Fix unreachable dead code in error event handling

[yurekami]

Summary

• Fixed unreachable dead code in _streaming.py where the error event check was incorrectly placed inside the thread.* event handling block
• Added regression tests for error event handling

Problem

In src/openai/_streaming.py, the check for sse.event == "error" was placed inside the if sse.event.startswith("thread.") block. Since "error" doesn't start with "thread.", this condition could never be true, making lines 67-79 (sync) and 170-182 (async) unreachable dead code.

# Before (dead code):
if sse.event and sse.event.startswith("thread."):
    data = sse.json()
    if sse.event == "error":  # <-- Can never be true!
        raise APIError(...)

This was identified as a regression from a previous refactoring commit.

Solution

Moved the error event check to be before the thread.* check, so it's evaluated independently:

# After:
if sse.event == "error":
    data = sse.json()
    raise APIError(...)

if sse.event and sse.event.startswith("thread."):
    # Handle thread.* events

Also simplified the error handling to correctly parse the ErrorObject structure from the Assistants API (which has message directly in data, not nested under data.error).

Test plan

• Added test_error_event_raises_api_error - verifies error events raise APIError with correct message
• Added test_error_event_with_missing_message - verifies default message is used when message field is missing
• All 24 streaming tests pass (including 4 new test cases for sync/async variants)

Fixes #2796

🤖 Generated with Claude Code