convez.ai
SkillIntegração & APIsAvançado
aldegad

Skill Hook Authoring — Guia de Interoperabilidade Cross-Runtime

aldegad/skill-hook-authoring

Uma fonte canônica para skills, hooks e plugins — compatível com 7 plataformas de agentes IA.

Ver no GitHub
InicialPluginScore · de 0 a 100

Repositório que centraliza a documentação técnica comparativa de sete plataformas de agentes IA (Claude Code, Codex, Grok, Hermes, Cursor, Kuma Studio, Antigravity) — skills, hooks, plugins, arquivos de instrução, sessões e billing. Fornece uma wiki atualizada diariamente com fatos citados em docs oficiais e uma metodologia para manter um source of truth único instalável em todos os runtimes sem drift.

Instalação

$npx convez-ai aldegad/skill-hook-authoring

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
12
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
v1.1
Issues abertas
0
Contribuidores
2

Casos de uso

  • Comparar capacidades de skills/hooks entre Claude Code, Codex, Grok, Hermes e Cursor
  • Automatizar instalação e sincronização de plugins em múltiplas plataformas de agentes
  • Documentar padrões de invocação, session resume e billing por runtime
  • Construir dashboards de gerenciamento cross-agent unificados

Veredito

Material de referência técnico robusto para times que operam múltiplos runtimes de agentes IA. Resolve problemas reais de interoperabilidade com fatos citados, não suposições — essencial para quem está construindo infraestrutura cross-plataforma, mas requer contexto avançado para aplicar.

README

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

English | 한국어 | 日本語 | 简体中文 | Español | Français | Deutsch


You don't run one AI agent anymore. Codex, Claude Code, Grok, Hermes, Antigravity, Cursor — each has its own idea of what a "skill" is, where hooks register, which instruction file it reads, how sessions resume, and how billing actually works. Ship the same tooling to all of them by hand, and they drift apart within a week. Ask "can I even do X on runtime Y?" and the answer is buried in seven different doc sites — or documented nowhere at all.

This repo is the map and the method:

  1. A compatibility wiki, refreshed daily from official vendor docs. Seven runtimes — Codex, Claude Code, Grok, Hermes, Antigravity CLI, Cursor, and Kuma Studio — compared across skills, hooks, plugins/extensions, project-instruction & memory files, CLI spawn (interactive vs headless), session resume, and billing.
  2. A methodology for shipping one repo-owned source of truth — skills, hooks, commands, scripts, references, assets, MCP/app wiring, plugin metadata — installed across every runtime without drift: one canonical package root, symlink installs, explicit retire/rename procedures, and a validation checklist.

What it answers

Questions you'd otherwise burn an afternoon on, answered with citations:

You askThe wiki answers
"Can I toggle a skill off on Grok or Hermes?"No official per-skill disable exists on either — so a toggle must move the skill out of the discovery root. That fact is the difference between shipping a real toggle and shipping a lie.
"Which instruction file does each agent read?"AGENTS.md vs CLAUDE.md vs GEMINI.md vs .hermes.md — who reads what, in which priority, and how to share one file across all of them.
"How do I resume a session from a script?"The Session Resume table: resume command, session store location, and session-id format for every runtime.
"Will my nightly claude -p cron cost me money?"The billing rows track exactly how headless/SDK usage is charged — including the 2026-06-15 Agent SDK credit change.
"Is /<skill-name> a thing on Codex?"No — Codex uses $<skill-name>. The Skill Invocation matrix records each runtime's real invocation token, so you stop assuming one token works everywhere.

This is the layer you build management tooling on. Kuma Studio's skill/hook toggle system — flip any skill or hook on/off across Claude, Codex, Grok, and Hermes from one GUI — was built directly on these facts: the wiki says where each runtime's real on/off switch is (skillOverrides, [[skills.config]], hook state keys), and where no switch officially exists, it says that too, so the tooling compensates deliberately instead of guessing. Whatever cross-agent dashboard, sync tool, or fleet manager you're building, this wiki is the ground truth it needs.

Why trust it

  • Every claim cites the vendor's own documentation. No vibes, no folklore, no "it worked on my machine".
  • Absence is recorded, not inferred. Where a runtime doesn't document a capability, the wiki says not documented instead of assuming parity — knowing a switch doesn't exist is as valuable as knowing where it is.
  • It re-verifies itself daily. A scheduled cloud agent re-fetches every source in docs/official-sources.json and opens a PR when the evidence changed; a deterministic gate auto-merges docs-only changes that pass the source check. Stale compatibility tables are how cross-runtime tooling rots — this one doesn't sit still.

What This Repo Owns

  • SKILL.md — the skill entrypoint and the authoring/interoperation methodology.
  • docs/compatibility-matrix.md — the cross-runtime comparison (Skills/Hooks/Plugins/Instructions, Session Resume, Skill Invocation).
  • docs/cli-invocation.md — per-runtime CLI spawn (interactive vs headless) and resume syntax.
  • docs/plugin-packaging.md — how plugin/extension packaging differs by platform.
  • docs/official-sources.json — the source manifest the daily refresh re-verifies.
  • docs/cloud-automation.md — the daily update automation, and why it runs in the cloud.
  • docs/kuma-studio-patterns.md — public Kuma Studio operating patterns.
  • CHANGELOG.md plus the git tag — the version record. History stays here, not in the doc bodies.

Daily Wiki Refresh

A daily agent keeps the wiki current: it reads docs/official-sources.json, fetches the official vendor URLs, and opens a pull request when the evidence changed. It never pushes to main.

The recommended path is Claude Routines — a scheduled Claude Code session that runs in Anthropic's cloud on a Claude subscription, with no API key and no GitHub Actions, and keeps running when your laptop is closed. Running the refresh locally via claude -p is discouraged: from 2026-06-15, Agent SDK / claude -p usage on eligible subscription plans draws from a separate monthly Agent SDK credit (per-user, refreshes monthly, does not carry over) instead of the interactive plan pool — and a local cron only fires while the machine is awake. See docs/cloud-automation.md.

  1. In Claude Code, run /schedule (or open https://claude.ai/code/routines).
  2. Point the routine at this repository and use the prompt in prompts/daily-official-doc-update.md.
  3. Schedule it once a day. It opens a PR, then runs an auto-merge gate that squash-merges docs-only changes passing the source check; anything else waits for your review (see docs/cloud-automation.md).

Codex users can run the same flow with a Codex App Automation instead. See docs/cloud-automation.md for both paths.

Local Checks

node scripts/check-official-sources.mjs --write-report

The script validates source IDs, allowed official hosts, URL reachability, the required source categories, and the SKILL.md line budget.

Ver o original no GitHub

Bom para

Relacionados