What problem does it solve? Running an AI agent unattended over a backlog of work is risky: the agent may pick stale or blocked items, invent work, lose track of stop conditions after context compaction, or silently edit its own rulebook. This Skill provides a disciplined driver loop that selects work only through a queue adapter, enforces mechanical stop conditions stored on disk, and journals every decision so long autonomous runs stay auditable and bounded. ## Core Features & Use Cases - Adapter-based queue selection: Picks the next work item through a neutral queue core with pluggable adapters (plan-md, github-issues, jira), applying blocker, trigger, owner, and lifecycle filters in order. - Run-level stop conditions: Stops the session on queue unreadability, runtime regression, kill switch, revalidation holds, consecutive escalations, budget exhaustion, or an empty queue — with state persisted in a run directory, not in memory. - Safety gating: Arms an unattended flag with an explicit allow-list before touching rulebook paths, verifies the flag after arming, and runs premise checks both before implementing an item and before shipping its prose. - Use Case: You have a PLAN.md backlog and want Claude Code to work through it overnight. Invoke this Skill at session start: it runs preflight checks, declares a run directory, selects items one at a time, escalates failures, and stops cleanly on a named condition with a full journal. ## Quick Start Start an autonomous run by asking the agent to use the loop skill to pick the next queue item and work through the backlog under the autonomy rules.
