dsh-llm-chatgpt-oauth

Use a ChatGPT Plus/Pro subscription as a DeepSeek Harness model provider through pi-ai's openai-codex OAuth route—without an OpenAI API key.

This package registers the OpenAI Codex model catalog exposed through ChatGPT OAuth. It does not turn the public api.openai.com route into subscription access, and it does not scrape browser sessions.

This is an independent community project, not an official OpenAI, DeepSeek, or pi-ai package. OpenAI, ChatGPT, and Codex are trademarks of their respective owners.

What it provides

• DeepSeek Harness provider route: openai-codex
• ChatGPT OAuth authentication through chatgpt.com/backend-api
• No OPENAI_API_KEY
• pi-ai's current Codex model catalog (for example gpt-5.4, gpt-5.5, and compatible gpt-5.6-* models)
• Existing Harness message conversion, tool calling, reasoning levels, replay metadata, attachments, and stream translation through PiAiAdapter
• Automatic OAuth access-token refresh
• Private, atomic credential persistence
• Optional bootstrap from an existing Codex CLI login
• A dedicated device-code login command
• An active OpenAI Codex (ChatGPT) provider row in DSH's Models settings page
• Live, settings-backed updates for operational options; custom headers are redacted on settings wires

Model availability and usage limits are controlled by the ChatGPT account and may change independently of this connector.

Requirements

• Node.js >=22.19.0
• pnpm on PATH (used by dsh plugin add)
• DeepSeek Harness compatible with @deepseek-ai/dsh-llm-pi-ai@0.1.0-rc.7
• @earendil-works/pi-ai@0.82.1
• A ChatGPT account with access to OpenAI Codex models

The connector intentionally extends a private-by-convention PiAiAdapter.current() construction seam so it can inject pi-ai's OAuth credential store. Its Harness and pi-ai peer versions are therefore pinned exactly; upgrade those pins only after the full suite and opt-in live test pass against the new versions.

Install & wire (one command)

For an immutable install, use the release commit SHA shown in the GitHub release notes:

dsh plugin --profile web add 'github:zhangxiubo/dsh-llm-chatgpt-oauth#<release-commit-sha>'

A version tag is a more readable alternative:

dsh plugin --profile web add 'github:zhangxiubo/dsh-llm-chatgpt-oauth#v0.1.0'

The connector is a DSH bundle, so either command installs the Git dependency and appends its declared dsh.bundle.patch to the profile's bundle stack—no manual cordis.patch.yml edit. Restart dsh web afterwards.

The repository ships directly runnable JavaScript and needs no install-time build, prepare script, or pnpm allowBuilds entry. This follows Harness's official GitHub plugin-install guidance, including its recommendation to pin a commit SHA. To test the latest development branch instead, omit the revision:

dsh plugin --profile web add github:zhangxiubo/dsh-llm-chatgpt-oauth

The package is marked private and includes a failing prepublishOnly guard to prevent accidental npm-registry publication; GitHub releases are the supported distribution channel.

Manual fallback

If you manage the profile yourself, install the package where the profile's Cordis loader resolves Node packages, then add one row to ${DSH_HOME:-$HOME/.dsh}/profiles/web/cordis.patch.yml:

- insert:
    - id: llm-chatgpt-oauth
      name: dsh-llm-chatgpt-oauth

The important requirement is that import('dsh-llm-chatgpt-oauth') resolves from the profile composition.

Authenticate

Dedicated OAuth login (recommended)

After installing the package:

dsh-llm-chatgpt-auth login

The command prints an OpenAI device-code URL and code, waits for authorization, and saves a dedicated connector credential. It never prints tokens.

Other commands:

dsh-llm-chatgpt-auth status
dsh-llm-chatgpt-auth logout

The credential is stored at:

${DSH_HOME:-~/.dsh}/llm-chatgpt-oauth/oauth.json

Bootstrap from Codex CLI

By default, when no dedicated credential exists, the connector reads:

${CODEX_HOME:-~/.codex}/auth.json

That makes an existing codex login immediately usable. Refreshed connector credentials are written to the connector's own file and never overwrite the Codex CLI file.

