convez.ai
SkillDesenvolvimento WebIntermediário
mrkhachaturov

Agent Harness Docs

mrkhachaturov/agent-harness-docs

Docs locais auto-sincronizadas para agentes de IA, agnóstico a indexadores

Ver no GitHub
InicialPluginScore · de 0 a 100

Espelho local de documentação para agentes de IA (Claude Code, OpenAI Codex, OpenCode) com sincronização automática a cada 3 horas. Oferece skills adaptáveis que funcionam com diferentes indexadores de busca (Miyo ou busca simples com grep/ls). Ideal para manter referência offline e integrada aos seus agentes de codificação.

Instalação

$npx convez-ai mrkhachaturov/agent-harness-docs

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
3
Forks
0
Acompanhando
1
Saúde
42%
Último commit
há 1 dia
Versão
v
Issues abertas
0
Contribuidores
5

Casos de uso

  • Manter documentação de agentes de codificação sincronizada localmente
  • Integrar referências de Claude Code, Codex e OpenCode em seus workflows
  • Buscar rapidamente documentação offline sem depender de internet

Veredito

Solução prática para devs que trabalham com múltiplos agentes de codificação e precisam de referência local sempre atualizada. A flexibilidade de indexadores e suporte a macOS via launchd torna útil para pipelines de desenvolvimento, embora tenha escopo restrito a três harnesses específicas.

README

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

Agent Harness Docs

Last Update Platform

Local docs mirror for AI coding agents — auto-synced, indexer-agnostic.

Today this repo collects documentation for three coding-agent harnesses:

More harnesses will be added over time (Cursor, Aider, etc.). The docs update upstream every 3 hours via GitHub Actions; an hourly launchd job mirrors them into flat folders under your home directory.

You decide how to consume them: drop a semantic-search tool on top, grep the folder, or just let your agent read files directly. The installer ships a small skill that adapts to whichever indexing tool you have.

Originally claude-code-docs. Forked from ericbuess/claude-code-docs and generalised to multiple harnesses + indexers.

What the installer sets up

PathContents
~/claude-code-docs/Claude Code docs (~140 flat .md files)
~/codex-docs/Codex docs (~80 flat .md files)
~/opencode-docs/OpenCode docs (~34 flat .md files, MDX-converted)
~/Library/Caches/agent-harness-docs-mirror/Cache clone of this repo
~/Library/Application Support/agent-harness-docs/sync.shSync script
~/Library/LaunchAgents/com.mrkhachaturov.agent-harness-docs.plistHourly launchd job
~/.claude/skills/claude-code-docs/SKILL.mdClaude Code skill (variant picked by installer)
~/.claude/rules/claude-code-docs.mdAlways-on hint for Claude Code
~/.agents/skills/codex-docs/SKILL.mdCodex skill (agentskills.io USER scope)
~/.agents/skills/opencode-docs/SKILL.mdOpenCode-docs skill (shared across all three agents)
~/.claude/skills/opencode-docsSymlink → ~/.agents/skills/opencode-docs (for Claude Code)
~/.config/opencode/opencode.json[c]Miyo MCP merged into the mcp block (when Miyo is detected)

Skills adapt to your indexer

At install time the script detects what's on your machine and installs the matching skill variant from skills/<docset>/<indexer>/SKILL.md:

DetectedVariantWhat the skill tells the agent
Miyo CLI is on $PATHmiyoUse mcp__miyo__search with folder_path: "<docset>"
NothingplainUse ls / grep / Read against the docs folder directly

If you install (or remove) an indexer later, just re-run install.sh — it's idempotent and will swap in the right variant.

Want another indexer? PR welcome — add <indexer> to INDEXERS in scripts/render_skills.py and extend the Jinja2 templates in skills/templates/ with the tool-specific blocks. Run python scripts/render_skills.py and commit both the template change and the rendered output.

Install

curl -fsSL https://raw.githubusercontent.com/mrkhachaturov/agent-harness-docs/main/install.sh | bash

macOS only for now (uses launchd). Requires git, rsync, and curl — no Python or Node on your machine. The MDX → Markdown conversion for OpenCode runs in CI; you only receive pre-converted .md files.

