OpenAI provider

Select an API base endpoint, exact model, API credential and the output mode verified for that connection. The normal base endpoint is https://api.openai.com/v1. Credentials are resolved by the host; do not put credentials in the endpoint URL, prompts or model-generated content.

Connection checking reads model metadata without generating text. It verifies model access, not Responses or JSON Schema compatibility. Explicit generation has a token budget and may incur provider charges. Select text, json or native_schema deliberately; unsupported combinations fail without downgrade.

Generation is unary, with storage/background execution disabled. No tools, retrieval, embeddings, execution, persistence or automatic retry are available. Refusal, truncation, timeout and cancellation remain failures. Missing usage means unavailable, not zero. Host schema/domain validation and approval remain necessary before a candidate is saved or executed.

Use HTTPS and optionally an approved CA certificate. TLS verification is always on. Insecure HTTP requires explicit configuration; redirects and cloud metadata endpoints are refused. Configure network egress directly because ambient proxy environment variables are ignored. The provider is optional in air-gap systems.

The bundled tests use mock servers. Live model/API behavior has not been tested.