[BUG] Cowork VM root partition 86% full on fresh image — only 940 MB free for user work

Summary

The Cowork VM ships with a 9.6 GB root partition, of which ~8.2 GB (86%) is consumed by pre-installed system packages before the user does anything. This leaves ~1.4 GB for user work, logs, caches, and session data — a margin that is quickly exhausted during normal use, especially with plugins installed. When the disk fills completely, Bash cannot execute (cannot create /tmp entries), making the VM effectively dead.

This consolidates findings from #22543, #29448, #30751, and #34602 with fresh diagnostic data captured from inside the VM.

Environment

• Platform: macOS, Claude Desktop (Cowork mode)
• VM: Ubuntu 22.04.5 LTS, kernel 6.8.0-94-generic x86_64
• Root disk: /dev/nvme1n1p1 — 9.6G total, 8.2G used, 1.4G free (86%)
• Session disk: /dev/nvme0n1 — 9.8G total, 1.3M used, 9.3G free (nearly empty)
• Plugins installed: productivity, engineering, design, product-management, cowork-plugin-management

Root Cause Analysis

Where the 9.6 GB goes

| Category | Size | % of disk | Mutable? |
|---|---|---|---|
| System packages (/usr) | ~6,400 MB | 67% | No |
| Snap (dual revisions) | ~1,400 MB | 15% | No |
| Logs + caches (/var) | ~650 MB | 7% | Grows over time |
| Boot + EFI | ~210 MB | 2% | No |
| Available for user work | ~940 MB | 10% | — |

Key findings

1. Massive pre-installed footprint (~6.5 GB immutable)

| Component | Size | Needed? |
|---|---|---|
| /usr/lib/firmware | 1,131 MB | VM uses virtual hardware — most blobs are dead weight |
| /usr/local/lib/python3.10/dist-packages | 717 MB | 181 pip packages, includes duplicate opencv (both opencv-python 115MB AND opencv-python-headless 81MB) |
| /usr/lib/modules | 599 MB | Kernel modules |
| /usr/lib/libreoffice | 319 MB | For docx/pptx |
| /usr/share/texlive | 286 MB | For PDFs |
| /usr/local/lib/node_modules_global | 192 MB | Global npm |
| /usr/lib/jvm | 180 MB | Java |

2. Snap dual-revision retention wastes ~400 MB

Each snap keeps two revisions (current + previous). LXD (184 MB × 2) is likely unnecessary in a sandboxed VM.

3. Plugin cache duplication (#30751)

Each session copies all installed plugins to /sessions/<name>/mnt/.local-plugins/. Never cleaned up. Over ~80+ sessions → disk full.

4. Per-session package reinstallation (compounding factor)

pip install / npm install writes to root partition (/usr/local/lib/). Packages don't persist across sessions, so users reinstall them every time. Each session burns tens to hundreds of MB on the already-tight root partition. For power users (agentic workflows with multiple sessions/day), this exhausts the disk in days, not weeks.

There is no way to install packages onto the /sessions volume (10 GB, nearly empty), persist packages across sessions, or pre-configure a session with required packages.

Symptoms

1. Total Bash failure — all commands fail with ENOSPC or cannot create /tmp. Complete loss of tool execution.
2. Cross-session contamination — affects ALL open Cowork sessions simultaneously (shared root filesystem).
3. No warning — first sign is a cryptic Bash failure. No proactive disk usage notification.
4. Catch-22 — when disk is full, the tools needed to diagnose (df, du) also can't run.

Remediation Timeline

What did NOT help (on its own)

• ❌ Restarting Claude Desktop (multiple attempts)
• ❌ Rebooting the computer (multiple attempts)
• ❌ Deleting old conversations from Claude Desktop UI alone
• ❌ Time (hours between attempts made no difference)
• ❌ Moving sessiondata.img to backup location (without deleting it)
• ❌ Continuing existing sessions after sessiondata.img reset

What DID help

Deleting older sessions in Cowork view + deleting sessiondata.img (not just moving it) + restarting Claude Desktop + opening a new session. All steps were required.

rm ~/Library/Application\ Support/Claude/vm_bundles/claudevm.bundle/sessiondata.img
# Restart Claude Desktop, open NEW session

Observation: Reclaimed disk space was roughly proportional to the number of sessions deleted — confirming session data accumulates and is never auto-cleaned.

