Fail fast on hard Gemini quota instead of spamming every key

A 429 with 'limit: 0' means the model is unavailable on the plan for every key
(e.g. free tier gives 0 image requests), so rotating through all keys just made
N redundant calls and still failed. Detect that case and stop after one call.
Surface a clear quota/billing message on the try-on task instead of a raw error.

app/services/gemini_keys.py

@@ -37,12 +37,23 @@ def _parse_keys() -> List[str]:
 
 
 def _should_rotate(resp: requests.Response) -> bool:
-    """True when the failure is plausibly key-related (so trying another key helps)."""
-    if resp.status_code in (401, 403, 429):
+    """True when the failure is plausibly key-related (so trying another key helps).
+
+    Rotating is only useful when a *different* key has spare capacity. A 429 that
+    reports a hard ``limit: 0`` means the model is not available on this plan for
+    ANY key (e.g. the free tier gives 0 image requests) — rotating just spams every
+    key for nothing, so we fail fast (1 call) in that case.
+    """
+    body = resp.text.lower()
+    if resp.status_code in (401, 403):
+        return True
+    if resp.status_code == 429:
+        if "limit: 0" in body:
+            return False  # hard cap / model not on this plan — rotating won't help
         return True
     if resp.status_code >= 500:
         return True
-    if resp.status_code == 400 and "api key" in resp.text.lower():
+    if resp.status_code == 400 and "api key" in body:
         return True
     return False
 

app/workers/tasks.py

@@ -272,8 +272,17 @@ def run_tryon(task_id: str):
         db.rollback()
         logger.error(f"Error in Virtual Try-On task {task_id}: {str(e)}", exc_info=True)
         if task:
+            err = str(e)
+            low = err.lower()
+            if "429" in err or "quota" in low or "resource_exhausted" in low or "limit: 0" in low:
+                friendly = (
+                    "Đã hết quota tạo ảnh của nhà cung cấp (model sinh ảnh không khả dụng "
+                    "trên gói miễn phí). Vui lòng bật billing hoặc đổi provider rồi thử lại."
+                )
+            else:
+                friendly = f"Tạo ảnh thử đồ thất bại: {err}"
             task.status = "FAILED"
-            task.error_message = f"Lỗi hệ thống local try-on: {str(e)}"
+            task.error_message = friendly
             db.commit()
         return False
     finally:

tests/test_tryon.py

@@ -200,6 +200,27 @@ def test_gemini_key_rotation_on_failure():
         assert "key=good-key" in calls[1]
 
 
+def test_gemini_no_rotation_on_hard_quota():
+    """A 429 with 'limit: 0' (model not on plan) must fail fast — exactly one call, no rotation."""
+    from app.services import gemini_keys as gk_module
+
+    calls = []
+
+    def do_request(key):
+        calls.append(key)
+        mock_resp = MagicMock()
+        mock_resp.status_code = 429
+        mock_resp.text = '{"error":{"code":429,"message":"...limit: 0, model: gemini-2.5-flash-preview-image"}}'
+        return mock_resp
+
+    with patch.object(gk_module.gemini_keys, "_keys", ["k1", "k2", "k3", "k4", "k5"]), \
+         patch.object(gk_module.gemini_keys, "_idx", 0):
+        resp = gk_module.gemini_keys.request(do_request)
+        # request() returns the 429 (caller raise_for_status), and only ONE key was tried.
+        assert resp.status_code == 429
+        assert len(calls) == 1
+
+
 def test_tryon_service_disabled_raises():
     """With no API key configured, generate() should raise a clear error (task -> FAILED)."""
     from app.services import tryon_service as ts_module