convez.ai
AgenteDesenvolvimento WebAvançado
anatomia-dev

Anatomia — Agent Harness para Claude Code

anatomia-dev/anatomia

Pipeline spec-driven multi-agente com análise automática de stack e auditoria de artefatos

Ver no GitHub
BomPluginScore · de 0 a 100

Orquestrador de agentes TypeScript que executa um pipeline de 5 estágios (scope, plan, build, verify, learn) para qualquer projeto, automatizando análise de código, detecção de stack e entrega auditável de artefatos com prova de cadeia. Escaneia projetos em 10 segundos e integra-se nativamente com Claude Code e Codex.

Instalação

$npx convez-ai anatomia-dev/anatomia

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
32
Forks
5
Acompanhando
0
Saúde
87%
Último commit
há 2 meses
Versão
v1.4.0
Issues abertas
46
Contribuidores
2

Casos de uso

  • Automatizar análise de codebase e detecção de padrões e convenções
  • Orquestrar múltiplos agentes Claude para desenvolvimento spec-driven
  • Auditar e rastrear mudanças de código com prova criptográfica de cadeia
  • Enriquecer contexto de projeto com setup assistido para agentes

Veredito

Ferramenta robusta e bem-arquitetada para equipes que já usam Claude Code e precisam de orquestração multi-agente com rastreabilidade. Pipeline de 5 estágios com scan automático de stack é diferenciador forte; requer Node 22+ e fluxo de trabalho estruturado. Ideal para projetos maiores com exigência de auditoria.

README

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

Anatomia

CI npm License: MIT

Anatomia is a CLI and agent harness for Claude Code and Codex. It scans your codebase — detecting your stack, conventions, and patterns — then runs every change through a five-agent pipeline that saves every artifact: scope, spec, contract, build report, and independent verification. Other harnesses are prompt libraries. This one has an engine.

Scan any project in 10 seconds

npx anatomia-cli scan .

No install. One command. See more examples →

Install

Install globally to use the ana command directly:

npm install -g anatomia-cli

Requires Node.js 22+. To update: npm update -g anatomia-cli

Quick start

ana init                      # generate context + agents
ana init commit               # persist to git (so teammates get it too)
ana doctor                    # verify installation is healthy
ana run setup                 # enrich with your team's knowledge (optional, recommended, ~10 min)
ana run                       # start working — "hey Ana"

Tell Ana what you want to build. It'll investigate the codebase, surface tradeoffs, and push back if the approach has problems. When the scope is right, it hands off to Plan, Build, and Verify.

init runs scan automatically and works standalone. The pipeline and setup run through Claude Code or Codex, and scan output works with any markdown-aware AI tool.

init commit commits to the artifact branch — staging, develop, or your pre-production branch if one exists, otherwise main. Check with ana config get artifactBranch before your first commit.

After updating the CLI, run ana init again to refresh scan data and skill detection. Your rules, context, and proof chain are preserved.

What it does

Scan + init

ana scan reads your project and detects framework, database, auth, testing, services, conventions, and patterns. In monorepos, scan identifies each surface (package or app) and detects per-surface commands. Re-running ana init refreshes scan data without overwriting your edits.

ana init writes that intelligence to files agents read:

  • ana.json — project config: build/test/lint commands, surfaces, artifact branch, branch prefix. Every agent and CLI command reads this.
  • scan.json — full structured scan data for agent consumption
  • CLAUDE.md and AGENTS.md — cross-tool project context
  • 5 core + 3 conditional skill templates with scan-driven Detected sections
  • 16 stack-specific gotchas with compound triggers

Setup (ana run setup) bridges the gap between what scan detects and what your team knows. A ~10 minute session that investigates your codebase, asks 2-3 questions, and writes enriched context. After setup, agents understand your product and decisions — not just your stack.

The pipeline

StageCommand               RoleProduces
Thinkana runThinking partner — scope, investigate, advise, push backscope.md
Planana run planArchitect — design + sealed contractspec.md + contract.yaml + plan.md
Buildana run buildBuilder — implement spec, prove it worksCode + tests + build_report.md
Verifyana run verifyFault-finder — reads spec and code, skips Build's reportverify_report.md
Learnana run learnProof analyst — runs between cyclesStronger skills and system improvements

Proof intelligence

Every pipeline run writes a proof chain entry — here's one:

Each entry adds to a proof chain. ana proof health tracks the trajectory across runs — first-pass verification rate, risks per run, hot spots where findings cluster, and what to fix next. When patterns recur, proof promote turns them into skill rules that reach the next build. proof audit groups active findings by file. proof stale flags findings whose files changed since discovery.

Commands

These are the commands you'll run directly:

ana scan [path]              # detect stack, conventions, patterns
ana init                     # generate context + agent definitions
ana init commit              # persist infrastructure to git
ana doctor                   # check project health
ana work status              # show pipeline state for active work
ana proof health             # quality trajectory dashboard
ana proof audit              # active findings grouped by file
ana proof <slug>             # display a proof chain entry

Pipeline agents invoke ~20 additional CLI commands — the toolbelt. When an agent saves an artifact, the CLI validates its structure, computes a content hash, and rejects malformed output. The agent can't skip a check. Full CLI reference →

Works with

Native pipeline support for Claude Code and Codex.

Scan output (AGENTS.md, CLAUDE.md) works with any AI tool that reads markdown.

Development

git clone https://github.com/anatomia-dev/anatomia.git
cd anatomia && pnpm install && pnpm build
cd packages/cli && pnpm vitest run

See CONTRIBUTING.md for extension guides and ARCHITECTURE.md for the module map.

This project is built with Anatomia. The .ana/ directory is the proof — every feature was scoped, planned, built, and verified through the same pipeline this tool installs for you.

Uninstalling

rm -rf .ana                        # all Anatomia data
rm .claude/agents/ana*.md          # Claude Code pipeline agents, if installed
rm -rf .codex/agents               # Codex agent manifests, if installed
rm AGENTS.md                       # generated project summary

Skill directories created by init live under .ana/skills/: coding-standards, testing-standards, git-workflow, deployment, troubleshooting, and conditionally ai-patterns, api-patterns, data-access. Remove the ones init created; keep any you added yourself.

If Anatomia created your CLAUDE.md, remove that too (git blame CLAUDE.md to check). If you ran ana init commit, revert: git revert <commit-hash>. Your source code is never modified.

npm uninstall -g anatomia-cli      # remove the CLI

License

MIT

Ver o original no GitHub

Bom para

Relacionados