convez.ai
SkillDevOps & InfraestruturaIntermediário
cskwork

Superhooks Skill

cskwork/superhooks-skill

Git hooks com agentes de IA para code-review e QA no seu laptop

Ver no GitHub
InicialPluginScore · de 0 a 100

Gera git hooks que delegam a agentes de IA (kiro-cli, codex, claude, agy) para code-review e QA com verditos estruturados, bloqueio de push em achados críticos e notificações desktop. Escopo pessoal em laptop, sem distribuição para equipe.

Instalação

$npx convez-ai cskwork/superhooks-skill

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
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 4 semanas
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Bloquear pushes com code-review automático antes de enviar
  • Detectar secrets e achados críticos no pre-commit/pre-push
  • QA distribuído para múltiplos repositórios com mesmo workflow
  • Integrar diferentes CLIs de IA (Claude, Codex, Kiro) sem reescrever bash

Veredito

Ferramenta sólida para automação local de gates de qualidade usando IA, com contrato bem definido de verditos e suporte a múltiplos agentes. Ideal para desenvolvedores que querem enforçar padrões no próprio laptop sem configuração repetitiva.

README

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

/superhooks

English | 한국어

git event → AI agent → structured verdict → block or alert.

Wire any git hook to a code-review/QA agent (kiro-cli, codex, claude, agy) without writing the bash. Activate once, then superhooks install <intent> in any repo.

Scope

This is a personal-laptop skill, not a team gate. The hooks fire on your pushes from your laptop. They do not distribute to other clones — git hooks are local by design (.git/ is not versioned; core.hooksPath is per-clone config; even committed hooks need each contributor to opt in, and any contributor can bypass with git push --no-verify).

For team-wide enforcement, run the same prompt in CI or a server-side pre-receive hook. The intents are designed to port — same VERDICT: contract, same agent CLIs, just a different trigger surface. See docs/DESIGN.md for the recommendation and why a "bootstrap on clone" pattern was deliberately not shipped.

What it does

A git hook for an AI agent has the same shape every time:

  1. detect the git event,
  2. decide if this invocation is in scope (branch / file filter),
  3. build a prompt with the right context (diff, range, …),
  4. run the agent — sync if blocking is meaningful, async otherwise,
  5. parse a structured verdict and act (block, alert, log, dispatch follow-up).

superhooks keeps that shape in one worker library and moves the variable parts into named intents. An intent is three files; adding a new one takes no code change.

Why

  • One review/QA pattern, many repos. Stop hand-rolling the same pre-push bash every time you want an AI gate.
  • The CLI you can swap (kiro-cli ↔ codex ↔ claude ↔ agy) is orthogonal to the hook contract — superhooks agent is a TUI for it.
  • Blocking on critical findings keeps developers honest without freezing every push: the verdict is a single grep-able line, missing or timed-out verdicts fail open with a loud banner.

Quick start

git clone <this repo> ~/Documents/superhooks-skill
ln -s ~/Documents/superhooks-skill ~/.kiro/skills/superhooks
ln -s ~/Documents/superhooks-skill/bin/superhooks       ~/.local/bin/superhooks
ln -s ~/Documents/superhooks-skill/bin/superhooks-agent ~/.local/bin/superhooks-agent

superhooks agent                                       # pick the active agent
cd /path/to/repo
superhooks install review-gate                         # gate all pushes
superhooks install secret-scan                         # block commits with secrets
superhooks status                                      # show what's installed

See SKILL.md for the agent-facing reference and docs/DESIGN.md for the architecture.

Verdict contract

The agent must end every report and its stdout with one line:

VERDICT: block (<N> critical: <reason>)           → push aborted
VERDICT: fix-required (<N> findings: ...)         → push proceeds, Sosumi banner
VERDICT: clean                                    → push proceeds, Glass banner
(no line / garbled / timeout)                     → push proceeds, Funk banner (fail-open)

Status

This skill grew out of a real-world prototype on a closed-source codebase (Spring Boot + Java APIs, Kiro CLI, Playwright). The version published here is the generic re-extraction. The shipped intents (review-gate, secret-scan, qa-dispatch) are language-agnostic; example prompts under examples/ cover Java/Spring, Node/TS, and Python/FastAPI repos.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados