Codex subscription (ChatGPT login)

Codex supports the official ChatGPT device-code flow (no API key needed — uses your ChatGPT account):

elanous login openai-codex      # opens browser + waits for code entry
elanous login status            # list providers with tokens on file
elanous login logout openai-codex   # forget tokens

Flow: request user code → print https://auth.openai.com/codex/device and a short code → you enter it in any browser → the CLI polls until you finish (max 15min, Ctrl+C cancels) → tokens persist to ~/.elanous/auth.json (chmod 0o600) and mirror to ~/.codex/auth.json so the official Codex CLI stays in sync.

Access tokens auto-refresh when within 120s of expiry. Refresh tokens rotate single-use per OpenAI's OAuth policy — elanous always writes the new pair to both locations so no manual sync is needed.

When OAuth tokens are present, the provider uses https://chatgpt.com/backend-api/codex as the base URL. When only an apiKey is set (fallback mode), it uses https://api.openai.com/v1. Both are recognized automatically.

Subscription first

The primary path for LLM access is a subscription, not an API key:

elanous login openai-codex             # ChatGPT device-code flow — no API key needed
elanous config set llm.provider openai-codex   # optional — auto already prefers it
elanous login status                   # which providers have tokens on file
elanous usage                          # quota and subscription state per account

With llm.provider = auto (2026-09-24 decision) every selection path picks the Codex subscription first; account rotation moves between signed-in Codex accounts by remaining quota, and when none is left the fallback follows llm.fallbackChain before any other keyed provider. The second line above is optional — it pins Codex explicitly.

⚠️ elanous usage may report unavailable (auth) for codex even when login succeeded and the harness runs fine — it queries the Codex CLI mirror (~/.codex/auth.json), a different axis from the token elanous itself uses (~/.elanous/auth.json). Do not read that line as "the credential does not work".

⚠️ The harness also shells out to gh for pull-request lookups. Without it, base-branch selection degrades to the default branch and says so (base-selection=pr-lookup-failed); nothing crashes, but install the GitHub CLI if you want the harness to open and track pull requests.

⛔ On Linux the harness additionally needs the Codex CLI itself. elanous login openai-codex stores a token for elanous's own LLM layer but prints, and means, this:

Codex CLI mirror was not created because this login response lacked its required fields; run codex login once to initialize ~/.codex/auth.json.

Measured on 2026-09-20 on WSL2 Ubuntu with the token present and codex absent: goal authoring, implement, gate and review all completed — and then the run died with codex app-server stdin drain timeout after 5000ms (src/acp/codex-app-server-client.ts), because the ACP delegate spawns the Codex CLI's app-server. The postmortem then classified the abort as quota-exhausted, since the Codex rotation snapshot is empty when the CLI mirror does not exist — a correlation the code itself labels as such, but one that reads as "you ran out of quota" when in fact the CLI was never installed.