// grow between batches. The view goes out as the same messages as in the
// request before, while their lines give the same text, and the new lines
// follow as one new message. So the end of the view of each earlier turn stays
-// a message end. On the ChatGPT backend, the first request of a turn with this
-// layout read the view of the turn before from the prompt cache in each of 10
-// turns, and with the view in one message it missed once in 5.
+// a message end. Stable message boundaries let the model reuse the unchanged
+// view prefix from its prompt cache instead of invalidating it when new lines
+// arrive.
//
// The first message starts with "<chat>\n". The closing tag goes into the state
// message after the view, so that no view message changes when lines come. The