TokenThresholdEventCompactor

An EventCompactor that compacts events once the most recent prompt reaches a token threshold, keeping a tail of recent events raw (tail retention).

It is intended to run intra-invocation, before each model call: when the most recently observed prompt token count reaches tokenThreshold, every event older than the last eventRetentionSize events (since the previous compaction) is summarized into a single event, while the most recent events are preserved raw to maintain continuity.

Safety net

The window never strands an open interaction: the window is shrunk to its longestSelfContainedPrefix.

Parameters