How it works

ConPTY / Unix pty          bytes + timestamps
  (portable-pty)  ─────────────────────────────▶  VT parser (vt100)
        ▲                                                │
        │ drives input,                                  │ screen snapshots
        │ waits on screen state                          ▼
   .tape script                                    frame timeline
                                                         │
                                                         ▼
                                          optimizer (dedupe, trim idle,
                                                     quantize, speed, tail)
                                                         │
                                                         ▼
                                            SVG emitter, one @keyframes

Escape sequences are never turned into SVG directly

The byte stream feeds a real terminal grid that understands cursor movement, scroll regions, the alternate screen and line wrapping, and each state of that grid is photographed into a frame.

That is the difference between recording a full screen TUI correctly and only recording echo correctly. A recorder that replays escape sequences into markup gets the second one right and the first one wrong.

One keyframes block, however many frames

The output stacks every frame vertically inside one clipped group and animates translateY using a single @keyframes block with step-end timing.

One CSS rule drives the whole animation no matter how many frames there are, and frames of different durations come for free, because each keyframe percentage is just placed where that frame's timestamp falls.

The platform boundary

Everything after the parser is platform independent, which is why the emitter is tested against fixed byte streams with no terminal involved at all. The pty is the only part that knows what operating system it is on.

src/capture/    spawning a pty, reading it, running a tape against it
src/term/       terminal grid to Frame, the platform independent boundary
src/optimize.rs dedupe, trim idle, quantize, speed, tail
src/raster.rs   webp, gif, apng and png output, one frame at a time through resvg
src/redact.rs   masking secrets and rewriting paths before anything is written
src/session.rs  saving and loading a capture, so it can be re-rendered
src/svg/        the emitter, themes, escaping
src/tape/       tape tokenizer and directives
themes/         one TOML file per theme

Raster output

Each format shares one frame loop. A frame is emitted as a single frame timeline, rasterized through resvg, and handed to an encoder along with the timing the animation already has.

APNG stores only the bounding box of what changed since the previous frame, using BlendOp::Source and DisposeOp::None. Writing every frame in full came to 1.5 MB on the reference recording, worse than the GIF; the dirty rectangle brings it to 96 KB.