@corbits/codex-provider

An Interchange inference provider for OpenAI Codex ("Login with ChatGPT"): OAuth constants and token mapping for @corbits/oauth-core, and a Responses-protocol adapter for Codex's ChatGPT backend built on @corbits/openai-responses. It does not implement login flow orchestration, token storage, or usage/quota reporting — a host composes those itself from @corbits/oauth-core directly.

Install

bun add github:corbitsdev/corbits-codex-provider

The package ships TypeScript source and needs no build step; Bun consumes it directly. @intx/inference and @intx/types are peer dependencies and resolve to the host's own copy.

Usage

import type { AdapterManifest } from "@intx/inference";
import { createDefaultScheduler } from "@intx/inference";
import { loadAdapterRegistry } from "@intx/inference/providers";
import type { InferenceSource } from "@intx/types/runtime";
import { withCodexContentTypeRepair } from "@corbits/codex-provider";

const manifest: AdapterManifest = [
  {
    provider: "codex",
    specifier: "@corbits/codex-provider",
    export: "createCodexResponsesAdapter",
  },
];

const adapters = await loadAdapterRegistry(manifest);
const dependencies = {
  adapters,
  fetch: withCodexContentTypeRepair(fetch),
  scheduler: createDefaultScheduler(),
};

// Quirks ride on the `InferenceSource` the harness resolves at call time,
// not on the manifest entry:
const source: InferenceSource = {
  id: "codex/1",
  provider: "codex",
  baseURL: "https://chatgpt.com/backend-api",
  apiKey: "",
  model: "gpt-5.5",
  quirks: {
    productName: "My Harness",
    environmentTagName: "my_harness_environment",
  },
};

codexOAuthConfig, exchangeCodexCode, and refreshCodexTokens plug directly into @corbits/oauth-core's startOAuthLogin and createTokenSession, as shown in that package's own README.

API

See src/index.ts for the full public surface and its TSDoc; the Codex quirks shape is CodexQuirks in src/quirks.ts.

Design notes

• Codex has no separate API-key credential path — only the ChatGPT OAuth subscription token is supported.
• A host's own operating prompt rides as the leading developer message via wrapCodexBridgeMessage. This package does not send an instructions field.
• CodexQuirks (productName, environmentTagName) has no default: there is no honest generic product name, so an absent quirks bag is a validation error, not a silently-generic adapter.
• createCodexResponsesAdapter is a plain (source, quirks?) AdapterFactory, loadable from an AdapterManifest entry like any other provider.
• withCodexContentTypeRepair exists because the backend omits content-type on some streamed responses despite a valid SSE body; it restores the header from the request's own accept commitment rather than guessing.
• Requests identify as the public Codex CLI (originator: codex_cli_rs, the CLI's own OAuth client id) because the ChatGPT backend only serves that client; the ChatGPT subscription token from that login is the only credential this package uses, and OpenAI closing that door is an outage this package cannot route around, not a bug in it.

Not supported

• No API-key credential path — Codex has none.
• No live model catalog (GET /codex/models) — a host that needs one fetches it itself.
• No usage/quota or rate-limit surfaces — out of scope for an inference provider adapter.
• No login flow orchestration or token storage — call @corbits/oauth-core directly with codexOAuthConfig.

License

LGPL-2.1-or-later.