@openharness/provider-chatgpt

ChatGPT/Codex OAuth model provider for OpenHarness. It lets SDK consumers build their own login UX, store a user's ChatGPT OAuth tokens, and pass a Vercel AI SDK language model into @openharness/core.

This package is experimental. It uses the local Codex-style ChatGPT OAuth flow and the ChatGPT Codex backend surface, not the public OpenAI Platform API billing path.

Install

npm install @openharness/provider-chatgpt @openharness/core

Usage

import { Agent } from "@openharness/core";
import {
  FileTokenStore,
  createChatGPTAuth,
  createChatGPTProvider,
} from "@openharness/provider-chatgpt";

const auth = createChatGPTAuth({
  tokenStore: new FileTokenStore("~/.my-harness/chatgpt-auth.json"),
});

const chatgpt = createChatGPTProvider({ auth });

const agent = new Agent({
  name: "coder",
  model: chatgpt("gpt-5.2-codex"),
});

Browser Callback Login

OpenHarness does not provide a CLI. Host apps can trigger the login flow wherever it fits their UX.

await auth.loginWithLocalCallback({
  port: 1455,
  openBrowser: true,
});

For custom callback handling:

import { parseAuthorizationCallback } from "@openharness/provider-chatgpt";

const flow = await auth.createAuthorizationFlow();

console.log(flow.authorizationUrl);

const { code, state } = parseAuthorizationCallback(callbackUrlOrQueryString);

await auth.exchangeCode({
  code: code!,
  state,
  expectedState: flow.state,
  codeVerifier: flow.codeVerifier,
  redirectUri: flow.redirectUri,
});

Device Login

For headless environments:

const flow = await auth.startDeviceFlow();

console.log(flow.verificationUrl);
console.log(flow.userCode);

await flow.pollUntilComplete();

Token Stores

Credential storage is pluggable:

import {
  FileTokenStore,
  MemoryTokenStore,
  OpenCodeTokenStore,
} from "@openharness/provider-chatgpt";

OpenCodeTokenStore reads and writes OpenCode's openai auth entry explicitly. The package does not silently read another tool's credentials.

You can provide your own store:

const tokenStore = {
  async get() {
    return tokens;
  },
  async set(next) {
    tokens = next;
  },
  async clear() {
    tokens = null;
  },
};

Request Behavior

The provider wraps @ai-sdk/openai Responses models and rewrites requests to the ChatGPT Codex backend:

https://chatgpt.com/backend-api/responses
https://chatgpt.com/backend-api/codex/responses

It refreshes OAuth tokens, injects the ChatGPT account id header, forces stateless store: false, requests encrypted reasoning continuity, and converts Codex SSE responses back to JSON for non-streaming AI SDK calls.

License

MIT