perpetuum

Perpetuum is a framework skill for setting up, supervising, and controlling a persistent three-layer "explore, execute, escalate" agent loop on a project. It targets tasks whose value grows with how many findings an agent can produce over hours or days, such as bug hunting, writing polish, style distillation, feed watching, and gap scanning, and it also manages the lifecycle of an already-running loop: inspect, pause, resume, stop, or nudge with a new instruction. What ships is an adaptable framework rather than fixed source code; the architecture is fixed but the prompts, trigger scripts, and contract files are templates rewritten per task.

The design separates a producer inner agent (Layer 1, fresh context per dispatch, so it cannot rubber-stamp prior behavior) from a judging middle agent (Layer 2, fresh per cycle) that commits, fixes, or escalates, above a persistent scheduler (`trigger.sh`, Layer 3) and an optional human/host monitor layer. This GAN-like discriminator/generator split prevents the self-certifying failure of simpler loops, with durable state kept in files and git rather than a long-lived conversation. The skill is organized as a router into references for setup, status, control, feedback, git-worktree parallelism, trigger authoring, design rationale, and an optional browser dashboard. It runs on top of cc-use and ships worked examples (adversarial testing, article polish, GitHub watcher, observability-gap scanning, style distillation) with plan/inbox/escalation contract files and trigger scripts.

Target users are developers and power users who want an autonomous, long-running agent that keeps producing findings across sessions while remaining inspectable and controllable, escalating genuine decisions to a human. Because it runs a persistent scheduler that autonomously dispatches shell-driven agent work, commits, and can act on GitHub, it grants meaningful autonomy and should be pointed at scoped projects with human escalation in place.