Xiaoman Claude
xiaoman-kb/xiaoman-claude
Multi-agent terminal assistant com equipes persistentes e isolamento git
Ver no GitHubAssistente 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-claude1 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
- 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
waitsupport for observation-friendly loopsEsccancellation 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
Esccancellation for the current supported foreground task without exiting the whole CLI
Architecture Overview
agent_loop.py: tiny entrypoint that boots the runtimexiaoman_agent/runtime.py: main orchestration loop, model calls, tool execution, streaming, and turn lifecyclexiaoman_agent/team.py: teammate lifecycle, inbox protocol, approvals, and team coordinationxiaoman_agent/tasks.py: task board, ownership, dependencies, and task statexiaoman_agent/worktrees.py: git worktree creation, binding, and isolated execution supportxiaoman_agent/ui.py: Rich-based terminal rendering, message flow, wait rendering, and input presentationxiaoman_agent/cancel.py: foreground cancellation state andEsclistener logicxiaoman_agent/skills.py: loading and exposing reusable skillsxiaoman_agent/compact.py: context compaction behaviorxiaoman_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:
agent_loop.pyxiaoman_agent/runtime.pyxiaoman_agent/team.pyxiaoman_agent/tasks.pyxiaoman_agent/worktrees.pyxiaoman_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.