pub const AUTO_COMPACT_BUFFER_SMALL: usize = 5_000;Expand description
Reserved headroom for small/proxy-bound windows (typical self-hosted GLM 65K). 5K leaves space for one streaming response + a round of tool results without forcing compaction so early it shrinks the usable session. Larger buffers (13K) on a 65K cap kick compaction at 52K — wasting the 12K immediately above where users do real work.