The assistant stopped dying on long jobs

It stopped mid-sentence

Ask the assistant for anything large and it could stop halfway through, leave a cut-off reply on screen, and show a connection error. The only way forward was to ask again and hope. Worse, you could not tell what it had already done, so asking again risked doing half of it twice.

What changed

• It estimates what the next piece of work will cost before starting it, using how long this turn's earlier pieces took, and stops cleanly rather than starting something that will not fit. A stop is a pause the panel continues from on its own.
• Every model call carries its own deadline, so a stalled call is ended by us rather than by the host cutting the connection.
• It says what it actually did, rather than what it intended to do.
• A bulk change that fails partway is rolled back as a unit, so you are never left with half of it applied.
• It remembers what it has already read, so a long conversation stops re-reading the same things and running out of room.