[jaeminbird]

Motivation and Context

Streaming an OpenAI chat completion with n>1 merges every choice into one. _ChatCompletionsStreamState (posthog/ai/openai/_streaming.py) reads only choices[0] of each chunk and appends to a single buffer, but each chunk carries one choice identified by choices[].index. The text interleaves ("HelloBonjour world le monde"), tool calls from different choices collide on index 0, and $ai_stop_reason comes from whichever choice finished last. The non-streamed path already keeps one item per choice.

How did you test it?

The stream state now collects content and tool calls per choice index and builds $ai_output_choices in the same shape as the non-streamed path; the stop reason is choice 0's, as in extract_openai_stop_reason. n=1 streams are unchanged.

• New test_streaming_chat_keeps_n_choices_apart fails on main (interleaved text) and passes with the fix; 3 reruns passed.
• The real openai SDK against a mock HTTP transport serving the documented n=2 SSE format: the streamed $ai_output_choices equals the non-streamed capture of the same answer.
• OpenAI tests: 108 pass. ruff format/check, mypy baseline and python -W error -c "import posthog" clean. Full suite: no new failures (19 failures and 2 collection errors are pre-existing Windows/Anthropic environment problems).

Checklist

• Tests added/updated. No public API change (make public_api_check not run).
• No docs change needed.
• Not run: a Linux run, a live OpenAI call.

Agent context

Autonomy: human-driven (agent-assisted). Kimi Code found the bug by comparing the streamed capture with the non-streamed one and wrote the fix and test; I reviewed and verified it.

[marandaneto]

The code looks good; no qualifying findings. Verdict: correct.

[marandaneto]

thanks @jaeminbird