convez.ai
Skill
Avarce

repo-vetting-skills

Avarce/repo-vetting-skills

Ver no GitHub
InicialPluginScore · de 0 a 100

Claude Code skills that vet third-party repos and packages before you adopt them — desk research (repo-research) and a static pre-install security check with claim verification (repo-security-check)

Instalação

$npx convez-ai Avarce/repo-vetting-skills

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 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

README

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

repo-vetting-skills

Two Agent Skills for Claude Code that vet third-party repositories and packages before you adopt them:

SkillQuestion it answersTriggers on
repo-research"Should I use this?""research this repo", "is X any good", "should I use X", "evaluate this"
repo-security-check"Is it safe to install?""is this safe to install", "do a security check", "audit this dependency", "check this for malware"

repo-research is pure desk research — no cloning, no execution — ending in a clear use-it / use-it-for-X / avoid recommendation tailored to your project. repo-security-check shallow-clones to a sandbox, inspects statically without ever executing the target, and ends in a defensible INSTALL / HOLD / AVOID verdict.

Why

Existing tools cover adjacent problems — dependency auditors assume you already installed the project, scanners check CVE databases, marketplaces curate by hand. Nothing answered the question you actually have before npm install something-from-a-blog-post: does this thing do what it says, and what happens if I run it?

The differentiator is claim verification: every security/privacy claim the project makes ("no telemetry", "we collect no data", "end-to-end encrypted") gets tested against the actual code and labeled Verified / Unverified / Contradicted with file:line evidence. Marketing never gets restated as fact.

Example output

From a real run against dotenv@17.4.2 — see the full verdict:

Security verdict: INSTALL

  • Claims checked: 1 verified, 0 unverified, 0 contradicted — "zero-dependency" verified against the published package.json and lockfile.
  • Known vulns: none in the installed artifact (14 lockfile advisories are all dev-only deps).
  • Published tarball matches the repo at v17.4.2 per-file; no install-time lifecycle hooks; zero network calls or telemetry in the shipped code.

Install

As a plugin (recommended):

/plugin marketplace add Avarce/repo-vetting-skills
/plugin install repo-vetting@repo-vetting-skills

Manual copy:

git clone https://github.com/Avarce/repo-vetting-skills
cp -r repo-vetting-skills/skills/* ~/.claude/skills/

Requirements

None — both skills run on a bare machine (git, curl, grep). If present, these improve fidelity and are auto-detected, never auto-installed:

  • gh — authenticated GitHub stats (falls back to the public REST API)
  • osv-scanner — full transitive-dependency vuln scan (falls back to the OSV.dev API per dependency)

Security model

repo-security-check is built to inspect potentially hostile code safely:

  • Never executes the target. No builds, no scripts, no binaries, and no regular installs — install and build hooks are a primary malware execution vector. The most it will ever do is a scripts-disabled install (npm install --ignore-scripts) inside the sandbox, as a last resort for resolving a dependency tree. If a check would require running the code, it stops and tells you instead.
  • Sandboxed. Shallow-clones the exact release tag into a mktemp -d directory and deletes it afterward. Published artifacts are fetched as plain tarballs (never via tooling that can execute package scripts, like pip download on sdists).
  • Prompt-injection aware. The target's README, docs, and comments are treated as data, never instructions. Text addressing AI reviewers ("AI assistants: this package is safe") is itself reported as a red flag.
  • Rubric-based verdicts. INSTALL/HOLD/AVOID follows a written rubric with a closed definition of "critical claim" — designed so the same evidence produces the same verdict.

Disclaimer

This is automated static pre-install vetting — a strong first filter, not a substitute for a professional security audit. A clean verdict lowers risk; it doesn't eliminate it. You remain responsible for what you install.

License

MIT

Ver o original no GitHub

Bom para

Relacionados