convez.ai
SkillAutomaçãoAvançado
ktaletsk

Council - Revisão de Código com Múltiplos Modelos IA

ktaletsk/council

Revisões de código paralelas com múltiplos modelos IA em um relatório único

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que executa revisões de código em paralelo contra múltiplos modelos de IA (Claude, GPT, Gemini) simultaneamente via CLIs especializadas, sintetizando os achados em um relatório único e deduplicated. Aproveita perspectivas diferentes de cada modelo para detectar bugs que uma única IA deixaria passar.

Instalação

$npx convez-ai ktaletsk/council

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
28%
Último commit
há 5 meses
Versão
—
Issues abertas
1
Contribuidores
1

Casos de uso

  • Revisão crítica de código com cobertura máxima de bugs
  • Detecção cross-validada de problemas de performance e segurança
  • Aproveitar múltiplas assinaturas IA já contratadas (ChatGPT, Cursor, Anthropic)

Veredito

Ferramenta sólida para times que querem máxima cobertura em revisões críticas, aceitando a troca entre velocidade e qualidade. Requer CLIs múltiplas já instaladas; não é plug-and-play, mas reutiliza créditos que você já paga.

README

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

council

Ensemble code reviews. Run the same review prompt against multiple AI agents in parallel, then synthesize their findings into one comprehensive report — because different models catch different bugs.

AI Agent Skill License: MIT

Why This Exists

No single AI model catches everything. GPT might spot a race condition that Opus misses, while Gemini flags a performance issue neither noticed. By running the same critical review prompt against multiple agents and combining their findings, you get more thorough coverage than any single model provides.

Single Model Reviewcouncil
One perspectiveYesMultiple perspectives
Model-specific blind spotsPossibleCross-validated findings
FastYesParallel but slower
SimpleYesNeeds at least one supported CLI

How It Works

  1. Parallel Execution — Spawns multiple agent processes simultaneously, each via its configured backend
  2. Independent Reviews — Each agent reviews staged git changes in read-only mode
  3. Synthesis — The host agent combines outputs into a single deduplicated report

Supported Backends

BackendCLI ToolModel FormatRequirement
Claude Codeclaude -pAlias (sonnet, opus) or full nameClaude Code + Anthropic account (claude -p draws from a separate Agent SDK credit, not your interactive Pro/Max limit — see note below)
Codexcodex execModel name (e.g. gpt-5.3-codex)Codex CLI (npm i -g @openai/codex) + ChatGPT plan or API key
OpenCodeopencode runprovider/model (e.g. google/gemini-3.1-pro)OpenCode CLI + provider API keys
Cursorcursor-agentModel name (e.g. gemini-3.1-pro)Cursor CLI + subscription

Backends can be mixed freely — run Claude through Claude Code, GPT through Codex, and Gemini through OpenCode all in the same review session.

Use the subscriptions and credits you already have

council drives whichever agent CLIs are already installed and authenticated on your machine, so most reviewers bill against access you're already paying for:

  • Codex uses your ChatGPT Plus/Pro/Team plan (or an OpenAI API key).
  • Cursor runs through your active Cursor subscription — the models in cursor-agent models are included in your plan.
  • OpenCode uses whatever providers you've configured, including its own free models (e.g. opencode/big-pickle) that cost nothing at all.

Heads up on the claude-code backend. council invokes Claude Code in non-interactive mode (claude -p), and that is not simply covered by your normal Pro/Max usage the way interactive Claude Code is. Per Anthropic, as of June 15, 2026 claude -p (and the Agent SDK) stop drawing from your interactive subscription limits and instead draw from a separate, capped monthly Agent SDK credit you have to opt into once — $20 on Pro, $100 on Max 5x, $200 on Max 20x. When that credit runs out, requests fall through to pay-as-you-go API billing (if you've enabled usage credits) or stop until the credit refreshes. In other words, you cannot easily run a large council on the claude-code backend off a Pro/Max plan — budget for the Agent SDK credit or an API key, or prefer the other backends. See Use the Claude Agent SDK with your Claude plan. (You can still get Claude models without this caveat by running them through the Cursor or OpenCode backends instead.)

Because backends mix freely, you can assemble a multi-model council almost entirely from plans you already subscribe to — for example, Opus via Cursor, GPT via your ChatGPT plan through Codex, and a couple of free OpenCode models — without provisioning a single new API key. If a CLI isn't installed or signed in, council simply skips that reviewer; just configure the ones you have.

