Responses API error handling reads error.message, but spec says message is top-level

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

• This is an issue with the Python library

Describe the bug

In the Responses API event handling logic, the SDK reads the error message from error.message. However, according to the official documentation, the message field is at the top level of the error object, not nested inside another error field.

SDK code (current behavior):

if sse.event == "error" and is_mapping(data) and data.get("error"):
    message = None
    error = data.get("error")
    if is_mapping(error):
        message = error.get("message")
    if not message or not isinstance(message, str):
        message = "An error occurred during streaming"

    raise APIError(
        message=message,
        request=self.response.request,
        body=data["error"],
    )

Official API response example:

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

Proposed fix: Change the error handling logic to read data["message"] (the top-level field) and data["error"]["message"] (for backward compatibility) instead of only data["error"]["message"].

To Reproduce

Code snippets

OS

Python version

Library version

Read the Responses API event handling logic in src/openai/_streaming.py around line 88, then compare it with the official streaming error response described in the issue. Done means extracting the top-level message while preserving support for the nested error.message form and retaining the existing APIError behavior.