Start here: prepare a Tidy debugging brief through H3Max.
Use the author artifact for tracing Tidy’s sorting behavior, or the matching Feishu archive for inspecting the reference. Tidy is fictional: this is a code-animated launch film, not desktop footage, SaaS verification, or evidence of H3Max performance.
The upstream project’s account of its author and model credits Claude Opus 5.5. Those sessions have not been independently verified. These commands are reader instructions, not our run log; the entry page does not establish local-terminal access.
LemoLab author-film frame at approximately 5.04 seconds, also the cover.
1. Diagnose the destination before changing the easing
Preserve the reference unchanged. Watch the story, then inspect cursor movement, reading holds, and sound. Compare the intended layout and dark-stage hierarchy with the documented relationship between elements and sorting cues.
The generator contains 264 mixed-category elements, not 264 application windows. Fictional product figures are not measured benchmarks.
Separate wrong destinations, mistimed cues, and obscured narration. Record the timestamp and source before editing. These are diagnostic categories, not reported failures.
2. Preserve an identifiable environment
Prepare Git, Node 20+, FFmpeg/ffprobe, and Python 3.11–3.13 or uv. Use a new directory, without overwriting existing work:
git clone https://github.com/lemomo-ai/lemo-opuscar.git tidy-case
cd tidy-case
export LEMO_OPUSCAR_HOME="$PWD"
sh plugin/skills/lemo-opuscar/scripts/setup.sh deps voice
sh tools/fetch.sh instruments vcsl
mkdir -p evidence
D=styles/dark-keynote/demo
git rev-parse HEAD > evidence/tidy-commit.txt
npm ls --depth=0 > evidence/tidy-dependencies.txt
Keep locks/Python versions; check fonts, source, Kokoro/transcription and VCSL samples. Separate reference and output. Setup is not acceptance.
3. Give the agent a testable reproduction brief
New prompt—not the author’s original instructions:
Reproduce and debug LemoLab's Room to Think from existing dark-keynote source.
Read commit, locks, STYLE.md, DEMO.md, demo/CREDITS, core/README.md, build.sh,
reference and assets. Report missing dependencies; otherwise follow the
documented build without extra approval gates.
Preserve fictional Tidy, dark staging, original mixed-category elements,
cursor-led sorting, shared sound cues and credits. The 264 items are not
264 windows. No real screenshots, fabricated benchmarks, desktop-success
claims or external media-generation APIs.
Inspect timeline.js, world.js, film.js, engine, voice lines, score and mix.
Verify 42 seconds/24 fps from source; record release differences without
silent retiming. Use absolute render(t), fixed seeds, stable identities
and destinations, and shared note/event tables. No wall-clock state.
Test forward, reverse and arbitrary seeks. Save repeated-time comparisons
separately so a later capture cannot overwrite the earlier evidence.
Check destinations, cue alignment, click, flights, landing, captions,
reading holds, narration, intentional silence and complete ending.
Save actual model/provider/effort/session records, prompts, dependencies,
diffs and genuine checks. Do not infer success from the author film,
publish, or identify the reproduction as an H3Max original.
Original-release frame at approximately 20.16 seconds; not a local debugging capture.
4. Build, isolate, and repeat the same check
Run the checked-in implementation of Tidy’s full build:
sh styles/dark-keynote/demo/build.sh
node core/render/still.mjs "$D" 0 14.9 16.45 26 41.9 \
--out evidence/tidy-stills
The script builds narration, score, events, checks, mix, subtitles, and a 24 fps/two-worker render, then muxes. Still arguments are seconds; inspect press/landing neighbors. Do not mix renderer flags.
Revisit a time after seeking elsewhere. Changed destinations suggest playback-history dependence; inspect shared state, not camera tricks. Restore missing fonts before adjusting coordinates.
Read the final ASR spot-check and listen: its failure need not stop the script. Resolve Chromium or missing voice/sample dependencies explicitly; never silently remove audio.
Repair only the observed destination, cue, caption or narration defect.
Preserve the fictional case and source timing. Show the responsible data
and code diff, repeat the original seek/frame/audio checks, and replay the
whole action. Do not conceal jumps with cuts or report untested fixes as passed.
5. Validate the complete artifact
OUT=styles/dark-keynote/dark-keynote.mp4
ffprobe -v error -count_frames -show_streams -show_format -of json "$OUT"
ffmpeg -v error -i "$OUT" -f null -
Unchanged 42-second/24 fps settings imply 1,008 picture frames. Check dimensions, rate, audio, and ending; watch and listen throughout. Decoding does not certify cue alignment, and web/local encodes need not be byte-identical.
LemoLab reference frame at approximately 34.02 seconds.
Archive source, inputs, versions, prompts, session records, edits, checks, and MP4. Retain the credited origins of Tidy’s lettering, samples, and voices: MIT code, OFL fonts, CC0-credited VCSL samples, and separate voice licenses.
Top comments (0)