[pcbeingused333]

A refused Chat Completions request (for example a refused structured output) ends with finish_reason: "stop", content: null and the text in message.refusal, or in delta.refusal when streamed. OpenAICompletionsProvider ignored the field, so the turn came back empty and was reported as "success": chat.chat(...) returned an empty string, a streamed call yielded nothing, and with a data_model the failure showed up as a missing JSON payload instead of a refusal.

The provider now yields streamed delta.refusal text, keeps the refusal as the turn's text when the message has no content, and reports finish_reason == "content_filter" for it, which makes the existing "filtered by the provider's content moderation policy" warning (or error, with a data_model) apply. Responses with content are unchanged, and so is a refusal-free stop. Databricks, Cloudflare, Portkey, Perplexity, Hugging Face and the other providers built on this class inherit the change.

Tests (tests/test_provider_openai_completions.py): a refused ChatCompletion gives the refusal text and content_filter; the same with a data_model raises the content-moderation error; and streamed chunks with delta.refusal fragments are yielded as text, merged, and end in a turn with the full text and content_filter. All three fail on main. The existing tests in the file still pass; the failures in tests/test_chat.py in my environment come from missing optional dependencies and are the same on main. No request was made to a live server.

This change was made with Claude Code; the tests above were run before opening the PR.

🤖 Generated with Claude Code