[pcbeingused333]

Problem

A non-streamed OpenAI chat completion that the model refused has message.refusal and an empty content. format_openai_response only read the text, tool calls and audio of the message, so the generation was captured with an empty $ai_output_choices. The streaming path already records a refusal delta as a refusal content block.

Changes

format_openai_response adds {"type": "refusal", "refusal": ...} to the message content when message.refusal is set, matching the streamed block. A pypi/posthog patch changeset is included.

Test strategy

• Unit: test_chat_response_keeps_refusal in posthog/test/ai/openai/test_openai_converter.py builds a ChatCompletion whose message has content=None and a refusal; it fails on master and passes here. The rest of that file passes (18 passed) and ruff format --check is clean on the touched directories.
• No end-to-end run against the OpenAI API was made.
• Edge cases considered: a message with both text and a refusal keeps both, in that order; a message without a refusal is unchanged.

🤖 Agent context

Autonomy: Fully autonomous

• Written with Claude Code (Claude Sonnet 5.5): the gap was found by comparing the streaming and non-streaming converters; the change, test and this description were produced by the agent. Agent-authored, needs human review.

🤖 Generated with Claude Code

🤖 Generated with Claude Code

[marandaneto]

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