convez.ai
AgenteAutomaçãoAvançado
craigeous

loom — Loop de Desenvolvimento Multi-Agent Spec-Driven

craigeous/loom

Orquestração multi-agent com avaliação cega e memória durável em arquivos

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent orquestrador que tece um ciclo de desenvolvimento com cinco papéis especializados (pesquisador, planejador, avaliador de plano, desenvolvedor, avaliador de código) usando avaliação cega e memória durável baseada em arquivos. Funciona como plugin Claude Code, detecta conformidade com convenções loom e retoma trabalho através de contextos resetados.

Instalação

$npx convez-ai craigeous/loom

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 1 semana
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Retomar desenvolvimento após contexto expirado sem perda de progresso
  • Implementar ciclos de código-review imparciais com avaliadores cegos
  • Otimizar uso de tokens atribuindo modelos menores a papéis específicos
  • Gerenciar projetos complexos com checkpoints e limites de escopo declarados

Veredito

loom é uma solução sofisticada e bem-arquitetada para problemas reais de desenvolvimento colaborativo com IA — context drift, revisão enviesada, consumo de tokens. Pronto para produção em M1, com roadmap claro e dogfooding da própria estrutura.

README

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

loom

loom weaves a development loop out of five specialist roles — researcher, planner, plan evaluator, developer, code evaluator — and drives them through a file-based, spec-driven process so that work survives context resets and is reviewed with blind, impartial rigor.

loom is a Claude Code plugin. You invoke it inside any repository; it detects how aligned that repo is with loom's conventions and either bootstraps, migrates, or resumes work. A thin orchestrator spawns each role as a cold agent on the model best suited to its job, hands off work through files in .docs/, and stops at the scope boundary or human checkpoint you declared.

Why

  • Context drift — long sessions lose the thread. loom keeps durable memory in .docs/ so any cold agent can resume from files alone.
  • Blind rigor — evaluators judge work without knowing who produced it, so review stays critical and free of self-approval.
  • Token-smart automation — each role runs on the cheapest model that can do its job well; only judgment-heavy roles use the strongest model.

Status

M1 complete; M2 (init modes & gate learning) in progress. The plugin is built, installed, validated, and run end-to-end (planner → blind plan-eval → developer + gate → blind code-eval → finalize). The /loom:run orchestrator, the five role agents, and the playbook all exist. The authoritative design lives in .docs/spec/ — start with 00-overview.md; decisions are in .docs/ADR/.

Repository layout

loom/                          # this repo = the loom project + its marketplace
├── .claude-plugin/marketplace.json   # lists the loom plugin (source ./plugins/loom)
├── plugins/loom/              # the shippable plugin
│   ├── .claude-plugin/plugin.json
│   ├── commands/              # /loom:run + one-off /loom:research, :plan, :eval-plan, :develop, :eval-code, :status, :init
│   ├── agents/                # researcher · planner · plan-evaluator · developer · code-evaluator
│   └── skills/loom-playbook/  # templates, rubrics, conventions, gates
└── .docs/                     # loom's OWN design memory (dogfooding) — not shipped

Install

/plugin marketplace add craigeous/loom     # or: /plugin marketplace add ./loom (local)
/plugin install loom@loom
/plugin validate ./loom                     # optional: check manifests + frontmatter

Then, inside any repo, run the orchestrated loop or a single role pass. Plugin commands are namespaced as /loom:<name>:

CommandWhat it does
/loom:run [scope]the orchestrator — detect state, take scope/gates, drive the roles
/loom:research <topic>one-off researcher pass
/loom:planone-off planner pass
/loom:eval-plan [artifact]one-off blind plan/research review
/loom:develop [slice]one-off developer pass
/loom:eval-code [slice]one-off blind code review
/loom:statusprint .docs/ state
/loom:initinitialize/align this repo to loom

loom operates on the current repo's .docs/.

loom dogfoods its own structure: this repository is managed by the very process loom implements.

Ver o original no GitHub

Bom para

Relacionados