What problem does it solve? Running an unattended backlog loop safely is hard: an autonomous agent might pick its own work, answer approval gates meant for humans, run concurrent writers, or trust a worker's claim of completion. This Skill encodes one disciplined kanban foreman iteration with hard boundaries so a bare /loop can run unattended without those failure modes. ## Core Features & Use Cases - Queue watch and dispatch: Monitors the backlog file via a checksum-polling watcher, reads the queue with moai todo list --json, and dispatches only the oldest operator-picked card to a single worker spawned with isolation: "worktree". - Evidence-based completion: Advances a card with moai todo done only after reading the worker's evidence file containing verbatim passing check output and named gaps; missing or stale evidence leaves the card picked. - Hard safety boundaries: Never adds or picks cards, never answers approval gates, never pushes or merges, runs only one write-capable worker at a time, and removes AskUserQuestion so the loop cannot stall on a prompt. - Use Case: An operator backgrounds a Claude Code session running bare /loop overnight; each iteration the foreman checks the queue, dispatches one picked card to an isolated worktree worker, collects its evidence on return, and writes a short report the operator reads on reattach. ## Quick Start Run a bare /loop in the project session to start unattended foreman iterations over the moai todo backlog queue.
