Cleanup

Claude Code Cleanup

Scan ~/.claude/ for deletable data and help the user reclaim disk space.

Design note: All classification, sizing, active-session detection, safe deletion, and before/after reporting are handled deterministically by the bundled scripts. You (the model) should not compute sizes, build tables, or decide what is safe — just run the scripts and echo their pre-rendered output verbatim.

Step 1 — Scan

bash "${CLAUDE_SKILL_DIR}/scripts/scan.sh"

The script emits a single JSON object. The fields you care about:

• markdown_table — a fully rendered markdown table (Directory / Size / Contents / Safe to delete), sorted by size descending, with Reclaimable total and ~/.claude/ total footer rows already included. Print this verbatim to the user.
• active_sessions_note — single-line note about protected active sessions. Print it below the table if the count is non-zero.
• reclaimable_human — total size of directories marked safe-to-delete (already shown in the table footer; reference it in prose if helpful).
• safe_dirs — array of directory names classified safe-to-delete. Used for -all mode and to build the selection prompt.
• directories[] — per-directory records (name, human, safe_to_delete) for constructing the AskUserQuestion options.

Do not recompute or reformat the table. Just echo it.

Step 2 — Select what to delete

Mode: /cleanup -all

If $ARGUMENTS contains -all, skip the checklist and ask a single confirmation:

"This will delete all safe directories listed above. Proceed? (yes/no)"

Use AskUserQuestion. On "yes", run delete.sh --all in Step 3.

Mode: /cleanup (no arguments)

Use AskUserQuestion to present a checklist. Build the options list from directories[] where safe_to_delete == "yes", in the same order they appear in the scan output (already sorted by size). Each option label should be <name>/ (<human size>).

Record the user's selected directory names — these are the arguments for Step 3.

Step 3 — Delete (deterministic)

Pass the selected names to delete.sh. It handles:

• Hard blocklist — never touched even if passed as args. Covers everything the official ~/.claude/ docs describe as user-owned: scripts/, skills/, agents/, commands/, output-styles/, rules/, memory/ (aka agent-memory/), ide/, plugins/, plus settings.json, settings.local.json, credentials.json, keybindings.json, CLAUDE.md, MEMORY.md, .claude.json, .mcp.json, .worktreeinclude.
• Active-session protection — reads sessions/*.json, checks each pid, preserves any UUID transcript/dir whose session is still running.
• projects/ special-case — preserves memory/, CLAUDE.md, settings*.json, and active session UUIDs; only deletes stale UUID .jsonl files and UUID dirs.
• Before/after sizing and the final report table.

# checklist mode — pass selected names as args
bash "${CLAUDE_SKILL_DIR}/scripts/delete.sh" projects debug telemetry

# -all mode
bash "${CLAUDE_SKILL_DIR}/scripts/delete.sh" --all

Step 4 — Report

delete.sh emits JSON. The only field you need to show the user is markdown_report — a fully rendered before/after table with a Total Memory Saved footer and the new ~/.claude/ size. Print it verbatim. Do not rebuild it.

If skipped is non-empty, print a one-line note: Skipped: <contents of skipped field>.

That's the entire flow: scan.sh → AskUserQuestion → delete.sh → echo markdown_report. No LLM-side computation, no table construction, no size math.