What problem does it solve? Long unattended agent loops degrade in quality because each iteration re-invokes the same conversation, so late iterations work through a context consumed by earlier ones. This Skill moves the loop outside the session: a host process spawns one fresh claude -p per iteration and classifies each outcome from files on disk rather than from the assistant's prose. ## Core Features & Use Cases - Fresh-context iterations: Each loop iteration spawns a new Claude process with a pinned session id, so iteration 20 is as sharp as iteration 1, and every iteration stays reopenable with claude --resume. - Deterministic outcome classification: A pure classifier reads run-state files, pipeline locks, halt-file timestamps, and a git progress oracle to label each iteration progress, done, halt, incomplete, error, or idle. - Budgets, observability, and notifications: Pre-spawn ceilings on iterations, cost, duration, and idle streaks; status/watch pure-reader views; end-of-run macOS or webhook notifications; optional dashboard status frames. - Use Case: Point it at a development roadmap overnight with --adapter develop-next --max-duration 8h --max-cost 20 and it works items sequentially until the frontier empties, a halt occurs, or a budget cap is reached. ## Quick Start Ask the agent to run the loop-supervisor dry-run for the develop-next adapter to print the plan and exact claude argv before starting an unattended roadmap run.
