bozz-agents
bozbornehoo09/bozz-agents
Context e review skills para agentes de IA mantê-lo projeto coerente
Ver no GitHubSuite 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-agents1 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
- 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 asBLOCK/FIX/SUGGEST, and returns a verdict ofREADY/READY-WITH-FIXES/REVISE. Every specialist runs under an anti-hallucination contract: each finding needs a verbatim quote and afile:linecitation, 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 editingTARGETSin the workflow). - Never edit
.claude/skills/or.agents/skills/by hand — they're markedlinguist-generatedand will be overwritten. Editskills/.
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:
- Copy
config/context-manifest.example.yamlto your project ascontext-manifest.yamland remap the layer paths. - Give
plan-review/code-reviewbriefs for your architecture. They ship as skeletons with a_brief-template.md; copy it into each skill'sreferences/(one brief per concern). The Prism project's real design briefs live in the Prism repo as a worked example — seedocs/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.