Recommendations

Short-term (image optimization)

1. Increase root partition to 15–20 GB — 9.6 GB leaves no margin
2. Remove dead weight from base image:

• /usr/lib/firmware (1.1 GB) — virtual hardware doesn't need firmware blobs
• lxd snap (370 MB with dual revision) — not needed in sandboxed VM
• /usr/share/doc (170 MB) — not needed at runtime
• Duplicate opencv-python-headless (81 MB)
• Set snap refresh.retain=1 to keep only one revision
• Potential savings: ~1.7 GB (86% → ~68% baseline)

1. Aggressive log rotation: journalctl --vacuum-size=50M (currently 234 MB), clean APT cache

Medium-term (session lifecycle)

1. Implement session cleanup — clean /sessions/<name>/ on session end (core fix for #30751)
2. Add disk usage monitoring — warn users at 90% before the catch-22 hits
3. Self-service reset button in Claude Desktop settings (instead of manual sessiondata.img deletion)

Long-term (architecture)

1. Move session data to session partition — /sessions volume (10 GB) is 99.99% empty while root is 86% full
2. Shared read-only plugin mounts instead of per-session duplication
3. Package persistence mechanism — persistent /usr/local overlay or requirements.txt-style auto-install

Diagnostic Data

<details>
<summary>Full <code>df -h</code></summary>

Filesystem       Size  Used Avail Use% Mounted on
/dev/nvme1n1p1   9.6G  8.2G  1.4G  86% /
tmpfs            2.0G     0  2.0G   0% /dev
tmpfs            392M  940K  391M   1% /run
/dev/nvme1n1p15  105M  6.1M   99M   6% /boot/efi
/dev/nvme0n1     9.8G  1.3M  9.3G   1% /sessions

</details>

<details>
<summary>Block devices (<code>lsblk</code>)</summary>

NAME         MAJ:MIN RM  SIZE RO TYPE MOUNTPOINTS
loop0          7:0    0 63.8M  1 loop /snap/core20/2682
loop1          7:1    0 63.8M  1 loop /snap/core20/2717
loop2          7:2    0 91.6M  1 loop /snap/lxd/38331
loop3          7:3    0 91.7M  1 loop /snap/lxd/38469
loop4          7:4    0 48.1M  1 loop /snap/snapd/25935
loop5          7:5    0 48.4M  1 loop /snap/snapd/26382
nvme1n1      259:0    0   10G  0 disk
├─nvme1n1p1  259:1    0  9.9G  0 part /           ← root (the problem)
├─nvme1n1p14 259:2    0    4M  0 part
└─nvme1n1p15 259:3    0  106M  0 part /boot/efi
nvme0n1      259:4    0   10G  0 disk /sessions   ← nearly empty

</details>

<details>
<summary>bindfs mounts (15 per session)</summary>

/dev/nvme0n1 on /sessions type ext4 (rw,nosuid,nodev,relatime)
bindfs on /sessions/.../mnt/.skills type fuse.bindfs (ro)
bindfs on /sessions/.../mnt/.cowork-lib type fuse.bindfs (ro)
bindfs on /sessions/.../mnt/.cowork-perm-req type fuse.bindfs (rw)
bindfs on /sessions/.../mnt/.claude type fuse.bindfs (rw)
bindfs on /sessions/.../mnt/uploads type fuse.bindfs (ro)
bindfs on /sessions/.../mnt/<workspace> type fuse.bindfs (rw)
bindfs on /sessions/.../mnt/.auto-memory type fuse.bindfs (rw)
bindfs on /sessions/.../mnt/.local-plugins type fuse.bindfs (ro)
+ 5 plugin .mcpb-cache bindfs mounts (rw)

</details>

<details>
<summary>Top pip packages by size (717 MB total)</summary>

115M  opencv_python.libs
 81M  opencv_python_headless.libs    ← both variants installed (mutually exclusive)
 77M  imageio_ffmpeg
 77M  cv2
 49M  onnxruntime
 45M  pandas
 32M  numpy
 30M  sympy
 27M  numpy.libs
 22M  matplotlib
 21M  fontTools
 14M  pillow.libs
 14M  cryptography
 13M  tabula

</details>

---

All diagnostics captured live from inside an active Cowork VM session on 2026-03-23. The VM was functional only because sessiondata.img had been manually deleted and regenerated hours prior.