What problem does it solve? Coordinating many AI agents against a shared task queue is error-prone: agents can double-claim work, blow up prompt-cache efficiency with per-task messages, or overload the permission system with unbounded concurrency. This Skill runs a disciplined drain loop over the ok-planner task tracker so a filed run executes to completion without those failure modes. ## Core Features & Use Cases - Tracker drain loop: Repeatedly calls tasks next and acts on its output — starting profile agents, running exec tasks, and stopping at done or waiting. - Cache-optimized dispatch: Every agent of one profile receives the identical fixed message ("Claim your task and finish it."), so the whole first request is one cached prefix per profile per run; task ids reach agents through the atomic tasks claim protocol instead. - Concurrency cap and usage accounting: Runs at most 8 agents at once across all profiles, stamps token usage on each task as its agent returns, and handles orphaned fork tasks via tasks retry. - Use Case: After a planning ceremony files a run of implementation and review tasks into .ok-planner/bin/tasks, invoke /execute-tasks to drain the run — it starts ok-opus, ok-sonnet, and ok-haiku agents as slots free up, runs exec tasks itself, and stops when no open task remains. ## Quick Start Run the /execute-tasks slash command to drain the currently selected task tracker run until it reports done or waiting.
