The Librarian

The Librarian — parts catalog & cut assembler

Runbook step 7, built & verified 2026-07-08. Replaces hand-maintained concat.txt files.

What it is

Two scripts in C:\AI turn the 63 loose render parts into a queryable library and

make new cuts a one-line command.

• librarian.py — scans the ComfyUI output dir, reads the workflow ComfyUI embeds in

each mp4's prompt metadata tag, and recovers full provenance per part. Writes

catalog.json. Provenance is deterministic, not guessed: the embedded

VHS_LoadImagesPath directory (sNNrM_c) gives source segment+run; the positive

CLIPTextEncode gives style+content; loader nodes give the model stack; the sampler

gives seed/steps/cfg. Recovered fields: segment, run, style, positive prompt, model,

lora+weight, controlnet type/strength/end, seed, steps, cfg, sampler, frames, fps,

resolution, duration, version.

• assemble.py — turns a cut file (JSON EDL) into a finished mp4: concat parts in

order, remux the original source audio.

Files

• catalog.json (this folder) — the machine-readable index. Regenerate anytime with librarian.py.
• cuts/.json — cut files (EDLs). cuts/.mp4 — assembled masters.

Cut file schema (cuts/<name>.json)

{
  "name": "v1_full",
  "audio": "~/Downloads/e_studios_source.mp4",
  "parts": ["RUNTMP_00005", "RUNTMP_00006", "..."]
}

audio optional (defaults to the source). parts is an ordered list of part_ids;

each resolves to a path via catalog.json (fallback: OUTDIR/<id>.mp4).

Commands

python librarian.py --print                          # rebuild catalog.json + human table
python librarian.py --query segment=21               # filter catalog (no rescan): see a segment's version stack
python librarian.py --query style=ligne model=LTX    # ints match ==, strings match substring; combine freely
python assemble.py cuts/v1_full.json                 # build a cut
python assemble.py --from-concat v1full/concat.txt v1_full   # import a legacy concat -> cut json
python assemble.py --new-v1 v1_from_catalog          # regenerate the v1 EDL from catalog metadata alone
python assemble.py --swap v1_full v1_lignev4 seg21=LIGNE_V4_00001   # clone a cut, swap part(s), assemble
python librarian.py --gaps                           # per-segment version count; the re-render queue

--swap keys are segNN (must be unique in the cut) or an existing PART_ID; multiple pairs allowed

in one call (audition several upgrades at once). This IS the "new cut in seconds" loop:

--query to find the part → --swap to drop it in.

Verified

• assemble.py cuts/v1_full.json reproduces the shipped master exactly: 40.099002s = 40.099002s.
• --new-v1 (catalog metadata only, no concat file) produces the identical 47-part ordering.

→ the concat.txt files are now redundant; the catalog is the source of truth.

How to make a new cut (the "in seconds" promise)

1. Render a new part (any pipeline) → it lands in OUTDIR with its workflow embedded.
2. python librarian.py → it appears in the catalog, provenance auto-recovered.
3. Copy a cut json, swap the part_id for the segment you're replacing, assemble.py it.

e.g. to swap seg21 to the v4 ligne claire: replace RUNTMP_00030 with LIGNE_V4_00001.

Coverage notes

• 63 parts cataloged, all 63 with recovered workflow; 56 mapped to segment/run, 61 to style.
• 7 "ext" parts (no segment): the LTX/Cog taste tests and early seg02/seg03 experiments —

correctly outside the 38-segment scheme.

• Model axis distinguishes: SD1.5+AnimateDiff (v1 stack), +LoRA, +IPA, lineart-CN vs canny-CN,

LTX-Video-2B. As cloud parts (Qwen/Wan) arrive with their own embedded or sidecar metadata,

extend parse_workflow/PREFIX_HINT in librarian.py.

Existing cuts (cuts/)

• v1_full — the shipped v1 master, imported from the legacy concat (40.099s).
• v1_from_catalog — proof cut: identical v1, regenerated from catalog metadata alone.
• v1_lignev4 — v1 with seg21 swapped to the ligne claire v4 lineart part (LIGNE_V4_00001).

Note: v4 was rendered on seg21's bass-macro frames, so this is a "v1 + ligne v4 upgrade"

demonstration at seg21, not the post-swap casting.

• v1_cast — the first casting-accurate cut. The decided seg09↔seg21 swap made real:

seg09 = LIGNE_S09_00001 (ligne claire Palantir, wordmark legible), seg21 = CONSTRUCT_S21_00001

(constructivist bass macros). The film as designed in E_CASTING.md, not as first-drafted.

Base for cloud upgrades → the maximal version.

Next (when wanted)

• Per-part sidecar .json for the non-ComfyUI parts (Cog diffusers, cloud renders) so their

provenance is as rich as the ComfyUI ones.

• --query sort/format options and a --gaps mode (segments with only a v1 part, no upgrade yet).