Delegate bounded coding tasks to Codex CLI subagent

Delegates a bounded coding, review, or verification task to a separate Codex CLI session - OpenAI's terminal coding agent. codex exec runs it non-interactively: it works autonomously in a sandbox, streams progress to stderr, and prints only the final message to stdout. Auth reuses the user's ChatGPT subscription, never an API key.

Delegate a self-contained coding task with clear success criteria (fix, feature, refactor, review), several independent parallel tasks at once, or when you want a second opinion or independent verification of your own changes. Do NOT delegate tasks that need conversation context you can't fully write into the prompt - Codex sees nothing of the conversation, so the prompt must carry the goal, relevant paths, constraints, and how to verify completion.

Preflight checks codex --version (install via npm i -g @openai/codex or brew install --cask codex if missing) and codex login status (exit 0 + "Logged in using ChatGPT" means ready; otherwise stop and tell the user to run codex login, a one-time browser OAuth - never read, print, or copy ~/.codex/auth.json). Launch:

OUT=$(mktemp /tmp/codex-out.XXXXXX)
codex exec \
  --cd /path/to/repo \
  --sandbox workspace-write \
  --output-last-message "$OUT" \
  "Full task prompt: goal, constraints, files to touch, definition of done." \
  </dev/null

</dev/null is mandatory when stdin isn't a real terminal, since Codex treats open stdin as extra context and waits forever for EOF. A long prompt can be piped via stdin instead, and the verbose stream can be backgrounded or wrapped in a subagent to stay out of the parent context; runs take minutes with no built-in timeout. Optional flags: -m <model> to override the model, --json for a JSONL event stream. Collect results with:

cat "$OUT"                            # final message = the deliverable
git -C /path/to/repo status --short   # see what Codex actually changed

Follow-up in the same session (cwd-filtered) uses codex exec resume --last "follow-up instruction" </dev/null.

Wraps OpenAI's Codex CLI (@openai/codex, installable via npm or Homebrew) and reuses the host's ChatGPT auth. Can be wired as a Cursor-native /codex subagent by adding ~/.cursor/agents/codex.md pointing at this skill. For parallel runs, only genuinely independent tasks should be parallelized with upfront file ownership so results merge cleanly - one git worktree per Codex run, never two in the same tree. Known failure modes: a hang with no output means stdin was left open (kill and relaunch with </dev/null); a non-zero codex login status means the user must run codex login (don't work around it); a ChatGPT plan rate limit should be reported to the user, never retried in a loop; a "Not a git repo" error needs --skip-git-repo-check or a git init; network is blocked inside the workspace-write sandbox by default unless the task needs installs or API calls, in which case pass -c sandbox_workspace_write.network_access=true. Never use --dangerously-bypass-approvals-and-sandbox.

Agents or developers who want to hand off a bounded coding, review, or verification task to an independent CLI agent, run several independent tasks in parallel, or get a second opinion on their own changes - one task per launch, splitting big jobs into multiple launches, and always reviewing Codex's diff before declaring the task done.