Multi-Agent Orchestration - Setup Guide¶
Source: cat-xierluo/legal-skills (32+ installs) Category: Engineering / Agent Orchestration Quality Tier: 🔵 Community
A comprehensive PM-style multi-agent orchestration framework for running parallel agent workers with isolated worktrees, tmux sessions, and structured Wave-based dispatch. Supports Claude Code, Codex, OpenCode, and custom CLI workers with provider-aware model routing, escape-prevention gates, and checkpoint-based monitoring.
Installation¶
npx skills add cat-xierluo/legal-skills --skill multi-agent-orchestration
Prerequisites¶
| Requirement | Details |
|---|---|
| Git | Worktree support for isolation |
| tmux | Session management for parallel workers |
| Agent CLI | Claude Code, Codex, OpenCode, or custom CLI |
| Bash | Script execution environment |
Key Capabilities¶
PM-Worker Architecture¶
Current session acts as PM (Project Manager): reads task sources, groups work, spawns isolated workers, monitors progress via checkpoint files, and merges results. Workers execute in isolated worktrees with their own branches.
Wave-Based Dispatch¶
Tasks are grouped into Waves - sets of parallel workers running against the same base ref. PM plans provider slots, concurrency limits, and risk tiers before launching. Supports 4-6 parallel workers for independent tasks, 1-3 for shared-dependency work.
Escape Prevention Gate¶
When user requires tmux/independent sessions, PM must pass a startup gate before touching business code: create worktree → start tmux → verify session alive → confirm STATUS.json appears → only then proceed.
Provider-Aware Routing¶
Worker backends include Claude Code (with provider registry for multi-provider routing), Codex, OpenCode, WorkBuddy/CodeBuddy, QoderWork CN, and custom CLI. Per-worker model selection, provider slot allocation, and concurrency management.
Checkpoint Protocol¶
Workers write STATUS.json, RESULT.md, and PATCH_SUMMARY.md to a Session Context directory. PM monitors via pm-monitor.sh or sentinel.sh for stale workers, blocked tasks, and completion signals.
Quick Start¶
# 1. Preflight dependency check
bash scripts/check-dependencies.sh
# 2. Create worktree and session
bash scripts/spawn-worker.sh \
--project /path/to/repo \
--branch feat/my-feature \
--session my-worker-1 \
--worker-backend claude-code \
--model "your-model" \
--verify-cmd 'npm run typecheck' \
--command "claude --permission-mode auto"
# 3. Verify worker alive (30-second postflight)
tmux has-session -t my-worker-1
tmux capture-pane -t my-worker-1 -p | tail -5
# 4. Send worker prompt via tmux
tmux send-keys -t my-worker-1 -l -- "Your task prompt here..."
tmux send-keys -t my-worker-1 Enter
# 5. Monitor progress
bash scripts/pm-monitor.sh --project /path/to/repo --once
Verification¶
# Check skill installed
npx skills list | grep multi-agent-orchestration
# Verify tmux available
tmux -V
# Check worktree support
git worktree list
Notes¶
- Comprehensive 8-section SKILL.md covering: boundaries, execution modes, role/backend selection, Wave orchestration, naming conventions, session context, worker prompt templates, and monitoring/intervention
- Personal routing preferences via
config/orchestration-personal.json(gitignored, not committed) - Default: same-host workers (Claude Code PM → Claude Code workers). Cross-tool only with explicit justification
- Supports lightweight mode (
--no-worktree) for non-git targets or folder-isolated workers - Worker prompt template at
templates/worker-prompt.mdwith Bootstrap-only and Full variants - Built-in sentinel monitoring with harness re-invoke on worker completion
- Quality tier 🔵 Community: 32 installs, but extremely comprehensive (v1.18.4) with extensive production battle-testing
Integration with CorpusIQ¶
CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.
This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.