disk-rotate

name	disk-rotate
description	Per-target retention + cleanup for noisy folders (claude-mem logs, bun cache, uv cache, old claude installer versions, session JSONLs, plugin marketplaces). Companion to /disk-audit (which is read-only). Dry-run by default; destructive only with explicit --apply --target X. Each target has a verification gate. Never schedules itself.
version	1.1
updated	"2026-05-06T00:00:00.000Z"
tags	["architecture","infrastructure","active"]
effort	medium
disable-model-invocation	true

/disk-rotate — Per-Target Retention + Cleanup

Usage

# Dry-run (default) — list candidates per target with policy + estimated recovery
/disk-rotate

# Per-target dry-run — focus on one bucket
/disk-rotate --target claude-mem-logs
/disk-rotate --target bun-cache
/disk-rotate --target uv-cache
/disk-rotate --target claude-versions
/disk-rotate --target session-jsonls
/disk-rotate --target plugin-cache

# Per-target destructive — explicit --apply per target, never global
/disk-rotate --target bun-cache --apply

No global --apply --all flag exists by design. Mike approves each target separately.

Targets + retention policy

Target	Policy	Verification gate (must pass before --apply)	Recovery
claude-mem-logs	Delete logs >2 days old	Sample 3 dates from each log → query mcp-search; require ≥1 observation per date	~250 MB
bun-cache	Wipe .bun/install/cache (canonical: bun pm cache rm)	Confirm no process has .bun/install/cache open via lsof	~900 MB
uv-cache	uv cache prune --force (lock blocks indefinitely on persistent MCP daemons)	None — chroma-mcp + claude-batch-mcp holders are expected, not a blocker	unknown; most cache is in-use install paths for persistent MCPs (~fixed cost)
claude-versions	Keep current (symlink target) + previous; remove older	readlink /root/.local/bin/claude → exclude active; lsof confirm older not in use	~250 MB per old version
session-jsonls	Delete .claude/projects/*.jsonl files >30 days old	List candidates; require Mike eyeball before --apply	~500 MB
plugin-cache	.claude/plugins/cache/ + .claude/plugins/marketplaces/ stale entries (>14 days)	List per-plugin; require Mike approval	~500 MB
claude-debug	.claude/debug/ + .claude/cache/changelog.md if >30 days	None — pure debug ephemera	~few MB
pm2-logs	Truncate (not delete) /root/.pm2/logs/*.log >10 MB	None — pm2 keeps writing to truncated files	~few MB rolling

Verification gates (universal — always run, regardless of target)

1. Disk free check — print df -h /root before AND after; if free space change is unexpected (less than 50% of estimate, or NEGATIVE), abort and report.
2. lsof open-file check — for any target involving file deletion, run lsof on candidate paths first; abort if anything has them open.
3. Process-still-alive check — for verifications that depend on background processes (claude-mem worker for claude-mem-logs, MCP servers for uv-cache), confirm the worker pid is alive AND idle (not mid-write to its destination DB).
4. AUDITS.md row — every --apply run appends a row to docs/AUDITS.md Run history table.

Pattern — what each target does in --apply mode

claude-mem-logs

1. List all .claude-mem/logs/claude-mem-*.log files; group by named date.
2. For each log named for date D where D ≤ today − 2:
   • Query mcp-search for ≥1 observation dated D.
   • If found → safe to delete; print path + size.
   • If not found → skip this log + warn ("autodream may not have processed; will retry next run").
3. Print combined recovery estimate. Wait for explicit Mike confirmation in chat before rm.
4. rm confirmed-safe logs.
5. Append AUDITS.md row.

bun-cache

1. lsof +D /root/.bun/install/cache → must be empty.
2. Confirm bun binary at /root/.bun/bin/bun (not in cache path).
3. Run bun pm cache rm.
4. Append AUDITS.md row.

uv-cache

Persistent-MCP framing (S204 lesson): chroma-mcp + claude-batch-mcp are long-lived MCP daemons. Claude Code spawns them at session start and they immediately re-attach to the uv cache. "Fresh session boot" does NOT release them. Plain uv cache prune blocks indefinitely on the cache lock these daemons hold. The 1.1 GB cache is the fixed cost of running these MCPs — most of it is reachable, in-use install paths, NOT prunable.

