fix(proxy): prevent router_settings_override leak when request carries api_key

litellm/proxy/route_llm_request.py

@@ -485,6 +485,33 @@ async def _route_request_single_attempt(  # noqa: ANN202  # returns unawaited pr
         # Map generationConfig to config parameter for Google GenAI compatibility
         if "generationConfig" in data and "config" not in data:
             data["config"] = data.pop("generationConfig")
+
+    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",
+            "content_policy_fallbacks",
+            "num_retries",
+            "timeout",
+            "model_group_retry_policy",
+            "routing_strategy",
+            "enable_tag_filtering",
+        )
+
+        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]
+
     if "api_key" in data or "api_base" in data:
         if llm_router is not None:
             return getattr(llm_router, f"{route_type}")(**data)
@@ -508,30 +535,7 @@ async def _route_request_single_attempt(  # noqa: ANN202  # returns unawaited pr
     elif "user_config" in data:
         return _route_user_config_request(data, route_type)
 
-    elif "router_settings_override" in data:
-        # 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",
-            "content_policy_fallbacks",
-            "num_retries",
-            "timeout",
-            "model_group_retry_policy",
-            "routing_strategy",
-            "enable_tag_filtering",
-        ]
-
-        # 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]
-
+    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)

tests/test_litellm/proxy/test_route_llm_request.py

@@ -1,8 +1,6 @@
-
 import pytest
 
 
-
 from typing import Final
 from unittest.mock import MagicMock
 
@@ -474,6 +472,45 @@ async def test_route_request_with_router_settings_override_preserves_existing():
     assert call_kwargs["timeout"] == 30
 
 
+@pytest.mark.asyncio
+async def test_route_request_with_router_settings_override_and_api_key():
+    """
+    Test that router_settings_override is merged and stripped from data
+    even when the request carries api_key or api_base.
+    """
+    data = {
+        "model": "gpt-3.5-turbo",
+        "messages": [{"role": "user", "content": "Hello"}],
+        "api_key": "sk-user-test-key",
+        "api_base": "https://api.example.com/v1",
+        "router_settings_override": {
+            "fallbacks": [{"gpt-3.5-turbo": ["gpt-4"]}],
+            "num_retries": 5,
+            "timeout": 30,
+            "model_group_retry_policy": {"gpt-3.5-turbo": {"RateLimitErrorRetries": 3}},
+            "routing_strategy": "least-busy",
+            "model_group_alias": {"alias": "real_model"},
+        },
+    }
+
+    llm_router = MagicMock()
+    llm_router.acompletion.return_value = "success"
+
+    response = await route_request(data, llm_router, None, "acompletion")
+
+    assert response == "success"
+    call_kwargs = llm_router.acompletion.call_args[1]
+    assert call_kwargs["api_key"] == "sk-user-test-key"
+    assert call_kwargs["api_base"] == "https://api.example.com/v1"
+    assert "router_settings_override" not in call_kwargs
+    assert call_kwargs["fallbacks"] == [{"gpt-3.5-turbo": ["gpt-4"]}]
+    assert call_kwargs["num_retries"] == 5
+    assert call_kwargs["timeout"] == 30
+    assert call_kwargs["model_group_retry_policy"] == {"gpt-3.5-turbo": {"RateLimitErrorRetries": 3}}
+    assert call_kwargs["routing_strategy"] == "least-busy"
+    assert "model_group_alias" not in call_kwargs
+
+
 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
@@ -1257,6 +1294,7 @@ async def test_route_request_read_through_disabled_without_store_model_in_db(mon
 
     assert table.find_many_wheres == []
 
+
 @pytest.mark.asyncio
 async def test_route_request_routing_group_name_passes_model_gate():
     from unittest.mock import AsyncMock, patch