Skip to main contentModule compaction
Source - archive
- Append the pre-compaction history to the archive on disk.
- compact
- Compact the message history via an LLM-generated summary.
- compact_split
- Hierarchical compaction used when primary
compact() fails. - count_tokens
- Count tokens in a single string using cl100k_base.
- estimate_tokens
- Count total tokens across all messages.
- should_compact
- Returns true if compaction should fire.
Fires when
estimate_tokens(messages) + tool_overhead_tokens >= compact_at_tokens.