Merge pull request #41056 from BerriAI/litellm_moderation_call_id_header

fix(proxy): resolve x-litellm-call-id from response metadata when routes omit call_id

litellm/proxy/common_request_processing.py

@@ -1571,6 +1571,9 @@ class ProxyBaseLLMRequestProcessing:
     ) -> dict:
         exclude_values: Final = {"", None, "None"}
         hidden_params = hidden_params or {}
+        resolved_call_id: Final = (
+            call_id or hidden_params.get("litellm_call_id") or (request_data or {}).get("litellm_call_id")
+        )
         timing_values: Final = _timing_values(
             hidden_params=hidden_params,
             logging_obj=litellm_logging_obj,
@@ -1598,7 +1601,7 @@ class ProxyBaseLLMRequestProcessing:
         classifier_cost: Final = _classifier_cost_from_request_data(request_data)
 
         headers: Final = {
-            "x-litellm-call-id": call_id,
+            "x-litellm-call-id": resolved_call_id,
             "x-litellm-model-id": model_id,
             "x-litellm-model-name": model_name,
             "x-litellm-cache-key": cache_key,

tests/test_litellm/proxy/test_common_request_processing.py

@@ -1402,6 +1402,51 @@ class TestProxyBaseLLMRequestProcessing:
         assert "x-litellm-key-spend" in headers_7
         assert float(headers_7["x-litellm-key-spend"]) == 0.001  # Should use original spend on error
 
+    @pytest.mark.parametrize(
+        ("hidden_params", "request_data", "expected_call_id"),
+        [
+            (
+                {"litellm_call_id": "call-from-hidden-params"},
+                {"litellm_call_id": "call-from-request"},
+                "call-from-hidden-params",
+            ),
+            ({}, {"litellm_call_id": "call-from-request"}, "call-from-request"),
+            ({"model_id": "m-1"}, {"litellm_call_id": "call-from-request"}, "call-from-request"),
+        ],
+    )
+    def test_get_custom_headers_call_id_falls_back_to_hidden_params_then_request_data(
+        self, hidden_params, request_data, expected_call_id
+    ):
+        mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth)
+        mock_user_api_key_dict.tpm_limit = None
+        mock_user_api_key_dict.rpm_limit = None
+        mock_user_api_key_dict.max_budget = None
+        mock_user_api_key_dict.spend = 0.0
+
+        headers = ProxyBaseLLMRequestProcessing.get_custom_headers(
+            user_api_key_dict=mock_user_api_key_dict,
+            hidden_params=hidden_params,
+            request_data=request_data,
+        )
+
+        assert headers["x-litellm-call-id"] == expected_call_id
+
+    def test_get_custom_headers_explicit_call_id_wins_over_fallbacks(self):
+        mock_user_api_key_dict = MagicMock(spec=UserAPIKeyAuth)
+        mock_user_api_key_dict.tpm_limit = None
+        mock_user_api_key_dict.rpm_limit = None
+        mock_user_api_key_dict.max_budget = None
+        mock_user_api_key_dict.spend = 0.0
+
+        headers = ProxyBaseLLMRequestProcessing.get_custom_headers(
+            user_api_key_dict=mock_user_api_key_dict,
+            call_id="explicit-call-id",
+            hidden_params={"litellm_call_id": "call-from-hidden-params"},
+            request_data={"litellm_call_id": "call-from-request"},
+        )
+
+        assert headers["x-litellm-call-id"] == "explicit-call-id"
+
     @pytest.mark.asyncio
     async def test_queue_time_seconds_is_set_in_metadata(self, monkeypatch):
         """

tests/test_litellm/proxy/test_proxy_server.py

@@ -12795,6 +12795,45 @@ async def test_moderations_reraises_proxy_exception_unwrapped():
     mock_logging.post_call_failure_hook.assert_awaited_once()
 
 
+@pytest.mark.asyncio
+async def test_moderations_response_carries_litellm_call_id_header():
+    from fastapi import Response
+
+    from litellm.types.utils import ModerationCreateResponse
+
+    call_id = "moderation-call-id-123"
+    moderation_response = ModerationCreateResponse(id="modr-1", model="omni-moderation-latest", results=[])
+    moderation_response._hidden_params = {"litellm_call_id": call_id, "model_id": "mod-deployment-1"}
+
+    async def fake_llm_call():
+        return moderation_response
+
+    async def passthrough_add_litellm_data(data, **kwargs):
+        return {**data, "litellm_call_id": call_id}
+
+    request = MagicMock()
+    request.body = AsyncMock(return_value=b'{"input": "hi"}')
+    fastapi_response = Response()
+    user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", spend=0.0)
+
+    with (
+        patch.object(proxy_server_module, "add_litellm_data_to_request", new=passthrough_add_litellm_data),  # test-quality-ok: the route reads this module global, no injection point
+        patch.object(proxy_server_module, "route_request", new=AsyncMock(return_value=fake_llm_call())),  # test-quality-ok: fakes the provider call so the response headers assembled by the real route are observable
+        patch.object(proxy_server_module, "proxy_logging_obj") as mock_logging,  # test-quality-ok: module global, no injection point
+    ):
+        mock_logging.pre_call_hook = AsyncMock(side_effect=lambda user_api_key_dict, data, call_type: data)
+        mock_logging.update_request_status = AsyncMock()
+        result = await proxy_server_module.moderations(
+            request=request,
+            fastapi_response=fastapi_response,
+            user_api_key_dict=user_api_key_dict,
+        )
+
+    assert result is moderation_response
+    assert fastapi_response.headers["x-litellm-call-id"] == call_id
+    assert fastapi_response.headers["x-litellm-model-id"] == "mod-deployment-1"
+
+
 @pytest.mark.asyncio
 async def test_init_agents_in_db_rebuilds_registry_under_agent_reconcile_lock(monkeypatch):
     from litellm.proxy.agent_endpoints.agent_registry import (