fix(ai): require 111 before alert cloud fallback

apps/api/src/core/config.py

@@ -522,11 +522,11 @@ class Settings(BaseSettings):
         ),
     )
     ALERT_OLLAMA_MODEL: str = Field(
-        default="gemma3:4b",
+        default="qwen3:14b",
         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."
+            "Ollama model used for incident/alert deep diagnosis. Alert cards "
+            "may wait for this model; Gemini remains a backup after GCP-A, "
+            "GCP-B, and 111 fail."
         ),
     )
     # 2026-03-29 ogt: ADR-036 Nemotron Tool Calling 整合

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

@@ -29,7 +29,7 @@ from src.services.model_registry import get_model_registry
 logger = structlog.get_logger(__name__)
 settings = get_settings()
 
-_GCP_SAFE_MODELS = {
+_GCP_LIGHTWEIGHT_MODELS = {
     "gemma3:4b",
 }
 
@@ -54,26 +54,34 @@ def _resolve_model_for_endpoint(
     context: dict | None,
 ) -> str:
     """
-    Keep GCP-A/B on the fast alert model unless explicitly allowed.
+    Keep non-diagnosis calls from polluting the GCP diagnosis lane.
 
-    The GCP hosts currently expose CPU-only Ollama. Loading 7B/14B/32B models on
-    that lane blocks synchronous alerts long enough to fall through to Gemini.
-    Heavy/deep workloads must use 111 or the future AwoooP Inference Gateway.
+    GCP-A/B are allowed to run the deep incident diagnosis model because the
+    alert goal is correctness and resolution, not the fastest Telegram card.
+    Accidental non-diagnosis workloads still fall back to the lightweight health
+    model so embedding/Hermes/background calls cannot occupy the same lane.
     """
     model_name = requested_model.strip()
     context = context or {}
     allow_gcp_heavy = bool(context.get("allow_gcp_heavy_model"))
+    task_type = str(context.get("task_type") or context.get("intent_hint") or "").lower()
+    is_deep_diagnosis = task_type in {"diagnose", "alert_deep", "incident_diagnosis"}
 
-    if _is_gcp_alert_lane(endpoint_url) and not allow_gcp_heavy and model_name not in _GCP_SAFE_MODELS:
-        alert_model = str(getattr(settings, "ALERT_OLLAMA_MODEL", "gemma3:4b")).strip() or "gemma3:4b"
+    if (
+        _is_gcp_alert_lane(endpoint_url)
+        and not allow_gcp_heavy
+        and not is_deep_diagnosis
+        and model_name not in _GCP_LIGHTWEIGHT_MODELS
+    ):
+        fallback_model = str(getattr(settings, "OLLAMA_HEALTH_CHECK_MODEL", "gemma3:4b")).strip() or "gemma3:4b"
         logger.warning(
-            "ollama_gcp_heavy_model_coerced",
+            "ollama_gcp_non_diagnosis_model_coerced",
             endpoint=endpoint_url,
             requested_model=model_name,
-            safe_model=alert_model,
-            task_type=context.get("task_type"),
+            safe_model=fallback_model,
+            task_type=task_type,
         )
-        return alert_model
+        return fallback_model
 
     return model_name
 

apps/api/src/services/ai_router.py

@@ -1142,6 +1142,10 @@ class AIRouterExecutor:
             _lf_trace_ctx = None
 
         errors: list[str] = []
+        attempted_providers: set[str] = set()
+        alert_requires_ollama_before_cloud = bool(
+            (context or {}).get("alert_requires_ollama_before_cloud")
+        )
 
         # 2026-04-27 Claude Sonnet 4.6: A2 INC-20260425 — DIAGNOSE fallback metric 追蹤
         # 透過 context.get("intent_hint") 判斷是否為 DIAGNOSE，避免改動 execute() 簽名
@@ -1191,13 +1195,31 @@ class AIRouterExecutor:
                 errors.append(f"{provider_name}: privacy_skip(non_local)")
                 continue
 
+            if alert_requires_ollama_before_cloud and provider.privacy_level == "cloud":
+                if "ollama_local" not in attempted_providers:
+                    errors.append(f"{provider_name}: blocked_until_ollama_local_attempted")
+                    logger.warning(
+                        "ai_router_cloud_blocked_until_ollama_local_attempted",
+                        provider=provider_name,
+                        provider_order=provider_order,
+                        attempted_providers=sorted(attempted_providers),
+                    )
+                    continue
+
             # 閘門 1: Circuit Breaker (per-provider, C2 修復)
             cb = self._get_circuit_breaker(provider_name)
             if cb.is_open():
-                errors.append(f"{provider_name}: circuit_open")
-                logger.warning("ai_router_circuit_open", provider=provider_name)
-                # 2026-04-27 Claude Sonnet 4.6: F6 — circuit_open 不設 _last_attempted_provider（未嘗試）
-                continue
+                if alert_requires_ollama_before_cloud and provider_name.startswith("ollama"):
+                    logger.warning(
+                        "ai_router_alert_ollama_circuit_bypassed",
+                        provider=provider_name,
+                        reason="alert_requires_ollama_before_cloud",
+                    )
+                else:
+                    errors.append(f"{provider_name}: circuit_open")
+                    logger.warning("ai_router_circuit_open", provider=provider_name)
+                    # 2026-04-27 Claude Sonnet 4.6: F6 — circuit_open 不設 _last_attempted_provider（未嘗試）
+                    continue
 
             # 閘門 2: Rate Limiter
             # 2026-04-02 Claude Code: Phase 24 B3 + C1 修復 — Rate Limiter (含 openclaw_nemo)
@@ -1217,6 +1239,7 @@ class AIRouterExecutor:
             sem = self._get_semaphore(provider_name)
             async with sem:
                 try:
+                    attempted_providers.add(provider_name)
                     result = await provider.analyze(prompt, context)
 
                     if result.success:

apps/api/src/services/openclaw.py

@@ -535,7 +535,7 @@ class OpenClawService:
             registry = get_model_registry()
             model_name = registry.get_model("ollama", "rca")
             if ollama_only:
-                model_name = getattr(settings, "ALERT_OLLAMA_MODEL", "gemma3:4b")
+                model_name = getattr(settings, "ALERT_OLLAMA_MODEL", "qwen3:14b")
             options = registry.get_provider_options("ollama")
             timeout_seconds = max(
                 float(settings.OPENCLAW_TIMEOUT),
@@ -1145,7 +1145,9 @@ class OpenClawService:
                 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")
+                    exec_context["ollama_model"] = getattr(settings, "ALERT_OLLAMA_MODEL", "qwen3:14b")
+                    exec_context["allow_gcp_heavy_model"] = True
+                    exec_context["alert_requires_ollama_before_cloud"] = True
 
                 result = await executor.execute(
                     prompt=prompt,