feat(mulmoterminal): clip for renderShapeScript (4.18.0), en and ja - #124
Merged
Merged
Conversation
The agent writes a desk, renders it, reads its own picture, then is asked for a chair — which it cannot place without looking at the desk. That is the whole feature, and it needs no mistake to happen on camera: an additive second step depends on seeing the first, whatever the model. The model is enlarged twice: once inside the Canvas with the app's own zoom, keeping the three panes, and once by the camera at the end of each scene, because at full width it was seven percent of the frame. Shot on codex/astra. The ship judgment passed all five gates. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
The pushed-in frames kept a twenty-pixel sliver of the cell beside the one being watched, with a status chip sliced down the middle — it read as a crop that missed rather than a chosen frame. The left edge now lands past the border. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
…already had The first cut resolved both beats on the Canvas — the 3D viewer, which shipped long before 4.18.0. The new thing is the PNG the agent renders for itself and reads, and that appeared only as two lines of tool output. The author said as much: the clip was hard to follow because the part being demonstrated was the part already there. Each beat now cuts from the terminal line naming the file to that file, full frame, captioned with what it is and where it came from. The filename on screen and the filename in the caption are the same string, because it is the same file. The Canvas closes the second beat as the result rather than the point. Narration rewritten to match; Codex reviewed it twice. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
The clip began with the request and the model being written, so nothing model-shaped was on screen until second eight — including the frame X shows in the timeline. The author caught it. A poster beat now opens on the sheet the agent read, and the two working beats follow it. The order in both is the real one: the agent sees the model, then you do. The Canvas closes each of them as the result. Narration rewritten for three beats and reviewed by Codex twice; the second pass caught that beat one's line ran longer than its own clip. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
…ropped terminal The push-in cut words off both ends of the lines it was meant to make readable, the top of the panel was empty, and nothing model-shaped was on screen until second eleven. The author caught it in the rendered file. The terminal section is now 2.5s at full frame — the request, the call, the line naming the file — and the seconds it gave up went to the sheet. Eight and a half of the beat's ten seconds now have a model in them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
…g between beats Two things the author saw in the render. The camera pushed in during each working beat, so every beat boundary threw the frame back out and crawled in again — a reset the viewer has no reason for. And the chair arrived at its full size on the last frame before the cut, so the thing the clip is building towards was gone the moment it existed. Every cut is full frame now; the app's own zoom is what makes the model big, which is the zoom the viewer can see a reason for. A silent beat holds the finished desk and chair for two and a half seconds before the end card. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
The Canvas updated to the desk-and-chair while the terminal was still being held, so the finished scene flashed on screen a beat before the sheet that was supposed to reveal it. The hold now ends at 31.4s of the take — the line naming the file is up, the Canvas still has the desk alone — and the chair arrives where it was meant to. The third beat also began mid-thought; it now opens by asking for the chair, which is the sentence a viewer needs to follow the turn. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
…on asks for Both decks were copied from a clip that predates the 2026-09-06 decision, so neither named a model and both fell back to the flash TTS. The author noticed before they shipped. Worth knowing while working nearby: several existing clips carry the model on the _ja deck and not on the en one, so they are still half flash. The English second beat now clears its clip by 0.53s, which is barely over the gate — if a later re-render of the TTS comes back longer, lengthen the sheet in chair.mp4 rather than cutting the line. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH
ystknsh
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
4.18.0 の
renderShapeScript（エージェントが自分の書いた 3D モデルを 4 方向の 1 枚に焼き、その画像を自分で読む）を紹介する短尺クリップを、英語版・日本語版で追加します。gpt-6-astra）のセル。「間違えて直す」ではなく「足す」構成にしたのは、修正型がエージェントの失敗に依存し、モデルが強いほど成立しなくなるため完成した動画
英語版:
mt-look-before-showing-en.mp4
日本語版:
mt-look-before-showing-ja.mp4
Items to Confirm / Review
check-beat-fit.pyに落ちます。落ちたらchair.mp4のシート区間を伸ばしてください（ナレーションは短くしない）User Prompt
制作メモ
mulmoterminal-videoskill 同梱のrecord-shapescript.mjs（この PR とは別リポジトリ）footage/2026-09-11/shapescript-g/🤖 Generated with Claude Code
https://claude.ai/code/session_01UxmPTsHfRprS89F9NF4jGH