dsh-turn-scratch

After every dsh turn, the workspace is littered with temporary files you cannot identify, do not dare delete, and end up manually cleaning after each conversation. dsh-turn-scratch exists to eliminate that friction: at the end of each turn it scans the entire workspace, moves transient files into an isolation folder under the user directory (never inside the workspace itself), lets code-level rules decide what is safe to keep or discard, and escalates genuinely ambiguous files to a single isolated model call. Actual deletion only happens when you explicitly purge the isolation folder, so the worst case is a file being quarantined and easily restored, never a wrong deletion.

It also provides a fixed-path staging area: mid-turn intermediate output is written straight into the staging directory so the workspace is clean from second one, then batch-moved into the per-turn isolation folder at turn end. The staging area is an additional source, not a replacement; the original full-workspace scan runs unchanged, so anything that slipped into the workspace without touching the staging path is still collected at turn-end. Cross-volume transfers follow a strict copy, verify file count and byte total, persist the record, and only then remove the original, meaning if any step fails the original remains intact.

Ideal for anyone who relies on dsh for daily work and wants the workspace to stay pristine without a manual cleanup ritual after every conversation. Zero configuration out of the box; adjust behaviour by adding a config block for turn-scratch in the profile cordis.patch.yml. The plugin is tiny, exposes five conversation-driven tools (status, restore, purge, mark, workdir), and requires no extra interface.