feat(ai-router): 推翻 A2 鐵律 — DIAGNOSE primary 改 Ollama 本地優先
Some checks failed
CD Pipeline / build-and-deploy (push) Failing after 2m26s
Some checks failed
CD Pipeline / build-and-deploy (push) Failing after 2m26s
統帥鐵律 2026-04-29:「主要優先用 111 主機的 Ollama」
+ feedback_ai_autonomous_direction.md:以本地免費 LLM 為主
+ feedback_ollama_111_only.md:Ollama 唯一主機 = 111
## 推翻 A2 (2026-04-27 INC-20260425) 的事實基礎
**舊事實**:Ollama = CPU-only deepseek-r1:14b @ 238s(不可用)
**新事實**:prod Ollama 111 = M1 Pro Apple Silicon GPU + qwen2.5:7b-instruct
VRAM 8.2GB 全載入,ctx 32k,實測 hi prompt 0.54s
**雲端全死**(2026-04-29 prod log 證據):
- OpenClaw 188:8088 → 500 Internal Server Error
- Gemini → 429 Too Many Requests(配額爆)
- Claude → 404 Not Found(model claude-3-haiku-20240307 過期)
**不推翻 A2 → 100% incident llm_failed → AI 自動修復永遠不啟動**
## 修改範圍(最小、安全、可驗證)
### ai_router.py
- `_diagnose_fallback_chain`: OLLAMA 第一順位(取代「永久排除」舊註解)
順序:[OLLAMA, OPENCLAW_NEMO, GEMINI, CLAUDE]
- `_intent_provider_overrides[DIAGNOSE]`: OPENCLAW_NEMO → OLLAMA
- 不動 _full_fallback_chain(避免影響 RESTART/SCALE/CONFIG/DELETE)
- 不動 _tool_calling_fallback_chain
- 不動 complexity_map(critic M2 留待後續)
### openclaw.py
- 注入 task_type="diagnose" 到 alert_context(critic C2 真根因)
- 修復 ai_providers/ollama.py:77 timeout 對齊問題:
- 有 task_type → OLLAMA_DIAGNOSE_TIMEOUT_SECONDS=200s
- 沒有 → OPENCLAW_TIMEOUT=30s(不夠 qwen2.5:7b 推理)
- prod log 看到 latency_ms=120014 的根因
- 用 dict(alert_context) 複製,不污染原 context
## Regression Test 同步更新(5 個)
A2 鐵律守門 test 全部反映新鐵律:
- test_p0_diagnose_routing.py::test_diagnose_override_is_ollama
(原 test_diagnose_override_is_openclaw_nemo)
- test_ai_router_diagnose_fallback.py::test_diagnose_fallback_chain_ollama_primary
(原 test_diagnose_fallback_chain_no_ollama)
- test_ai_router_diagnose_fallback.py::test_diagnose_route_primary_is_ollama
(原 test_diagnose_route_fallback_chain_excludes_ollama)
- test_ai_router_diagnose_fallback.py::test_diagnose_route_sync_primary_is_ollama
(原 test_diagnose_route_sync_fallback_chain_excludes_ollama)
- test_ai_router_diagnose_fallback.py::test_build_fallback_chain_for_intent_diagnose_with_ollama_primary
(原 test_build_fallback_chain_for_intent_diagnose_no_ollama)
- test_ai_router_failover_integration.py::test_router_uses_failover_for_diagnose_ollama_primary
(原 test_router_does_not_use_failover_for_openclaw_nemo)
每個 test docstring 都記載歷史脈絡 + 推翻原因。
## 驗證
- 1608 unit tests 全綠
- LLM 路徑 16 個 test 全綠(含 6 個 A2 守門 test 更新版)
- complexity_scorer / failover_manager / intent_classifier 不受影響
## 期望 prod 行為(部署後驗證)
incident 進入 → DIAGNOSE intent → primary OLLAMA (qwen2.5:7b on M1 Pro GPU)
失敗才 fallback → OpenClaw 188 → Gemini → Claude
Ollama 用 200s timeout(之前 30s 不夠)
→ AI 自動修復終於可以啟動,不再 100% llm_failed
## 已知債(後續處理)
- models.json:21 ollama.default 仍是 deepseek-r1:14b(critic C1,但 prod 已自動 route 到實載 model)
- complexity 4/5 仍寫死 gemini/claude(critic M2)
- Gemini API key 在 prod log 明文(需輪換 + sanitize)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -253,16 +253,27 @@ class AIRouter:
|
||||
(AIProviderEnum.OLLAMA, self._ollama_default),
|
||||
]
|
||||
|
||||
# 2026-04-27 Claude Sonnet 4.6: A2 INC-20260425 — DIAGNOSE 移除 Ollama (CPU 238s 不可用)
|
||||
# 根因: INC-20260425 NIM timeout 後 fallback 到 Ollama deepseek-r1:14b(CPU 238s),
|
||||
# 造成二次 timeout,統帥批准 A+B 雙修,本任務 A2 將 Ollama 從 DIAGNOSE chain 移除。
|
||||
# chain 順序: OPENCLAW_NEMO(主力) → GEMINI(第一備援) → CLAUDE(第二備援)
|
||||
# Ollama 刻意排除: CPU-only 實測 238s,絕對不可用於 DIAGNOSE(INC-20260425 血的教訓)
|
||||
# 2026-04-29 ogt + Claude Code: 推翻 A2 鐵律(INC-20260425 事實基礎已過期)
|
