| name | codex-work |
| description | Delegate a task's implementation wholesale to OpenAI Codex (`codex-run work`) — the whole requirements document in one long autonomous run — keeping policy, orchestration, and acceptance in the main loop. Use at the Implement step of the Orchestrator loop, for any multi-part build. |
| argument-hint | [task-description] |
Codex work
Dispatch the Implementer: codex-run work runs OpenAI Codex through the whole plan in one autonomous run, committing as it goes (danger-full-access by default). codex-run --help lists the flags.
codex-run work -C <repo> --log /tmp/codex-work-<slug>.jsonl < brief.md
codex-run work --resume <id> -C <repo> < findings.md
- Brief = the
/define plan-file path plus only what Codex can't infer from the repo: in-session decisions, the branch, scope boundaries. Don't restate ~/.codex/AGENTS.md or the wrapper's preamble (Implementer seat, safety rails, report format).
- Branch off the default branch first. A repo rooted at
$HOME (this dotfiles repo) must be delegated as a git worktree — the wrapper refuses a writable run there.
- Run it in the background (a run can outlast foreground timeouts) and tail
--log; never edit the checkout while a run is active, and give parallel runs separate worktrees.
- Exit
3 = turn finished but STATUS is not COMPLETE → --resume to finish. Other non-zero = turn failed → read stderr and the log tail; partial edits may remain (git status). --resume (id in the stderr banner) keeps the session's context — prefer it to a cold re-delegate, but never resume a session another run is still writing.
- stdout is the final report: STATUS line plus the requirement-by-requirement audit that Accept judges.
