convez.ai
SkillDevOps & InfraestruturaAvançado
gzb1128

Skill Forge

gzb1128/skill-forge

Plugins e skills para agentes Claude gerenciarem repos como runtime

Ver no GitHub
InicialPluginScore · de 0 a 100

Marketplace de plugins Claude Code para forjar ambientes de runtime de agentes. Agrupa skills reutilizáveis em documentação agent-first, workflows de qualidade de código, commit disciplinado e customização OpenCode, permitindo que agentes leiam, julguem e executem de forma confiável.

Instalação

$npx convez-ai gzb1128/skill-forge

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
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Scaffolding de documentação agent-first e manutenção de conhecimento de equipe
  • Code review automatizado, gates de commit e loops autônomos de fix
  • Customização de configuração OpenCode e hydration de metadados de modelos

Veredito

Skill Forge endereça um gap real: estruturar repositórios como runtime legível para agentes. A abordagem de versioning automático por commit SHA e o catálogo de skills bem-definido são sólidos, mas a documentação em Makefile e a curva de adoção são íngremes para equipes em estágio inicial de automação.

README

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

skill-forge

License: MIT

skill-forge is a Claude Code plugin marketplace for forging agent runtime environments. It distills reusable skills into plugins covering repository knowledge, code quality, commit discipline, autonomous fix loops, and OpenCode configuration.

Core idea: Human at the helm. Agents execute. The repo is the agent's runtime — knowledge, rules, and workflows must be shaped into forms agents can reliably read, judge, and execute.

Keywords: claude-plugin, claude-code, agent-skills, agent-harness, openai-harness-engineering, opencode, code-review, clean-commit

Quick Start

# 1. Add the Skill Forge marketplace
claude plugin marketplace add gzb1128/skill-forge

# 2. Install the plugins you need
claude plugin install agent-docs@skill-forge
claude plugin install code-quality@skill-forge
claude plugin install opencode-customize@skill-forge

# 3. In your target repo, ask Claude:
#    "bootstrap agent docs"       -> scaffold Agent-First docs
#    "review my changes"          -> quality review on local diff
#    "commit this"                -> gated commit with impact message
#    "loopfix"                    -> autonomous review-fix loop
#    "hydrate model config"       -> fill OpenCode custom model parameters

Plugin versions are resolved to git commit SHA. Every push produces a new installable version — no manual semver maintenance. Run claude plugin update <plugin>@skill-forge (or wait for auto-update) to pull the latest.

Plugins

PluginPurposeSkills
agent-docsScaffold Agent-First documentation and maintain non-derivable team knowledgebootstrap-agent-docs, learn, remember
code-qualityTurn code review, commit gates, diff cleanup, and fix loops into repeatable agent workflowsquality-reviewer, clean-commit, diff-cleanup, loopfix
opencode-customizeCustomize OpenCode configuration, especially hydrating custom provider model metadatahydrate-opencode-models

Skill Catalog

agent-docs

SkillTypePurpose
bootstrap-agent-docsmodel-invokedGenerate AGENTS.md, docs/_templates/, docs/rules/, and per-category INDEX.md in a target repo
learnmanual skill (/agent-docs:learn)Persist non-obvious session insights to the right AGENTS.md — verified and approval-gated
remembermanual skill (/agent-docs:remember)Audit AGENTS.md knowledge for staleness, duplication, and misplacement

The template payload used by bootstrap-agent-docs lives at plugins/agent-docs/templates/ and resolves at runtime via ${CLAUDE_PLUGIN_ROOT}/templates/. No separate repo clone needed.

code-quality

SkillTypePurpose
quality-reviewermodel-invokedStructured quality pass on local changes: three-pass review, diff hygiene, lint, tests, caller check
clean-commitmodel-invokedRun quality gates (via quality-reviewer) before committing, with messages that explain WHY
diff-cleanupmodel-invokedRemove AI-generated bloat (slop comments, dead code, defensive noise, redundant logic) from a feature branch diff
loopfixmodel-invokedAutonomous review-fix loop: reviewer subagent finds issues, main agent triages and fixes, repeat until clean

opencode-customize

SkillTypePurpose
hydrate-opencode-modelsmodel-invokedLook up model metadata from the Models.dev catalog and map it to OpenCode custom provider model config

What agent-docs Scaffolds

When you ask Claude to "bootstrap agent docs" in a target repo, the plugin creates:

your-repo/
├── AGENTS.md                          # root entry for agents (~100-line navigation)
└── docs/
    ├── codemaps/INDEX.md              # architecture maps: concept -> file path
    ├── design/INDEX.md                # design specs: YYYY-MM-DD-<topic>-design.md
    ├── plans/INDEX.md                 # implementation plans: YYYY-MM-DD-<feature>.md
    ├── rules/                         # engineering standards
    │   ├── INDEX.md
    │   ├── non-derivability.md        # non-derivability principle
    │   ├── document-conventions.md
    │   └── openai-harness-engineering.md
    ├── troubleshoot/INDEX.md          # symptom-indexed troubleshooting
    ├── runbooks/INDEX.md              # deterministic operational procedures
    ├── lib/INDEX.md                   # third-party library usage notes
    ├── verify/INDEX.md                # dry-run verification flows
    └── _templates/                    # templates for new docs (copy & customize)
        ├── codemap.md
        ├── design.md
        ├── plan.md
        └── subpackage-AGENTS.md

Practices

PracticeMeaning
Repo as record systemKnowledge agents can't see doesn't exist. Critical constraints must not live only in chat logs or external docs.
Progressive disclosureAGENTS.md provides the entry navigation, docs/codemaps/*.md points to components, source code carries the details.
INDEX per categoryEvery docs/*/ directory has an INDEX.md with a "When to Use" column so agents can triage at a glance.
Non-derivability filterRecord only what cannot be derived from source code, git history, or existing docs.
Maps, not encyclopediasCodemaps maintain concept-to-path tables only — they link to source, never copy code.
Date-prefixed designs/plansYYYY-MM-DD-<topic>-design.md keeps designs and plans browsable in chronological order.

Full rationale: openai-harness-engineering.md.

Development

This repo is both the marketplace catalog and the plugin source. See AGENTS.md for the development workflow, local verification commands, and SHA-based versioning policy.

Common commands:

make validate
make test-skills-link
make test-skills-status
make test-skills-unlink

License

This project is licensed under the MIT License.

Ver o original no GitHub

Bom para

Relacionados