AI Vibe Coding Toolkit
ASMN-96/ai-agents-skills-toolkit
Governança documentada para agentes de IA com proveniência e validação de skills
Ver no GitHubKit de governança para workflows de agentes de IA, com padronização de proveniência de código, limites de ativação e validação de métodos reutilizáveis. Permite escalar engenharia assistida por IA sem introduzir código externo não verificado, mantendo rastreabilidade e controle em pipelines como Codex e Claude Code.
Instalação
npx convez-ai ASMN-96/ai-agents-skills-toolkit1 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
- Padronizar comportamento de agentes de IA entre projetos
- Gerenciar proveniência e rastreabilidade de métodos externos
- Implementar validação e aprovação explícita de workflows de IA
- Escalar engenharia assistida com controle de qualidade
Veredito

Toolkit robusto e bem documentado para times que precisam de governança rigorosa em workflows de agentes de IA. Exige maturidade técnica e disciplina operacional, mas oferece rastreabilidade, validação e controle que faltam em abordagens ad-hoc — essencial para ambientes regulated ou de alto risco.
README
A tradução para o português deste README ainda está na fila.
AI Vibe Coding Toolkit
AI Vibe Coding Toolkit is a documentation-first governance repository for AI coding-agent workflows. It standardizes reviewed methods, source provenance, validated agents, activation boundaries, and controlled sync artifacts so teams can scale high-quality AI-assisted engineering without copying raw external runtime behavior.
Public release status
- Current-tree blockers are 0.
- Remaining findings are limited to owner-decision blockers, history-only blockers, safe guardrail/scanner evidence, and false positives.
- Public package validation is not whole-repo publication readiness.
v0.2.3is the current controlled AI Vibe Coding Toolkit release for agent-assisted governance after full resource refresh, zero passive source holds, toolkit activation hardening, leak-scan check-mode validation, and release gates.- Public-facing release status is based on observed validation evidence; external submissions and publication channels are separate approval-gated decisions.
- See
STATUS.mdfor the current boundary snapshot andMIGRATION.mdfor version migration notes. - This repository intentionally has no root
package.json; run directnode scripts/...commands from the repository root.
One-paragraph summary
The toolkit defines how to discover, evaluate, and operationalize reusable AI coding-agent methods without directly introducing unverified external code. It can support workflows around Codex, Claude Code, local project agents, and similar assistants while keeping runtime changes intentional and repeatable through verified metadata, explicit approvals, and observable validation.
Who it is for
- engineering leads who need predictable agent behavior across projects,
- platform teams that need provenance and review standards,
- security and governance owners requiring evidence-linked approvals,
- startup/product operators preparing for controlled pilots and future releases.
Problem it solves
- unstructured agent onboarding from external repositories,
- hidden assumptions between prompts, artifacts, and runtime behavior,
- weak source provenance and stale method risk,
- inconsistent documentation and validation language across teams.
Core concepts
- Skills: reviewed external method artifacts represented as supply-chain inputs, not default runtime activators.
- Agents: scoped bundles of rules, prompts, and workflows built from approved skills.
- Profiles: context templates that adapt validated artifacts to stack, risk level, and operating mode.
- Registries: metadata indexes for methods, tools, routing, and governance assets.
- Validators: command-gated checks for runtime consistency, package surface rules, and public/private safety policy.
- Source records: explicit provenance records (license, freshness, trust review, extraction limits).
Current canonical runtime is 5 skills + 12 repo-local agent files. Agent file presence, compiled fallback presence, registry recommendation, inline fallback use, and actual spawned-agent proof are separate facts.
Quick start
- Read
AGENTS.mdandREADME.md. - Read
docs/ROLLOUT_MATURITY_AND_PUBLIC_RELEASE_READINESS.mdanddocs/NO_FAKE_VALIDATION_POLICY.md. - Run required validation commands (below).
- Edit only in scope and report warnings before PR.
How to use with AI coding agents
Use mode-aligned prompts:
- planning-only review before implementation,
- controlled implementation for scoped edits,
- release review for merge-readiness.
Typical flow:
- Align scope and do-not-touch constraints.
- Implement documentation/workflow-only changes.
- Run validation commands.
- Open a PR with explicit blockers and remaining risk.
For real projects, treat the toolkit as an AI coding-agent governance and evidence layer. Select or recommend the 5 canonical skills and 12 repo-local project agent lenses explicitly where this runtime is used, report TOML file presence, compiled fallback presence, inline fallback use, and actually spawned agents separately, and use project-owned checks before proposing new tools.
Validation commands
node scripts/validate-public-package.mjsnode scripts/ai-toolkit/validate-codex-runtime.mjs
Optional, when release context is requested:
node scripts/validate-toolkit.mjsnode scripts/ai-toolkit/run-toolkit-evals.mjsgit diff --checkgit status --short
Only report checks that were actually executed.
There is no dependency install step for the toolkit itself. Do not run npm install, create a root package manifest, or activate hooks unless a separate owner-approved task changes that architecture.
External-facing status
Public-facing status:
- Canonical runtime is 5 skills and 12 agents.
- Public package validation can pass while still not proving whole-repo publication readiness.
v0.2.3is a controlled toolkit release, not a claim of higher maturity, enterprise or production certification, automatic installs, or broad cross-runtime active support.- External submissions, marketplace listings, package publication, and broader runtime support remain separate approval-gated actions.
Limitations
- Not a product runtime.
- Does not replace project application logic.
- Does not automatically activate external skills or tools.
- Public package validation is not full-release proof.
Contribution path
- Open an issue with clear scope and expected evidence.
- Propose docs and workflow edits in a PR.
- Include validation output and unresolved blockers.
- Keep changes limited to governance and documentation artifacts.
License
This repository uses the root LICENSE.