dsh-codex-oauth

Use your OpenAI Codex subscription (ChatGPT Plus/Pro) inside DeepSeek Harness — via OAuth, the same way the official Codex CLI and other harnesses do.

The upstream harness's multi-provider adapter deliberately withholds openai-codex because Codex authenticates through ChatGPT OAuth, and that adapter holds no credential store and runs no login flow. This community plugin supplies both pieces as an installable bundle: a file-backed OAuth credential store, a /codex login human command, and a codex provider route registered on the public LLM seam.

• Built on the published seam packages (@deepseek-ai/dsh-llm, @deepseek-ai/cordis) — no fork, no core change.
• pi-ai's provider-owned Codex OAuth flows handle the wire protocol: browser login with a local callback server, headless device-code login, and automatic refresh under a cross-process credential-store lock.
• Tokens live in $DSH_HOME/codex-oauth.json (0600, owner-only directory), the same place the CLI bin and the harness plugin both read.

Requirements

• A ChatGPT Plus or Pro subscription. (A plain OpenAI platform API key does not work — subscription access is bound to your ChatGPT account, not an API key.)
• DeepSeek Harness installed (npx @deepseek-ai/dsh web or a source checkout).

Install

One command installs the bundle into the web profile (it writes the one-time pnpm build approvals and runs dsh plugin add for you):

npx --yes https://github.com/birat-chapagain/dsh-codex-oauth/releases/download/v0.1.6/dsh-codex-oauth.tgz install

Then restart dsh web and run /codex login once.

Manual alternatives (same effect, both use prebuilt artifacts with no build permission):

dsh plugin --profile web add https://github.com/birat-chapagain/dsh-codex-oauth/releases/download/v0.1.6/dsh-codex-oauth.tgz
# or, from git (pin a commit for reproducibility: github:…/…#<sha>):
dsh plugin --profile web add github:birat-chapagain/dsh-codex-oauth

pnpm 11.22+ hard-fails when any transitive dependency has an unapproved build script — pi-ai's tree carries two (@google/genai, protobufjs, both unused by the Codex route). The one-command installer approves exactly those packages, repairs their set this to true or false placeholders, and preserves every unrelated allowBuilds value, including explicit denials. A manual install that ends with ERR_PNPM_IGNORED_BUILDS just needs this one-time snippet in the profile's pnpm-workspace.yaml:

allowBuilds:
  '@google/genai': true
  protobufjs: true

(If pnpm prints different exact keys in its error, use those — the printed keys are authoritative.)

Expected peer-dependency warning

pnpm may report @deepseek-ai/cordis, @deepseek-ai/dsh-llm, or @deepseek-ai/dsh-invariants as missing peers. This bundle is a Harness plugin: Harness supplies those packages from the host installation, while the profile intentionally sets autoInstallPeers: false. pnpm checks only the profile's package graph and cannot see the host packages that Harness makes available when it loads plugins.

The warning alone is not an installation failure. Do not add duplicate Cordis, LLM, or invariants packages to the profile to silence it; duplicate host packages can create separate plugin contexts or service instances. Confirm the installation with dsh --profile web --dump-config, then start dsh web; investigate the warning only if either command fails or pnpm names a different missing package.

The profile manifest ends up listing the bundle after @deepseek-ai/dsh-base; verify the composed tree without booting:

dsh --profile web --dump-config

Log in

Logging in is a human command, not a model tool — it never enters a prompt.

Web UI

Type /codex login in the chat input. A browser window opens on the ChatGPT authorization page; complete it, and the command reports when the token is stored. Use /codex logout and /codex status to manage it. Device login needs instructions while authentication is pending, but a human command returns only one final result, so /codex login device immediately directs you to the CLI command below instead of starting a flow whose code the UI cannot show.

Headless / CLI

The bundle also ships a dsh-codex-oauth bin that runs outside the harness (the headless profile has no command plane):

npx dsh-codex-oauth login                 # browser flow (desktop)
npx dsh-codex-oauth login --method device # device-code flow (headless)
npx dsh-codex-oauth status
npx dsh-codex-oauth logout

Device flow prints a one-time code plus the OpenAI device-verification URL; enter the code on any device, and the CLI waits until you authorize and stores the token in the same file the harness reads.

Use Codex

The plugin registers provider route codex with the Codex catalog models (gpt-5.x-codex and friends, from the installed pi-ai catalog). Select codex / a Codex model in the Web model picker, or set the default for a headless profile in the profile's cordis.patch.yml:

- id: agent-default-model
  config:
    provider: codex
    model: gpt-5.4

Per-session selection in the Web UI needs no patch. Provider, model, and capabilities resolve through the same LLM seam as shipped providers; prompts, tools, persistence, and history replay behave identically.

Configuration

Override in a later patch layer (profile cordis.patch.yml replaces this row's whole config):

- id: codex-oauth
  config:
    provider: codex
    transport: sse

Security notes

• The store document is written atomically with 0600 permissions under a 0700 directory, and a group/world-readable document is refused on POSIX. It holds your ChatGPT OAuth tokens — treat it like an API key.
• The harness process and its tool subprocesses run as your user; like the upstream credentials document, this file is not hidden from tools the model can drive. Do not point the model's workspace at your Harness home.
• Only https URLs issued by the login flow are ever handed to the browser opener. A missing or failing OS opener is reported without terminating the harness; the CLI still prints the URL for manual use.
• The login flow is pi-ai's provider-owned implementation (authorization-code + device-code against chatgpt.com); this plugin answers its interaction prompts and stores the result.

How it works

• src/store.ts — FileCredentialStore, a persistent pi-ai CredentialStore with serialized read-modify-write (dsh-atomic-write).
• src/auth.ts — login/status/logout over pi-ai's openai-codex OAuth provider.
• src/adapter.ts — CodexAdapter extends LlmAdapter (from @deepseek-ai/dsh-llm), registered with ctx.llm.registerAdapter(['codex'], …); stream() resolves/refreshes auth through pi-ai, enforces provider-idle timeout, and aborts SDK work when its consumer stops.
• src/convert.ts — request/stream vocabulary conversion, adapted from @deepseek-ai/dsh-llm-pi-ai (MIT, © DeepSeek AI) with image attachment support and provider-native replay state omitted.
• src/index.ts — the Cordis function plugin (name/inject/Config/apply); registers the adapter and, when the composition mounts ctx.commands, the /codex command.

Limitations

• Text only. Model metadata advertises only text input, and image content is refused with UNSUPPORTED_CONTENT before any provider request.
• No browser Models-page card. Configuration happens through the patch layer and the picker lists the route through the adapter registry; login is the bin or /codex command, not the credentials page.
• No provider-native replay state. Historical assistant messages replay as provider-neutral content (correct, but without signature/cache reuse).
• Browser login assumes a desktop browser. Machines without one run dsh-codex-oauth login --method device in a terminal.
• One browser login at a time. The OAuth callback uses one local port; wait for one browser flow to finish before starting another. The store lock separately serializes credential writes.

Development

npm install
npm test        # builds lib/ then runs vitest (unit + Loader composition + built-bin smokes)

The composition test boots the real dsh-llm service and this plugin through the Cordis Loader with only the pi-ai SDK mocked, and the bin tests exercise the built artifact under plain Node.

License

MIT. The conversion modules in src/convert.ts are adapted from @deepseek-ai/dsh-llm-pi-ai (MIT, © DeepSeek AI).