fix(ai): stabilize GCP Ollama alert lane

apps/api/models.json

@@ -25,7 +25,7 @@
         "log_anomaly": "deepseek-r1:14b",
         "nemoclaw": "deepseek-r1:14b",
         "playbook_draft": "qwen3:14b",
-        "code_review": "qwen2.5-coder:32b",
+        "code_review": "qwen2.5-coder:7b",
         "embedding": "bge-m3:latest",
         "rag_generate": "qwen3:14b",
         "image_analysis": "minicpm-v:latest",
@@ -175,7 +175,7 @@
       },
       "pr_code_review": {
         "phase": 32,
-        "model": "qwen2.5-coder:32b",
+        "model": "qwen2.5-coder:7b",
         "timeout_seconds": 120,
         "purpose": "Gitea PR 自動審查"
       },

apps/api/src/core/config.py

@@ -520,6 +520,14 @@ class Settings(BaseSettings):
             "then local 111 before cloud backup providers such as Gemini."
         ),
     )
+    ALERT_OLLAMA_MODEL: str = Field(
+        default="gemma3:4b",
+        description=(
+            "Ollama model used for incident/alert diagnosis. Keep this separate "
+            "from the general RCA model so alert cards stay on the fast local "
+            "lane before Gemini backup is considered."
+        ),
+    )
     # 2026-03-29 ogt: ADR-036 Nemotron Tool Calling 整合
     NVIDIA_API_KEY: str = Field(
         default="",

apps/api/src/jobs/ai_slo_watchdog_job.py

@@ -46,6 +46,7 @@ _DEDUP_TTL_SEC = 3600    # 同一告警 1 小時內不重複發送
 _TG_SILENCE_THRESHOLD = 2    # PENDING telegram_message_id IS NULL 告警門檻
 _FLYWHEEL_SUCCESS_MIN = 0.30  # 執行成功率下限
 _STUCK_ANALYSIS_THRESHOLD = 3  # Agent Debate 失敗導致卡住的告警門檻
+_TRUST_DRIFT_META_MIN_RATIO = 0.20  # 低於此比例只記治理事件，不升 Meta System
 
 # 2026-05-03 ogt + Claude Opus 4.7 — feedback_silencing_alerts_recurring_violation
 # 啟動寬限期：30 分鐘內可 skip「資料還沒到」噪音；超過寬限期仍空 = 真資料管線斷，必須告警
@@ -210,7 +211,8 @@ async def _check_once() -> None:
         from src.services.governance_agent import get_governance_agent
         trust_result = await get_governance_agent().check_trust_drift(emit_alert=False)
         drifted = trust_result.get("drifted", 0)
-        if drifted > 0:
+        drift_ratio = float(trust_result.get("drift_ratio") or 0.0)
+        if drifted > 0 and drift_ratio >= _TRUST_DRIFT_META_MIN_RATIO:
             auto_deprecated = trust_result.get("auto_deprecated", 0)
             kept = trust_result.get("kept", 0)
             violations.append(
@@ -219,6 +221,13 @@ async def _check_once() -> None:
             )
             # 2026-05-05 ogt W6 修復：移除動態 low_count，避免 count 微變繞過 dedup
             violation_codes.append("W6:trust_drift")
+        elif drifted > 0:
+            logger.info(
+                "watchdog_w6_trust_drift_below_meta_threshold",
+                drifted=drifted,
+                drift_ratio=round(drift_ratio, 3),
+                threshold=_TRUST_DRIFT_META_MIN_RATIO,
+            )
     except Exception as e:
         logger.warning("watchdog_w6_trust_drift_check_failed", error=str(e))
 

apps/api/src/services/ai_providers/ollama.py

@@ -77,7 +77,7 @@ class OllamaProvider:
             client = await self._get_client()
 
             registry = get_model_registry()
-            model_name = registry.get_model("ollama", "rca")
+            model_name = str((context or {}).get("ollama_model") or registry.get_model("ollama", "rca")).strip()
             options = registry.get_provider_options("ollama")
 
             # P0 2026-04-04 Claude Code: per-task timeout（Option C 分情境）
@@ -112,7 +112,13 @@ class OllamaProvider:
             tokens = data.get("eval_count", 0) + data.get("prompt_eval_count", 0)
             latency = (time.perf_counter() - start) * 1000
 
