Large language model context compaction strategies for AI Harnesses.