Skip to content

SKILL.md

Latest commit

 

History

History
146 lines (115 loc) · 5.19 KB

SKILL.md

File metadata and controls

146 lines (115 loc) · 5.19 KB
Raw
name vm-cleanup
description Clean up disk space in the Claude working environment (Cowork session VM or Claude Code sandbox). Use this skill when the user says "clean VM", "clean disk", "free disk space", "disk full", "ENOSPC", "no space left", "clean session", "session cleanup", "disk cleanup", or when Bash commands fail with disk-full errors. Also use proactively before heavy operations like npm install or pip install in long sessions.
metadata
version platforms
0.1.0
claude-code
cowork

VM / Session Disk Cleanup

Recover from or prevent disk-full situations in Claude's working environment.

Context

Both Cowork session VMs and Claude Code sandboxes have limited disk space. Long sessions with package installs, builds, and cached data can fill the disk completely. When full, ALL Bash commands fail with ENOSPC errors because the shell can't create temp files — a deadlock where you can't even diagnose the problem.

Cleanup Strategy

Run cleanups in phases — smallest operations first so they succeed even when disk is critically low.

Phase 1: Emergency (needs almost zero free space)

Run these first when disk is completely full or nearly full:

# Remove Python bytecode (tiny files, frees inodes)
find /sessions -name __pycache__ -type d -exec rm -rf {} + 2>/dev/null
find /sessions -name "*.pyc" -delete 2>/dev/null
find . -name __pycache__ -type d -exec rm -rf {} + 2>/dev/null
# Clear temp files
rm -rf /tmp/*.tmp /tmp/*.log /tmp/tmp* /tmp/pip-* /tmp/npm-* 2>/dev/null
# Clear apt/package manager cache
rm -rf /var/cache/apt/archives/*.deb 2>/dev/null
apt-get clean 2>/dev/null

Phase 2: Medium cleanup (clear package caches)

# Clear pip cache
pip cache purge 2>/dev/null
rm -rf ~/.cache/pip 2>/dev/null
# Clear npm cache
npm cache clean --force 2>/dev/null
rm -rf ~/.npm/_cacache 2>/dev/null

Phase 3: Heavy cleanup (removes reinstallable items)

# Remove node_modules (recreate with npm install)
find /sessions -name "node_modules" -type d -exec rm -rf {} + 2>/dev/null
find . -name "node_modules" -type d -exec rm -rf {} + 2>/dev/null
# Remove build artifacts (recreate with build command)
find /sessions \( -name ".next" -o -name "dist" -o -name "build" \) -type d -exec rm -rf {} + 2>/dev/null
find . \( -name ".next" -o -name "dist" -o -name "build" \) -type d -exec rm -rf {} + 2>/dev/null
# Remove global npm packages
rm -rf ~/.npm-global/lib/node_modules/* 2>/dev/null

Execution Steps

  1. Check current disk usage: df -h /
  2. Run Phase 1 commands (always)
  3. Check if space was freed — if yes and enough, stop here
  4. Run Phase 2 commands
  5. Run Phase 3 commands if more space is still needed
  6. Final disk check: df -h /
  7. Report before/after free space to the user
  8. If disk was over 80% full, recommend running this at session start

When Disk Is Completely Full (Deadlock)

If even Phase 1 Bash commands fail with ENOSPC:

In Claude Code: Suggest the user exit and restart Claude Code. The sandbox environment resets on restart.

In Cowork — important architecture note:

Cowork has two separate disks that fill independently:

  1. Mac filesystem (mounted as /sessions/.../mnt/) — the user's actual folder on their computer. Has lots of free space typically.
  2. VM internal disk (everything else: /tmp, ~/.cache, /var/cache/apt, etc.) — a fixed-size overlay filesystem. This is what hits ENOSPC.

When Bash is broken, try these in order:

Step 1 — Clean Mac-side workspace with Desktop Commander: Use mcp__Desktop_Commander__start_process to run rm -rf on the Mac host for large dirs in the user's workspace folder. Common culprits:

rm -rf "/Users/<username>/Documents/Claude/<project>/node_modules"

Find the real Mac path via mcp__Desktop_Commander__start_process with du -sh ~/Documents/Claude/**/node_modules 2>/dev/null | sort -rh | head. Note: Desktop Commander cannot reach VM-internal paths like /var/cache/apt/.

Step 2 — Scan VM internal disk with Glob: Glob can read VM-internal paths even when Bash is broken. Scan for large caches: /var/cache/apt/archives/**, /tmp/**, ~/.cache/**.

Step 3 — allow_cowork_file_delete scope: This tool only enables deletion for paths inside the mounted user workspace (the mnt/ folder). It cannot grant access to VM-internal system paths.

Step 4 — If VM internal disk is the problem: There is no way to clean the VM's internal disk (apt cache, pip cache, npm cache) from outside the VM when Bash is broken, Docker is not installed on the Mac, and Desktop Commander cannot reach those paths. Recommend restarting the Cowork app — this resets the VM entirely. The user's files in their workspace folder are safe and unaffected.

Prevention Tips

Share these with the user when relevant:

  • Run cleanup before heavy installs (npm install, pip install, apt install)
  • Use pip install --no-cache-dir to avoid caching
  • Remove node_modules from old projects before starting new ones
  • For recurring maintenance, pair with a scheduled task that runs every 2 hours
  • In Cowork, the VM internal disk is the bottleneck — the Mac disk rarely fills up