1. lsof +D /root/.cache/uv → list holders for the record. chroma-mcp + claude-batch-mcp expected; this is NOT a blocker.
2. Run uv cache prune --force (override the cache-lock wait; still safe with active processes — only removes unreachable objects).
3. Verify recovery via du -sh /root/.cache/uv before/after; expect modest delta (KB-MB range, not GB).
4. If recovery <10 MB: don't re-run until tools are uninstalled. The cache is fixed cost.
5. Append AUDITS.md row.

claude-versions

1. readlink -f /root/.local/bin/claude → resolve current.
2. List all /root/.local/share/claude/versions/* files.
3. Sort by version semver; keep the active version + the previous 1; mark older for delete.
4. lsof each candidate; abort if anything has them open.
5. Print delete list with sizes. Wait for explicit Mike confirmation.
6. rm confirmed list.
7. Append AUDITS.md row.

session-jsonls

1. List .claude/projects/-root-GodWorld/*.jsonl files.
2. Group by mtime; mark files >30 days old.
3. Print candidate list (path + size + age + first-line preview).
4. Wait for explicit Mike eyeball + confirmation. These contain session transcripts; the destructive call belongs to Mike, not the verification gate.
5. rm confirmed list.
6. Append AUDITS.md row.

plugin-cache

1. List .claude/plugins/cache/*/ + .claude/plugins/marketplaces/*/ directories.
2. Mark dirs whose mtime is >14 days as stale.
3. Wait for Mike confirmation per dir (or accept blanket "delete all marked" with explicit --blanket-stale).
4. rm -rf confirmed dirs.
5. Append AUDITS.md row.

claude-debug

1. rm /root/.claude/debug/* if mtime >30d.
2. Append AUDITS.md row. Lower-ceremony than other targets — debug files are pure ephemera.

pm2-logs

1. For each log under /root/.pm2/logs/*.log larger than 10 MB:
   • truncate -s 0 <path> (pm2 holds an open fd; truncating preserves the fd while clearing content).
2. Append AUDITS.md row.

What it does NOT do

• Does not schedule itself. Always manually invoked.
• Does not auto-delete anything in dry-run mode (the default).
• Does not touch:
  • .claude-mem/chroma/ (the vector DB itself — back up before ANY destructive action)
  • .claude-mem/claude-mem.db (the digested observations DB)
  • GodWorld/ working tree (use git commands for that)
  • .git/ directories anywhere
  • System paths (/usr, /var, /opt, /etc)
  • Backups (/root/GodWorld/backups/ — managed by scripts/backup.sh rotation)

Companion skills

• /disk-audit — read-only inventory + classification + report. Run this FIRST to identify candidates; then /disk-rotate --target X against the surfaced bucket.
• /md-audit — existence-staleness for .md files. Sometimes surfaces orphan markdown that becomes a /disk-rotate target.
• /health — lightweight check; will print disk-pressure signal if >85%.

Plan + acceptance

Plan file: this skill is the operationalization of the S203 cleanup pattern that recovered ~1.2 GB by running canonical commands manually (bun pm cache rm, manual rm of stale logs + claude versions). The skill encodes those gates so future runs are deterministic + Mike-supervised.

Future extensions (not built):

• --target playwright-binaries / --target puppeteer-binaries (~1 GB combined) — requires audit of which MCPs use them; if MCP is uninstalled, browser binary is dead weight.
• --target node-modules-stale — for GodWorld project deps not touched in N days, surface for manual bun install --clean cycle.
• --target chroma-checkpoint — copy .claude-mem/chroma/ to /tmp/<DATE>/ before any chroma-touching destructive action elsewhere.

Tracking

Every --apply run appends a row to docs/AUDITS.md Run history. That table is the single index of cleanup actions across all audit skills.