Responses streaming: response.function_call_arguments.done events lack the required 'name' field, failing strict event-union validation #3472

Confirm this is an issue with the Python library and not an underlying OpenAI API

• This is an issue with the Python library (or spec drift between the SDK models and the live API)

Describe the bug

ResponseFunctionCallArgumentsDoneEvent declares name: str as required (both in v2.38.0 and in the current generated model on main), but the live Responses API's streamed response.function_call_arguments.done events do not include a name key.

Observed payload from a streamed responses.create call with function tools (gpt-5.4):

{
  "type": "response.function_call_arguments.done",
  "arguments": "{...}",
  "item_id": "fc_0e7fa7b0f70886a0006a4a7f734bd881929f7d0eae45f46f97",
  "output_index": 2,
  "sequence_number": 10
}

Strict validation of the ResponseStreamEvent union therefore fails on every such event with missing: name (plus the ~100 other enumerated union-member errors). Nothing crashes — construct_type (openai/_models.py:621 → validate_type:843 → _validate_non_model_type:890) tries strict validation inside a try/except and falls back to lenient construction, so the event flows through with name unset — but:

1. Anyone running pydantic validation instrumentation (e.g. logfire's pydantic plugin) sees a warning with ~101 enumerated errors per function-call round, and
2. event.name on these events is silently None despite the type declaring str, so type-checked user code reading it is misled.

Since the current main model (Stainless-generated) still requires name, this looks like drift between the OpenAPI spec and the live API rather than a stale pin — either the API should emit name on response.function_call_arguments.done (it's genuinely useful; today one must correlate via item_id back to the output_item.added event to know which function completed), or the model field should be Optional[str].

For prior art on this bug class: #2311 (a mis-mapped streaming event type, since fixed).

To Reproduce

1. Call client.responses.create(stream=True, tools=[...]) with a prompt that triggers a function call (observed on gpt-5.4).
2. Validate incoming SSE event dicts strictly against ResponseStreamEvent (or run any pydantic validation instrumentation).
3. Every response.function_call_arguments.done event fails strict validation with missing: name for ResponseFunctionCallArgumentsDoneEvent.

Code snippets

# The event payload above, validated strictly:
from openai.types.responses import ResponseFunctionCallArgumentsDoneEvent
ResponseFunctionCallArgumentsDoneEvent.model_validate({
    "type": "response.function_call_arguments.done",
    "arguments": "{}",
    "item_id": "fc_x",
    "output_index": 0,
    "sequence_number": 1,
})  # pydantic.ValidationError: name Field required

OS

Linux (Docker)

Python version

Python 3.14.0

Library version

openai 2.38.0 (model on current main also declares name: str required)