huashu-agent-swarm_skill

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