The llama.cpp b9978 release addresses a checkpointing issue that negatively impacted agentic workloads. Previously, every agent turn created a new checkpoint regardless of the minimum step spacing, which collapsed the coverage window.
- Context rewinds, common in tool-calling loops, erased all checkpoints and forced a full reprocess.
- The fix ensures closely-spaced checkpoints from prior tasks are evicted to keep the window wide.
- This optimization allows long agent sessions to remain fast.
The update resolves the checkpoint collapse issue, improving performance for multi-turn agentic applications.