test(proxy): add real router regression test and clean up routine comments

litellm/proxy/route_llm_request.py

@@ -488,13 +488,7 @@ async def _route_request_single_attempt(  # noqa: ANN202  # returns unawaited pr
 
     has_router_settings_override: Final = "router_settings_override" in data
     if has_router_settings_override:
-        # Apply per-request router settings overrides from key/team config
-        # Instead of creating a new Router (expensive), merge settings into kwargs
-        # The Router already supports per-request overrides for these settings
         override_settings: Final = data.pop("router_settings_override")
-
-        # Settings that the Router accepts as per-request kwargs
-        # These override the global router settings for this specific request
         per_request_settings: Final = (
             "fallbacks",
             "context_window_fallbacks",
@@ -507,7 +501,6 @@ async def _route_request_single_attempt(  # noqa: ANN202  # returns unawaited pr
         )
 
         if isinstance(override_settings, dict):
-            # Merge override settings into data (only if not already set in request)
             for key in per_request_settings:
                 if key in override_settings and key not in data:
                     data[key] = override_settings[key]
@@ -536,7 +529,6 @@ async def _route_request_single_attempt(  # noqa: ANN202  # returns unawaited pr
         return _route_user_config_request(data, route_type)
 
     elif has_router_settings_override:
-        # Use main router with overridden kwargs
         if llm_router is not None:
             return getattr(llm_router, f"{route_type}")(**data)
         else:

tests/test_litellm/proxy/test_route_llm_request.py

@@ -511,6 +511,41 @@ async def test_route_request_with_router_settings_override_and_api_key():
     assert "model_group_alias" not in call_kwargs
 
 
+@pytest.mark.asyncio
+async def test_route_request_with_router_settings_override_and_api_key_real_router():
+    import litellm
+
+    router: Final = litellm.Router(
+        model_list=[
+            {
+                "model_name": "gpt-3.5-turbo",
+                "litellm_params": {
+                    "model": "openai/gpt-3.5-turbo",
+                    "api_key": "fake",
+                    "mock_response": "real-router-success",
+                },
+            }
+        ]
+    )
+    data = {
+        "model": "gpt-3.5-turbo",
+        "messages": [{"role": "user", "content": "Hello"}],
+        "api_key": "sk-user-test-key",
+        "router_settings_override": {
+            "num_retries": 3,
+            "timeout": 15,
+        },
+    }
+
+    coroutine = await route_request(data, router, None, "acompletion")
+    response = await coroutine
+
+    assert response.choices[0].message.content == "real-router-success"
+    assert "router_settings_override" not in data
+    assert data["num_retries"] == 3
+    assert data["timeout"] == 15
+
+
 def test_gated_mock_params_cover_mock_router_testing_params():
     """``GATED_MOCK_PARAM_NAMES`` is hardcoded to avoid a cyclic import
     against ``litellm.types.router``. This test guards against drift — if a