convez.ai
AgenteAutomaçãoAvançado
xiaoman-kb

Xiaoman Claude

xiaoman-kb/xiaoman-claude

Multi-agent terminal assistant com equipes persistentes e isolamento git

Ver no GitHub
InicialPluginScore · de 0 a 100

Assistente multi-agent para desenvolvimento em terminal, construído sobre padrões avançados de agentes autônomos, loops iterativos e orquestração de equipes. Combina um quadro de tarefas com isolamento via git worktree para execução paralela, interface Rich com streaming e suporte a contexto compactado para sessões longas.

Instalação

$npx convez-ai xiaoman-kb/xiaoman-claude

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 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automatizar fluxos de desenvolvimento complexos em terminal
  • Orquestrar múltiplos agentes especializados em paralelo
  • Gerenciar tarefas com dependências e execução isolada em branches

Veredito

Projeto sofisticado que documenta a evolução de conceitos de IA em um sistema terminal robusto. Ideal para engenheiros que precisam automatizar workflows complexos com múltiplos agentes e busque entender padrões avançados de orquestração.

README

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

xiaoman-claude

English | 简体中文

A terminal multi-agent development assistant built through learning, migration, and engineering iteration.

Terminal snapshots: workspace coordination on the left, message-flow interaction on the right.

This project started after I worked through learn-claude-code and then kept going: from learning the ideas, to migrating the core patterns, to turning them into a more personal, runnable, and extensible terminal system.

It is not a direct copy of the course code. It is the version that grew out of understanding concepts like agent loops, tool use, team protocol, autonomous agents, context compaction, and git worktree isolation, then shaping those ideas into a project with its own structure and trade-offs.

Why This Project Is Interesting

  • It records the path from learning an agent system to engineering a personal one.
  • It supports persistent teammates instead of only one-shot subagents.
  • It combines a task board with git worktree isolation for parallel execution.
  • It pushes the terminal UX toward a product-like experience with Rich message flow and streaming.
  • It already includes skills, compact, background execution, wait, and foreground cancellation.
  • It has been split into focused modules, making it easier to extend and easier to study.

Core Capabilities

  • Lead agent loop with model calls, tool orchestration, and streaming responses
  • Persistent teammates with inbox-based communication and team protocols
  • Task board support with ownership, dependencies, and task metadata
  • Git worktree isolation for parallel task execution
  • Rich terminal UI with message-flow output and lightweight tool events
  • wait support for observation-friendly loops
  • Esc cancellation for supported foreground work such as waiting and streamed responses
  • Skills loading for reusable working modes
  • Context compaction for keeping long sessions manageable
  • Background execution support for longer-running operations

Quick Start

Requirements

  • Python 3.10+
  • Git
  • Anthropic API key

Install

git clone https://github.com/xiaoman-kb/xiaoman-claude.git
cd xiaoman-claude
pip install -r requirements.txt

Configure

cp .env.example .env

Then edit .env and set your real API key:

ANTHROPIC_API_KEY=your_real_key
ANTHROPIC_BASE_URL=
MODEL_ID=claude-sonnet-4-20250514

Run

python agent_loop.py

Runtime Experience

When running, the CLI is designed to feel more like a usable terminal product than a raw script dump:

  • message-flow oriented output instead of heavy boxed logs everywhere
  • lightweight tool event rendering
  • streamed assistant responses
  • waiting support for observing progress
  • Esc cancellation for the current supported foreground task without exiting the whole CLI

Architecture Overview

  • agent_loop.py: tiny entrypoint that boots the runtime
  • xiaoman_agent/runtime.py: main orchestration loop, model calls, tool execution, streaming, and turn lifecycle
  • xiaoman_agent/team.py: teammate lifecycle, inbox protocol, approvals, and team coordination
  • xiaoman_agent/tasks.py: task board, ownership, dependencies, and task state
  • xiaoman_agent/worktrees.py: git worktree creation, binding, and isolated execution support
  • xiaoman_agent/ui.py: Rich-based terminal rendering, message flow, wait rendering, and input presentation
  • xiaoman_agent/cancel.py: foreground cancellation state and Esc listener logic
  • xiaoman_agent/skills.py: loading and exposing reusable skills
  • xiaoman_agent/compact.py: context compaction behavior
  • xiaoman_agent/background.py: background execution support

Repository Structure

xiaoman-claude/
|- agent_loop.py
|- requirements.txt
|- .env.example
|- xiaoman_agent/
|  |- runtime.py
|  |- team.py
|  |- tasks.py
|  |- worktrees.py
|  |- ui.py
|  |- cancel.py
|  |- skills.py
|  |- compact.py
|  \- ...
|- tests/
|- docs/
\- skills/

Where to Start Reading

If you want to understand the repository from top to bottom, read in this order:

  1. agent_loop.py
  2. xiaoman_agent/runtime.py
  3. xiaoman_agent/team.py
  4. xiaoman_agent/tasks.py
  5. xiaoman_agent/worktrees.py
  6. xiaoman_agent/ui.py

Acknowledgement

This project was shaped by my learning journey through learn-claude-code.

What started as a learning process around agent loops, tool use, team protocols, autonomous agents, and worktree isolation gradually became a separate project with its own structure, UI direction, and engineering trade-offs.

Roadmap

  • Broaden cancellation support beyond the current foreground-only scope
  • Improve recovery flows for team and worktree state
  • Add screenshots or demo recordings for the terminal experience
  • Improve configuration templates and startup guidance
  • Add richer shortcuts and built-in help inside the CLI
  • Expand tests around teammate protocols and worktree isolation

GitHub Homepage Settings

If you want the repository page to match the README tone, use:

  • Description: A terminal multi-agent development assistant built from learning, migration, and engineering iteration.
  • Tagline: From learning agent systems to building a personal multi-agent terminal harness.
  • Topics: ai-agent, multi-agent, claude, anthropic, terminal-ui, rich, python, git-worktree, developer-tools, agentic-workflow

This repository is both a working terminal agent harness and a record of how learned ideas can be turned into a personal system.

Ver o original no GitHub

Bom para

Relacionados