驗證)
- 核鑰 UX (長按授權 + 風險分級)
- Terminal Telemetry (Sentry 整合)

### P0-P2 修復
- P0: Singleton → FastAPI Depends 依賴注入
- P1: Zod Schema 升級 (7 個驗證 Schema)
- P1: 錯誤分類碼聚合 (Sentry fingerprint)
- P2: Slow Query 監控 (5s 警告 / 10s 嚴重)

### 測試
- test_terminal_service.py: 54 項測試全通過
- 意圖分類: 42 個測試案例 (9 種 IntentType)

### 文檔
- ADR-031: SSE 架構實作紀錄
- ADR-032: GenUI 渲染實作紀錄
- Skills: v1.9 (後端 Terminal 章節)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 18:04:12 +08:00
OG T
ecdcb6110e fix(api): 修復 Sentry Approval 創建參數 (P2) 
ApprovalDBService.create_approval() 不接受 approval_id 參數
ID 由 Service 自動生成，返回後從 ApprovalRequest.id 取得

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 00:18:12 +08:00
OG T
e5ded3b3f2 feat(phase19): OmniTerminal + GenUI + Hybrid SSE 架構實作 (Wave 0-2) 
Phase 19 OmniTerminal MVP 完成:
- Wave 0: Backend (Hybrid SSE POST→GET 架構)
- Wave 1: Frontend (OmniTerminal 狀態機 + GenUI Registry)
- Wave 2: UI 組件 (8 個 GenUI 動態卡片)

ADR 文檔:
- ADR-031: OmniTerminal SSE 架構
- ADR-032: GenUI 動態渲染框架
- ADR-033: K3s HA 架構設計

GenUI 組件:
- GenUIRenderer, K8sPodStatusCard, SentryErrorCard
- MetricsSummaryCard, IncidentTimelineCard
- TraceWaterfallCard, ApprovalCard, NuclearKeyButton

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-28 00:17:26 +08:00
OG T
a5ff57ddc3 fix(api): 修復 Sentry Approval 欄位對齊 ApprovalRequestBase 
- ApprovalRequestCreate 使用正確欄位 (action, description, blast_radius...)
- BlastRadius 改用 Model 實例而非不存在的 enum
- 移除未使用的 DryRunCheck import
- 原始欄位移至 metadata

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 23:14:24 +08:00
OG T
74734f5b8a fix(api): 修復 SentryService.check_dedup Redis import 
- get_redis_pool → get_redis (正確函數名稱)
- Phase 10.2.1 E2E 測試發現

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 23:03:31 +08:00
OG T
7456492482 fix(api): 註冊 Sentry Webhook Router (Phase 10.2.1) 
- 新增 sentry_webhook_v1 import
- include_router 註冊 /api/v1/webhooks/sentry/* 路由
- 修復 Sentry Alert Rule → AWOOOI 連線

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 16:13:04 +08:00
OG T
2b069818af refactor(api): Sentry dedup 邏輯移至 Service 層 (leWOOOgo 模組化) 
Phase 10.2.1 - 2026-03-27 台北時區
- 將 check_sentry_dedup() 從 Router 移至 SentryService.check_dedup()
- Router 層禁止直接存取 Redis (遵循 leWOOOgo 積木化原則)
- 保持 10 分鐘 TTL 去重窗口

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 15:04:53 +08:00
OG T
138ef0c2db fix(api): 修復 7 個 Lint 錯誤 (unused imports + zip strict + dict comprehension) 
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 14:42:47 +08:00
OG T
177563f513 fix(api): 告警收斂不重複發送 Telegram 
問題: 相同 fingerprint 的告警收斂時，仍會重複發送 Telegram
修復: 收斂告警只更新 hit_count，跳過 Telegram 推送
影響: /alerts + /alertmanager 兩個端點

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 14:21:22 +08:00
OG T
7720551b8c fix(api): 修復 Telegram 訊息 INC-INC- 重複前綴 
問題: TelegramMessage.format() 中的 incident_id 生成邏輯
當 approval_id 已是 "INC-xxx" 格式時仍添加 "INC-" 前綴

修復: 檢查 approval_id 是否已有 INC- 前綴，避免重複

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 10:28:18 +08:00
OG T
f1b037bb0c refactor(api): playbook_rag.py 模組化改造 (P1 違規修復) 
修復 P1 違規:
- Line 29: Service 直接 import Redis → Repository Pattern
- Line 156: 自建 httpx.AsyncClient → DI 注入

變更:
- 新增 IEmbeddingCacheRepository Protocol (interfaces.py)
- 新增 EmbeddingCacheRepository 實作 (embedding_repository.py)
- PlaybookRAGService 改用 DI 注入 http_client + embedding_cache
- get_playbook_rag_service() 改為 async factory
- PlaybookService 改用 lazy initialization

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 10:07:30 +08:00
OG T
abc21c735e feat(api): P1 Telegram 按鈕優化 - 稍後/靜默 
新增按鈕:
- ⏰ 稍後 (snooze): 延遲 30 分鐘後再提醒
- 🔕 靜默 1h (silence): 同類資源告警靜默 1 小時

實作細節:
- telegram_gateway.py: 新增 _handle_snooze/_handle_silence
- decision_manager.py: 發送前檢查 silence 狀態
- Redis Key: telegram_snooze:{approval_id}, telegram_silence:{resource_name}
- Skill 03 v1.5 → v1.6

