Tidy scratch — date-aware GC for the scratch notebook

A read → propose → execute sweep. This is the only ccx skill that proposes destructive actions; save-state stays a pure read-only dashboard. The sweep is dry-run by default — it prints a plan and touches nothing until you confirm. If the scratch dir is not version-controlled (the usual setup), a delete is permanent. When in doubt: archive (move), don't delete.

Live data (injected at load)

Project + config: !bun "${CLAUDE_PLUGIN_ROOT}/scripts/where.ts"

Now: !date "+%Y-%m-%d %H:%M %Z"

Folder scan — date-aware, stalest first (newest-file age per folder; STATE=NO = no handoff doc): !bun "${CLAUDE_PLUGIN_ROOT}/scripts/scan.ts"

Open PRs (a matching open PR keeps a thread alive): !gh pr list --limit 30 2>/dev/null || echo "(gh unavailable / not a GitHub repo)"

Then do this

1. Classify every folder from the scan. With ticket_system: none there is no external "closed" oracle — only age + PR signals exist, so stay patient and conservative. age = the scan's age_days; thresholds are defaults you may adjust per item when presenting the plan:

   • KEEP (active) — age < 3 (protects an active spike), OR an open PR that matches the thread's topic/branch.
   • ARCHIVE (likely done) — has a STATE doc AND age > 30 AND no open PR. (Longer threshold than a ticket-oracle setup would use: age alone must be more patient.)
   • ADOPT (unfiled work) — STATE = NO AND files > 3 — too much accumulated work to be throwaway; propose creating a STATE doc for it (offer /ccx:start-thread), never delete in this pass.
   • DELETE (throwaway) — STATE = NO AND age > 14 AND files ≤ 3 (no handoff value + cold + tiny).
   • DELETE (empty) — files = 0.
   • FLAG → default KEEP (unsure) — anything that fits no rule above, and every kind: hub reference note. Never auto-propose deleting a folder that has a STATE doc.

2. Print the PLAN — execute NOTHING yet. One table, stalest first:

   item	class	proposed action	why	last edit (age)

   End with a one-line tally: N delete · M archive · K adopt · rest keep.

3. Confirm. Ask: apply all, a subset (named), or none. Wait for the answer.

4. Execute only what was approved, echoing each action (paths from the injected config block):

   • archive: mkdir -p <scratch_root>/<archive_dir> && mv <scratch_root>/<slug> <scratch_root>/<archive_dir>/, then date-stamp: printf '\n_Archived %s by ccx tidy-scratch._\n' "$(date +%Y-%m-%d)" >> <scratch_root>/<archive_dir>/<slug>/<state_basename>
   • delete: rm -rf <scratch_root>/<slug> — permanent if the notebook isn't version-controlled.

5. Reindex. Run /ccx:save-state so the INDEX reflects the cleanup. Then a ≤3-line summary: deleted / archived counts and the new active-thread count.

Safety rails (non-negotiable)

• Dry-run is mandatory — never delete, move, or prune before the step-3 confirmation.
• Hard rm is only ever proposed for folders with no STATE doc AND ≤3 files. Anything carrying a STATE — or carrying real bulk (>3 files) — is archived or adopted, never deleted; its work survives.
• When unsure, archive, don't delete — without a ticket system there is no authoritative "this is finished" signal, only age.
• Reads the clock every run (scan.ts + date) — no date is ever hardcoded, not even the model's notion of "today". Run it next month and the ages come out next-month-correct.
• Archived threads drop off the save-state dashboard (scans skip the archive dir). A graph view may still show them until the archive dir is excluded from its filter — one-time setup, out of scope for this per-run GC.