Linux / Windows installers are planned. For now, the doc-fetch and render scripts work cross-platform; only install.sh is mac-specific.

After install, if you're on the miyo variant: open Miyo, add the three doc folders, label them claude-code-docs, codex-docs, and opencode-docs. Restart your agent so the skills/rules/MCP load.

Use

Claude Code:

/claude-code-docs how do hooks interact with permissions?
/claude-code-docs explain "sandbox environments"
/claude-code-docs path "permissions" "sandboxing"

Codex (skill is auto-discovered via metadata — no rule file needed):

/codex-docs how do I configure AGENTS.md for nested directories?
/codex-docs explain "skills vs plugins"
/codex-docs path "MCP" "hooks"

OpenCode (works in any of the three agents — opencode itself, Claude Code, or Codex — since the skill is shared):

/opencode-docs how do I add a custom MCP server?
/opencode-docs explain "primary agents vs subagents"
/opencode-docs path "skills" "permissions"

Each skill scopes its search to its own indexed folder so results never cross-contaminate.

How it stays current

  • Upstream (this repo, GitHub Actions): every 3 hours, three fetchers (scripts/fetch_claude_docs.py, scripts/fetch_codex_docs.py, scripts/fetch_opencode_docs.py) read the upstream sources and update docs/claude-code/, docs/codex/, and docs/opencode/. The OpenCode fetcher additionally pipes each MDX file through a Node remark-mdx converter so the result is plain Markdown.
  • Local (your machine, launchd): every hour, the sync script pulls the latest cache clone and rsyncs all three subfolders into your home folders.

Update

curl -fsSL https://raw.githubusercontent.com/mrkhachaturov/agent-harness-docs/main/install.sh | bash

Idempotent — re-running just re-pulls the cache, re-wires the launchd job, and re-picks the indexer variant. Docs themselves refresh every hour without manual action.

Uninstall

curl -fsSL https://raw.githubusercontent.com/mrkhachaturov/agent-harness-docs/main/uninstall.sh | bash

Removes all installed files and the launchd job. Does not remove Miyo MCP registrations on agent CLIs / configs — those may be used for other things. To remove them too:

claude mcp remove miyo
codex  mcp remove miyo
# opencode: edit ~/.config/opencode/opencode.json[c] and delete the
# "miyo" entry under the "mcp" object.

Layout choice — three skills, two-and-a-half folders

Each agent scans its own conventional skill locations:

  • Claude Code → ~/.claude/skills/
  • Codex → ~/.agents/skills/ (per agentskills.io USER scope)
  • OpenCode → ~/.config/opencode/skills/ (native), plus both ~/.claude/skills/ and ~/.agents/skills/ (OpenCode reads all three)

claude-code-docs and codex-docs live in each harness's native folder — automatic isolation, no leakage between agents.

opencode-docs is shared across all three harnesses (any of them might be asked OpenCode questions). The canonical SKILL.md is at ~/.agents/skills/opencode-docs/ — read natively by OpenCode and Codex — and exposed to Claude Code via a symlink at ~/.claude/skills/opencode-docs.

This is the pattern future general-purpose skills should follow.

Maintainer workflow

To add/edit a skill variant: edit the Jinja2 template in skills/templates/, then:

python scripts/render_skills.py             # render all combinations
python scripts/render_skills.py --check     # verify on-disk matches templates

Commit both the template change and the rendered output. CI runs --check and fails the build if they're out of sync.

For the OpenCode pipeline (MDX → Markdown), Node deps live in scripts/package.json + scripts/package-lock.json (committed). CI runs npm ci inside scripts/ before invoking fetch_opencode_docs.py. To test locally:

cd scripts && npm ci && cd ..
python scripts/fetch_opencode_docs.py       # populates docs/opencode/

License

Documentation content belongs to its respective authors — Anthropic (Claude Code), OpenAI (Codex), and the OpenCode project (OpenCode). Installer, skills, fetchers, and renderer are open source.

Ver o original no GitHub

Bom para

Relacionados