convez.ai
SkillAutomaçãoAvançado
MercuriusDream

Agent Estate

MercuriusDream/agent-estate

Loop autônomo com memória persistente para Claude Code — zero config

Ver no GitHub
InicialPluginScore · de 0 a 100

Loop autônomo para Claude Code que executa trabalho contínuo sem intervenção manual. Usa hooks nativos do Stop event para interceptar saídas, mantém memória persistente em ledger e detecta rate limits automaticamente. Ideal para projetos que requerem ciclos longos de desenvolvimento autossustentável.

Instalação

$npx convez-ai MercuriusDream/agent-estate

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
4
Forks
1
Acompanhando
0
Saúde
28%
Último commit
há 8 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Desenvolvimento de projetos complexos com múltiplos ciclos autônomos
  • Construção de aplicações que requerem testes extensivos e iteração contínua
  • Tarefas de longa duração que precisam de handoff entre contextos

Veredito

Implementação técnica sólida de autonomia para Claude Code, usando hooks nativos sem wrappers externos. Prova de conceito validada (287 ciclos no Vibrowser), mas exige conforto com shell scripting e debugging de comportamentos autônomos — não é para fluxos simples.

README

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

Agent-Estate

IT'S THE FREE REAL AGENT ESTATE

Autonomous work loop for Claude Code. No end condition. Persistent memory. Zero config.

Used it to build a from-scratch C++ web browser — 287 autonomous cycles, 100+ hours, 3,374 tests, zero failures, <10 manual steerings.

Why Agent Estate?

  • Native hooks, not a wrapper — uses Claude Code's Stop event to intercept exits. No bash wrapper, no tmux, no external process manager
  • Zero config — no PRD files, no task definitions, no project setup. Just /agent-estate:start
  • Persistent memory — ledger tracks everything across sessions. "Tell The Next Claude" handoff for cross-context memory. Context doesn't reset between cycles
  • Auto-stop when done — --done flag lets Claude stop itself when the task is complete. Or run perpetual (default)
  • Rate limit resilience — detects 429/529 errors, waits, retries. Never crashes, never loses state
  • ~400 lines total — the entire plugin. 3 directories, 7 files. Nothing to configure, nothing to break

Install

Claude Code Plugin (recommended)

/plugin marketplace add MercuriusDream/agent-estate
/plugin install agent-estate@MercuriusDream

npx (skills.sh)

npx skills add MercuriusDream/agent-estate

Manual

git clone https://github.com/MercuriusDream/agent-estate.git
mkdir -p ~/.claude/plugins/local
ln -s "$(pwd)/agent-estate" ~/.claude/plugins/local/agent-estate
chmod +x agent-estate/scripts/setup-estate.sh
chmod +x agent-estate/hooks/stop-hook.sh

Requires jq:

brew install jq        # macOS
sudo apt install jq    # debian/ubuntu

Usage

/agent-estate:start                                       # full autonomy, perpetual
/agent-estate:start build a web server with tests         # guided prompt, perpetual
/agent-estate:start --done build a web server with tests  # auto-stops when done
/agent-estate:status                                      # check cycle, mode, stats
/agent-estate:stop                                        # manual stop

How It Works

start → read ledger → pick task → do the work → update ledger → try to exit
          ↑         stop hook intercepts → blocks exit → re-injects prompt ↲

The Hook

hooks/stop-hook.sh runs on every Claude Stop event:

  1. No state file? → exit normally
  2. done: true in frontmatter? → remove state file, exit (auto-stop)
  3. Rate limited (429)? → wait 60s, retry
  4. API overloaded (529)? → wait 30s, retry
  5. Otherwise → increment cycle, re-inject prompt, block exit

The Ledger

.claude/agent-estate.md — the persistent brain. Every Claude reads it, every Claude updates it.

SectionPurpose
Current StatusPhase, focus, momentum, cycle count
Session LogWhat happened each session
Worked ThingsTable of everything done with files touched
Future WorksPrioritized task queue
StatisticsCycles, tests, bugs fixed, features shipped
Tell The Next ClaudeSacred handoff message — letter to yourself with amnesia

Work Priority

Built into the protocol, in order:

  1. Broken — failing tests, bugs, errors
  2. Incomplete — half-finished features, TODOs, FIXMEs
  3. Missing — no tests, no error handling
  4. Ugly — code smells, dead code
  5. Slow — performance issues
  6. New — features that improve the project
  7. Creative — surprise the user

Battle-Tested

Built Vibrowser — a from-scratch C++ browser engine:

Autonomous cycles287 across 123 sessions
Runtime100+ hours, <10 manual steerings
Tests3,374 — zero failures
Features2,440+ shipped, 124 bugs fixed
Lines168,360+
EnginesCSS (370+ properties), JS (QuickJS), CORS/CSP/TLS
ResultRenders real websites off the live internet
Fun factclaude.ai 403'd the browser Claude built

Files

agent-estate/
├── .claude-plugin/plugin.json   # plugin manifest
├── commands/
│   ├── start.md                 # /agent-estate:start
│   ├── stop.md                  # /agent-estate:stop
│   └── status.md                # /agent-estate:status
├── hooks/
│   ├── hooks.json               # registers stop hook
│   └── stop-hook.sh             # core engine
├── scripts/
│   └── setup-estate.sh          # creates state file
├── SKILL.md                     # skill definition
├── AGENTS.md                    # agent instructions
└── README.md

Tips

  • Long sessions — Claude Max recommended for hours-long autonomous runs
  • Guiding prompts — focus without being restrictive
  • Ledger size — if it gets large (300KB+), Claude will auto-summarize older cycles
  • Resuming — ledger persists after stop. Start again and Claude picks up from the handoff
  • Full autonomy — no prompt = Claude explores and decides what's most valuable

License

AGPL-3.0

Author

MercuriusDream (Wooseok Sung)

Ver o original no GitHub

Bom para

Relacionados