fix(responses): stop internal metadata reaching strict upstreams

When proxy-internal fields leaked into the provider-facing metadata
parameter (including empty {}), strict OpenAI-compatible /v1/responses
backends returned HTTP 400. Only forward caller metadata: prefer
litellm_metadata.requester_metadata on proxy routes and drop metadata
entirely when no caller-supplied values exist.

Fixes BerriAI/litellm#35780

litellm/responses/utils.py

@@ -231,14 +231,30 @@ class ResponsesAPIRequestUtils:
             )
             non_default_params["previous_response_id"] = decoded_previous_response_id
 
-        if "metadata" in non_default_params:
-            from litellm.utils import add_openai_metadata
+        litellm_metadata = params.get("litellm_metadata")
+        if not isinstance(litellm_metadata, dict):
+            litellm_metadata = {}
 
-            converted_metadata: Final = add_openai_metadata(non_default_params["metadata"])
-            if converted_metadata is not None:
+        requester_metadata = litellm_metadata.get("requester_metadata")
+        metadata_in_params = non_default_params.get("metadata")
+        is_proxy_internal_metadata = any(key.startswith("user_api_key_") for key in litellm_metadata)
+
+        metadata_source: dict | None = None
+        if isinstance(requester_metadata, dict):
+            metadata_source = requester_metadata
+        elif not is_proxy_internal_metadata and isinstance(metadata_in_params, dict):
+            metadata_source = metadata_in_params
+
+        if metadata_source is not None:
+            from litellm.utils import get_requester_metadata
+
+            converted_metadata: Final = get_requester_metadata(metadata_source)
+            if converted_metadata:
                 non_default_params["metadata"] = converted_metadata
             else:
                 non_default_params.pop("metadata", None)
+        else:
+            non_default_params.pop("metadata", None)
 
         return cast(ResponsesAPIOptionalRequestParams, non_default_params)
 

tests/test_litellm/responses/test_responses_utils.py

@@ -132,6 +132,59 @@ class TestResponsesAPIRequestUtils:
         assert result["max_output_tokens"] == 100
         assert result["prompt"] == {"id": "pmpt_456"}
 
+    def test_get_requested_response_api_optional_param_strips_empty_metadata(self):
+        """Regression #35780: empty metadata dict must not reach upstream."""
+        result = ResponsesAPIRequestUtils.get_requested_response_api_optional_param(
+            {"temperature": 0.7, "metadata": {}}
+        )
+        assert "metadata" not in result
+
+    def test_get_requested_response_api_optional_param_strips_internal_metadata(self):
+        """Regression #35780: proxy/router internal metadata must not reach upstream."""
+        result = ResponsesAPIRequestUtils.get_requested_response_api_optional_param(
+            {
+                "temperature": 0.7,
+                "metadata": {
+                    "model_group": "test-model",
+                    "deployment": "openai/gpt-4o",
+                    "model_info": {"id": "dep-1"},
+                    "user_api_key_user_id": "user-1",
+                },
+                "litellm_metadata": {
+                    "user_api_key_team_id": "team-1",
+                    "tags": ["internal-tag"],
+                },
+            }
+        )
+        assert "metadata" not in result
+
+    def test_get_requested_response_api_optional_param_preserves_caller_metadata(self):
+        """Caller-supplied string metadata should still be forwarded."""
+        result = ResponsesAPIRequestUtils.get_requested_response_api_optional_param(
+            {
+                "temperature": 0.7,
+                "metadata": {"customer_id": "cust-123", "campaign": "spring"},
+            }
+        )
+        assert result["metadata"] == {"customer_id": "cust-123", "campaign": "spring"}
+
+    def test_get_requested_response_api_optional_param_uses_requester_metadata(self):
+        """Proxy snapshots caller metadata under litellm_metadata.requester_metadata."""
+        result = ResponsesAPIRequestUtils.get_requested_response_api_optional_param(
+            {
+                "temperature": 0.7,
+                "metadata": {
+                    "model_group": "test-model",
+                    "user_api_key_user_id": "user-1",
+                },
+                "litellm_metadata": {
+                    "requester_metadata": {"customer_id": "cust-456"},
+                    "user_api_key_team_id": "team-1",
+                },
+            }
+        )
+        assert result["metadata"] == {"customer_id": "cust-456"}
+
     def test_decode_previous_response_id_to_original_previous_response_id(self):
         """Test decoding a LiteLLM encoded previous_response_id to the original previous_response_id"""
         # Setup