As for the summary, I agree that it's a big gap. I'm not trying to make a better summary: it's not a separate summarizer rewriting the session afterwards, it's the session itself writing while it's still active.

"Asking the old one again" works as long as it's alive. In my case it's exactly the opposite — either I've hit the limit, or I've switched to a different model. Then there's no one to ask, and whatever wasn't recorded is gone.

And "cheap" is measured from the wrong side. The conversion is cheap, but the transcript doesn't go into a file — it goes into a live session and spends its turn every time it moves.