Codex Worker for Claude Code

A shared external worker that lets Claude Code consult Codex — as an advisor or a second reviewer — without launching Codex from inside Claude Code's own process tree. See worker/README.md for why this separation matters and how the request/response protocol works.

Claude Code
    |
    | repo-local filesystem IPC (<repo>/.codex-worker/)
    v
codex-worker            (started independently, own terminal)
    |
    v
codex exec --sandbox read-only

Components

• worker/ — the shared worker (codex-worker.sh) and its per-type handlers (advisor, review). One process serves one repo.
• skills/codex-advisor/ — client skill for consulting Codex as an independent technical advisor.
• skills/codex-review/ — client skill for a second-model review of specs, architecture, API designs, or code.
• test/ — test/run-all.sh runs all suites (protocol validation, e2e, cancellation, process-boundary, etc.).
• docs/superpowers/plans/ — milestone specs (M1–M4) and the production cutover runbook.

Quick start

cd /path/to/target/project
codex-worker              # independent terminal, not from Claude Code

codex-advisor medium <<'EOF'
Your question here
EOF

codex-review spec path/to/file.md medium

Development

test/run-all.sh

Production installation (~/.claude/skills/, ~/.local/) is handled by install.sh, which wraps the scripts in docs/superpowers/plans/2026-08-12-codex-worker-m4-scripts/:

./install.sh

Run it yourself, from your own terminal — not something Claude Code runs on your behalf, since the writes land outside this repo. Safe to re-run; back up/rollback follow the cutover runbook.

License

MIT License — see LICENSE.

Written by

Eason Kang in/eason-kang-b4398492/