Session summary

Phase 2 of the sitemap-health-check Friday run. Sam picked option 2 from that run's self-audit: the Cowork Linux sandbox kept running out of disk, which crippled crawling. He asked for a skill plus a frequently-running task to clean and fix it, and to run it now.

What was done

Built the sandbox-disk-hygiene skill. It diagnoses disk use, purges regenerable caches (pip, npm, uv, typescript, __pycache__), reclaims tmp directories left behind by finished Cowork sessions, quarantines anything ambiguous (reversible, 7-day retention), logs each run to mnt/Projects/_logs/sandbox-disk-hygiene/, and self-audits. It never touches any mnt path (Sam's real synced files).

Skill file saved to mnt/Projects/sandbox-disk-hygiene/SKILL.md. It could not be installed live into mnt/.skills/skills/ because that directory is read-only from a scheduled-task session.

Drafted two scheduled tasks, fully specified, in mnt/Projects/sandbox-disk-hygiene/: scheduled-tasks-spec.md, task-prompt-recurring.txt, task-prompt-run-now.txt.

• Recurring: taskId sandbox-disk-hygiene, cron 0 2,8,14,20 * * * (every 6 hours, offset to land before the 05:00 image-audit and 23:00 content windows).
• One-time: taskId sandbox-disk-hygiene-run-now, the immediate fix. It fires in a fresh session whose temp root is clean, so Bash works even when the volume is wedged.

Why the session ended / blockers

Two environment limits hit, neither breakable from this session type:

1. Scheduled tasks cannot be created from inside a scheduled-task session ("Cannot create scheduled tasks from within a scheduled task session"). This run is the sitemap-health-check task, so the two tasks could not be created here.
2. mnt/.skills/skills/ is read-only from a scheduled-task session, so the skill could not be installed live.

This session's own sandbox is also ENOSPC-wedged: Bash cannot start because it cannot create its per-call temp dir. Cleanup could not run inline. That is exactly why the fix is delivered as a task: a scheduled task spawns a fresh session with a clean temp root.

Root cause for future agents

When the shared sandbox volume hits 100%, the Bash tool wrapper fails at mkdir /sessions/<id>/tmp/claude-NNNN. The session cannot run the commands it would need to clean up. Prevention beats cure: keep the volume below the wedge point with frequent cleanup.

Recommended next actions

1. From any regular (non-scheduled) Cowork session: create both tasks using the specs in mnt/Projects/sandbox-disk-hygiene/scheduled-tasks-spec.md, and copy the skill folder into mnt/.skills/skills/sandbox-disk-hygiene/. Both are one-step actions.
2. Consider a one-line "load sandbox-disk-hygiene first" preflight on disk-heavy tasks (screaming-frog-crawl, image-audit, injury-content-series). image-audit already has a disk-cleanup preflight; this skill can standardize it.
3. After the recurring task has a few runs logged, review mnt/Projects/_logs/sandbox-disk-hygiene/ for a recurring offender path and tune thresholds if needed.

Handoff for a new agent