Playground

Change the script. Watch the camera.

This runs the real op set and the real camera model against a small app, in this tab. It does not make a file: there is no ffmpeg in a browser. It shows you the thing that is hard to picture from a JSON file, which is where the camera goes and how it gets there.

take.jsonl

Run and record captures this tab while the script runs, then plays the result back here so you can watch it before deciding whether you want the file. Your browser asks first and shows its own recording indicator throughout. It is the camera, not the render: the binary films at twice the resolution, draws the cursor and the backdrop, motion interpolates a slow capture, and writes H.264.

Platform

Picking one writes a frame line into the script, which is how the binary is told the same thing.

The camera is derived from the ops, never described by them. A type leans further left than a click, because a line of text grows away to the right as it is written and the beginning of it is the part you want to keep.

Things worth trying

Delete the wait before stop_recording. The script is refused before it runs, because that last zoom would be cut off.
Put "typewriter_ms": 120 on a type and watch the camera keep up rather than chase.
Point a click at #nope. You get the line number and the selector, which is what a CI log needs to be useful.
Remove start_recording entirely.
Pick iPhone screen recording and run it again: the same script, filmed as the phone's own recording, with the app in its phone layout.

What the binary adds

An MP4. A real Chromium at a device scale factor of 2, so text is sharp rather than upscaled. The backdrop and the rounded plate. A cursor that is drawn rather than captured. And any URL at all, rather than the one app bundled here, because it drives its own browser instead of borrowing yours.

cargo install --git https://github.com/thisisisheanesu/kaviri
kaviri record --script take.jsonl --out take.mp4