The Replay Is Not a Copy Anymore

Until recently, a reviewer watching a recorded session was watching a reconstruction, a separate interface maintained on its own and built to look like the one the candidate used.

What changed

Piece by piece, the replay was rebuilt to use the real, live interface components instead of a lookalike. The chat transcript is now the actual chat bubble the candidate saw. Clicking a file named in one of the agent's tool steps opens it in the editor, as it does live, and the buttons for keeping or rejecting an agent's edit never appear on a recorded turn. A multi-step agent's work (its plan, its tool calls, its reasoning) now folds and expands the same way in playback as it did live, instead of showing a flattened summary. The problem panel and its tabs are the real problem panel, including a "Rules" tab that used to sit there doing nothing in the replay and now behaves exactly as it does for a live candidate. The problem description goes through the same rendering the candidate's screen used, so its text, bullet points, and any mockup image sit where they sat for the candidate.

The agents window had drifted furthest, because its replay drew every message as a plain bubble with no turn grouping at all. It now uses the live window's own pieces, so messages group into turns the way they did for the candidate, each stamped with how long the agent worked and how many steps it took.

Each of these landed as its own small change: the chat bubble first, then the problem panel, then the turn-folding behavior. Each piece was checked against the real interface before the next one started, which makes the result easier to trust than one large rewrite would be. When the turn grouping was pulled out of the live window for the replay to share, the live window's existing tests passed unchanged.

Why it matters

A reconstruction is a second product that has to be kept in step with the first one, forever, by hand. Every time the live interface changes, somebody has to remember to update the replay's copy of it too, and sooner or later the copy falls behind. Once it does, the reviewer can see something the candidate never saw, which is the one thing a replay has to avoid.

Reusing the real interface removes the second product entirely. There is one implementation, used live and used in playback, and the replay is a recording of it rather than a description of it. When the live interface improves, the replay improves for free, on the same day.

The multi-step agent view shows why this is worth the effort. Detail that only exists in the live product, like an agent's expandable plan and tool calls, is detail a recruiter loses the moment they switch to reviewing a recording, unless the recording is built from the same pieces.

Where it points

Every other page that shows a candidate's work back to a reviewer should use the live components too, even where a lookalike already comes very close.