OpenAI
OpenAI model access in Coven through Codex, including login, doctor checks, and credential boundary.
1 min read
OpenAI model access currently flows through Codex.
npm install -g @openai/codex
codex login
coven doctor
coven run codex "fix the failing tests"Boundary
Codex owns OpenAI provider credentials and model calls. Coven owns the session record, PTY supervision, attach/replay behavior, and event log.
Coven does not read OpenAI API keys, OAuth refresh tokens, or Codex auth files. If Codex auth expires, run codex login and retry.
How OpenAI failures surface under Coven
Because Codex owns the provider relationship, OpenAI problems appear as harness behavior, never as Coven credential errors:
| Symptom | Likely cause | Fix |
|---|---|---|
coven doctor reports Codex missing | codex is not on PATH | npm install -g @openai/codex, then re-run coven doctor. |
coven run codex ... fails with pty_spawn_failed | Executable found but could not launch | Check that codex runs standalone; reinstall if it crashes at startup. |
| Session starts, then exits with auth errors in the log | Expired or missing OpenAI login | Run codex login, then relaunch. The failed session stays replayable. |
| Session runs but hits rate limits or model errors | OpenAI account limits or model availability | Resolve in the OpenAI account or Codex config; Coven has no quota controls. |
In every case the session record and event log survive: coven attach <session-id> replays exactly what Codex printed, which is usually the fastest way to see the provider's own error text.
What Coven records — and never reads
| Data | Coven's posture |
|---|---|
| Session record, status, exit code | Recorded in the ledger. |
| Codex terminal output, including any error text | Recorded as append-only events. |
| OpenAI API keys, OAuth tokens, Codex auth files | Never read, proxied, or stored. |
| Model selection | coven run --model forwards a model id to Codex's native --model flag; provider configuration and credentials stay with Codex. |
Passing --model is argv forwarding, not provider access: Coven strips any openai/ prefix and hands the bare id to Codex. Coven still treats the harness as a black box that writes output.