agents-system-setup

agents-system-setup

A multi-runtime plugin that bootstraps, updates, improves, or replicates a complete multi-agent system across GitHub Copilot CLI, Claude Code, OpenCode, and OpenAI Codex CLI — from a single skill, with a Canonical IR for bidirectional replication and parallel-aware orchestration baked in.

What it generates

• AGENTS.md at repo root with Directory Architecture, Agent Roster (with parallel-safety waves), Capability Matrix.
• Orchestrator + N subagents (3–50, sized to scope) emitted in the right format for every selected runtime, with a fan-out clause so parallel-safe subagents always run in one wave.
• AGENT-TEAMS.md for Claude Code projects when the roster benefits from peer-to-peer teammates (3+ independent concerns).
• Project-scoped skills under each runtime's conventional path.
• Curated plugin / MCP recommendations from vendor-official catalogs, every recommendation tagged [Tier · Vendor] and opt-in per item.
• Mandatory MCP approval gate — no MCP config is ever written without explicit user approval.
• Cross-OS scripts (.sh + .ps1), .gitattributes for line-ending safety.

Modes

Mode	When to use
init	Brand-new repo, no agent artifacts
update	Existing artifacts, regenerate managed blocks non-destructively
improve	Audit existing system → propose checklist of targeted fixes → opt-in apply
replicate	Port agents/skills/MCP from one runtime to another (Copilot ↔ Claude ↔ OpenCode ↔ Codex) via Canonical IR

Install — per runtime

Each runtime has a different install mechanism. The repo ships the right manifest for each.

GitHub Copilot CLI

copilot
> /plugin install ytthuan/agents-system-setup
> /agents-system-setup

Reads the root plugin.json and exposes the bundled skill as /agents-system-setup.

Claude Code

claude
> /plugin install ytthuan/agents-system-setup
> /agents-system-setup:agents-system-setup

Reads .claude-plugin/plugin.json. Skills are namespaced as /<plugin>:<skill>. Source: https://docs.anthropic.com/en/docs/claude-code/plugins.

OpenAI Codex CLI

codex plugin marketplace add ytthuan/agents-system-setup
codex
> /plugins        # browse and install agents-system-setup
> @agents-system-setup

Reads .codex-plugin/plugin.json and the marketplace descriptor at .agents/plugins/marketplace.json. Source: https://developers.openai.com/codex/plugins/build.

OpenCode

OpenCode plugins are JS/TS hooks — not skill bundles — so install by clone-and-copy:

git clone https://github.com/ytthuan/agents-system-setup.git
cd agents-system-setup
./scripts/install-opencode.sh project   # or "global"

Windows / cross-platform PowerShell:

git clone https://github.com/ytthuan/agents-system-setup.git
cd agents-system-setup
pwsh ./scripts/install-opencode.ps1 -Scope project

This places the skill at .opencode/skills/agents-system-setup/ (or ~/.config/opencode/skills/ for global).

Usage

Once installed, invoke the skill — no arguments needed; it auto-detects mode:

/agents-system-setup

Or be explicit:

/agents-system-setup init
/agents-system-setup update
/agents-system-setup improve
/agents-system-setup replicate

Parallel subagents & Claude Code agent teams

The generated orchestrator always fans out parallel-safe subagents in one wave (multiple Task calls in a single response), then awaits before the next wave. Parallel-safety is computed automatically from the Directory Architecture — see parallelism reference.

For Claude Code, when 3+ subagents are independent and would benefit from peer-to-peer challenge, the generator additionally emits AGENT-TEAMS.md with the opt-in env var (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1), settings snippet, suggested teammate roster, and a token-cost warning. Source: https://docs.anthropic.com/en/docs/claude-code/agent-teams.

Why a plugin (not a bare skill)?