fix(responses): count custom-tool and MCP argument deltas in the streamed usage fallback

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/responses/streaming_iterator.py

@@ -356,7 +356,7 @@ class BaseResponsesAPIStreamingIterator:
                     _delta: Final = getattr(openai_responses_api_chunk, "delta", None)
                     if isinstance(_delta, str):
                         self._generated_content += _delta
-                elif _event_type == ResponsesAPIStreamEvents.FUNCTION_CALL_ARGUMENTS_DELTA:
+                elif _event_type in _TOOL_ARGUMENTS_DELTA_EVENTS:
                     _args_delta: Final = getattr(openai_responses_api_chunk, "delta", None)
                     if isinstance(_args_delta, str):
                         self._generated_tool_arguments += _args_delta
@@ -1358,6 +1358,15 @@ def _usage_as_model(usage: object) -> ResponseAPIUsage | None:
         return None
 
 
+_TOOL_ARGUMENTS_DELTA_EVENTS: Final = frozenset(
+    {
+        ResponsesAPIStreamEvents.FUNCTION_CALL_ARGUMENTS_DELTA,
+        ResponsesAPIStreamEvents.CUSTOM_TOOL_CALL_INPUT_DELTA,
+        ResponsesAPIStreamEvents.MCP_CALL_ARGUMENTS_DELTA,
+    }
+)
+
+
 def _estimate_usage_from_text(
     model: str,
     request_input: object,

tests/test_litellm/responses/test_streaming_iterator.py

@@ -816,3 +816,31 @@ async def test_completed_event_survives_a_failing_usage_estimate():
 
     assert yielded
     assert iterator.completed_response.response.usage is None
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "tool_delta_event_type",
+    ["response.custom_tool_call_input.delta", "response.mcp_call_arguments.delta"],
+)
+async def test_completed_event_without_usage_counts_tool_input_deltas(tool_delta_event_type):
+    """Custom-tool and MCP argument deltas feed the streamed usage fallback the
+    same way function_call_arguments deltas do."""
+    response = _responses_api_response_without_usage()
+    iterator = _make_iterator(
+        sse_events=[
+            _sse_event({"type": tool_delta_event_type, "delta": '{"query": "weather in sf"}'}),
+            _sse_event({"type": "response.completed", "response": {}}),
+        ],
+        logging_obj=_logging_obj_stub(),
+        config=_mock_config_with_completed_response(response),
+        request_data={"input": "what is the weather in san francisco"},
+    )
+
+    async for _ in iterator:
+        pass
+
+    usage = iterator.completed_response.response.usage
+    assert usage is not None
+    assert usage.output_tokens > 0
+    assert usage.total_tokens == usage.input_tokens + usage.output_tokens