Merge 34eb909d3b into b5a72c2042

litellm/litellm_core_utils/litellm_logging.py

@@ -5185,27 +5185,35 @@ class Logging(LiteLLMLoggingBaseClass):
             (ResponseCompletedEvent, ResponseIncompleteEvent, ResponseFailedEvent),
         ):
             ## return unified Usage object
-            if isinstance(result.response.usage, ResponseAPIUsage):
-                set_response_cost_in_hidden_params(
-                    result.response,
-                    provider_reported_usage_cost(
-                        result.response.usage, self.model_call_details.get("custom_llm_provider")
-                    ),
-                )
-                transformed_usage: Final = ResponseAPILoggingUtils.transform_response_api_usage_to_chat_usage(
-                    result.response.usage
-                )
+            # `result.response` may be a plain dict on the streaming /v1/responses
+            # spend-log path (#29913) because ResponseCompletedEvent allows extra
+            # fields; a bare `.usage` access raises AttributeError there and no
+            # spend row is written at all.
+            resp: Final = result.response
+            resp_dict: Final = (
+                cast(dict[str, object], resp)  # cast-ok: the event allows extra fields, so a plain dict can land here
+                if isinstance(resp, dict)
+                else None
+            )
+            usage: Final = resp_dict.get("usage") if resp_dict is not None else getattr(resp, "usage", None)
+            if isinstance(usage, ResponseAPIUsage):
+                if resp_dict is None:
+                    set_response_cost_in_hidden_params(
+                        resp,
+                        provider_reported_usage_cost(usage, self.model_call_details.get("custom_llm_provider")),
+                    )
+                transformed_usage: Final = ResponseAPILoggingUtils.transform_response_api_usage_to_chat_usage(usage)
                 # Set as dict instead of Usage object so model_dump() serializes it correctly
-                setattr(
-                    result.response,
-                    "usage",
-                    (
-                        transformed_usage.model_dump()
-                        if hasattr(transformed_usage, "model_dump")
-                        else dict(transformed_usage)
-                    ),
+                new_usage: Final = (
+                    transformed_usage.model_dump()
+                    if hasattr(transformed_usage, "model_dump")
+                    else dict(transformed_usage)
                 )
-            return result.response
+                if resp_dict is not None:
+                    resp_dict["usage"] = new_usage
+                else:
+                    setattr(resp, "usage", new_usage)
+            return resp
         elif isinstance(result, InteractionsAPIStreamingResponse):
             return self._assemble_completed_interaction_response(result)
         else:

tests/unit/litellm_core_utils/test_litellm_logging.py

@@ -5271,6 +5271,37 @@ def test_get_assembled_streaming_response_returns_result_for_streaming():
     assert assembled is result
 
 
+def test_get_assembled_streaming_response_handles_dict_response():
+    """ResponseCompletedEvent.response may hold a plain dict on the streaming
+    /v1/responses spend-log path (#29913). A bare `.usage` access raised
+    AttributeError there, so no LiteLLM_SpendLogs row was written at all."""
+    import datetime
+
+    from litellm.types.llms.openai import ResponseCompletedEvent
+
+    logging_obj = _make_logging_obj(stream=True)
+    event = ResponseCompletedEvent.model_construct(
+        type="response.completed",
+        response={
+            "id": "resp-dict-1",
+            "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15},
+        },
+    )
+
+    assembled = logging_obj._get_assembled_streaming_response(
+        result=event,
+        start_time=datetime.datetime.now(),
+        end_time=datetime.datetime.now(),
+        is_async=False,
+        streaming_chunks=[],
+    )
+
+    assert assembled == {
+        "id": "resp-dict-1",
+        "usage": {"input_tokens": 10, "output_tokens": 5, "total_tokens": 15},
+    }
+
+
 def test_streaming_success_handler_includes_vertex_ai_metadata_in_standard_logging():
     """Assembled streaming responses should include Vertex AI metadata in logging payload."""
     import datetime