fix: raise APIError for top-level Responses API error events during streaming

[mayuriphad]

Summary

Fixes #2487.

Stream.__stream__ / AsyncStream.__stream__ in src/openai/_streaming.py only raise an APIError when the SSE event body has a nested "error" key:

if is_mapping(data) and data.get("error"):
    ...
    raise APIError(..., body=data["error"])

But per the Responses API streaming error spec, the error event carries its fields (type, code, message, param, sequence_number) directly on the event body -- there is no nested "error" key:

{"type": "error", "code": "...", "message": "Something went wrong", "param": null, "sequence_number": 1}

Because data.get("error") is None for these events, the condition never matches, so the event is silently passed to process_data(...) as if it were ordinary stream data instead of raising a clear APIError. Depending on the expected type being cast to, this can surface as a confusing Pydantic validation error deep in the stack instead of the server's actual error message.

Fix

In both the sync Stream and AsyncStream __stream__ methods, also treat sse.event == "error" as an error condition, falling back to the top-level event body for the message/body when there's no nested "error" key (kept for backwards compatibility with any caller still relying on that shape):

if is_mapping(data) and (data.get("error") or sse.event == "error"):
    ...
    error = data.get("error") if is_mapping(data.get("error")) else data
    message = error.get("message") if is_mapping(error) else None
    ...
    raise APIError(
        message=message,
        request=self.response.request,
        body=data.get("error") if data.get("error") is not None else data,
    )

This is a minimal, targeted change -- no unrelated refactors.

Tests

Added regression tests in tests/test_streaming.py (for both Stream and AsyncStream, via the existing sync parametrization):

• test_response_error_event_raises_api_error -- a top-level response.error-shaped SSE error event raises APIError with the server's message.
• test_response_error_event_without_message_uses_default -- the same event without a message field raises APIError with the existing default message.

Both tests fail on main (DID NOT RAISE APIError) and pass with this fix.

$ pytest tests/test_streaming.py -q
24 passed

ruff check and ruff format --check pass on both changed files.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 698bbc5c9b

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Preserve Assistant error events

When this shared stream path is used by the Assistants endpoints, synthesize_event_and_data=True causes a raw SSE event: error body to be wrapped into the documented AssistantStreamEvent shape (event == "error", data: ErrorObject). This new condition raises before that wrapping, so an Assistants stream error that consumers previously could handle via iteration or on_event now aborts as APIError; please scope the new top-level handling to Responses streams or skip it when synthesizing event/data. This is an unrelated exported-stream behavior change from a handwritten core helper.

Useful? React with 👍 / 👎.