0012. What we adopt from Diffusion Studio

Context

Diffusion Studio's editor shares our thesis and ships. Before building cutcut's editor UI we studied its write-back path, runtime and agent surface (in a comparison page since removed). The question was whether to change course.

Decision

Keep cutcut's file model: the resolver, sequence tracks and anchors, time references against the manifest, the patch log and the fold. Their design validates the write-back idea and shows the cost of absolute placement and of a live-mount export.

Adopt five things:

1. Fold at gesture end. One source write per history group, when the group closes; the model always reads a file that is at most one gesture behind the preview. This supersedes the open policy question in 0003. Their 120 ms debounce, which can write a drag several times, is rejected.
2. Editing policy in the authoring prompt. Brief first; A-roll, then B-roll, then captions; no editorial cuts without a target length; pauses inside clips rather than gaps; 0.3–0.5 s between sentences, at most 0.8 s of tail, never mid-word; verify after every stage.
3. A check equivalent. Most of what their structural check finds is already a resolve error (missing assets, trims past the source, dangling anchors, zero durations). Two lint rules close the gap: sequence-gap (black frames: a gap on the first sequence track) and cut-inside-word (a numeric trim inside a transcript word).
4. A capture equivalent. PlayerCore.captureSheet(times) renders a labelled contact sheet from renderFrame, so an agent or a test looks at the result without exporting.
5. Scrub-to-keyframe while dragging and decoder handoff on split as player follow-ups; recorded, not built.

Not adopted, with reasons: absolute placement (every trim becomes arithmetic the agent does by hand); stamping ids into the user's file (the file the model handed over should be the file that runs); overwriting expression props (we lock or bind them); unrolling a loop on the first edit (we bake as an undoable action); view state in the source (noise in diffs); an ECS runtime and a live-mount export (determinism becomes the author's problem).

Consequences

• The editor package, when built, closes a history group by folding.
• The prompt and the guide carry editing policy, not only file conventions.
• lint needs the manifest for cut-inside-word; without one the rule is silent.
• captureSheet is part of the player's public surface and of the e2e suite.