convez.ai
SkillDevOps & InfraestruturaAvançado
weblue

Inspector Gadget

weblue/inspector-gadget

Setup modular e otimizado para IA com economia de tokens

Ver no GitHub
InicialPluginScore · de 0 a 100

Ferramenta de setup em camadas para ambientes de IA (Claude, Codex, Pi). Automatiza bootstrapping, gerenciamento de extensões/skills, auth e switching de API keys. Inclui compressão de tokens, subagentes e acesso remoto via Tailscale.

Instalação

$npx convez-ai weblue/inspector-gadget

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
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Bootstrapping rápido de múltiplos harnesses (Claude, Codex, Pi)
  • Gerenciamento centralizado de skills e perfis de subagentes
  • Compressão de tokens (60-90%) via RTK em fluxos de trabalho
  • Configuração de acesso remoto seguro para equipes

Veredito

Ferramenta robusta para quem precisa orquestrar múltiplos agentes IA em produção, com foco em economia de tokens e modularidade. Curva de aprendizado íngreme, mas payoff real para arquiteturas complexas.

README

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

agent-jacket

Layered installer for LLM coding setups. Every layer optional — pick harnesses, bring your own models, stop at whatever depth you need. Token savings built in throughout.

┌─ 5  remote access ──── server + clients over tailscale/ssh/tmux
├─ 4  workers & skills ─ subagent profiles, shared skills, sandboxed sub-agent runs (pi)
├─ 3  token savings ──── rtk wiring, prompt rules, output caps
├─ 2  models & auth ──── bring your own: subscriptions or API keys
└─ 1  harnesses ──────── any of: claude · codex · pi

1 — Harnesses

./install.sh                  # interactive
./install.sh claude pi        # any subset, or 'all'
./install.sh --force pi       # update existing install (re-seed, backups kept)
claudecodexpi
Prompt (base + add-on)~/.claude/CLAUDE.md~/.codex/AGENTS.md~/.pi/agent/AGENTS.md
rtk wiringhook + settingsrules fileextension
Shared skills✓✓✓
Worker profiles~/.claude/agents/——
Extensions + pi-only skills——~/.pi/agent/
Secrets deny-list + output capssettings.json——

2 — Models & auth

Nothing provider-pinned. Subscriptions (Claude Pro/Max, ChatGPT/Codex, pi /login OAuth) or API keys (installer offers a Pi key flow, warns before billing an API key a subscription would cover). Model choice surfaces everywhere: harness defaults, subagent model:/effort: frontmatter, Pi thinking levels, per-run model param.

3 — Token savings

rtk installed everywhere (pinned + sha256-verified on linux), wired into each harness — CLI output compressed 60–90% before it hits context:

$ rtk gain
  #  Command                Count  Saved   Avg%   Impact
 1.  rtk grep -n TODO           7    412  84.0%   ████████░░

Plus shared prompt rules (docs over code-reading, summaries over raw output, targeted edits) and caps: MAX_MCP_OUTPUT_TOKENS, effort routed by role.

4 — Workers & skills

Claude subagents (coder, reviewer, tester, architect, security) ~10 lines each, cost routing in frontmatter; main session orchestrates. Shared skills: stop-slop, grill-me.

Pi-only: the sandcastle skill gates sandcastle_run — headless Pi agent, same auth (auth.json mounted read-only), in a Docker container on an isolated git worktree; changes return as commits on a sandcastle/* branch. Sub-agent calls only; synchronous work runs directly on the host. Once per repo (install.sh offers to run this for a repo of your choice):

npx @ai-hero/sandcastle init --agent pi && npx @ai-hero/sandcastle docker build-image

5 — Remote access

laptop ──tailscale──▶ server ── tmux "agent-jacket" (pi, shell)
  └─ cmux workspaces            sshd (key-only, hardened)
./server.sh             # server: verify tailscale + ssh, start tmux session
./client-install.sh     # client: tailscale + cmux + ssh config + keys
./connect.sh            # client: open pi + shell workspaces in cmux

cmux = client-side macOS terminal, driven by connect.sh. Server needs only tmux — any SSH client works.

Maintenance

./maintain.sh           # report: rtk savings, junk, prompt bloat, stale branches
./maintain.sh --prune   # also delete junk + stale backups

Layout

install.sh            layers 1–4 (--force to update)
server.sh             layer 5, server side
client-install.sh     layer 5, client side
connect.sh            layer 5, cmux workspaces
maintain.sh           hygiene (report / --prune)
system-prompt.md      shared base prompt
agents/               harness add-ons · pi-only skills · worker profiles
extensions/pi/        sandcastle, copy-all, usage
skills/               stop-slop, grill-me
tools/rtk/            rtk reference
docs/TODO.md          open + resolved items
Ver o original no GitHub

Bom para

Relacionados