fix(code-review): keep Gemini fallback opt-in

apps/api/src/services/local_code_review_service.py

@@ -20,6 +20,7 @@ import structlog
 
 from src.core.config import get_settings
 from src.services.model_registry import get_model
+from src.services.ollama_endpoint_resolver import resolve_ollama_endpoint
 
 logger = structlog.get_logger(__name__)
 settings = get_settings()
@@ -75,18 +76,24 @@ class LocalCodeReviewService:
                 redis = None
 
             diff_size = len(diff.encode())
-            use_gemini = diff_size > _MAX_DIFF_BYTES
+            allow_cloud_fallback = settings.LOCAL_CODE_REVIEW_ALLOW_GEMINI_FALLBACK
 
-            if use_gemini:
+            if diff_size > _MAX_DIFF_BYTES and allow_cloud_fallback:
                 result = await self._review_with_gemini(pr_id, repo, title, diff)
             else:
+                if diff_size > _MAX_DIFF_BYTES:
+                    logger.info(
+                        "pr_review_large_diff_using_ollama_truncated",
+                        pr_id=pr_id,
+                        diff_size_bytes=diff_size,
+                        cloud_fallback_enabled=allow_cloud_fallback,
+                    )
                 result = await self._review_with_ollama(pr_id, repo, title, diff)
-                if result is None:
-                    # Ollama 失敗 → fallback Gemini
+                if result is None and allow_cloud_fallback:
                     result = await self._review_with_gemini(pr_id, repo, title, diff)
 
             if result is None:
-                return None
+                result = self._cloud_fallback_disabled_result(pr_id, repo, title)
 
             result["diff_size_bytes"] = diff_size
 
@@ -116,7 +123,7 @@ class LocalCodeReviewService:
         try:
             http = await self._get_http()
             resp = await http.post(
-                f"{settings.OLLAMA_URL}/api/generate",
+                f"{resolve_ollama_endpoint('code_review')}/api/generate",
                 json={
                     "model": _MODEL_OLLAMA,
                     "prompt": prompt,
@@ -138,6 +145,14 @@ class LocalCodeReviewService:
     async def _review_with_gemini(
         self, pr_id: str, repo: str, title: str, diff: str
     ) -> dict[str, Any] | None:
+        if not settings.LOCAL_CODE_REVIEW_ALLOW_GEMINI_FALLBACK:
+            logger.warning(
+                "pr_review_gemini_fallback_disabled",
+                pr_id=pr_id,
+                repo=repo,
+            )
+            return None
+
         try:
             from src.services.openclaw import get_openclaw
             openclaw = get_openclaw()
@@ -156,12 +171,37 @@ class LocalCodeReviewService:
             logger.error("pr_review_gemini_failed", pr_id=pr_id, error=str(e))
         return None
 
+    def _cloud_fallback_disabled_result(
+        self,
+        pr_id: str,
+        repo: str,
+        title: str,
+    ) -> dict[str, Any]:
+        logger.warning(
+            "pr_review_cloud_fallback_skipped",
+            pr_id=pr_id,
+            repo=repo,
+            title=title,
+            reason="LOCAL_CODE_REVIEW_ALLOW_GEMINI_FALLBACK=false",
+        )
+        return {
+            "review_text": (
+                "⚠️ Code Review：GCP-B/Ollama 審查未完成，"
+                "已依成本策略跳過 Gemini fallback。"
+            ),
+            "issues_count": 1,
+            "model": _MODEL_OLLAMA,
+            "provider": "ollama_unavailable",
+            "cloud_fallback_skipped": True,
+        }
+
     async def _save_to_db(
         self, pr_id: str, repo: str, title: str, diff_size: int, result: dict
     ) -> None:
         try:
-            from src.db.base import get_db_context
             from sqlalchemy import text
+
+            from src.db.base import get_db_context
             async with get_db_context() as db:
                 await db.execute(
                     text("""
@@ -206,7 +246,7 @@ class LocalCodeReviewService:
             try:
                 http = await self._get_http()
                 resp = await http.post(
-                    f"{settings.OLLAMA_URL}/api/generate",
+                    f"{resolve_ollama_endpoint('code_review')}/api/generate",
                     json={
                         "model": _MODEL_OLLAMA,
                         "prompt": prompt,

apps/api/src/services/ollama_endpoint_resolver.py

@@ -0,0 +1,101 @@
+"""
+Ollama endpoint resolver for non-critical workload placement.
+
+ADR-110 gives AWOOOI three Ollama endpoints. This resolver is intentionally
+small: it chooses the preferred endpoint by workload class, while health-aware
+failover remains owned by ollama_failover_manager.py.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Literal, Protocol
+
+from src.core.config import settings
+
+OllamaWorkloadType = Literal[
+    "interactive",
+    "healthcheck",
+    "batch",
+    "embedding",
+    "rag",
+    "code_review",
+    "shadow",
+    "canary",
+    "local_required",
+    "privacy_sensitive",
+    "dr",
+]
+
+_GCP_B_PREFERRED_WORKLOADS = {
+    "batch",
+    "embedding",
+    "rag",
+    "code_review",
+    "shadow",
+    "canary",
+}
+
+_LOCAL_PREFERRED_WORKLOADS = {
+    "local_required",
+    "privacy_sensitive",
+    "dr",
+}
+
+
+class _OllamaSettings(Protocol):
+    OLLAMA_URL: str
+    OLLAMA_SECONDARY_URL: str
+    OLLAMA_FALLBACK_URL: str
+
+
+@dataclass(frozen=True)
+class OllamaEndpointSelection:
+    url: str
+    provider_name: str
+    workload_type: OllamaWorkloadType
+    reason: str
+
+
+def resolve_ollama_selection(
+    workload_type: OllamaWorkloadType = "interactive",
+    *,
+    config: _OllamaSettings | None = None,
+) -> OllamaEndpointSelection:
+    """Return the preferred Ollama endpoint for a workload class."""
+    cfg = config or settings
+    primary = cfg.OLLAMA_URL
+    secondary = cfg.OLLAMA_SECONDARY_URL
+    fallback = cfg.OLLAMA_FALLBACK_URL
+
+    if workload_type in _GCP_B_PREFERRED_WORKLOADS and secondary:
+        return OllamaEndpointSelection(
+            url=secondary,
+            provider_name="ollama_gcp_b",
+            workload_type=workload_type,
+            reason="gcp_b_batch_lane",
+        )
+
+    if workload_type in _LOCAL_PREFERRED_WORKLOADS and fallback:
+        return OllamaEndpointSelection(
+            url=fallback,
+            provider_name="ollama_local",
+            workload_type=workload_type,
+            reason="local_privacy_or_dr_lane",
+        )
+
+    return OllamaEndpointSelection(
+        url=primary,
+        provider_name="ollama_gcp_a",
+        workload_type=workload_type,
+        reason="primary_interactive_lane",
+    )
+
+
+def resolve_ollama_endpoint(
+    workload_type: OllamaWorkloadType = "interactive",
+    *,
+    config: _OllamaSettings | None = None,
+) -> str:
+    """Return only the preferred Ollama base URL."""
+    return resolve_ollama_selection(workload_type, config=config).url