convez.ai
AgenteAutomaçãoIntermediário
DavvidC

AI Dev Team CC

DavvidC/ai-dev-team-cc

Equipe de IA para dev completo: spec → código → review → testes

Ver no GitHub
InicialPluginScore · de 0 a 100

Loop de desenvolvimento com quatro agentes (PO, Dev, Reviewer, QA) orquestrado pelo Claude Code CLI, sem necessidade de API key. Automatiza o ciclo completo de especificação, implementação, revisão e testes usando sua assinatura Claude Code existente.

Instalação

$npx convez-ai DavvidC/ai-dev-team-cc

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
28%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automatizar ciclos de desenvolvimento end-to-end sem intervenção manual
  • Validar implementações com testes e linters antes da revisão
  • Iterar rapidamente em features com feedback de múltiplos agentes
  • Manter histórico limpo com branching git automático

Veredito

Solução bem arquitetada para orquestração de agentes em fluxos de dev real, aproveitando CLI do Claude sem custos extras. Recomendado para times que querem automação de qualidade com segurança (sandboxing git + branches isoladas).

README

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

ai-dev-team-cc

A four-agent AI development loop powered by the Claude Code CLI — no Anthropic API key needed. Uses your existing Claude Code subscription.

Product Owner → Developer → [Validator] → Reviewer ┐
                                              QA    ├─ parallel
                            └─────────── PO judges ┘
└──────────────── repeats until approved or MAX_ITERS ──────┘

How it works

  1. Product Owner — inspects the codebase, writes a spec with testable acceptance criteria
  2. You — review the spec and choose y to proceed, n to abort, or edit to revise it
  3. Developer — reads files, implements the spec, commits to a feature branch
  4. Validator (optional) — runs your linter/tests as a hard gate before review
  5. Reviewer + QA — run in parallel; Reviewer reads the diff, QA starts the app and tests every criterion
  6. PO — reads the actual changed files and QA results, returns a JSON verdict
  7. If approved → delivered to you with the branch ready to merge. If not → feedback is compressed and fed back (up to MAX_ITERS times, then escalates to you)

Requirements

  • Claude Code CLI installed and logged in
  • Python 3.8+
  • Project should be a git repo (for branching + rollback)
claude --version   # verify it works

Usage

export APP_DIR=/path/to/your/project
python3 agent_team.py "Add a /health endpoint that returns uptime"

# Let the PO choose what to work on
python3 agent_team.py

# With optional overrides
export START_CMD="npm run dev"       # how QA starts the app
export VALIDATE_CMD="npm run lint"   # hard gate after Dev
python3 agent_team.py "Add pagination"

Configuration

VariableDefaultDescription
APP_DIRscript directoryRoot of the project
START_CMD(auto-detected)Command to start the app for QA
VALIDATE_CMD(none)Linter/test command run after Dev as a hard gate
MAX_ITERS3Max cycles before escalating to you
CLAUDE_BINclaudePath to the Claude Code CLI binary

What happens with git

  • A branch ai-dev/task-{timestamp} is created before any work starts
  • The Dev agent commits directly to that branch
  • On approval you get instructions to review and merge
  • On failure the branch survives so you can inspect or delete it
# Review what the agents did
git diff main..ai-dev/task-1234567890

# Merge if happy
git merge ai-dev/task-1234567890

# Discard if not
git branch -D ai-dev/task-1234567890

Works with any project type

The PO inspects the project to understand what it is and how to run it. No config needed for common stacks.

# Python/FastAPI
export APP_DIR=~/my-api && python3 agent_team.py "Add rate limiting"

# Node/Express
export APP_DIR=~/my-app && export START_CMD="npm start" && python3 agent_team.py "Add dark mode"

# Anything else
export APP_DIR=~/any-project && python3 agent_team.py "your task"

Safety

  • Agents run scoped to APP_DIR only
  • --dangerously-skip-permissions is used so the loop runs unattended — point APP_DIR at a branch copy, not production
  • Git branching means main is never touched directly
  • MAX_ITERS caps runaway loops

vs. industry tools

ThisDevin / SWE-agentCrewAI
CostExisting CC planSeparate billingAPI billing
Git branching✅✅❌
Parallel agents✅❌❌
Validation gate✅✅❌
Human checkpoints✅✅❌
Browser access❌✅❌
Ver o original no GitHub

Bom para

Relacionados