Capture Keeper

Capture Keeper is a Uliya Skill for UGREEN AI NAS that runs the full lifecycle of a screenshot stream: sweep, archive, index, audit, purge, recall. It exists for one reason — a NAS is the only machine that should be allowed to read your screenshots, because it never sends them anywhere.

THE PROBLEM

My desktop screenshots folder held 1,889 files. They were filed, sorted, and named by the minute they were captured, and completely unrecoverable: I could not find the error message I captured last March, and I could not tell you which of those 1,889 images contained a bank balance, a one-time code, or an API key sitting in a terminal window.

That second part is the real problem. Screenshots are the only folder on a machine that is simultaneously private and full of secrets, and it grows by itself. Cloud OCR is not an option for that corpus — uploading a screenshot of a recovery seed to a third-party service is the exact harm you are trying to prevent. So the corpus stays unmanaged. A NAS changes the maths: always on, plenty of storage, and a local model.

WHAT IT DOES

Sweep — inventory captures across authorized roots (synced phone gallery, Downloads, Telegram exports, Desktop). Read-only, and reports what exists before touching it.
Archive — file them under a stable contract: fixed-capacity batches of 256, names derived from capture time, collision-safe rollover. Previewed first, applied only after approval.
Index — read new captures on-device so they become searchable by what you remember ("the error from March"), not by filename.
Audit — flag captures carrying credentials, recovery phrases, payment data, IDs, medical records or private conversations. Evidence is masked in the report, and nothing moves without approval.
Purge — report storage per batch, near-duplicate clusters and retention candidates. Proposes only.
Recall — answer natural-language questions over the index and return ranked paths with a reason.

Reruns are safe by construction: a digest ledger means a capture is never filed twice and never indexed twice, and a rejected plan changes nothing.

HOW IT IS BUILT

A declarative Uliya Skill: SKILL.md plus seven Markdown references, no scripts, no runtime, no binaries. Orchestration runs through registered expert agents (file, document, photo, knowledge) over supported UGOS CLI capabilities, with a local model for OCR and recall.

The archive contract is not invented for a demo. It is the convention I already run on my own machine on a weekly timer: 8 batches, 256 captures each, mtime-derived names, preview mode, collision rollover. Porting it into a declarative state machine meant rewriting the reasoning behind each rule rather than the code, because the Skill has no code.

THE HARD PARTS

Deciding what the Skill must refuse to do. Deleting is the obvious feature and the wrong one: a Skill that audits secrets and also deletes files is one approval mistake away from being the incident. So Purge proposes and the human acts in the file manager.

Precision in the audit. A detector that cries "credential" at every QR code trains the user to ignore it. Findings carry confidence tiers, only structural matches are called leaks, evidence is masked, and a dismissed finding is remembered instead of re-raised.

STATUS

Built and running: the sweep, batching, naming and collision logic, proven on 1,889 captures across 8 batches. Written: pipeline, archive contract, leak rules, capability map, test plan and limits register. Not built yet: NAS execution, indexing at scale, leak thresholds, gallery intake. Verified instead: every capability named in the Skill is mapped to catalog 2.2.0, and the package passes the official UGOS static checker with zero errors and zero warnings. Untested: OCR latency on iDX6011, shared-folder permissions — all of which the assigned testing day resolves, with a declared degraded mode for every stage.

WHAT'S NEXT

Multi-user permission scoping so a shared NAS can audit each member's captures against their own consent boundary. Video frame extraction. A second audit pass for rotated and handwritten captures, which are the current blind spot.

The Round 1 deliverables are the proposal PDF and the Pre-test Skill Package (ZIP).