Disk Space Triage: four layers, a ceiling that never moves

Every free-up-space walkthrough on the internet says the same thing: find the biggest folder and delete it. That advice has a hard ceiling nobody computes, and it fails on the first real case - a 53G disk being asked for 50G free.

This skill answers three questions in one read-only pass, in this order:

1. What is eating the drive? Not a leaderboard of the ten biggest folders, but an attribution: which of four layers each one sits in.
2. What can this disk free up at most? A number, not an adjective: capacity minus what may not be touched minus headroom.
3. Is the number you were told to hit achievable? If it is not, say so plainly, because "not achievable" is the conclusion that stops a fourth round of cleanup.

The four layers, and why the distinction is the whole product:

• L1_FLOOR - touch it and the machine will not boot. Windows, Program Files*, pagefile, hiberfil, and Windows\Installer. That last one is 3-8G, looks like the obvious win, and silently destroys every MSI uninstall and repair path. It is not junk.
• L2_LOCKED - held by a running process; a forced delete corrupts it. Close the program or reboot first.
• L3_RENEWABLE - regenerates itself the next time it is used. This is the layer to clear first, because getting it wrong costs a cache re-download.
• L4_MOVABLE - data and cache whose path a link can keep unchanged. Move only if cleaning was not enough, and never before the arithmetic. Anything the classification table does not match goes to UNKNOWN and is treated as do-not-touch, with the reason written out. Guessing here costs somebody else's data.

On the machine this was measured on: Windows 11.2G + Program Files 3.1G + running software puts untouchable at 16G on a 53G disk. The user asked for 50G free. The physical ceiling is about 33G; it was brought to 31G free, which is nearly the floor of what the hardware can give. The correct answer to "50G" was never a better cleanup.

It deletes nothing, moves nothing, creates no link, changes no permission. Apart from the report you explicitly name with --json / --md, it writes nothing at all - and it reads no file contents, only names, sizes and attributes, so it cannot leak what is inside your files.

Includes scripts/triage.py (standard library only, no subprocess, no network call of any kind) and a report guide explaining what each line means and why the total disagrees with Explorer.

What it does not do: it will not clean anything, and it will not move anything. That is the paid companion skill's job, and this one exists to tell you whether that job is worth starting. It also does no partition work - when the target is physically impossible, the advice is add or replace a disk, not resize one.

One pointer, since you are already here

Everything above is a refusal. This skill's job is to say "the number you were asked for is not achievable on this disk" and then stop - no encouragement, no fourth round of cleanup. That is the whole product.

My paid skill makes the same refusal against a different file: your SKILL.md. Prompt Contract Audit runs 23 machine rules plus one human criterion, and every rule prints a verdict instead of an opinion - a references/*.md the body cites that does not exist, a frontmatter name that does not match its folder, a document claiming "read-only" while the code opens a file for writing, a package declaring "offline" while a script opens a socket. It carries eleven incidents in references/INCIDENTS.md, each with the original words and the root cause, because every rule in it started as one of those.

You can score it yourself before believing any of that. It ships 20 fixtures:

python tests/_make_cases.py && python scripts/baseline.py
recall = 1.000   precision = 1.000   evidence assertion = 2/2 passed

Standard library only, no network call at runtime - the same two conditions as this skill.

If you have published something that works on your machine and nowhere else, that is the audit for it. Search for Prompt Contract Audit on Agensi.