Importance: High. Disk exhaustion intermittently blocks Bash and any Bash-dependent scheduled task.
Status: Task repoint DONE. First live run PENDING a one-time Cowork desktop app restart.
Sam asked to repoint the existing folder-cleanup scheduled task so it cleans whatever folder keeps making Cowork run out of space, run it daily, and free space without deleting skills, routines, tasks, etc. Design, run, test.
Diagnosis (this session + prior): the culprit is the Cowork VM's own guest filesystem, not any host folder. The guest disk is a dynamically-expanding VHDX that grows as sessions write scratch and never auto-shrinks. Host C: is healthy at 109 GB free, so host cache cleanup was never the fix. VHDX files (host path C:\Users\SAguiar\AppData\Roaming\Claude\vm_bundles\claudevm.bundle\) probed 2026-05-23: rootfs.vhdx ~9.74 GB, sessiondata.vhdx ~9.78 GB, both at cap.
Actions taken:
folder-cleanup (taskId unchanged). Old job was the host-folder quarantine routine, already superseded by the GitHub Actions pipeline. New job: Cowork VM Disk Cleanup.0 3 * * *). Task stays enabled.fstrim, and flags VHDX compaction if a disk stays above 12 GB./sessions/*/mnt/ (mnt/SAguiar Sam's real folder, mnt/.skills ALL skills, mnt/uploads, mnt/.remote-plugins), scheduled task/routine definitions (host-side in Documents\Claude\Scheduled\, out of the guest by construction), TASKS/todo/handoff files, credentials, and any active or recent session. This satisfies "free space without deleting skills, routines, tasks."Codex\folder-cleanup-runs\ROLLBACK-folder-cleanup-task-2026-05-23.md (full prior cron, description, and prompt to restore exactly).Codex\folder-cleanup-runs\SUMMARY-vm-cleanup-repoint-2026-05-23.md.Attempted a live run. Blocked by an ENOSPC deadlock: the guest filesystem is currently 100% full, so Bash cannot create its temp dir and will not start. No in-guest cleanup tool runs without Bash, and Desktop Commander runs on the Windows host and cannot reach inside the guest VM. This is a one-time chicken-and-egg condition, not a flaw in the new task. The new task includes a self-recovery branch for exactly this state.
Work complete except the first live run, which cannot execute until the guest disk is unblocked once. Surfacing that decision to Sam rather than taking a destructive action on the live VM.
/sessions data and frees enough room for Bash to start. After that single unblock, the daily 3 AM task self-maintains and the deadlock should not recur.Optimize-VHD, VM briefly offline) to reclaim host-side space. Compaction is intentionally not automated.