Read the exact conversation a run had

The live feed shows previews clipped at about two thousand characters. That is the right size for watching a run happen and the wrong size for working out what went wrong afterwards. The full copy of everything the model was handed back existed the whole time, with no way to read it.

The run page now has the transcript. Items come out exactly as they were stored, with their position and their timestamp, and the page decides how to display them. A tool result the platform has never seen before still reaches your screen.

A run whose agent never engaged shows an empty transcript rather than an error. The run exists. Its conversation just never started.

Where a run shares a conversation with earlier runs, an unscrolled transcript shows you the most recent part of it, so you get your own turns rather than somebody else's from last week.