@irisrun/provider-openai

A vendor-neutral, replay-safe model adapter. The OpenAI model_call performer sits behind Iris's portable model port — a peer of @irisrun/provider-anthropic behind the same port — so swapping providers leaves the agent and its byte-identical journal replay unchanged. No provider is baked into the core. (Faithful record-replay of the captured reply, not a claim that the model is deterministic.)

What it is

A direct adapter for the OpenAI Chat Completions API over the built-in fetch — zero runtime dependencies. It exposes a buffered performer (openaiModelPerformer) and a streaming one (openaiStreamingModelPerformer) that hold the same content === join(deltas) reconcile invariant as the Anthropic adapter and pass the same shared conformance suite. The real adapter calls OpenAI only when OPENAI_API_KEY is set; tests use a fake fetch / fake model.

Use it

Selected automatically from an openai/… model prefix in your Agentfile:

# agent.yaml
model: openai/gpt-x

…or serve it explicitly with iris serve --model openai.

This adapter speaks the OpenAI Chat Completions protocol, so --base-url (or IRIS_MODEL_BASE_URL) on iris run / serve / chat points it at any compatible endpoint — Groq, Together, Fireworks, OpenRouter, DeepSeek, Mistral, xAI, vLLM, Ollama, LM Studio, Azure OpenAI. Which are replay-safe vs known-divergent is a conformance-tested matrix in @irisrun/provider-compat (iris providers --matrix).

See docs/Models & providers for the shared port, the conformance suite, and the compatibility matrix.

Part of Iris — own, portable, verifiable state.