The recording

A recording, not a description.

Everything that actually happened in the app, kept and replayable — so nobody has to reconstruct it from memory or retype it into a ticket.

DOM, network, console

Scrub the session back, freeze any frame and annotate it. The requests and the logs sit on the same timeline as the clicks that caused them.

DOM replay — the page as it was, not a video of it — you can stop on a frame and mark it up.
Network waterfall — every request with its timing and status, lined up against the clicks.
Console — errors and logs captured on the same timeline, so a failure has its cause beside it.

The agent does not read the raw event stream. It reads a compact digest, which is what makes a whole test session fit inside a single prompt.

The measurement — a 55-minute, 6,010-event recording comes out at roughly 76,000 tokens.
One prompt — that is small enough to hand over whole, rather than in summarised fragments.
Not the raw log — the digest is the interface; the full stream stays on disk for replay.

The tester stops transcribing what they did, and the developer stops asking for steps. Both are already in the recording.

No transcription — “write the repro steps” stops being a task anybody has to do.
Cannot-reproduce ends — the frozen frame, the failing request and the branch all arrive together.
It works without agents — as a plain session recorder and annotation canvas; the agent loop is optional.