convez.ai
SkillDevOps & InfraestruturaAvançado
raccrompton

Claude Codex Skills

raccrompton/claude-codex-skills

Revisão de código e specs com Codex CLI via ChatGPT subscription

Ver no GitHub
InicialPluginScore · de 0 a 100

Duas skills para Claude Code que integram a CLI Codex (ChatGPT subscription, sem custo de API) no fluxo de desenvolvimento. Oferece revisão soft de specs/plans e uma gate obrigatória em PRs compatível com merge automático.

Instalação

$npx convez-ai raccrompton/claude-codex-skills

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Revisar especificações e planos antes de aprovação
  • Implementar gate obrigatório de revisão em pull requests
  • Integrar verificação de código sem custo de API OpenAI

Veredito

Solução engenhosa que contorna limitações do Codex oficial ao usar CLI local com auth ChatGPT e reportar status via GitHub. Exige setup prévio robusto (Codex CLI, gh CLI, integração GitHub) mas oferece verdadeira gate obrigatória sem custo de API — ideal para equipes com ChatGPT subscription ativa.

README

A tradução para o português deste README ainda está na fila.

claude-codex-skills

Two Claude Code skills that wire the Codex CLI in as a reviewer in your development workflow — running on your ChatGPT subscription (no OpenAI API cost), read-only.

  • consulting-codex — a soft second opinion on a drafted spec or plan before it reaches your approval gate. Iterates the artifact to convergence (≤3 rounds) and hands back a digest. Never blocks.
  • codex-pr-review — a hard PR gate. Reviews the PR diff at submission time, posts a codex-review commit status, and is compatible with gh pr merge --auto. Fails closed when a repo requires the status.

Both designed to pair with the superpowers brainstorming / writing-plans workflow, but neither depends on it.

Why

Codex's cloud GitHub integration is advisory only — it always posts a COMMENTED review, so it can never be a required status check. And the official openai/codex-action needs an OpenAI API key (paid), not your ChatGPT plan.

These skills sidestep both: they run the local codex CLI (authed in chatgpt mode = subscription, $0 API) and, for the PR gate, report the verdict to GitHub as a commit status you can require in branch protection. Codex never runs in CI, so there's no API key and no token-refresh fragility.

Prerequisites

  • Codex CLI installed and signed in with ChatGPT: codex login (verify ~/.codex/auth.json shows "auth_mode": "chatgpt").
  • gh CLI authed with repo scope (the PR gate posts commit statuses via gh api .../statuses).
  • Claude Code.

Install

Copy the skills into your Claude Code skills directory:

git clone https://github.com/raccrompton/claude-codex-skills
cp -r claude-codex-skills/skills/consulting-codex  ~/.claude/skills/
cp -r claude-codex-skills/skills/codex-pr-review    ~/.claude/skills/

Then tell Claude when to invoke them by adding a section to your global ~/.claude/CLAUDE.md — see CLAUDE-snippet.md for a ready-to-paste block.

Using codex-pr-review as a real merge gate

The skill posts a codex-review commit status on every PR. Posting a status does not block a merge by itself — a repo opts in by making codex-review a required check:

gh api -X PUT repos/<owner>/<repo>/branches/<default>/protection --input - <<'JSON'
{ "required_status_checks": { "strict": false,
    "contexts": ["codex-review"] },
  "enforce_admins": false, "required_pull_request_reviews": null,
  "restrictions": null }
JSON

(Merge any existing required contexts into that array — don't drop your CI checks.) Once required, gh pr merge --auto waits for codex-review and the merge fails closed: no status posted ⇒ the PR parks.

Rollout order matters: make sure the skill has posted a codex-review status on at least one PR before adding it to required contexts — otherwise every PR blocks forever waiting for a status nothing produces.

Protected vs unprotected repos

The skill's final merge step is protection-state-aware: on a protected repo it uses gh pr merge --auto --squash (GitHub waits for required checks); on an unprotected repo it does not use --auto (which would merge immediately, before CI) and instead polls CI then merges by hand. So the Codex review runs everywhere, but the merge mechanic adapts to the repo.

Trust model

codex-pr-review is self-attesting: the same agent writes the code, runs Codex, judges the findings, and posts the status. That's appropriate for a solo dev trusting their own agent — the PR comment is the durable record. It is not an independent review; pair it with human review or the cloud Codex integration if you need that.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados