Okayus Skills
okayus/okayus-skills
Skills de produção para agentes: sandbox Docker, Git relay e Cloudflare Workers
Ver no GitHubColeção de skills prontos para produção compatíveis com Claude Code, GitHub Copilot e agentes Agent Skills. Inclui padrões testados para Docker sandbox seguro, integração Git/GitHub sem credenciais no sandbox, e deploy automático em Cloudflare Workers com D1.
Instalação
npx convez-ai okayus/okayus-skills1 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
- Executar código de agentes em ambiente isolado com firewall iptables
- Automatizar push e PR no GitHub sem expor credenciais ao sandbox
- Deploy contínuo de Workers + D1 no Cloudflare com GitHub Actions
Veredito

Conjunto maduro e bem documentado de skills para segurança e automação de agentes. Resolve problemas reais (isolamento npm, credenciais, CI/CD) com padrões comprovados em produção, mas exige conhecimento avançado de Docker, systemd e GitHub Apps.
README
A tradução para o português deste README ainda está na fila.
okayus-skills
Agent Skills for Claude Code / GitHub Copilot / other Agent Skills-compatible agents.
Extracted from production-ready patterns used across my personal projects (Cloudflare Workers + D1 + Discord integrations).
Installing
Pick a skill and install it with the GitHub CLI skill extension:
# Project scope (recommended for project-specific patterns)
gh skill install okayus/okayus-skills <skill-name> --agent claude-code --scope project
# User scope (available in every project)
gh skill install okayus/okayus-skills <skill-name> --agent claude-code --scope user
Pin to a version (optional):
gh skill install okayus/okayus-skills <skill-name>@v1.0.0 --agent claude-code
Available skills
| Skill | What it does |
|---|---|
claude-code-docker-sandbox | Set up a Docker Compose-isolated dev environment where npm install and agent-run code execute behind a default-deny iptables egress firewall, never touching the host — to contain npm supply-chain attacks without depending on VS Code. Reuses Anthropic's published devcontainer Dockerfile + init-firewall.sh but drives them with a plain docker-compose.yml so any host editor (nvim, etc.) works via bind mount. Covers egress allowlist tuning (fatal vs non-fatal OPTIONAL domains — real Statsig endpoints and your production host; statsig.anthropic.com no longer exists), keeping Claude Code current inside the firewall (startup npm i -g — the native updater's host is blocked by the allowlist), the DISABLE_* env split that lets the /model picker show flag-gated models, the container-scope bypassPermissions default for autonomous loops, the in-container Claude Code OAuth "paste code" flow, the telemetry-domain DNS trap that silently kills the container, host-side git hygiene (keep credentials out of the boundary), the non-interactive scaffolding traps (sv/create-* prompt stalls), mounting a host skills repo read-only via a gitignored docker-compose.override.yml, the -f-flag trap that silently unloads that override, and the dig retry + ipset -exist robustness for flaky DNS and shared anycast IPs. Ships copy-paste templates for all four files. |
sandboxed-agent-git-relay | Let a sandboxed coding agent get its work onto GitHub as branches + PRs — and optionally all the way to merge — without any credential ever entering the sandbox: a host-side systemd timer scans claude/* branches in the bind-mounted repo, mints a 1-hour GitHub App installation token, pushes exact refspecs, opens PRs idempotently, and squash-merges only when the HEAD commit carries a Relay-Merge: yes trailer (CI-green is enforced server-side by the main ruleset; early attempts get 405 and retry next tick). Policy — prefix-only, no force, never main, merge-by-explicit-signal — lives outside the boundary. Covers why a GitHub App beats deploy keys / GITHUB_TOKEN / PATs, the App-ID-vs-Installation-ID trap, the gh credential-helper hijack, the systemd AccuracySec surprise, squash-residue resurrection (three-dot diff), and reading PR/CI status from the sandbox with unauthenticated curl (gh refuses to run tokenless). |
cloudflare-workers-deploy-skeleton | Walking-Skeleton setup for a single-Worker SPA + API + Cron on Cloudflare Workers with D1 and GitHub Actions auto-deploy. Covers the 3-layer SPA routing dance, wrangler.jsonc template, deploy.yml, and the well-known setup pitfalls (D1 token scope, pnpm deploy collision, RP_ID locking). |
cloudflare-workers-builds-keyless-deploy | Deploy Cloudflare Workers from GitHub with zero Cloudflare credentials stored in GitHub (no CLOUDFLARE_API_TOKEN in Actions secrets) using Workers Builds — Cloudflare's git-connected CI/CD. Covers the traps that cost hours: the default build token lacking D1 Edit (migrations fail silently), Root directory hiding in the Advanced-settings accordion, preview builds sharing the production D1 binding (turn non-production builds off), and Workers Builds not waiting for GitHub CI — guarantee "main = CI-green" with a branch ruleset instead. |
cloudflare-mcp-claude-tooling | Wire the Claude Code project tooling for a sandboxed Cloudflare project: a docs-only Cloudflare MCP (docs.mcp.cloudflare.com), a committed .claude/settings.json allowlist that denies git commit/push as an execution-level guard for the "git on the host" rule, the egress-firewall domains the docs MCP needs, and the grill-with-docs design step. Covers WHY docs-only — the bindings/builds/observability MCP servers overlap with wrangler and their OAuth callback is fragile in a container (vs Claude Code login's paste-code flow) — and the API-token (Bearer) escape hatch if you ever need an account server. |
cloudflare-cron-to-discord | Cron Trigger → Discord Webhook pattern with the pure-function-then-boundary architecture. Covers UTC→JST pure conversion, build/post separation with throw-free error handling, vitest mock testing, dev/prod webhook naming for contamination detection, and the /__scheduled dev caveat with @cloudflare/vite-plugin. |
cloudflare-d1-drizzle-migration | Safely run drizzle-kit migrations on Cloudflare D1 without losing data. Covers the silent PRAGMA foreign_keys=OFF incompatibility (D1 ignores it, so table-rebuild migrations cascade-delete child rows), the phased NULLABLE → backfill → NOT NULL column migration pattern for live databases, the mandatory pre-deploy backup + post-deploy row count check runbook, and 3 meta-lessons on environment parity, test blind spots, and destructive-change design review. |
cloudflare-workers-e2e-playwright | Wire Playwright e2e tests against a Cloudflare Workers + Vite + Hono app without falling into the two silent traps: strict CSP blocking Vite's HMR inline preamble (preventing React mount on page.reload()) and wrangler dev --config resolving D1 state to a different empty sqlite. Covers why to target the build artifact via wrangler dev instead of vite dev, why --persist-to .wrangler/state is mandatory, the WebAuthn virtual authenticator recipe (avoiding DEV_BYPASS_USER_ID so register/login wiring is actually tested), the third-party-OAuth seeded-session seam for apps without WebAuthn (seed a real session row + inject the cookie — no DEV_BYPASS/test-login route, since OAuth has no virtual authenticator), running the whole suite credential-free inside a Docker sandbox by baking Chromium at image-build time (two sandbox-only traps: the unsafe ratelimit binding and a localhost bind both hang wrangler dev), and the narrow "3 specs only" test scope philosophy. |
cloudflare-d1-weekly-backup-via-pr | Set up an automated weekly Cloudflare D1 backup workflow via GitHub Actions that exports the production database and opens a PR adding the SQL dump to the repo. Covers the "GitHub Actions is not permitted to create or approve pull requests" repo-setting gotcha (every fresh repo's first PR-creating workflow trips it), the cron schedule UTC-vs-local timezone conversion, the wrangler d1 export path-relativity quirk in pnpm monorepos, the explicit "commit backups to git" decision with its trade-offs and migration triggers, and reusing the existing CLOUDFLARE_API_TOKEN (no new secrets). |
cloudflare-api-token-permissions | Map a Cloudflare CI deploy auth error (code: 10000 / 7403 / 9106) to the missing API token permission, and the permission matrix per binding (Workers Scripts / D1 / R2 / KV / Queues / Vectorize / Hyperdrive / Workers AI / Browser Rendering). Covers the in-place permission edit workflow that keeps the token value (so the GitHub Secret stays untouched), the Edit Cloudflare Workers template's silent omission of D1 / R2 / Queues, the diagnostic flow when multiple permissions are missing cumulatively, and how to identify which token CI is actually using. |
cloudflare-workers-bot-scan-defense | Make a Cloudflare Workers app resilient to bot scans that arrive within minutes of HTTPS publication via CT Log enumeration (/.env, /.git/config, /admin, /wp-login.php probing). Covers the mental model (which paths actually cost you money — most are absorbed by the edge cache via SPA fallback), the audit recipe (curl -I checking for cf-cache-status: HIT), the narrow set of unauthenticated routes that need rate limiting (auth begin/verify), the exact wrangler.jsonc observability + ratelimits config, the IP-keyed Hono middleware pattern using CF-Connecting-IP, the verification flow (wrangler versions view + bundle grep + Workers Observability), and the documented eventual-consistency caveat that makes synthetic burst tests look like the limiter is broken even when correctly wired. Includes the binding-vs-WAF-rule decision matrix. |
cloudflare-workflows-for-long-tasks | Migrate a Cloudflare Worker post-response task off ctx.waitUntil() onto a WorkflowEntrypoint when the task can exceed the 30-second waitUntil wall-clock cap (Vision LLM, slow third-party APIs, multi-step orchestration). Covers the symptom-to-fix path when waitUntil() tasks did not complete within the allowed time shows up in wrangler tail and DB rows get stuck in running / pending, the exact wrangler.jsonc workflows binding, the class export pattern alongside default app, the step.do() retry / persist / serializable-output rules, the production cleanup runbook for orphaned rows, and the gotchas (Workflow<Params> is a @cloudflare/workers-types global not a cloudflare:workers import; bytes can't ride between steps; Workflow logs surface in the Dashboard, not wrangler tail). |
Repo layout
okayus-skills/
├── README.md
├── LICENSE
└── skills/
└── <skill-name>/
├── SKILL.md # metadata + core instructions
└── references/ # on-demand deep docs, code templates, diagnostics
Follows the Agent Skills specification skills/*/SKILL.md convention.
Versioning
Each release is tagged with semver (v0.1.0, v1.0.0, ...). Installing without @VERSION resolves to the latest release.
Skill-specific breaking changes bump the repo's major version. Additive refinements (new references, clarifications) bump the minor or patch version.
Validating locally
Skills are validated with gh skill publish --dry-run:
cd ~/dev/okayus-skills
gh skill publish --dry-run
Checks: name matches directory, required frontmatter, etc. See spec.
License
MIT.