Reaper

Reduce maintenance burden per unit of effort. Prefer removing a whole unnecessary thing over polishing its parts; do not optimize for deleted line counts.

Choose the workflow

When called from another workflow, run Discover within the caller's scope and return candidates; let the caller choose reporting, edits or cleanup PRs within its authorization. Do not start standalone delivery or worktree cleanup implicitly. For a standalone sweep, discover, execute, deliver and review, then improve from use.

Discover

• Work within the requested scope and budget. If unspecified, choose one bounded area for the first pass. Inspect existing work and ownership; do not propose reaping another agent's unfinished work.
• Read relevant prior outcomes in wiki/rule-provenance.md. Search cheaply for abandoned code, superseded explainers, completed or obsolete specs and repeated knowledge. Investigate the strongest candidates first; age and absent references are clues, not sufficient reasons to delete.
• For code, trace callers, public entry points, configuration and indirect use. For prose, identify its remaining purpose and unique knowledge. Identify useful knowledge that must move to its canonical home before removing a duplicate. Honor local protections, including immutable raw sources. Route agent procedures and task-local steps according to information-home; do not preserve expired instructions as wiki knowledge merely because they were unique to the deleted artifact.
• For unfinished specs, compare their assumptions and intended outcome with later decisions and implemented behavior. Open checkboxes do not establish continuing value. Propose deletion when the direction is obsolete; if remaining intent is unclear, ask whether it is still wanted rather than carrying it forward by default.
• For tests, identify the meaningful failure each catches. Scrutinize tests that mirror implementation, assert incidental details or duplicate coverage. Propose deleting tests without a useful purpose or simplifying setup and assertions while preserving valuable regression checks. Explain what protection remains or is deliberately retired; a passing reduced suite alone does not establish adequate coverage.
• Return concrete candidates with evidence, remaining users or knowledge, required reference repairs and checks. Separate supported deletions from unresolved intent questions. Identify generators or workflows that would recreate the waste.
• Stop at the budget, when a coherent cleanup is identified, or when further hunting offers little likely value. Report a no-op honestly; do not manufacture deletions.

Execute

• Use the caller's chosen delivery route; for standalone sweeps, use a cleanup PR. Inspect existing changes and isolate conflicting work before editing. Delete coherent units, preserve identified knowledge and repair references. Fix generators within scope or report follow-up. Verify affected behavior, links and rendered content.
• Have an independent agent review implemented cleanup, including cleanup made by a reviewer. Follow Deliver and review for that check without requiring a new PR when the caller chose edits to an existing one.

Worktrees

• When worktrees are in scope, inventory them and check owning chat/agent activity, running jobs, PR state, uncommitted and untracked work, unpushed commits and valuable ignored files. Treat age or a merged branch as a lead, not proof of abandonment; leave uncertain ownership for clarification. Exclude the active checkout and primary worktree.
• Have the review agent check proposed cleanup and recovery evidence before acting. Ensure unique work is recoverable, including needed ignored files that archive snapshots omit. Prefer the host's recoverable archive operation for managed worktrees; use normal Git removal for others. Do not force removal to bypass unexplained state. Verify archival/removal succeeded.
• Keep local paths and recovery details in chat, not the public PR. Describe the proposed operation and review outcome alongside the PR; distinguish that operation from its Git diff. During discovery validation, defer archival until the evaluator accepts the result so fresh runs can inspect the same worktrees.

Deliver and review

• For a standalone sweep, use the pull-request skill to open a draft cleanup PR. State what lost its purpose, the supporting evidence and the checks performed.
• Have an independent review agent inspect the actual diff and surrounding repo, with the requested scope and verification results. Ask it to look for lost behavior or knowledge, missed indirect users, broken references and unsupported deletion claims. Require concrete findings; no findings is a valid result.
• Address findings and have the reviewer check the revised diff when fixes change its conclusions. Keep the PR draft if material questions remain or independent review is unavailable. Mark ready only after verification and review; report their limits and leave merging to the user. Keep agent findings in chat unless posting review comments is explicitly authorized.

Improve from use

• Validate discovery with fresh runs from the skill and the same scope, without prior candidate lists or hoped-for deletions in the agent's context. Keep expected targets with the evaluator; compare them with the result afterward. Improve the general method and rerun from the same baseline rather than coaching the cleanup agent toward individual targets. Keep correctness fixes from review distinct from evidence that Reaper discovered something independently.

• Record consequential outcomes in wiki/rule-provenance.md: the search method, evidence of useful yield or wasted effort, false positives, and later restorations when known. Link the cleanup PR; do not create a second deletion ledger.

• Use these outcomes to change search order and remove ineffective instructions, including Reaper's own. Add or strengthen instructions only under the repository's ablation rule; do not infer success solely from passing checks or a quiet review.

• Include justified skill changes in the reviewed diff. Add reusable scripts only when repeated use demonstrates that they save effort. Do not expand deletion authority or weaken review requirements as a self-improvement.