fix(ai): keep GCP Ollama lane on safe models

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

@@ -29,6 +29,54 @@ from src.services.model_registry import get_model_registry
 logger = structlog.get_logger(__name__)
 settings = get_settings()
 
+_GCP_SAFE_MODELS = {
+    "gemma3:4b",
+}
+
+
+def _normalized_url(value: str | None) -> str:
+    return (value or "").rstrip("/")
+
+
+def _is_gcp_alert_lane(endpoint_url: str) -> bool:
+    """Return true for the CPU-only GCP-A/B synchronous alert lane."""
+    endpoint = _normalized_url(endpoint_url)
+    return endpoint in {
+        _normalized_url(getattr(settings, "OLLAMA_URL", "")),
+        _normalized_url(getattr(settings, "OLLAMA_SECONDARY_URL", "")),
+    }
+
+
+def _resolve_model_for_endpoint(
+    *,
+    requested_model: str,
+    endpoint_url: str,
+    context: dict | None,
+) -> str:
+    """
+    Keep GCP-A/B on the fast alert model unless explicitly allowed.
+
+    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.
+    """
+    model_name = requested_model.strip()
+    context = context or {}
+    allow_gcp_heavy = bool(context.get("allow_gcp_heavy_model"))
+
+    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"
+        logger.warning(
+            "ollama_gcp_heavy_model_coerced",
+            endpoint=endpoint_url,
+            requested_model=model_name,
+            safe_model=alert_model,
+            task_type=context.get("task_type"),
+        )
+        return alert_model
+
+    return model_name
+
 
 class OllamaProvider:
     """
@@ -77,7 +125,13 @@ class OllamaProvider:
             client = await self._get_client()
 
             registry = get_model_registry()
-            model_name = str((context or {}).get("ollama_model") or registry.get_model("ollama", "rca")).strip()
+            endpoint_url = self._endpoint_url()
+            requested_model = str((context or {}).get("ollama_model") or registry.get_model("ollama", "rca")).strip()
+            model_name = _resolve_model_for_endpoint(
+                requested_model=requested_model,
+                endpoint_url=endpoint_url,
+                context=context,
+            )
             options = registry.get_provider_options("ollama")
 
             # P0 2026-04-04 Claude Code: per-task timeout（Option C 分情境）
@@ -89,7 +143,6 @@ class OllamaProvider:
             else:
                 read_timeout = float(settings.OPENCLAW_TIMEOUT)
 
-            endpoint_url = self._endpoint_url()
             response = await client.post(
                 f"{endpoint_url}/api/generate",
                 json={

apps/api/src/services/nvidia_provider.py

@@ -40,6 +40,7 @@ from src.models.nvidia import (
 from src.services.langfuse_client import (  # 2026-03-29 ogt: P1-1 Langfuse 整合
     LangfuseTraceContext,
 )
+from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
 
 logger = structlog.get_logger(__name__)
 settings = get_settings()
@@ -822,8 +823,8 @@ class NvidiaProvider:
                     NVIDIA_REQUESTS_TOTAL.labels(status="error", tool_name="chat").inc()
                     import traceback
                     logger.warning(
-                        "nvidia_chat_failed", 
-                        error=str(e), 
+                        "nvidia_chat_failed",
+                        error=str(e),
                         error_type=type(e).__name__,
                         stacktrace=traceback.format_exc()
                     )
@@ -845,7 +846,7 @@ class OllamaToolProvider:
     取代 NVIDIA 雲端 NIM。延遲從 44s 降至 ~5s。
 
     模型: llama3.1:8b (tool calling 最穩定的 8B 模型)
-    Endpoint: OLLAMA_URL/v1/chat/completions (OpenAI 相容格式)
+    Endpoint: local tool lane /v1/chat/completions (OpenAI 相容格式)
     """
 
     def __init__(self) -> None:
@@ -872,10 +873,14 @@ class OllamaToolProvider:
     ) -> list[ToolCallValidationResult]:
         return [tc for tc in tool_calls if self.is_high_risk_tool(tc.tool_name)]
 
+    def _base_url(self) -> str:
+        """Tool-calling/Hermes models stay off the GCP alert lane."""
+        return resolve_ollama_endpoint("hermes").rstrip("/")
+
     async def health_check(self) -> bool:
         try:
             client = await self._get_client()
-            base_url = settings.OLLAMA_URL.rstrip("/")
+            base_url = self._base_url()
             resp = await client.get(f"{base_url}/api/tags", timeout=5.0)
             return resp.status_code == 200
         except Exception:
@@ -892,7 +897,7 @@ class OllamaToolProvider:
         """Ollama /v1/chat/completions tool calling"""
         start_time = time.perf_counter()
         model = model or settings.OLLAMA_TOOL_MODEL
-        base_url = settings.OLLAMA_URL.rstrip("/")
+        base_url = self._base_url()
         url = f"{base_url}/v1/chat/completions"
 
         # 轉換 tools 為 dict 格式（同 NvidiaProvider）
@@ -988,7 +993,7 @@ class OllamaToolProvider:
     async def chat(self, prompt: str, model: str = "", temperature: float = 0.7, max_tokens: int = 512) -> str:
         """簡單 chat（非 tool calling 路徑，保持 INvidiaProvider 相容）"""
         model = model or settings.OLLAMA_TOOL_MODEL
-        base_url = settings.OLLAMA_URL.rstrip("/")
+        base_url = self._base_url()
         try:
             client = await self._get_client()
             resp = await client.post(
@@ -1010,7 +1015,7 @@ _provider: NvidiaProvider | None = None
 _ollama_tool_provider: OllamaToolProvider | None = None
 
 
-def get_nvidia_provider() -> "NvidiaProvider | OllamaToolProvider":
+def get_nvidia_provider() -> NvidiaProvider | OllamaToolProvider:
     """
     取得 Tool Calling Provider 單例。
     USE_OLLAMA_TOOL_CALLING=True (預設) → OllamaToolProvider (本機，~5s)

apps/api/src/services/openclaw.py

@@ -534,6 +534,8 @@ class OpenClawService:
             # 從 ModelRegistry 取得模型配置
             registry = get_model_registry()
             model_name = registry.get_model("ollama", "rca")
+            if ollama_only:
+                model_name = getattr(settings, "ALERT_OLLAMA_MODEL", "gemma3:4b")
             options = registry.get_provider_options("ollama")
             timeout_seconds = max(
                 float(settings.OPENCLAW_TIMEOUT),