Skip to content
This repository was archived by the owner on Oct 31, 2026. It is now read-only.

Releases

v0.1.6

Choose a tag to compare

@github-actions github-actions released this 23 Oct 09:51

Full Changelog: v0.1.5...v0.1.6

v0.1.5

Choose a tag to compare

@github-actions github-actions released this 21 Oct 10:39

Keeper v0.1.5

Five features, five bumps. The release where every platform builds — including Intel macOS.

What's new

Schema-enforced verdicts. Keeper no longer asks Gemma for clean JSON — it hands the model a server-side JSON schema (Ollama structured outputs) so the verdict structurally cannot come back as anything but {keep, reason, confidence}. Old Ollama versions that don't understand schemas are handled with a silent retry.

Confidence escalation. A verdict the model isn't sure about (confidence < 0.6) gets one richer second look — more copies shown, longer excerpts, an explicit "look again carefully" nudge. Still unsure? Keeper says so and falls back to the transparent rule engine instead of bluffing. The spinner's countdown grows honestly when a second look happens.

--json. Machine-readable verdicts on stdout, progress on stderr — pipe straight into jq:

keeper ~/Downloads --json | jq '.decisions[] | select(.method=="model")'

Clutter report. Every run now ends with the payoff number:

DRY RUN — nothing was moved (8 loser(s) would be quarantined · 37.3 KB of duplicate clutter).

Perceptual image matching. The classic clone — a screenshot saved as photo.png, then re-exported as photo.jpg — is two different files with zero readable text. Keeper now fingerprints what images look like (64-bit difference hash) and clusters visually near-identical pictures across PNG/JPEG/WebP. Cheap hash before anything expensive; corrupt images and missing Pillow step aside instead of crashing.

Upgrading

curl -fsSL https://raw.githubusercontent.com/atif-1402/keeper/main/install.sh | bash

Or grab the binary for your platform below. Windows: pipx install git+https://github.com/atif-1402/keeper.

The fine print, as always

Dry-run by default. Never deletes — losers are quarantined with a manifest and restore.sh, and --restore brings every file back byte-for-byte. Nothing leaves your machine: the model runs on your own Ollama, and the only network hop is localhost:11434.


59 tests · MIT · Ko-fi · full changelog: v0.1.0...v0.1.5

v0.1.0

Choose a tag to compare

@atif-1402 atif-1402 released this 20 Oct 16:07

First public build of Keeper — a folder full of contract.pdf, contract_final.pdf, contract_final_v2(1).pdf, and contract_FINAL (copy).pdf should not be your problem anymore.

Keeper clusters version-clone files by actual content (not filename), asks a local open-weight model which copy is the keeper and why, and quarantines the rest.

What you get

  • Dry-run by default — nothing moves until you pass --apply
  • Never deletes — losers go to _keeper_quarantine/ with a manifest and restore.sh; --restore puts everything back byte-for-byte
  • 100% offline — verdicts come from a local Ollama model (gemma3:1b default, --model to swap); nothing leaves your machine
  • Rules engine fallback — byte-identical, unreadable, and no-model clusters are decided instantly from metadata, so Gemma only runs when it actually matters
  • Honest spinner — one self-updating line with a live clock and a self-measured ETA instead of a wall of progress output

Install

curl -fsSL https://raw.githubusercontent.com/atif-1402/keeper/main/install.sh | bash

Windows: pipx install git+https://github.com/atif-1402/keeper

Requirements

  • Ollama with an open-weight model: ollama pull gemma3:1b (~800 MB, runs on 4 GB RAM). Without Ollama, Keeper still works on the rule engine.