convez.ai
SkillAutomaçãoAvançado
joehz2007

Agent Foundry

joehz2007/Agent-Foundry

Author once, publish to many: factory para AI coding assets multiplataforma

Ver no GitHub
InicialPluginScore · de 0 a 100

Factory para autoria de assets de IA (skills, agents, commands) uma única vez e publicação em múltiplos runtimes (Claude Code, Codex, outros). Transforma prompts em código modular com adapters que compilam para diferentes plataformas.

Instalação

$npx convez-ai joehz2007/Agent-Foundry

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 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Criar skills reutilizáveis para múltiplos agent tools sem duplicação
  • Padronizar estrutura de agents, skills e commands em um projeto
  • Compilar assets neutros para Claude Code, Codex e outros runtimes
  • Implementar code review automatizado específico por stack (Spring Boot, Vue, etc)

Veredito

Projeto bem estruturado para equipes que precisam serializar e reutilizar AI coding assets em múltiplas plataformas. A abstração 'prompt como source code' é elegante, mas requer overhead arquitetural — recomendado para casos com múltiplos runtimes alvo.

README

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

Agent Foundry

A small factory for authoring AI coding assets once and publishing them to multiple agent tools: Claude Code skills/agents/commands, Codex-style skills, and other runtimes.

Core idea:

Prompt is source code. Skill is a module. Command is an entrypoint. Eval is a test. Adapter is a compiler.

Project layout

agent-foundry/
├── registry.yaml              # Asset index
├── artifacts/                 # Neutral source format
│   ├── agents/
│   ├── skills/
│   └── commands/
├── schemas/                   # JSON schema for source format
├── scripts/                   # Build and validation tools
├── evals/                     # Test prompts / expected behavior
├── docs/                      # Authoring rules and platform notes
├── templates/                 # Optional platform templates
└── dist/                      # Generated files, ignored by git if desired

Quick start

Install dependencies:

python3 -m pip install -r requirements.txt

Validate all artifacts:

python3 scripts/validate.py

Build all targets:

python3 scripts/build.py --target all

Build a single target:

python3 scripts/build.py --target claude-code
python3 scripts/build.py --target codex
python3 scripts/build.py --target pi

Generated output appears under dist/<target>/.

Asset types

  • agent: a role-based executor with boundaries, responsibilities, and output contract.
  • skill: a domain/workflow capability with methodology, references, and optional scripts.
  • command: a short user-facing entrypoint for a repeatable task.

Included richer skill examples:

  • code-review-spring-boot: technical implementation review for Java/Kotlin Spring Boot backends.
  • code-review-vue: technical implementation review for Vue/TypeScript frontends.
  • zentao-execution-rd-audit: audits ZenTao execution plans into RD baseline reports and JSON.
  • zentao-execution-code-review: orchestrates PR-based task-alignment review from the RD baseline and delegates technical review to stack-specific skills.

See docs/writing-artifacts.md for the source format.

Ver o original no GitHub

Bom para

Relacionados