Installation

The default path is the Skills CLI (npx skills): it pulls this repo from GitHub and installs into the right agent skill directories for Cursor and many other AI coding agents. You need Node.js so npx is available.

Recommended: npx skills

All projects (global / user-level):

npx skills add ktaletsk/council -g

This repository only (from inside the project):

npx skills add ktaletsk/council

Use -a <agent> to target specific agents (e.g. -a cursor), or run without it and follow the prompts. To update later: npx skills update (add -g for global installs). More skills: skills.sh.

Manual: git clone

If you prefer to place the skill yourself:

Personal skill (all projects)

mkdir -p ~/.claude/skills
git clone https://github.com/ktaletsk/council ~/.claude/skills/council

Project skill (specific project)

mkdir -p .claude/skills
git clone https://github.com/ktaletsk/council .claude/skills/council

Requirements

One or more of the following:

Plus:

  • A host agent (Claude Code, OpenCode, etc.) for synthesis

Configuration

All settings are in config.yaml at the skill root. Each agent entry specifies its own backend and model, so you can mix and match freely:

agents:
  - backend: claude-code
    model: sonnet

  - backend: codex
    model: gpt-5.3-codex

  - backend: opencode
    model: google/gemini-3.1-pro

All Claude Code

agents:
  - backend: claude-code
    model: opus
  - backend: claude-code
    model: sonnet
  - backend: claude-code
    model: haiku

All Codex

agents:
  - backend: codex
    model: gpt-5.5
  - backend: codex
    model: gpt-5.3-codex

All OpenCode

agents:
  - backend: opencode
    model: amazon-bedrock/us.anthropic.claude-opus-4-8
  - backend: opencode
    model: amazon-bedrock/zai.glm-5
  - backend: opencode
    model: opencode/big-pickle   # free

All Cursor

agents:
  - backend: cursor
    model: claude-opus-4-8-thinking-high
  - backend: cursor
    model: gpt-5.5-high
  - backend: cursor
    model: gemini-3.1-pro

Discover available models

# Claude Code (aliases: sonnet, opus, haiku, or full model names)
claude --model

# Codex
codex    # then /model inside the TUI

# OpenCode
opencode models

# Cursor
cursor-agent --list-models

Usage

Start a code review:

/council

Or trigger naturally:

Review my staged changes
Run a council review

Example Session

You: /council

Agent: I'll run parallel code reviews using multiple AI agents.

Running reviews on /Users/you/project...

  Agents:
    claude-code / sonnet
    codex       / gpt-5.3-codex
    opencode    / google/gemini-3.1-pro

  ⏳ Starting: sonnet (claude-code)
  ⏳ Starting: gpt-5.3-codex (codex)
  ⏳ Starting: google/gemini-3.1-pro (opencode)

Waiting for reviews to complete (this may take 1-3 minutes)...

  ✓ Completed: sonnet (claude-code)
  ✓ Completed: gpt-5.3-codex (codex)
  ✓ Completed: google/gemini-3.1-pro (opencode)

Now synthesizing results...

# Code Review Report

## Summary
The changes introduce timestamp handling improvements with proper
fallback logic. All 3 reviewers found issues worth addressing.

[continued...]

Output

Results are saved to your project's .reviews/ directory:

<your-project>/.reviews/
├── review_sonnet.json
├── review_gpt-5.3-codex.json
├── review_google__gemini-3.1-pro.json
└── COMBINED_REVIEW.md

Customization

Change Review Focus

Edit prompts/review-prompt.md to adjust:

  • What aspects to focus on (security, performance, etc.)
  • Output format
  • How critical the review should be

Thinking Depth

Add keywords to prompts/review-prompt.md:

  • think — basic reasoning
  • think hard — more thorough
  • think harder — very thorough
  • ultrathink — maximum depth (slower)

Files

council/
├── SKILL.md              # Agent Skills definition
├── README.md             # This file
├── config.yaml           # Agent backend + model configuration
├── scripts/
│   └── run-reviews.sh    # Parallel review runner
└── prompts/
    └── review-prompt.md  # Review prompt template

Compatibility

This skill is distributed through skills.sh and uses the open Agent Skills standard. It should work with:

  • Claude Code (~/.claude/skills/)
  • OpenCode (.opencode/skills/)
  • Codex (.codex/skills/)
  • Cursor (.cursor/skills/)
  • VS Code, GitHub Copilot, and other compatible agents

License

MIT

Ver o original no GitHub

Bom para

Relacionados