orchestrator_skill

Overview

This skill is an automated multi-agent orchestrator that spawns CLI subagents in parallel, coordinates state via MCP Memory, and monitors progress end-to-end. It is built for production monorepos and handles planning, execution, verification, retries, and result collection. The orchestrator enforces verification gates and uses retry and clarification-debt policies to maintain quality.

How this skill works

On request it decomposes work into tasks, creates a session and task board in memory, and spawns subagents as independent CLI processes (gemini -p ... --approval-mode=yolo) without exceeding MAX_PARALLEL. It polls progress at configured intervals, runs mandatory verification (oh-my-ag verify) for each completed agent, applies retry logic on failures, and collects result files into a final summary. Memory reads/writes go through configurable MCP tools and ownership rules ensure clear file responsibilities.

When to use it

• When a feature requires multiple specialized agents working in parallel
• When you want end-to-end automated execution without manual agent management
• When delivering full-stack changes spanning backend, frontend, mobile, and QA
• When the user requests “run automatically” or “run in parallel”
• When reproducible session metrics and postmortems are required

Best practices

• Keep MAX_PARALLEL small for resource-constrained CI runners and increase in dedicated environments
• Provide structured task specifications so planners can decompose reliably
• Never bypass the verification gate—automated verify is mandatory for acceptance
• Attach relevant workspace context and previous artifacts to reduce clarification debt
• Monitor CD (Clarification Debt) and trigger RCA if CD thresholds are exceeded

Example use cases

• Implement a full-stack feature: spawn backend, frontend, mobile, QA agents in parallel and collect results
• Run a multi-component migration across services with coordinated verification and retries
• Automated release smoke test: spawn test agents for API, UI, and mobile then aggregate pass/fail reports
• Perform scoped refactor across repo modules where each subagent handles a specific package
• Generate a consolidated developer handoff with lessons-learned if clarification debt is high