repo-cleanup-loop — the stale-vs-valuable classifier

Inspects local and remote branches, PRs, commits, and worktrees, and reports what's stale versus what's valuable-but-orphaned. Never deletes uncertain work, discards uncommitted changes, or closes someone else's PR without confirmation — and at L1, never deletes anything at all.

Category: A · Repo maintenance · Heartbeat: schedule (weekly) · Level: L1 (report-only) · Cost: Low

Source: Loop #10, "The repository cleanup loop," from Forward Future's Loop Library, original prompt by Matthew Berman.

Quickstart

1. Read LOOP.md — the six parts, the prompt, the limits, the three stops,
   and the source's explicit confirmation rule.
2. repo-cleanup-loop-state.md is your spine — already committed with an
   empty per-item register.
3. Claude Code: point a weekly Routine at .claude/skills/loop-task/SKILL.md.
   OpenCode: rename opencode.json.example → opencode.json, add the
   AGENTS.md line, wire skills/loop-task.md into a cron + wrapper script.
4. Rehearse once, by hand, before the schedule ever fires:
   "Run the repo-cleanup-loop skill now, exactly as the schedule would."
5. Grade the rehearsal with the loop-verifier agent — it spot-checks that
   every "safe-to-remove" claim actually has evidence, and that nothing
   uncertain or unowned was misclassified. PASS → arm the schedule. FAIL →
   fix the skill, never the report.

What's in this kit

Loop Ready checklist

• Provable success condition — the source's own spec: valuable work recovered, remaining state intentional, checked against evidence-backed classification
• Run limit — 15 runs (added by this library)
• Spine written first, committed empty
• Report-only (L1) start — writes only the report + spine + log, deletes nothing
• Human gate placed — the source's own explicit confirmation rule for uncertain work and unowned PRs
• One log line per beat, no silent runs
• Kill switch — loop-pause-all in loop-budget.md

How this relates to the course

This kit uses Forward Future's Loop #10 nearly unmodified. Its own explicit constraint — "do not delete uncertain work, discard uncommitted changes, or close someone else's pull request without confirmation" — is this course's maker-never-grades-itself and human-gate principles made concrete for a genuinely risky operation: cleanup work is exactly where an overconfident loop does irreversible damage. This library ships every kit at L1 report-only regardless of the catalog's target level (kit-state.md; CLAUDE.md rule 4), so this kit goes further than the source's own rule — it never deletes anything at all until a human has watched a real run and promoted it.

---

Full attribution: resources/sources.md.