I auto compact around 200k tokens both due to this and because the cached read cost really escalates when sessions have more tokens than that (too short and you pay a lot in per-compact re-reading of state)