Keeper v1 — AI-assisted culling for photos and videos

[thisiscsim]

Summary

• Pivots the repo from the Aperture video-studio scaffold to Keeper: import an SD card / phone dump into a local date-sorted library, auto-cull with explainable verdicts (blur/exposure/black-frame/accidental-clip via pure-JS CV over ffmpeg-piped grayscale), group bursts (pHash + capture time) with best-pick suggestions, pair RAW+JPEG and Live Photos, and search the whole library in natural language via local CLIP embeddings.
• Adds the three-tier AI model: deterministic local pass → budgeted GPT-5.5 vision judge over borderline items only (downscaled thumbs, repaired-then-validated output) → agent skills (/cull-shoot, /find-media, /organize-library) working the same catalog through JSON CLIs. Every user override feeds taste.json (exemplars + threshold tuning + standing rules injected into future runs).
• Rebuilds the app around the new contract: SQLite catalog (node:sqlite) validated by the new @keeper/schema zod package, brokered by a spawned catalog service; renderer on the inherited tokens/UI kit (lazy date grid, review queues with evidence crops, loupe with burst strip, keyboard-first culling with undo); export writes Lightroom-readable XMP sidecars; the only deletion path is a two-step empty-rejects to the OS Trash.
• Removes the video-editor surface (Remotion, timeline, EDL, style library, TTS/ElevenLabs, bundled music) and rewrites README/AGENTS/CHANGELOG, .cursor/rules, and skills for Keeper.

Screenshots / recordings

Dev-booted twice with a smoke library (grid, queues, loupe render clean); no recordings attached — reviewer can npm run dev against any folder of photos.

Test plan

• npm run typecheck passes (all workspaces)
• npm test passes — 89 tests: schema hostile-input, quality/pHash/grouping/taste/XMP/LLM-repair/catalog, renderer store (verdicts/undo/selection/search order)
• npm run build passes (resolves all CSS/asset imports)
• Manually exercised end to end against generated fixtures: import (dedupe + checksums) → CV verdicts (blurry/black/accidental-clip correctly routed to sure-reject) → burst grouping → CLIP embedding + semantic search ranking → verdict CLI → taste.json exemplar recording → export with XMP sidecars → catalog-service RPC (queues, search, rejects flow)

Checklist

• Catalog rows / taste.json validate against @keeper/schema (hostile-input tests included)
• No secrets committed (app/.env.local, keys) and no large/uploaded media
• CHANGELOG.md updated under "Unreleased"
• Restart note: this touches main/preload/engine scripts — npm run dev must be restarted after pulling (HMR only covers the renderer). Node >= 22.13 required (node:sqlite).

Made with Cursor