Multi Agent Coordination Skill

multi-agent-coordination-skill

A [Claude Code Skill](https://docs.claude.com/en/docs/claude-code/skills) that sets up a coordination scaffold for projects where several agent sessions — each holding a distinct, continuing role/zone — share one repository over an extended period.

The scaffold is platform-agnostic: a role is held by whatever session reads its file and makes commits, so Claude Code, Gemini via Antigravity, Cursor or Aider can each hold one. See [`docs/ru/CROSS_PLATFORM_BRIDGE.md`](docs/ru/CROSS_PLATFORM_BRIDGE.md).

That is enforced by the file layout rather than left to good intentions: the project's rules live in **`AGENTS.md`**, the cross-vendor convention governed since December 2025 by the Linux Foundation's [Agentic AI Foundation](https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation), and `CLAUDE.md` is a four-line adapter that imports it and adds the Claude Code specifics (`claude -n `, `claude agents --json`, `.claude/rules/` globs). Claude Code reads `CLAUDE.md` and not `AGENTS.md`, which is exactly why the adapter exists.

It does not compete with a tool's own multi-agent features. Claude Code's agent teams keep their shared task list in `~/.claude/`, scoped to one session and deleted when it ends; this scaffold is the durable half — in git, on every machine, readable by any tool.

It's not a framework or a running service: there's no server, no live process registry, no database. It's a set of markdown templates and two small stdlib-only Python scripts, distilled from a real multi-role project that ran this paradigm for weeks and fixed the specific problems it hit along the way. Those problems — and why the fix looks the way it does — are written up in [`references/rationale.md`](references/rationale.md).

What's in it

• coordination/ templates — role zone files, an append-only decision journal (QUESTIONS.md), cross-role request log (HANDOFFS.md), project-wide a