A Claude Code plugin providing a reusable skill for
driving the GitHub Codex bot (chatgpt-codex-connector) through a complete
review → address → resolve loop on a pull request until it reports no major issues.
The Codex review loop has non-obvious mechanics that are easy to get wrong:
- Codex signals its outcome across three GitHub channels (issue comments, PR reviews, inline comments).
- A clean pass is a top-level issue comment ("Didn't find any major issues"), not silence — a poller watching only inline comments hangs indefinitely.
- Availability detection needs an App-authorized token to list installed apps, so a plain user token can't prove a connector is present on a fresh repo.
This plugin captures all of that once, so any project can @codex review reliably.
/plugin marketplace add djm204/codex-review
/plugin install codex-review
Requires authenticated gh and jq, plus the GitHub Codex
connector installed on the target repo.
The codex-review-loop skill activates when a PR needs a Codex review driven to
resolution. It uses the bundled codex-review-loop.sh for the mechanical steps
(detect / trigger / poll / classify) and leaves judgement — is a finding real, how to fix
it — to the agent.
# the script is usable directly too:
skills/codex-review-loop/codex-review-loop.sh --help| Skill | Docs | What it does |
|---|---|---|
codex-review-loop |
docs · SKILL.md | Drives the GitHub Codex (@codex) review loop on a PR until it reports no major issues. |
bash skills/codex-review-loop/tests/run.sh # 14 dependency-free testsMIT