version: "1.0.0" name: clean description: Reclaim RAM, VRAM, and disk - kill orphaned dev servers, unload idle models, purge package caches - without touching your editors, terminals, or this Claude Code session. disable-model-invocation: true allowed-tools: ["Bash"] argument-hint: "[dust|sweep|scrub|wipe|nuke]"

superclean clean

Reclaim resources. This skill has side effects: it kills processes and deletes files. Never skip the dry run, and never act without the user's confirmation.

The ladder

Tiers are additive. Each includes everything lighter.

$ARGUMENTS may name a tier. If it is empty, default to sweep, which is the tier that handles the common complaint (leaked processes and pinned VRAM) without touching caches.

Procedure

1. Dry run first. Always.

uvx superclean-cli <tier> --dry-run --json

2. Show the user exactly what it would do. Name the processes it would kill with their RAM, the models it would unload with their VRAM, the caches it would purge with their sizes, and the total it would reclaim. Prose, not a JSON dump.

3. Ask for confirmation. Then, and only then:

uvx superclean-cli <tier> --yes --json

--yes is what makes it non-interactive; without it the CLI prompts and will hang here.

4. Report what was actually reclaimed, using the reclaimed field of the result. If a step failed or was skipped, say so. Do not round a failure up into a success.

Hard rules

• Never run `nuke` on your own initiative. It is destructive (Docker reset,

Windows.old) and the CLI requires the user to type NUKE. If they ask for it, run the dry run, show them, and tell them to run superclean nuke themselves in their own terminal. Do not pass --i-know for them.

• Never pass `--force-unlock`. It exists to let a run proceed alongside a live one.

Nothing good comes of an agent using it.

• Do not manually `pkill`/`taskkill` anything as a substitute or a follow-up. That is

the entire failure mode superclean exists to prevent: pkill -f node kills the orphaned dev server, and also the editor's language servers, the MCP servers, and this session's own runtime.

Why this is safe to run from inside Claude Code

superclean's perimeter is fail-closed: anything it cannot confidently classify is treated as protected. It shields editors, terminals, multiplexers, interactive shells, AI coding tools and the Ollama daemon by name; any process whose command line reveals it is an MCP server or agent runtime; every descendant of anything protected, transitively; and its own entire ancestor chain, which includes this session. Kills are re-validated against the process start time immediately before signalling, so a recycled PID is skipped.

Run uvx superclean-cli protected to show the user the full shield list.