界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
开发与工程 · SKILL.md

vibecoding-general-swarm

vibecoding-general-swarm

【开发与工程】General-purpose coding orchestration. MANDATORY for ANY coding task not covered by vibecoding-webapp-swarm. Skip ONLY if the task matches vibecoding-webapp-swarm or is entirely non-coding.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: vibecoding-general-swarm description: > General-purpose coding orchestration. MANDATORY for ANY coding task not covered by vibecoding-webapp-swarm. Skip ONLY if the task matches vibecoding-webapp-swarm or is entirely non-coding.

Vibecoding General — Multi-Agent Coding Orchestration

Design-first, multi-agent workflow for building general software projects. Covers Python tools, data pipelines, ML systems, APIs, games, bots, CLI apps, mobile apps, and any non-frontend coding task.

Core Principles

  1. Spec-first. A specification document (SPEC.md) is written before any implementation. It is the single source of truth for architecture, modules, interfaces, and data flow.
  2. Spec fidelity. Subagents implement SPEC.md faithfully — exact interfaces, exact module boundaries, exact data formats. No unilateral changes.
  3. Main agent owns init, merge, integration. Subagents implement and commit on their branches.
  4. Parallelism by modules & features. Each subagent owns one or more modules/features. Group related work into one agent to minimize cross-agent dependencies.
  5. Git worktrees isolate work. Each subagent works in its own git worktree to avoid conflicts. The shared repo is the coordination hub.
  6. Interface contracts are sacred. Define function signatures, data schemas, and file formats in the spec. Subagents implement to these contracts so modules integrate cleanly.
  7. Test before merge. Each subagent runs tests for their module before committing. Main agent runs integration tests after merge.

Mode Selection

ConditionMode
Task has 3+ distinct modules or componentsMode A (multi-agent)
Task involves both infrastructure and application logicMode A (multi-agent)
Task is a full system (API + workers + data pipeline, etc.)Mode A (multi-agent)
Task is a single script, tool, or focused featureMode B (single agent)
Task is a bug fix or small enhancementMode B (single agent)
When in doubtMode B (single agent)

Worktree Reference

Git worktrees allow multiple agents to work on the same repo simultaneously without conflicts.

Setup (each subagent)

cd /mnt/agents/output/project                         # shared repo
git worktree add $HOME/work-<branch> <branch>     # create local worktree
cd $HOME/work-<branch>                            # work here

Rules

  • Each subagent MUST use a unique worktree path ($HOME/work-<branch>).
  • Never run git worktree prune — it destroys other agents' worktree entries.
  • Never work directly in /mnt/agents/output/project — always use a worktree.
  • After committing in a worktree, the main agent merges from the shared repo.

Cleanup (main agent, after all merges complete)

cd /mnt/agents/output/project
git worktree remove $HOME/work-<branch>   # optional, per branch

File Layout

/mnt/agents/output/
├── SPEC.md                        # Project specification (single source of truth)
├── info.md                        # Research findings (if applicable)
├── project/                       # Shared git repo (coordination hub)
│   ├── .git/
│   └── ...                        # Project structure per SPEC.md
│
$HOME/work-<branch>/               # Local worktree (unique per subagent)
└── ...                            # Same structure, isolated working copy
我的收藏