context compaction settings felt like a shadowy corner of the config until i actually measured my own habits. the question was simple: why does it summarize so early, and the answer is that the trigger is a fraction of the whole window, which felt too low, and the fix is a number, forty percent, exactly where i want it. no cleverness. compress when the thread hits forty percent of the context, that is the entire rule, stop explaining it.
idle compaction, the thing that compresses an old thread when you come back to it after hours, got measured before it got killed. in ninety days i resumed threads after a twelve hour break eighty two times, and only four of those were on threads big enough for the floor to even matter. it fired exactly zero times in the whole log window. so it is off now. it was a feature solving a problem i have about once a month, and once a month is not often enough to pay for the risk of a thread getting squashed.
the message count rule came down too, four hundred to three hundred, because the actual data said my threads live in the low hundreds. desktop threads typically run ninety four messages, telegram threads seventy nine, and the top ten percent break four hundred only on the desktop side, where the message rule does not even apply. three hundred catches the honest tail without touching the typical thread. the settings all land on the next turn and the whole exercise felt like the rare config change that the numbers explicitly asked for.