convez.ai
SkillDevOps & InfraestruturaIntermediário
Starr-del

ShipSafe — Gate de Segurança para Deploy

Starr-del/ShipSafe

Detecta erros de deploy que quebram produção — antes de fazer push

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill que analisa estaticamente seu projeto Python antes do deploy para prevenir falhas de produção: databases expostos por RLS ausente, migrações irreversíveis, secrets commitados, cutover de DNS incorretos, buracos de CORS. Roda 100% offline na sua máquina, sem credenciais e sem dependências externas.

Instalação

$npx convez-ai Starr-del/ShipSafe

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
v0.1.0
Issues abertas
0
Contribuidores
1

Casos de uso

  • Validar segurança de RLS em bancos Supabase antes de lançar
  • Auditar secrets e chaves privadas commitados no repositório
  • Validar configurações de CORS que vazam requisições autenticadas
  • Verificar reversibilidade de migrações SQL antes do deploy
  • Validar registros DNS antes de cutover de domínios customizados

Veredito

ShipSafe preenche uma lacuna real: detecta falhas de deploy específicas (RLS exposto, migrações irreversíveis, secrets vazados) que outras ferramentas não cobrem. Design offline-first é correto para uma tool que inspeciona segredos. Prático para times que já usam Claude Code.

README

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

shipsafe

The deploy-safety gate for Claude Code — catch the failures that take production down, before you ship. Exposed databases from missing Row Level Security, irreversible migrations that can't be rolled back, secrets committed to the repo, DNS cutovers that black-hole your domain, CORS holes that leak authenticated requests. shipsafe statically analyzes your project for these before you push.

This isn't a general dev-productivity pack. It's focused on one thing: the specific mistakes that turn a deploy into an outage or a breach. Several of these checks — Supabase RLS auditing, migration reversibility, DNS cutover validation, CORS credential leaks — aren't covered by any other Claude skill that exists today.

Runs entirely on your machine. Every check is stdlib-only Python 3.8+ — no installs, no network calls, no credentials read. Your code never leaves your laptop, which for a tool that inspects your secrets and database schema is the only acceptable design.

Why this exists

Most deploy tooling for Claude Code operates after the failure — it fetches logs from your account to explain what already broke. shipsafe is the other half: deterministic checks that run on the repo itself, the way a linter does, to stop the preventable failures from happening at all.

And it deliberately covers the failures the rest of the ecosystem skips. There are good productivity packs for commits, changelogs, and general linting. There was no focused gate for deploy-time safety — the database left world-readable because RLS was never enabled, the DROP COLUMN that can't be undone, the service_role key shipped in a client bundle. That's the gap this fills.

The 10 skills

The first five are the deploy-safety core — the checks that prevent outages and breaches, several of which exist nowhere else. The rest cover reproducibility and build hygiene.

SkillCatchesRun it when
migration-reversibilityMissing Supabase RLS (public database), irreversible SQL, deploy-window breakers, lock-taking indexesAny release with schema changes
secret-preflightCommitted keys/tokens/private keys; decodes JWTs and escalates Supabase service_role keys specificallyBefore any push to a shared/public repo
dns-cutoverWrong/missing/stale records for Vercel, Netlify, CF Pages, GH Pages — generate them, then validate before you flipPointing a custom domain
cors-auditorWildcard+credentials, reflected origins, null-origin allowlists — the configs that leak authenticated requestsAny API serving a browser frontend
ci-workflow-doctorpull_request_target secret exfiltration, unpinned actions, missing deploy concurrencyEditing GitHub Actions
rollback-readinessComposite 0–100 deploy-safety score across all checksBefore every push — the capstone
deploy-log-doctorClassifies failed build logs (12 failure signatures incl. Next.js prerender)A build just failed
env-driftEnv vars missing / undocumented / leaked to client bundles; hardcoded localhost; server env in 'use client'First deploy, "undefined in prod"
lockfile-doctorMultiple lockfiles, manager mismatch, git/file/wildcard deps"Works on my machine"
bundle-weightHeavyweight deps with native replacements, duplicate-purpose librariesSlow builds, slow cold starts

Each skill = SKILL.md (instructions Claude loads on demand) + a tested CLI script. Every script supports --json for structured output and uses consistent exit codes: 0 clean, 1 findings, 2 usage error.

