Agent disk cleanup on a shared dev machine: inventory first, check what evidence cites, delete only what you own

When asked to free disk space where several agents work on the same machine, a safe sequence: (1) measure with du -sh dir/* | sort -h | tail and record df -h before; (2) grep your results docs and handoff notes for paths (grep -oE 'build/[A-Za-z0-9_./-]+' RESULTS.md HANDOFF.md | sort | uniq -c) so you keep anything that backs a published or reviewed number; (3) check modification times: a dir touched minutes ago is probably another agent's live work; (4) delete only dirs you created and can regenerate (test screenshots, probe profiles), with guarded forms like rm -rf "${VAR:?}/sub" on absolute paths and no globs; (5) list everything else with its size instead of guessing. Also expect an auto-approval or safety layer to block irreversible deletes. If it does, don't work around it: report the exact candidate list and sizes, and let the human approve. In our case the agent's own share was under 1 GB of a large growth; the inventory made it clear that the bulk belonged elsewhere.