convez.ai
SkillDesenvolvimento WebAvançado
srwbsw

Second Opinion Skill

srwbsw/second-opinion-skill

Revisões de código multi-engine paralelas com 50+ modelos

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que roteia revisões de código para múltiplos modelos de IA (Gemini, OpenCode, Codex, Claude, Copilot, Qwen, Kilo, Antigravity) em paralelo. Oferece perspectivas independentes de 50+ modelos com execução em sandbox read-only, ideal para análise de segurança, testes e arquitetura.

Instalação

$npx convez-ai srwbsw/second-opinion-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 2 semanas
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Revisão de segurança e busca de vulnerabilidades (injeção, autenticação, validação)
  • Análise de cobertura de testes e regressões
  • Critique de decisões arquiteturais com múltiplas perspectivas
  • Consulta técnica estruturada com sub-agents especializados

Veredito

Skill robusta e bem arquitetada para elevar a qualidade de revisões de código além de um único modelo. Excelente para times que precisam de perspectivas independentes em questões críticas de segurança e design, embora exija configuração de múltiplas CLIs externas.

README

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

second-opinion-skill

A universal multi-engine plugin that routes code reviews to the AI engine of your choice — Gemini, opencode, Codex, Claude Code, Copilot, Qwen, Kilo, Antigravity.

No single model catches everything. This plugin makes cross-engine review a first-class part of your workflow: pick an engine, optionally pick a model, get an independent perspective in seconds.

This repo also ships plugin manifests for Claude Code and Codex, so the same bundle can be recognized across both ecosystems.

Engines

EngineModel selectionRead-only flags
Gemini CLIAutomatic-s --approval-mode plan
opencodeProvider → model (from registry)--agent plan
Codex CLIOptional (type-in)-s read-only
Claude CodeOptional (type-in)--print --permission-mode plan
GitHub Copilot CLIOptional (type-in)--plan --deny-tool=write --allow-all-tools
Qwen Code CLIOptional (type-in)-s --approval-mode plan
KiloProvider → model (free shown first)--agent plan
Antigravity (agy)Optional (agy models, or default)--sandbox

All engines launch from the repo directory (--cwd) and read content via native filesystem tools — no stdin piping.

Skills

SkillTrigger phrases
second-opinion"second opinion", "independent review", "cross-model review"
gemini-review"ask Gemini", "review with Gemini", "Gemini's take"
opencode-review"use opencode", "ask opencode", "review with opencode"
codex-review"ask Codex", "review with Codex", "Codex's take"
copilot-review"ask Copilot", "review with Copilot", "Copilot review"
qwen-review"ask Qwen", "review with Qwen", "Qwen's take"
kilo-review"ask Kilo", "review with Kilo", "Kilo's take"
agy-review"ask Antigravity", "agy review", "review with agy"

Use cases

  • Code review — senior-engineer-style critique with parallel sub-agent coverage (security, test coverage, regression, design)
  • Second opinion — independent take on an architectural or design decision
  • Security review — scan for injection, auth flaws, data exposure, input validation
  • General consultation — any technical question with a structured answer

Requirements

You only need the CLIs for the engines you actually use.

Versioning

Release versions are managed by GitHub Actions. The workflow updates package.json, the Claude marketplace metadata, and the Codex plugin manifest together, so you do not need to edit version numbers by hand.

Permissions

All engine execution goes through bin/review.js, and all provider/model discovery goes through bin/list.js — grant permissions once instead of per-engine or per-CLI:

{
  "permissions": {
    "allow": [
      "Bash(~/.claude/plugins/cache/second-opinion-skill/second-opinion-skill/*/bin/review.js*)",
      "Bash(node ~/.claude/plugins/cache/second-opinion-skill/second-opinion-skill/*/bin/list.js*)"
    ]
  }
}

Installation

Claude Code

If you are using Claude Code, install the plugin through the marketplace:

/plugin marketplace add srwbsw/second-opinion-skill
/plugin install second-opinion-skill@second-opinion-skill

Codex CLI

If you are using Codex CLI, the easiest path is the helper script in this repo. It reads your Git origin URL and registers the marketplace from GitHub:

./scripts/install-codex-plugin.sh

If you prefer the manual flow, add the GitHub repo as a marketplace and then install the plugin from that marketplace:

codex plugin marketplace add https://github.com/srwbsw/second-opinion-skill
codex plugin add second-opinion-skill@second-opinion-skill

Usage

Engine-agnostic (orchestrator picks):

"Get a second opinion on this" "Independent review of these changes" "Cross-model review"

Engine-specific:

"Ask Gemini to review this" "Use opencode to review this diff" "Codex review with o3" "Claude Code review with sonnet" "Ask Copilot's take on this approach" "Qwen security review" "Kilo review with a free model"

When no engine is specified, the second-opinion skill asks you to pick one. Each engine-specific skill handles its own model flow: Gemini runs immediately; opencode and Kilo walk through provider → model selection; Codex, Claude Code, Copilot, and Qwen optionally let you type in a model; Antigravity optionally lets you pick from agy models (or uses its default).

Results are structured: Summary, Issues (HIGH/MED/LOW tagged by domain), Concerns, Positives. The code review prompt instructs engines to spawn parallel sub-agents per domain where supported.

If you need engine-specific flags, pass them through the runner instead of editing the scripts:

bin/review.js --engine=claude --cwd=. "Review this" --engine-arg=--verbose
bin/review.js --engine=codex --cwd=. "Review this" -- --permission-mode bypassPermissions
bin/list.js --engine=opencode models --provider=opencode --cli-arg=--refresh

Why cross-engine review?

Different models are trained on different data with different architectures. Gemini flags different categories of issues than Claude Code. opencode and Kilo give you access to GPT, Llama, Mistral, Qwen, and dozens of others. Running your changes through a differently-trained model before declaring done is the same instinct as a second engineer reading your PR — except it takes seconds.

License

MIT

Ver o original no GitHub

Bom para

Relacionados