OpenAI

Also the fallback when introspection finds nothing and --placeholder-task is in effect — the placeholder makes an OpenAI chat call.

{{PROVIDER_ASSERTS}} substitution

Optional env vars (do NOT assert; document in # TODO comments)

• OPENAI_ORG_ID — only needed for orgs that scope keys per org.
• OPENAI_BASE_URL — only needed when pointing at a self-hosted / Azure-compatible endpoint. If the call-site uses OpenAI(base_url=...), the wrapped function already handles this.

Adapter notes

• Sync chat completion: openai.OpenAI().chat.completions.create(model=..., messages=[...]) returns a ChatCompletion object. Extract the text via .choices[0].message.content.
• If the user’s function returns the raw ChatCompletion, wrap with a .choices[0].message.content extractor in task_fn (the evaluator expects a string output, not the full SDK object).
• Async variants (AsyncOpenAI, acreate): wrap the async call with asyncio.run(...) inside a sync task_fn. See Workflow step 2.5d “Signature adaptation”.

Common gotchas

• Rate limits — pass --jobs 1 or --jobs 2 if the experiment hits 429s.
• gpt-5.4-turbo deprecations: the user’s function may reference a model that has been retired. Don’t silently rewrite the model name — surface as a WARNING: if gpt-3.5-turbo or similar is detected.