||||
# 統帥鐵律 2026-04-29: 「主要優先用 111 主機的 Ollama」
|
||||
# 統帥鐵律 feedback_ai_autonomous_direction.md: 以本地免費 LLM 為主
|
||||
# 統帥鐵律 feedback_ollama_111_only.md: Ollama 唯一主機 = 111
|
||||
#
|
||||
# 推翻原因:
|
||||
# 舊事實 (2026-04-27 A2): Ollama = CPU-only deepseek-r1:14b @ 238s(不可用)
|
||||
# 新事實 (2026-04-29): prod Ollama 111 = M1 Pro Apple Silicon GPU + qwen2.5:7b-instruct
|
||||
# VRAM 8.2GB 全載入,ctx 32k,實測 hi 0.54s
|
||||
# 雲端全死: OpenClaw 188 → 500,Gemini → 429 配額爆,Claude → 404 endpoint 過時
|
||||
# 不推翻 → 100% incident llm_failed → AI 自動修復永遠不啟動
|
||||
#
|
||||
# 配套修改:
|
||||
# - IntentType.DIAGNOSE override: OPENCLAW_NEMO → OLLAMA
|
||||
# - openclaw.py 注入 task_type="diagnose" 讓 Ollama 用 200s timeout
|
||||
# - test_p0_diagnose_routing.py / test_ai_router_diagnose_fallback.py 同步更新
|
||||
self._diagnose_fallback_chain: list[tuple[AIProviderEnum, str]] = [
|
||||
(AIProviderEnum.OPENCLAW_NEMO, self._openclaw_nemo_default),
|
||||
(AIProviderEnum.GEMINI, self._gemini_default),
|
||||
(AIProviderEnum.CLAUDE, self._claude_default),
|
||||
# OLLAMA 永久排除於此 chain: CPU 238s, INC-20260425, 統帥授權 A2
|
||||
(AIProviderEnum.OLLAMA, self._ollama_default), # 主:本地免費,統帥鐵律
|
||||
(AIProviderEnum.OPENCLAW_NEMO, self._openclaw_nemo_default), # fallback 1
|
||||
(AIProviderEnum.GEMINI, self._gemini_default), # fallback 2
|
||||
(AIProviderEnum.CLAUDE, self._claude_default), # fallback 3
|
||||
]
|
||||
|
||||
# Tool Calling 專用 Fallback 鏈 (ADR-036)
|
||||
@@ -288,11 +299,12 @@ class AIRouter:
|
||||
IntentType.RESTART: None, # 依複雜度
|
||||
IntentType.SCALE: None, # 依複雜度
|
||||
IntentType.CONFIG: None, # 依複雜度 (但 HIGH 會升級)
|
||||
# 2026-04-16 ogt: 恢復 DIAGNOSE → OPENCLAW_NEMO
|
||||
# 2026-04-12 patch 改為 None(複雜度路由)後落入 Rule 6 → Ollama deepseek-r1:14b
|
||||
# deepseek-r1:14b CPU 需 238s → timeout → degraded 20% → 全部「待分析」
|
||||
# OPENCLAW_NEMO = via 188:8088 → NVIDIA NIM,實測 2-27s,JSON 輸出可靠
|
||||
IntentType.DIAGNOSE: AIProviderEnum.OPENCLAW_NEMO,
|
||||
# 2026-04-16 ogt: 恢復 DIAGNOSE → OPENCLAW_NEMO(已過期,見 2026-04-29 推翻)
|
||||
# 2026-04-29 ogt + Claude Code: 推翻 A2,DIAGNOSE → OLLAMA(統帥鐵律本地優先)
|
||||
# 舊事實過期:M1 Pro GPU + qwen2.5:7b 0.54s(不再 CPU deepseek 238s)
|
||||
# OPENCLAW_NEMO 188:8088 現況 500 → 不可用
|
||||
# 雲端全死 → 必須回到本地 Ollama 主推理
|
||||
IntentType.DIAGNOSE: AIProviderEnum.OLLAMA,
|
||||
# 輔助意圖
|
||||
IntentType.DELETE: AIProviderEnum.CLAUDE, # CRITICAL → 強制 Claude
|
||||
IntentType.ROLLBACK: None, # 依複雜度
|
||||
|
||||
@@ -920,10 +920,20 @@ class OpenClawService:
|
||||
# require_local=True 對 DIAGNOSE 只會讓所有 provider 被 privacy_skip → 永遠失敗
|
||||
require_local = decision.intent in (IntentType.CODE_REVIEW,)
|
||||
|
||||
# 2026-04-29 ogt + Claude Code: 注入 task_type 讓 Ollama 用正確 timeout
|
||||
# 根因: ai_providers/ollama.py:77 讀 context["task_type"] 決定 timeout
|
||||
# - "diagnose"/"force_local" → OLLAMA_DIAGNOSE_TIMEOUT_SECONDS=200s
|
||||
# - 其他/未注入 → OPENCLAW_TIMEOUT=30s(不夠 qwen2.5:7b 推理)
|
||||
# webhooks alert_context 從未注入 task_type → Ollama fallback 永遠 30s timeout
|
||||
# 對齊 decision.intent 後 Ollama fallback 真正能跑完
|
||||
exec_context = dict(alert_context) if alert_context else {}
|
||||
if decision.intent == IntentType.DIAGNOSE:
|
||||
exec_context["task_type"] = "diagnose"
|
||||
|
||||
result = await executor.execute(
|
||||
prompt=prompt,
|
||||
provider_order=provider_order,
|
||||
context=alert_context,
|
||||
context=exec_context,
|
||||
cache_ttl=3600,
|
||||
require_local=require_local,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user