ACE — Equipe de Engenharia Colaborativa com Agentes
dwang2/ACE
Equipe de dev multi-especialista orquestrada via agentes Claude + GitHub
Ver no GitHubSistema de subagentes especializados (PM, arquiteto, desenvolvedor, UX, DevOps) que colaboram nativamente com Claude Code para entregar software via GitHub. Cada agente tem seu papel definido e acessa uma base de conhecimento compartilhada, com disciplina de tokens integrada.
Instalação
npx convez-ai dwang2/ACE1 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
Casos de uso
- Planejamento e design de arquitetura de software com ADRs
- Implementação de features com testes e code review automatizados
- Configuração de CI/CD e pipelines de release
- Documentação técnica e API reference sincronizadas com código
- Triagem de issues e gestão de saúde da comunidade
Veredito

ACE é uma abordagem sólida para orquestração de agentes em projetos complexos, com ênfase em economia de tokens e convenções claras. Ideal para equipes que querem automatizar fluxo de engenharia, mas exige curva de aprendizado e disciplina de processo.
README
A tradução para o português deste README ainda está na fila.
ACE — Agentic Collaborative Engineering
A Claude Code-native development team of specialized subagents that collaborate on software delivery through GitHub. Designed to be Pro-survivable — token discipline is built into every agent and KB.
ACE = Agentic Collaborative Engineering. The agents are the specialists; GitHub is the shared workspace; you are the orchestrator who decides what happens when.
Read first
kb/USAGE.md explains how to use ACE without burning your Pro budget. Request tags ([lite], [design], [hotfix]), when to /compact, when to use Sonnet vs Opus, and rough budget math.
Team Composition
| Agent | Role | Primary Outputs |
|---|---|---|
| maintainer | Community health, issue triage, contributor experience | Issues, CONTRIBUTING.md, ROADMAP.md, SECURITY.md |
| architect | System design, technical decisions, public API stability | ADRs, design docs, API contracts, threat models |
| developer | Implementation, testing, code review | Pull requests, unit/integration tests |
| tech-writer | Documentation, API reference, changelogs | README, guides, CHANGELOG, release notes, docstrings |
| devops | CI/CD, release automation, dependency hygiene | GitHub Actions workflows, release pipeline, Dependabot config |
How It Works
- You drive the conversation. You ask Claude Code to do something ("plan a new feature", "implement issue #42", "set up CI").
- Claude Code delegates to the right agent. Agents are invoked automatically based on the
descriptionfield in their frontmatter, or you can explicitly request one (use the architect to review this design). - Agents read the shared KB. Every agent loads the relevant files from
kb/to stay aligned on conventions, tech stack, and process. - Agents use shared skills. Cross-cutting capabilities (GitHub Issues, ADRs, PR workflow) live in
.claude/skills/so every agent invokes them the same way. - GitHub is the source of truth. Issues track work, PRs track changes, Projects track flow, Actions run CI/CD.
Directory Layout
.
├── README.md # This file
├── kb/ # Shared knowledge base (read by all agents)
│ ├── USAGE.md # Token discipline — read first
│ ├── 00-team-charter.md # Mission, ways of working
│ ├── 01-tech-stack.md # Languages, frameworks, services
│ ├── 02-github-conventions.md # Branches, labels, PR/issue templates
│ ├── 03-definition-of-ready-done.md # Quality gates
│ ├── 04-coding-standards.md # Style, testing, review
│ ├── 05-security-and-compliance.md # Threat modeling, secrets, SAST
│ ├── 06-glossary.md # Shared vocabulary
│ ├── languages/ # Per-language deep references
│ │ ├── README.md # Index + when to read each
│ │ ├── python.md # Idioms, testing, async, layout
│ │ ├── go.md # Idioms, testing, concurrency, layout
│ │ ├── typescript.md # Idioms, testing, async, layout
│ │ └── rust.md # Idioms, testing, async, layout
│ └── principles/ # Cross-cutting design principles
│ ├── README.md # Index + when to read each
│ ├── solid.md # SRP, OCP, LSP, ISP, DIP
│ ├── domain-driven-design.md # Ubiquitous language, contexts, aggregates
│ ├── clean-architecture.md # Layering, dependency rule
│ ├── twelve-factor.md # 12-factor app methodology
│ ├── api-design.md # HTTP/REST, gRPC, async APIs
│ └── distributed-systems.md # Idempotency, retries, sagas, consistency
├── .claude/
│ ├── agents/ # Subagent definitions
│ │ ├── maintainer.md
│ │ ├── architect.md
│ │ ├── developer.md
│ │ ├── tech-writer.md
│ │ └── devops.md
│ └── skills/ # Cross-cutting skills
│ ├── github-issues/SKILL.md
│ ├── adr-authoring/SKILL.md
│ ├── pr-workflow/SKILL.md
│ ├── ci-pipeline/SKILL.md
│ └── release-engineering/SKILL.md
└── .github/
└── workflows/
└── ci.yml # Reference CI workflow
Install
# Zero-install — always pulls the latest version
npx ace-agents@latest init
# Or install globally once
npm install -g ace-agents
ace init
Run inside any project directory. ACE copies .claude/, kb/, and .github/workflows/ci.yml into the current directory and prints what it did.
ace init --no-ci # skip the CI workflow
ace init --force # overwrite existing files
ace update # re-sync after upgrading ace-agents; skips locally modified files
ace update --force # overwrite everything including locally modified files
ace --version
Clone the repo and copy the three directories into your project root:
git clone --depth 1 https://github.com/dwang2/ACE /tmp/ace
cp -r /tmp/ace/ace/.claude /tmp/ace/ace/kb /tmp/ace/ace/.github YOUR_PROJECT/
Quick Start
- Run
npx ace-agents@latest initin your project root. - Open Claude Code in that repo.
- Tell it what you want. Examples:
Maintainer, triage the three open issues and label any good first issues.Architect, propose an approach for idempotent retry handling and note the semver implications.Developer, implement issue #42 and open a PR.Tech writer, update the README and write a getting-started guide for the new CLI.DevOps, add a multi-OS CI matrix and a release workflow that publishes to PyPI.
Customizing for Your Project
- Update
kb/01-tech-stack.mdwith your actual languages, frameworks, and services. - Update
kb/02-github-conventions.mdwith your org's branch naming, labels, and templates. - Tune agent prompts in
.claude/agents/*.mdif your team has specific role boundaries. - Add domain skills under
.claude/skills/for anything specific (e.g., adatabase-migrationskill).