@irisrun/provider-compat

Own, portable, verifiable state — even across "OpenAI-compatible" endpoints. A conformance-verified registry that turns "OpenAI-compatible" from a loose claim everyone makes into a tested, replay-safe guarantee only Iris can make.

What it is

A model reply in Iris is a recorded journal effect that must replay byte-identically. So the canonicalization layer — where divergent "compatible" endpoints quietly differ (finish_reason, usage, streaming terminators, IDs) — is exactly where "compatible" breaks, and the one place that has to care is Iris. This package is the registry that makes the claim precise:

• COMPAT_MATRIX — OpenAI-protocol and Anthropic-protocol endpoints, each tagged replay-safe (point --base-url at it with its key and a recorded session replays byte-identically) or known-divergent (the response canonicalizes, but reaching it needs auth/URL/transport adaptation — Azure's api-key header, Bedrock's SigV4, Vertex's OAuth — named in note).
• entriesByProtocol(p), findEntry(id) — accessors.
• renderCompatMatrix() — a deterministic markdown table, used by iris providers --matrix and pinned to docs/providers.md by a drift guard.

Every entry is run through the matching adapter in CI (tests/provider-compat-matrix.test.ts) and must canonicalize to the stable ModelCallResult, or be flagged here. The matrix is conformance-tested data, never a doc that rots into a false promise.

"Replay-safe" is faithful capture + canonicalization of the model's reply — not a claim that the endpoint is deterministic.

Usage

iris providers --matrix     # print the rendered matrix

import { COMPAT_MATRIX, findEntry } from "@irisrun/provider-compat";
const groq = findEntry("groq");   // { protocol: "openai", baseUrl: "…/chat/completions", … }

Point a portable image at any replay-safe endpoint at deploy time:

iris serve ./image --model openai --base-url https://api.groq.com/openai/v1/chat/completions

Part of Iris — own, portable, verifiable state.