convez.ai
SkillProdutividadeAvançado
x1han

Light RIP

x1han/light-rip

Loop Review-Implement-Plan adaptativo com subagentes para coding workflows

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill leve para fluxos de codificação com Claude Code e Codex que implementa um loop de Review-Implement-Plan adaptativo. Oferece três tiers (Tiny, Medium, Large) de orquestração com subagentes para revisão, planejamento e implementação de código, escalando automaticamente conforme a complexidade e risco da tarefa.

Instalação

$npx convez-ai x1han/light-rip

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
18
Forks
2
Acompanhando
1
Saúde
28%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Implementação de features com verificação automática de qualidade
  • Refatoração de código complexo com múltiplas camadas de revisão
  • Projetos de risco elevado que exigem planejamento e review especializados

Veredito

Skill bem arquitetada para intensificar workflows de codificação em Claude Code e Codex, com escalabilidade inteligente baseada em risco. Requer instalação manual em diretório de skills e configuração de hooks, adequada para equipes que precisam de governança de código automatizada sem overhead processual.

README

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

🪦 Light RIP

Light RIP is an agent skill package, not a software application.

RIP names the three safeguards in the skill: Review, Implement, and Plan. In execution, the lightweight loop runs as Plan -> Implement -> Review; the name keeps the three pieces memorable without turning them into a heavyweight process.

Light RIP has exactly three tiers:

  • Tiny: the main session plans, implements, verifies, and self-reviews. No subagents.
  • Medium: the main session plans, implements, and verifies, then a reviewer subagent reviews the diff. The main session fixes review findings.
  • Large: planner, implementer, and reviewer are all subagents. The main session coordinates and fixes review findings.
flowchart LR
    A["User prompt"] --> B{"Coding task size?"}
    B -->|"Tiny"| C["Main session: plan + implement + verify + self-review"]
    B -->|"Medium"| D["Main session: plan + implement + verify"]
    D --> E["Reviewer subagent"]
    B -->|"Large"| F["Planner subagent"]
    F --> G["Implementer subagent"]
    G --> H["Reviewer subagent"]
    E --> I["Main session fixes findings"]
    H --> I
    C --> J["Done"]
    I --> K["Final verify"]
    K --> J

Risk upgrades the tier: risky tiny becomes medium, risky medium becomes large, and risky large stays large while reviewers focus on the risk area.

Review findings always return to the main session. For nontrivial P1 fixes, do one short reviewer subagent re-review of the changed lines.

flowchart LR
    A["Subagent review"] --> B{"P0/P1 findings?"}
    B -->|"Yes"| C["Main session fixes"]
    C --> D["Relevant verification"]
    D --> E{"Nontrivial P1 fix?"}
    E -->|"Yes"| A
    E -->|"No"| F["Complete"]
    B -->|"No"| F

Installation

Do not install it as an app, service, Python package, or normal project checkout. Install it by placing the light-rip folder in your agent's skills directory, then mount the required UserPromptSubmit reminder hook.

Installation is incomplete until the hook is mounted.

Claude Code

Claude Code documentation and community examples use $HOME/.claude/skills or ~/.claude/skills. Claude Code does not define a standard $CLAUDE_HOME variable.

Clone directly into the Claude Code skills directory, then run the required hook setup from that installed copy:

mkdir -p "$HOME/.claude/skills"
git clone https://github.com/x1han/light-rip "$HOME/.claude/skills/light-rip"
cd "$HOME/.claude/skills/light-rip"
python hooks/install_claude_hook.py

This writes or updates:

  • ~/.claude/settings.json

Restart Claude Code after installing or updating the skill.

Codex

If CODEX_HOME is unset, Codex normally uses $HOME/.codex.

Clone directly into the Codex skills directory, then run the required hook setup from that installed copy:

CODEX_HOME="${CODEX_HOME:-$HOME/.codex}"
mkdir -p "$CODEX_HOME/skills"
git clone https://github.com/x1han/light-rip "$CODEX_HOME/skills/light-rip"
cd "$CODEX_HOME/skills/light-rip"
python hooks/install_codex_hook.py

This writes or updates:

  • $CODEX_HOME/hooks.json
  • $CODEX_HOME/config.toml, ensuring [features] hooks = true

Restart Codex after installing or updating the skill.

Hook Behavior

The required UserPromptSubmit hook is non-blocking. For every prompt, it injects reminder.md as additional context so the agent remembers:

  • factual claims need current evidence
  • observations, inferences, and recommendations should stay separate
  • tiny code edits do not need Light RIP
  • medium and large coding tasks should use Light RIP
  • risky tasks should use the stronger review path

Files

  • SKILL.md: the skill instructions
  • reminder.md: the context injected by the hook
  • hooks/light_rip_reminder.py: the shared hook command
  • hooks/install_codex_hook.py: Codex hook setup
  • hooks/install_claude_hook.py: Claude Code hook setup
Ver o original no GitHub

Bom para

Relacionados