Millhouse
Knatte18/millhouse
Orquestração de tarefas de código com subagentes Claude em paralelo
Ver no GitHubSistema de orquestração de tarefas para Claude Code que automatiza o ciclo completo de desenvolvimento: triagem de issues, planejamento, implementação paralela e revisão de código com subagentes IA. Cada tarefa roda em seu próprio worktree Git isolado, com um repositório wiki centralizando o índice de tarefas.
Instalação
npx convez-ai Knatte18/millhouse1 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
Casos de uso
- Automação end-to-end de desenvolvimento de features a partir de issues GitHub
- Paralelização de múltiplas tarefas de codificação com revisão automática
- Gerenciamento de fluxo de trabalho em repositórios com CI/CD integrado
Veredito

Ferramenta sofisticada de orquestração que tira proveito real da natureza multi-agente do Claude Code. Exige setup robusto (Git 2.35+, wiki repository, configuração YAML) e conhecimento de arquitetura de CI/CD, mas resolve um problema legítimo: paralelizar tarefas de desenvolvimento com controle humano em checkpoints críticos.
README
A tradução para o português deste README ainda está na fila.
Millhouse
Millhouse (MH) is a task orchestration system for Claude Code. It manages the full lifecycle of coding tasks — from triaging GitHub issues to merging finished code — using AI subagents for design, planning, implementation, and review.
Each task runs in its own isolated git worktree. A separate wiki repository holds the task index. Claude subagents do the heavy lifting; the operator approves at key checkpoints.
Inspiration
Millhouse builds on ideas from three projects:
- claude-code-plugins by Craig Motlin — task tracking and skill plugins for Claude Code
- autoboard by Willie Tran — autonomous agent orchestration patterns
- skills by Matt Pocock — Claude Code skill conventions
General flow
GitHub issues
│
▼
mill-ghissues-to-tasks Triage open issues into grouped wiki tasks
│
▼
mill-spawn Claim a task, create a git worktree + branch
│
▼
mill-start Discuss the problem; produce discussion.md
│
▼
mill-plan Autonomous plan-writing: batched DAG of cards
│
▼
mill-go Build loop: implementer + reviewer per batch
│ Self-fix rounds, holistic review after all batches
▼
mill-finalize / mill-merge Squash-merge to main (direct or via PR)
│
▼
mill-cleanup Remove worktree, branch, and portal
Worktree isolation
Every task lives on its own branch in its own worktree under wts/<slug>/. A junction-based portal (portals/<slug>/) gives skills a stable path to the task's _mill/ state directory (status, plan, reviews, briefs) without hard-coding worktree paths.
AI subagents
mill-go is a lean orchestrator — it reads only status.md and plan metadata. The actual work is done by two subagent types:
- Implementer (
mill:mill-implementer) — reads a batch of cards, writes code, runs the verify command, commits. - Reviewer (
mill:mill-reviewer) — reviews diffs, returns a structured verdict (APPROVE / REQUEST_CHANGES with BLOCKING / NIT findings).
The Builder loops until all batches approve, then optionally runs a holistic review across the full diff.
Wiki
The task index lives in a sibling git clone (../wiki/). Skills never write directly to the wiki directory — all mutations go through the wiki daemon client (wiki/_client.py), which holds a write lock and commits atomically.
Configuration
mill-config.yaml in the hub root controls reviewer models, batch sizes, dispatch mode, and pipeline behaviour. A local override at .millhouse/config.local.yaml is gitignored.
Key settings:
llm:
claude:
dispatch: agent # agent | psmux
pipeline:
auto_merge: true # merge automatically after mill-go
auto_report: true # file self-report after each run
roles:
implementer:
model: haiku # fast/cheap model for implementation
code-review:
holistic:
reviewer: sonnethigh # holistic reviewer model
Requirements
- Claude Code with the mill plugin installed (
claude plugin add <path>) - Python 3.11+ (managed by the plugin's virtual environment)
ghCLI authenticated (gh auth login)- Git 2.35+ (for
git worktree)