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 09:50:28 +08:00
OG T
79b526b472 fix(api): P0 統一 Stream Key 為 awoooi:signals 
修復 Producer/Worker/Webhooks 使用不同 Stream Key 導致訊息無法消費

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 09:33:13 +08:00
OG T
e34b0f2e9a fix(api): Telegram 去重 + 修復 INC-INC-INC- 重複前綴 
- 加入 Redis 去重機制 (10 分鐘 TTL)
- 修復 approval_id 重複添加 INC- 前綴

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 09:27:40 +08:00
OG T
d1409fc4c0 fix(api): 增加 LLM 超時到 120/180s (llama3.2:3b CPU 約 2-3 分鐘) 
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 09:23:35 +08:00
OG T
a3a02398c3 fix(api): 增加 LLM 分析超時 30/60s (Ollama CPU 模式) 
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-27 09:11:39 +08:00
OG T
855b7f5593 fix: 修復 Telegram 轟炸問題 
問題: 765ee39 的修改導致 COMPLETED 狀態下 incident 未解決時
會建立新 decision，每次 poll 都觸發 Telegram 發送

修復: COMPLETED 狀態直接返回 existing_token，不建立新 decision

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 22:24:03 +08:00
OG T
3256142d29 feat(api): ADR-030 Phase 5 持續學習迴圈 
從執行結果中學習，持續優化決策：

1. learning_service.py - 持續學習服務
   - process_execution_result(): 處理執行結果
   - process_human_feedback(): 處理人工反饋
   - 自動調整信任度 (成功+1 / 失敗歸零)
   - 更新 Playbook 統計
   - 成功案例自動萃取 Playbook

2. approval_execution.py - 整合學習觸發
   - 執行成功後觸發學習
   - 執行失敗後觸發學習
   - _trigger_learning(): 非阻塞呼叫學習服務

學習流程:
執行完成 → LearningService.process_execution_result()
  ├─ 成功: TrustEngine +1 分 + Playbook 統計更新
  └─ 失敗: TrustEngine 歸零 + 記錄失敗原因

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 22:19:41 +08:00
OG T
ce7f8a1b23 feat(api): ADR-030 Phase 4 自動執行機制 
實作低風險操作自動執行策略：

1. auto_approve.py - 自動執行策略服務
   - AutoApprovePolicy: 評估是否可自動執行
   - 條件: LOW 風險 + 信任分數 >= 5 + Playbook 成功率 >= 95%
   - CRITICAL 永遠不自動執行
   - 完整審計追蹤

2. trust_engine.py - 新增 singleton
   - get_trust_manager(): 取得全域 TrustScoreManager

3. decision_manager.py - 整合自動執行 (Tier 3 紅區)
   - Step 5 加入 AutoApprovePolicy 判斷
   - 條件滿足時跳過 Telegram，直接執行
   - _auto_execute(): 自動執行邏輯
   - 失敗時 fallback 到人工審核

流程:
Incident → 分析 → AutoApprovePolicy 評估
  ├─ 可自動執行 → 直接執行 → 完成
  └─ 需人工審核 → Telegram 通知 → 等待批准

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 22:13:10 +08:00
OG T
17ee8838be revert: 還原 Telegram + CD 到正常狀態 
還原檔案到 d071019 版本：
- decision_manager.py: 移除 Redis dedup 邏輯
- telegram_gateway.py: 還原 INC- 前綴邏輯
- cd.yaml: 移除 selector immutable 處理和 Token injection

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 22:10:52 +08:00
OG T
3c034526a5 feat(api): ADR-030 Phase 3 Playbook RAG 向量搜尋 
實作 Playbook 語意搜尋能力：

1. playbook_rag.py - RAG 向量服務
   - Ollama nomic-embed-text 生成 embedding
   - Redis 儲存向量 (JSON 格式)
   - 餘弦相似度搜尋
   - 混合搜尋 (Vector 60% + Jaccard 40%)

2. playbook_service.py - 整合 RAG
   - extract_from_incident 後自動建立向量索引
   - get_recommendations 支援混合搜尋
   - RAG 失敗時 fallback 到純 Jaccard

功能:
- embed_text(): 文字向量化
- embed_playbook(): Playbook 向量化
- search_similar(): 向量相似度搜尋
- hybrid_search(): 混合搜尋

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 22:08:15 +08:00
OG T
60e9538889 feat(api): ADR-030 Phase 2 診斷資料收集強化 
實作智能自動修復系統的資料收集層：

1. k8s_diagnostics.py - K8s 診斷服務
   - Pod Events/Logs/ResourceUsage 收集
   - CrashLoopBackOff/OOM/ImagePull 偵測
   - 非同步並行收集 + 錯誤容忍

2. diagnosis_aggregator.py - 診斷聚合器
   - 整合 K8s + SignOz + Expert Rules
   - DiagnosisContext 提供結構化 LLM Prompt
   - DiagnosisSignal 信號分析

3. decision_manager.py - 決策引擎整合
   - Step 2.5 加入診斷收集
   - 傳遞 diagnosis_context 給 LLM

4. openclaw.py - LLM Prompt 增強
   - 整合 K8s/SignOz 深度診斷上下文
   - 支援 diagnosis_signals 摘要

ADR-030 架構: 診斷先行，根因分析，非盲目重啟

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 21:55:50 +08:00
OG T
bb6151cf44 revert: 移除 Telegram Redis dedup 邏輯 
原因: dedup 邏輯導致 Telegram 完全無法發送
保留: INC- 前綴修復 (approval_id = incident.incident_id)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
 2026-03-26 21:53:39 +08:00