b — Runtime de Disciplina para Agentes de IA
Instale uma linha, ganhe 6 verificações transparentes no seu AI coding agent
Ver no GitHubFramework que instala 6 ferramentas de verificação (bground, banchor, bsmell, bratch, bwatch, bspector) como skills nativos em Claude Code, Codex e Cursor. Funciona como guardrail automático do agent loop, capturando desvios de tarefa, padrões de deflexão e regressões sem intervenção humana.
Instalação
npx convez-ai bvasilenko/b1 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
- Validar factualidade de claims em sessões Claude Code
- Detectar drift de tarefas e padrões de mock-instead-of-fix
- Monitorar regressões e mudanças de assinatura em código novo
- Escanear vulnerabilidades em bundles de skills antes de instalar
- Manter agentes IA alinhados à missão do projeto
Veredito

Solução elegante e opinionada para hardening de agentes IA com zero overhead operacional. Requer confiança na heurística das 6 ferramentas, mas economiza debug em larga escala.
README
A tradução para o português deste README ainda está na fila.
b
Discipline runtime for your AI coding agent, installed in one line. After npm install -D @booga/b, every Claude Code session, Codex run, or Cursor pass picks up 6 small tools that fire transparently: bground verifies factual claims; banchor checks task drift; bsmell flags deflection patterns (mock-instead-of-fix etc.); bratch catches regression-signature changes; bwatch consults outward trackers; bspector scans new tools before install. Verdicts return as directives the agent acts on: no operator intervention. The capability shape is closed; the prompt library that powers each verdict evolves continuously via empirical-lift evaluation.
Entry point for the b-* prompt lookup tools. Installs the tools, registers each as a Claude Code skill, Codex AGENTS template, or Cursor rule so the agent loop discovers and invokes them, and updates installed tools on demand.
One-line install
npm install -D @booga/b
That single line:
- Pulls
bplus the 6 prompt lookup tools (bground, banchor, bsmell, bratch, bwatch, bspector) onto your machine. - Auto-registers all 6 with every supported substrate by default: Claude Code, Codex, and Cursor.
- Scope is project-local for a project install (writes to
<project>/.claude/skills/,<project>/AGENTS.md,<project>/.cursor/rules/). - Scope is user-wide for a global install (
npm install -g @booga/b).
Opt-out
- Skip the auto-register step entirely: set
B_SKIP_AUTO_INSTALL=1beforenpm install. - Skip specific substrates: set
B_SKIP_TARGETS=cursor,codexbeforenpm install.
You can still run the install CLI by hand later:
b install --target=claude
b install --target=codex
b install --target=cursor
b uninstall --target=claude
Tools
Shipped:
- bGround - prompt lookup for claim grounding
- bAnchor - prompt lookup for task-to-mission anchoring
- bSmell - prompt lookup for deflection-pattern smells
- bRatch - prompt lookup for regression-signature checks
- bWatch - prompt lookup for external-tracker findings
- bSpector - prompt lookup for vulnerability patterns in agent skill bundles
Shared dependency:
- bCore - types and helpers linked at build time by every tool above
License
MIT.