home / skills / alchaincyf / huashu-skills / huashu-agent-swarm

huashu-agent-swarm skill

safe

This skill enables a git-swarm workflow where multiple autonomous agents collaboratively develop by pulling tasks, coding, and pushing in parallel.

npx playbooks add skill alchaincyf/huashu-skills --skill huashu-agent-swarm

Review the files below or copy the command above to add this skill to your agents.

Files (10)

SKILL.md

3.6 KB

---
name: huashu-agent-swarm
description: 多Agent蜂群并行协作，纯git自组织，适合大型项目开发。当用户提到"蜂群模式"、"多agent"、"并行开发"、"agent swarm"时使用。
---

# Infinite Agent Loop - 无限Agent蜂群模式

> 受Nicholas Carlini用16个Claude实例自主构建C编译器的启发。
> 没有master agent，纯git自组织，每个agent独立认领任务、写代码、推送。

## 触发条件

当用户提到「蜂群模式」「多agent并行」「infinite loop」「agent swarm」「启动蜂群」时使用此技能。

## 前置要求

- tmux（`brew install tmux`）
- claude CLI（已安装）
- git 仓库（已有或新建）

## 使用流程

### Step 1: 描述项目

用户告诉我：
- 项目目录路径（必须是git仓库）
- 项目目标和总体描述
- 初始任务列表（或让agent自行拆解）
- agent数量（默认8个）
- 代码规范和测试命令

### Step 2: 初始化项目

```bash
bash SKILL_DIR/scripts/setup_project.sh <项目目录>
```

这会在项目中创建：
- `AGENT_PROMPT.md` - 从模板生成，需要我根据用户需求定制
- `TASKS.md` - 初始任务清单
- `current_tasks/` - 任务认领目录
- `agent_logs/` - 日志目录

然后我根据 `references/agent-prompt-template.md` 定制 `AGENT_PROMPT.md`，填入项目具体信息。

### Step 3: 启动蜂群

```bash
bash SKILL_DIR/scripts/start_swarm.sh <agent数量> <项目目录>
```

这会：
1. 为每个agent创建 git worktree（共享.git对象库，不浪费磁盘）
2. 创建 tmux session，每个pane一个agent
3. 每个agent进入无限循环：pull → 认领任务 → 执行 → push → 下一个

### Step 4: 打开观测台

```bash
python3 SKILL_DIR/scripts/dashboard.py <项目目录> 8420
```

浏览器打开 http://localhost:8420，可以：
- 实时查看所有agent状态、git log、任务进度
- 查看每个agent的最新日志
- 输入框直接发送指令给agent（写入HUMAN_INPUT.md）
- 一键停止所有agent

也可以用命令行监控：
```bash
# 终端状态
bash SKILL_DIR/scripts/status.sh <项目目录>

# 发送指令
bash SKILL_DIR/scripts/send_input.sh <项目目录> "你的指令"

# 直接进入tmux观察
tmux attach -t swarm-<项目名>
```

### Step 5: 停止

```bash
bash SKILL_DIR/scripts/stop_swarm.sh <项目目录>
```

自动停止所有agent + 合并分支 + 清理worktrees。

## 核心机制

### Git自组织协调
- 每个agent通过 `current_tasks/*.lock` 文件认领任务
- 通过 `TASKS.md` 了解全局进度
- 通过 `git log` 了解其他agent的工作
- 冲突由agent自行解决

### Git Worktree隔离
- 不用多份clone，用 `git worktree` 实现隔离
- 所有worktree共享同一个 `.git` 对象库
- 每个agent在自己的worktree独立工作

### 无限循环
- 每个agent完成一个session后自动开始下一个
- 通过 `git pull` 获取其他agent的最新成果
- 通过 sleep 间隔避免API限流

## 关键配置

| 参数 | 默认值 | 说明 |
|------|--------|------|
| agent数量 | 8 | 可在启动时指定 |
| sleep间隔 | 5秒 | agent_loop.sh中可调 |
| 模型 | claude-opus-4-6 | agent_loop.sh中可调 |

## 风险和应对

| 风险 | 应对 |
|------|------|
| API限流 | sleep间隔 + 可调agent数量 |
| 合并冲突 | AGENT_PROMPT指导小粒度commit |
| 死循环无用功 | 日志监控 + 停止条件 |
| 磁盘空间 | stop_swarm.sh自动清理 |
| 成本失控 | 可在AGENT_PROMPT中限制session数 |

---

> **花叔出品** | AI Native Coder · 独立开发者
> 公众号「花叔」| 30万+粉丝 | AI工具与效率提升
> 代表作：小猫补光灯（AppStore付费榜Top1）·《一本书玩转DeepSeek》

Overview

This skill runs a multi-agent swarm for parallel project development using pure git self-organization. It launches multiple independent agents that claim tasks, work in isolated git worktrees, push changes, and continuously loop until stopped. The setup is lightweight, tmux-driven, and designed for large codebases where concurrent work and autonomous tasking speed up delivery.

How this skill works

I initialize a repository scaffold (agent prompt, task list, task locks, logs) and spawn N agents, each in its own git worktree and tmux pane. Each agent repeatedly pulls, claims an unlocked task via lock files, edits code/tests, commits and pushes, then sleeps before the next loop. A dashboard and CLI tools provide live telemetry, per-agent logs, manual inputs, and coordinated shutdown/merge routines.

When to use it

You want parallel autonomous development across many small tasks
Project is in a git repo and can be partitioned into independent subtasks
You need continuous background contribution from many agents (e.g., writing tests, refactors)
You want reproducible, self-organizing multi-agent workflows without a central master
You need observable, controllable long-running agent loops for large feature sets

Best practices

Prepare a clear AGENT_PROMPT with coding standards, test commands, and merge rules
Start with a moderate agent count (default 8) and increase based on CPU, API quotas, and repo size
Keep tasks small and well-scoped to reduce merge conflicts and wasted cycles
Use the dashboard to monitor logs and stop the swarm if unproductive loops appear
Tune sleep interval and session limits to manage API rate limits and costs

Example use cases

Large project refactor where many files can be improved independently (linting, types, small API changes)
Bulk test generation and repair across a codebase by autonomous agents
Creating content pipelines: generate outlines, draft files, and collect assets in parallel
Prototype a feature set by splitting features into discrete tickets and letting agents iterate autonomously
Continuous improvement: agents detect TODOs and incrementally implement or document them

FAQ

How do agents avoid stepping on each other?

Agents use lock files in current_tasks to claim tasks and git worktrees to isolate changes; they pull frequently and resolve conflicts via prompts in the agent prompt.

Can I control or send commands to agents while running?

Yes. The dashboard and send_input script write HUMAN_INPUT for agents to pick up; you can also attach to the tmux session to interact directly.

What are the main resource risks?

API rate limits, cost from model usage, and disk growth from worktrees. Mitigate by lowering agent count, increasing sleep intervals, and using stop/cleanup scripts.