-            logger.info("ollama_provider_success", response_length=len(result), tokens=tokens, latency_ms=round(latency, 1))
+            logger.info(
+                "ollama_provider_success",
+                response_length=len(result),
+                tokens=tokens,
+                latency_ms=round(latency, 1),
+                model=model_name,
+            )
             return AIResult(
                 raw_response=result,
                 success=True,
@@ -158,7 +164,7 @@ class OllamaProvider:
         total_tokens = 0
         messages: list[dict] = [{"role": "user", "content": prompt}]
         registry = get_model_registry()
-        model_name = registry.get_model("ollama", "rca")
+        model_name = str((context or {}).get("ollama_model") or registry.get_model("ollama", "rca")).strip()
         options = registry.get_provider_options("ollama")
         task_type = (context or {}).get("task_type", "")
         if task_type in ("diagnose", "force_local"):
@@ -321,9 +327,9 @@ class Ollama188Provider(OllamaProvider):
             registry = get_model_registry()
             # 嘗試取 ollama_188 專屬設定，fallback 到 ollama 預設
             try:
-                model_name = registry.get_model("ollama_188", "rca")
+                model_name = str((context or {}).get("ollama_model") or registry.get_model("ollama_188", "rca")).strip()
             except Exception:
-                model_name = getattr(settings, "OLLAMA_HEALTH_CHECK_MODEL", "qwen2.5:7b-instruct")
+                model_name = str((context or {}).get("ollama_model") or getattr(settings, "OLLAMA_HEALTH_CHECK_MODEL", "qwen2.5:7b-instruct")).strip()
 
             try:
                 options = registry.get_provider_options("ollama_188")
@@ -364,6 +370,7 @@ class Ollama188Provider(OllamaProvider):
                 tokens=tokens,
                 latency_ms=round(latency, 1),
                 endpoint=fallback_url,
+                model=model_name,
             )
             return AIResult(
                 raw_response=result,

apps/api/src/services/governance_agent.py

@@ -165,9 +165,11 @@ class GovernanceAgent:
             auto_deprecated=len(auto_deprecated_ids),
             kept=len(kept_ids),
         )
+        drift_ratio = len(drifted) / total if total > 0 else 0.0
         return {
             "checked": total,
             "drifted": len(drifted),
+            "drift_ratio": drift_ratio,
             "auto_deprecated": len(auto_deprecated_ids),
             "kept": len(kept_ids),
         }

apps/api/src/services/openclaw.py

@@ -1142,6 +1142,8 @@ class OpenClawService:
                 exec_context = dict(alert_context) if alert_context else {}
                 if decision.intent == IntentType.DIAGNOSE:
                     exec_context["task_type"] = "diagnose"
+                if self._is_incident_alert_context(alert_context):
+                    exec_context["ollama_model"] = getattr(settings, "ALERT_OLLAMA_MODEL", "gemma3:4b")
 
                 result = await executor.execute(
                     prompt=prompt,

k8s/awoooi-prod/06-deployment-api.yaml

@@ -73,6 +73,12 @@ spec:
               value: "http://192.168.0.110:11436"  # 2026-05-04 ogt: GCP-B secondary via 110 nginx proxy（11436 → 34.21.145.224:11434）
             - name: OLLAMA_FALLBACK_URL
               value: "http://192.168.0.111:11434"  # 2026-05-04 ogt: 111 兜底（K8s 內網直連，GPU RTX）
+            - name: ALERT_AI_ALLOW_CLOUD_FALLBACK
+              value: "true"  # Gemini 只作 GCP-A → GCP-B → 111 全失敗後的備援
+            - name: ALERT_AI_ENFORCE_OLLAMA_FIRST
+              value: "true"  # 告警診斷強制先走 GCP-A → GCP-B → 111
+            - name: ALERT_OLLAMA_MODEL
+              value: "gemma3:4b"  # 2026-05-05 Codex: qwen3:14b 告警 JSON prompt 會拖到 504
             - name: OPENCLAW_DEFAULT_MODEL
               value: "qwen2.5:7b-instruct"
             - name: OPENCLAW_TIMEOUT