fix(proxy): carry used_client_oauth_token onto failure spend rows for litellm_metadata routes

litellm/proxy/hooks/proxy_track_cost_callback.py

@@ -83,6 +83,10 @@ _CAPTURED_IDENTITY_CALL_TYPES: Final[frozenset[str]] = frozenset(
         str(CallTypes.aretrieve_batch),
     )
 )
+_FAILURE_ROW_KEYS_LIFTED_FROM_LITELLM_METADATA: Final[tuple[str, ...]] = (
+    "standard_logging_guardrail_information",
+    "used_client_oauth_token",
+)
 
 
 def _proxy_spend_writer() -> DBSpendUpdateWriter:
@@ -195,13 +199,13 @@ class _ProxyDBLogger(CustomLogger):
         existing_metadata.update(_metadata)
 
         litellm_metadata_bucket: Final = request_data.get("litellm_metadata")
-        if (
-            isinstance(litellm_metadata_bucket, dict)
-            and "standard_logging_guardrail_information" not in existing_metadata
-        ):
-            guardrail_info: Final = litellm_metadata_bucket.get("standard_logging_guardrail_information")
-            if guardrail_info is not None:
-                existing_metadata["standard_logging_guardrail_information"] = guardrail_info
+        existing_metadata.update(
+            (key, litellm_metadata_bucket[key])
+            for key in _FAILURE_ROW_KEYS_LIFTED_FROM_LITELLM_METADATA
+            if isinstance(litellm_metadata_bucket, dict)
+            and key not in existing_metadata
+            and litellm_metadata_bucket.get(key) is not None
+        )
 
         if "litellm_params" not in request_data:
             request_data["litellm_params"] = {}

tests/test_litellm/proxy/hooks/test_proxy_track_cost_callback.py

@@ -160,6 +160,43 @@ async def test_async_post_call_failure_hook_does_not_clobber_guardrail_info_in_m
         assert metadata["standard_logging_guardrail_information"] == metadata_bucket_info
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("used_client_oauth_token", [True, False])
+async def test_async_post_call_failure_hook_carries_used_client_oauth_token_from_litellm_metadata(
+    used_client_oauth_token: bool,
+):
+    """
+    /v1/messages and /v1/responses stamp the proxy's own fields into request_data["litellm_metadata"]
+    and leave request_data["metadata"] to the caller's native metadata, so a failed request on those
+    routes wrote a spend row whose used_client_oauth_token was null instead of the stamped value
+    """
+    logger = _ProxyDBLogger()
+    request_data = {
+        "model": "claude-sonnet-5",
+        "messages": [{"role": "user", "content": "Hello"}],
+        "metadata": {"user_id": "anthropic-native-metadata"},
+        "litellm_metadata": {"used_client_oauth_token": used_client_oauth_token},
+        "proxy_server_request": {"request_id": "test_request_id"},
+    }
+
+    with patch(
+        "litellm.proxy.db.db_spend_update_writer.DBSpendUpdateWriter.update_database",
+        new_callable=AsyncMock,
+    ) as mock_update_database:
+        await logger.async_post_call_failure_hook(
+            request_data=request_data,
+            original_exception=Exception("rate limited"),
+            user_api_key_dict=UserAPIKeyAuth(api_key="test_api_key"),
+        )
+
+    call_kwargs = mock_update_database.call_args[1]["kwargs"]
+    assert call_kwargs["litellm_params"]["metadata"]["user_id"] == "anthropic-native-metadata"
+    payload = get_logging_payload(
+        kwargs=call_kwargs, response_obj={}, start_time=datetime.now(), end_time=datetime.now()
+    )
+    assert json.loads(payload["metadata"])["used_client_oauth_token"] is used_client_oauth_token
+
+
 @pytest.mark.asyncio
 async def test_async_post_call_failure_hook_bills_guardrail_cost_on_blocked_request():
     """LIT-5651: a request blocked by a guardrail never reaches the LLM, but the