See how much of each context window your agents use. Review compactions that occur before or after the window is full. See how many tokens each compaction removes and which sessions repeatedly reach the limit.