We just completely reworked our compaction algorithm to fit models with limited context window.
it is much more reliable and intelligent. Perfect for
- Local models
- Frontier models with super long running uninterrupted tasks (we tested a case with 90 compactions and 500+ steps)