convez.ai
SkillDevOps & InfraestruturaAvançado
bozbornehoo09

bozz-agents

bozbornehoo09/bozz-agents

Context e review skills para agentes de IA mantê-lo projeto coerente

Ver no GitHub
InicialPluginScore · de 0 a 100

Suite de skills para agentes de IA coding que gerenciam contexto e conhecimento de projetos. Inclui skills de escrita (update-*) para capturar decisões, arquitetura e documentação, e skills de revisão (docs-review, code-review, etc.) com validação anti-alucinação por citação direta. Ferramenta local compatível com Claude Code e Codex via padrão aberto Agent Skills.

Instalação

$npx convez-ai bozbornehoo09/bozz-agents

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
42%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
3

Casos de uso

  • Manter documentação de projeto sincronizada com código em evolução
  • Revisar código, planos e skills com findings validados (BLOCK/FIX/SUGGEST)
  • Capturar decisões arquiteturais e regras do projeto automaticamente
  • Stress-test conhecimento do projeto contra si mesmo

Veredito

Ferramenta sólida e bem-estruturada para manter metadados de projeto vivos em fluxos de IA coding. A anti-alucinação contract (citação + file:line) é um diferencial técnico legítimo. Requer instalação manual e customização de context-manifest, não é plug-and-play, mas compensa em projetos sérios.

README

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

bozz-agents

Context-management and review skills for AI coding agents.

Two families of Agent Skills that keep a project's knowledge — its decisions, architecture, rules, and docs — accurate and coherent as the code evolves, then stress-test that knowledge against itself:

  • update-* skills — the write side. A discover-and-dispatch orchestrator (update-context) plus eight narrow skills, one per authoritative content layer (decisions, strategy, architecture, rules, AI tooling, open questions, orientation, work tracker). Run at the end of a session to capture what changed in the right place.
  • review skills — the check side. A four-member review family (docs-review, skill-review, plan-review, code-review). Each fans out to 3–4 specialist briefs in parallel, reports findings as BLOCK/FIX/ SUGGEST, and returns a verdict of READY / READY-WITH-FIXES / REVISE. Every specialist runs under an anti-hallucination contract: each finding needs a verbatim quote and a file:line citation, or it is dropped.

It's a local developer tool. Nothing imports it; you install the skills into your agent (user-level) and they're available across every project you open.

Format & portability

Everything here is authored to the open Agent Skills standard (SKILL.md folders), released by Anthropic and adopted across 30+ tools. The same files work in Claude Code and Codex with no converter — what differs between tools is only the directory each one reads skills from. There is no dependency on any third-party skill-manager CLI; a sync tool is an optional convenience you can add, not a requirement.

Layout

bozz-agents/
├── .claude-plugin/
│   ├── plugin.json          # Claude Code plugin manifest
│   └── marketplace.json     # so the repo installs via /plugin marketplace add
├── skills/                  # canonical SKILL.md skills (flat, tool-agnostic)
│   ├── context-up/          #   session-start scoped context loader (read-only)
│   ├── update-context/      #   end-of-session orchestrator
│   ├── update-decisions/  …  update-work-tracker/   # 9 content-layer skills
│   ├── docs-review/         #   review skills; specialist briefs live in
│   ├── skill-review/        #   each skill's references/ subdirectory
│   ├── plan-review/         #   (skeletons — add briefs for your own
│   └── code-review/         #    architecture; a template brief is included)
├── config/
│   └── context-manifest.example.yaml   # the one file you customize
├── docs/                    # concepts.md, customizing.md, prism-case-study.md
├── codex/                   # Codex install/usage notes
└── AGENTS.md                # cross-tool orientation

Install

Claude Code (plugin + marketplace)

/plugin marketplace add bozbornehoo09/bozz-agents
/plugin install bozz-agents@bozz-agents

Skills then invoke as /bozz-agents:docs-review, /bozz-agents:update-context, and so on. Updates are a git pull away (or /plugin marketplace update).

For a quick local try without the marketplace, symlink the skills into your user skills directory:

ln -s "$PWD/skills"/* ~/.claude/skills/

Codex

Codex discovers skills under .agents/skills/. This repo's .agents/skills/ is generated by CI from skills/ (see below), so a clone already carries a Codex-ready tree — point Codex at it (repo, user, or admin scope). Details in codex/README.md.

How the skills are distributed

skills/ is the only hand-authored tree — spec-compliant SKILL.md folders. The per-tool directories are build artifacts: a GitHub Actions workflow (.github/workflows/generate-tool-skills.yml) runs scripts/generate-tool-skills.sh on every push that touches skills/**, mirrors it into .claude/skills/ and .agents/skills/, and commits them back. The loop is broken by a path filter — the generated commit only touches the generated dirs, which aren't in the trigger paths.

  • Claude gets the skills via the plugin/marketplace (above), or from the generated .claude/skills/ if you clone and open the repo directly.
  • Codex / Cursor / others read the generated .agents/skills/ (add more targets by editing TARGETS in the workflow).
  • Never edit .claude/skills/ or .agents/skills/ by hand — they're marked linguist-generated and will be overwritten. Edit skills/.

Reuse this in your own repo: copy scripts/generate-tool-skills.sh and the two workflows. PRs validate the canonical tree against the spec; pushes regenerate.

Customize for your project

The skills ship with sensible defaults (the conventional docs/decisions/, docs/architecture/, .claude/rules/, CLAUDE.md layout). To adapt:

  1. Copy config/context-manifest.example.yaml to your project as context-manifest.yaml and remap the layer paths.
  2. Give plan-review / code-review briefs for your architecture. They ship as skeletons with a _brief-template.md; copy it into each skill's references/ (one brief per concern). The Prism project's real design briefs live in the Prism repo as a worked example — see docs/prism-case-study.md.

Full guide: docs/customizing.md. The pattern itself is explained in docs/concepts.md, and the original Prism instantiation in docs/prism-case-study.md.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados