Expand description
LLM-driven context compaction.
When the transcript grows large, Compactor::compact asks the model to
summarize the older portion into a single system message, preserving key
decisions, paths, and outcomes. The agent then continues with the summary
plus recent messages, staying within the context window.
Compaction is disabled by default (threshold = usize::MAX). Enable
it via AgentBuilder::compactor(...).
Structsยง
- Compactor
- Configuration for LLM-driven transcript compaction.