Install

# Claude Code — copy the skills you want (or all of them):
mkdir -p .claude/skills
cp -r skills/* .claude/skills/

# Or use any script standalone, no Claude required:
python3 skills/rollback-readiness/scripts/readiness.py .

Quick start

# the capstone check — run before pushing
python3 skills/rollback-readiness/scripts/readiness.py .

# a deploy just failed — paste the log
pbpaste | python3 skills/deploy-log-doctor/scripts/diagnose_log.py -

# about to point your domain
python3 skills/dns-cutover/scripts/dns_cutover.py plan yourdomain.com vercel

Tested

tests/run_tests.py runs 105 assertions across all 10 scripts — positive cases (dirty fixtures must be flagged) and negative cases (clean fixtures must pass silently):

python3 tests/run_tests.py

Built against documented failures

The checks aren't hypothetical — they map to the failure classes documented across thousands of real AI-built deployments: Supabase service_role keys and other secrets shipped in frontend bundles, tables exposed because Row Level Security was never enabled, hardcoded localhost URLs, env vars that exist locally and nowhere else, and irreversible migrations run against production. secret-preflight decodes JWTs offline and escalates service_role keys specifically; migration-reversibility audits RLS on Supabase projects; env-drift catches client-bundle leaks, localhost URLs, and server env reads inside 'use client' components.

Design rules

  1. Offline always. If a check would need the network, it doesn't belong here.
  2. Stdlib only. pip install is itself a deploy step that can fail.
  3. Precision over recall. A preflight tool that cries wolf gets deleted. Clean fixtures must pass with zero findings; suppression comments (secret-preflight: allow) exist for the rest.
  4. Every finding teaches. Output explains why it breaks deploys, not just that it matched.

Prior art, honestly

Two kinds of adjacent tools exist, and this fits between them:

Live deploy tools (the official Vercel plugin, log-fetching skills) operate against your account after a failure. Use those to debug what already broke; use this to prevent what's preventable.

Productivity skill packs — most relevantly pluginpool, which ships ten stdlib-only Claude plugins (commit-narrator, changelog-forge, env-lint, secret-guard, deps-doctor, and more). pluginpool is excellent for everyday developer workflow. The overlap with shipsafe is narrow and intentional: a handful of checks (env, secrets, deps) appear in both, but pluginpool's focus is git/workflow productivity, while shipsafe is focused entirely on deploy-time safety — and covers RLS auditing, migration reversibility, DNS cutover, CORS credential leaks, and CI exfiltration that general packs don't. If you want both convenience and a deploy gate, run both; they don't conflict.

Disclaimer

This software is provided "as is", without warranty of any kind, per the MIT License. By using it you accept that:

  • These are static, heuristic checks. They can and will produce false negatives (real problems missed) and occasional false positives. A clean run is not a guarantee that your deployment is safe, secure, or free of exposed secrets, and must not be treated as one.
  • This is not a security audit, not penetration testing, and not legal, security, or compliance advice. Static analysis cannot detect runtime issues such as broken authentication logic or data exposed through legitimate endpoints. For anything handling real user data, get a proper security review.
  • You are solely responsible for your deployments, your data, your keys, and the consequences of acting (or not acting) on this tool's output. The authors and contributors accept no liability for any damages, losses, breaches, or downtime arising from use of this software.
  • Findings about third-party platforms (Vercel, Netlify, Supabase, GitHub, Cloudflare) reflect their behavior at time of writing and may be outdated. Always confirm against the platform's current documentation.

MIT licensed.

Ver o original no GitHub

Bom para

Relacionados