Objective
Clean up the Cowork Linux sandbox, which was reporting out-of-space and stalling bash. Build a reusable skill so the next session does not relearn the diagnosis.
/ was 9.3G of 9.6G (98%). /sessions was 9.3G of 9.8G (100%)./var/log/journal, 974M, root-owned/var/tmp/ms-playwright, 956M, root-owned/var/lib/snapd, 430M, root-owned/var/lib/apt and /var/cache/apt, ~434M combined, root-owned/sessions/, each owned by a different UIDsudo is unavailable inside Cowork (no new privileges flag is set)./tmp had 83 files but all owned by other UIDs.Skills/sandbox-cleanup/ with:
SKILL.md (trigger phrases, ordered playbook, safety rules, slash command /sandbox-cleanup)scripts/sandbox_diagnose.sh (non-destructive disk report with short timeouts so it does not stall when the disk is full)scripts/sandbox_clean.sh (reclaims what the session user owns, honestly reports skips, recommends harness reset if usage stays >= 95%)references/permission-map.md (UID and ownership map for every relevant path)Skills/SKILL-INDEX.md.Work complete.
The sandbox's near-full state on / and /sessions cannot be resolved by the session agent under any circumstance. The disk is shared with 376 stale sessions the unprivileged session user cannot touch, plus root-owned /var caches. Only a harness-level recycle reclaims real bytes. Worth a thumbs-down on the next visibly-slow session so Anthropic gets a signal.
/sessions accumulates across sessions./sandbox-cleanup whenever disk pressure appears.