# FIX-AI-FOLDERS — Re-sort plan

## Problem

~17K files under `/Volumes/miniShare1/Pictures/Images/AI/` were sorted by an older
scheme. Two scripts currently create `AI/{model}/{lora}/` folders with **different
sanitizers**, fragmenting the tree:

| | `organize-images` (`clean()`) | `rename-ai-snaps` (`sanitize_dir_name`) |
|---|---|---|
| `Pony XL\trserkanWildpony_v20.safetensors` | `Pony XL\trserkanWildpony_v20/` or `Pony XL-trserkanWildpony_v20/` (its `\\`→`-` fix only catches *double* backslashes) | `trserkanWildpony_v20/` (strips path prefix, `_`-sanitizes) |
| spaces | kept (`FLUX Handsome Male Model_1.0/`) | → underscores |

Legacy artifacts also exist: `None/` and `unknown/` folders (at Images root and inside
`AI/`), plus loose files at the `AI/` root.

Additional blockers:

- `rename-ai-snaps` **skips all `schmeeve-AI-*` files** — 17,178 of 17,255 files in
  `AI/`. As-is it would touch almost nothing.
- `rename-ai-snaps` is PNG-only (fine: `AI/` is 99.97% PNG; 5 jpgs reported, not moved).
- All 30 sampled `AI/` files still carry full ComfyUI `prompt` metadata — a
  metadata-driven re-sort is viable.
- `Unsorted/` sweep: 0 of 120 sampled files (jpg/png/webp) have ComfyUI metadata —
  they are web downloads, not ComfyUI outputs. Only files with real `prompt`
  metadata qualify for moving.
- `nas_mount_base()` does not know `/Volumes/miniShare1`.
- `~/.local/bin/organize-images` is stale (older than repo copy).

## Requirements (user decisions)

1. **Scope:** Everything — `AI/`, `None/`, `unknown/`, plus sweep `Unsorted/` and
   `Screenshots/` for PNGs with real ComfyUI metadata.
2. **Filenames:** Keep existing names, **except** generic names (`ComfyUI_*` etc.)
   which get a proper keyword-based rename.
3. **Architecture:** Re-separate concerns — `rename-ai-snaps` renames only (the
   folder-moving added in `e85fce1 "model first"` was a mistake), `organize-images`
   does all sorting. Fix its sanitizer; sync the stale `~/.local/bin` copy.

## Plan

### 1. `rename-ai-snaps` → rename-only

- Strip the move/folder logic (`output_root`, `mkdir`, `shutil.move`, model/lora
  folder computation). Keep metadata extraction, keyword scoring, and
  interactive/batch renaming **in place** (filename changes only, never directory
  changes).
- Default: skip `schmeeve-AI-*` as today. New flag `--include-renamed` for
  corpus-wide renames later.

### 2. `extract-ai-meta` → shared, recursive metadata engine

- Add `-r` recursive walk; relative-path keys instead of bare basenames.
- Port the robust extraction from `rename-ai-snaps` (`_strip_model_name`,
  `UNETLoader`/extra checkpoint classes, generic `*lora_name*` fields). It currently
  only knows `CheckpointLoaderSimple` and keeps backslash path prefixes — the cause
  of `Pony XL\…` folders.
- Keep its connected-LoRA check (excludes disconnected LoRA nodes).
- Add a `prompt_present` flag per file so `organize-images` can classify AI files
  **without** `mdls`, dimension heuristics, or filename guesses.

### 3. `organize-images` → all sorting, canonical scheme

- **Canonical sanitizer** (single implementation; fixes the `\\` bug and space
  handling): strip path prefix → spaces → `_` → non `[a-zA-Z0-9_.-]` → `_` →
  collapse. `Pony XL\trserkanWildpony_v20.safetensors` → `trserkanWildpony_v20`
  everywhere.
- **New `--resort` mode:** recursively walk `AI/` (including `schmeeve-AI-*` files),
  compute canonical `AI/{model}/{lora}/` per file, move only what's misplaced.
  **Filenames never touched.** Files already correct are skipped (idempotent).
  Missing model → `unknown/`; no loras → `no-lora/`.
- **Sort `None/` and `unknown/`:** pass as targets; files with metadata route into
  `AI/{model}/{lora}/`, files without go to `AI/unknown/`.
- **Sweep `Unsorted/` + `Screenshots/`:** only files with actual ComfyUI `prompt`
  metadata move into `AI/{model}/{lora}/`.
- **Generic-name hook:** before moving files matching `ComfyUI_*`/hash-style names,
  invoke `rename-ai-snaps -n` on them (its skip-rule already protects
  `schmeeve-AI-*`).
- **Collision policy:** destination exists → append `_1`, `_2`… (never overwrite,
  never delete).
- **Performance:** one metadata scan → JSON manifest in a temp file; move logic
  reads the manifest (no per-file `mdls`/`sips` subprocess storms).
- Keep dry-run default, `-e` to execute, `--limit` for batching.

### 4. Verification

- Full dry-run first: summary counts per destination folder + sample moves.
- Spot-check that `Pony XL\…`, `Pony XL-…`, `None/`, loose-root files land
  correctly; confirm already-correct files are untouched; re-run dry-run
  post-execute to confirm ~0 remaining moves (idempotency proof).

### 5. Deploy + docs

- Copy updated scripts to `~/.local/bin/` (fixes stale `organize-images`) and
  `~/Dropbox/bin/`.
- Add a short "Image pipeline" section to `AGENTS.md` documenting the restored
  separation: `rename-ai-snaps` = rename, `organize-images` = sort,
  `extract-ai-meta` = shared engine.

## Untouched

- `Unsorted/` web downloads (no metadata = no move), `Screenshots/` non-AI files,
  non-PNGs in `AI/` (5 jpgs — reported, not moved).
- Nothing is ever deleted; worst case is a `_1` suffix.

## Note

With `schmeeve-AI-*` names kept as-is, a handful may not match what today's keyword
scorer would produce — but placement will be correct, and any folder can be
re-renamed later with `rename-ai-snaps --include-renamed`.