For independent token lifecycle, use the dedicated login command and disable bootstrap in ${DSH_HOME:-~/.dsh}/settings.yaml:

llm-chatgpt-oauth:
  bootstrapFromCodexCli: false

Select the provider

After the post-install restart, the Models settings page lists OpenAI Codex (ChatGPT) as an active provider. Select a concrete model in the composer model picker, or set the default for new agents in ${DSH_HOME:-~/.dsh}/settings.yaml:

agent-default-model:
  provider: openai-codex
  model: gpt-5.4
  reasoningEffort: high

The current DSH Models editor has curated forms only for its built-in DeepSeek and pi-ai settings namespaces. This third-party OAuth row is therefore provider visibility/status; use the login CLI for OAuth and the llm-chatgpt-oauth: settings section below for advanced connector options.

Optional plugin configuration

Set connector options in ${DSH_HOME:-~/.dsh}/settings.yaml; valid changes apply live and refresh both the adapter metadata and Models provider directory:

llm-chatgpt-oauth:
  displayName: OpenAI Codex (ChatGPT)
  streamIdleTimeoutMs: 300000
  bootstrapFromCodexCli: true
  # Optional explicit paths:
  # codexAuthFile: /path/to/codex/auth.json
  # credentialFile: /path/to/dsh/oauth.json
  # Optional provider request headers (redacted as secret on settings wires):
  # headers:
  #   X-Custom-Header: value

The same fields may be set as the composition base by targeting llm-chatgpt-oauth in the profile's cordis.patch.yml; settings values override that base. OAuth access and refresh tokens never enter either file. Custom headers are validated, redacted from settings wire responses, and cannot override authorization, cookie, proxy-authorization, or HTTP transport headers.

To disable the route without uninstalling the package:

- id: llm-chatgpt-oauth
  disabled: true

Architecture

1. The plugin registers one Harness route, openai-codex, plus a configurable-provider directory entry addressed by the llm-chatgpt-oauth settings namespace.
2. CodexOAuthAdapter subclasses PiAiAdapter and builds pi-ai's Models collection with an OAuth credential store.
3. Settings changes replace the profile snapshot and the switching credential-store delegate without replacing the active adapter registration.
4. The provider comes from pi-ai's openaiCodexProvider() and uses https://chatgpt.com/backend-api.
5. pi-ai resolves and refreshes the OAuth credential, then supplies the access token to its Codex Responses transport.
6. The inherited PiAiAdapter converts Harness messages to pi-ai context and converts streamed pi-ai events back to Harness chunks.

The connector does not place access or refresh tokens in DSH settings, model configuration, logs, or the Git repository.

Development

npm install
npm run verify

The ordinary test suite never contacts OpenAI and does not require credentials.

Run the opt-in live test only on an authenticated machine:

RUN_LIVE_CODEX=1 npm run test:live

Choose another model with:

RUN_LIVE_CODEX=1 CODEX_MODEL=gpt-5.5 npm run test:live

Inspect the package payload before creating a GitHub release:

npm pack --dry-run

GitHub releases

This project is distributed directly from GitHub rather than the npm registry. package.json is intentionally marked private and carries a registry-publication denial hook; its package name still identifies the Cordis module when pnpm installs the Git dependency.

Canonical repository links are declared in package.json. Release maintainers should follow RELEASING.md, publish an annotated version tag, and include its immutable commit SHA in the GitHub release notes.

Security and limitations

• Never commit ~/.codex/auth.json or the connector's oauth.json.
• Dedicated device-code login is recommended over importing an existing Codex CLI grant.
• The credential store serializes refreshes inside one process; running several connector processes against the same credential file is not currently supported.
• Provider/model access and subscription accounting are controlled by OpenAI.
• This project depends on pi-ai's OpenAI Codex OAuth provider and the current Harness PiAiAdapter implementation; upstream changes may require an update.

See SECURITY.md for reporting guidance.

License

MIT