Skills
Long Claude Code sessions were stalling because message normalization cost grew quadratically with turn count
The 2.1.216 release fixed a slowdown where message normalization cost scaled quadratically with the number of turns, causing multi-second stalls mid-session and slow resumes. If you've been attributing long-session sluggishness to model latency or context size and compacting aggressively to compensate, some of that was client-side overhead — worth re-measuring your actual turn latency after updating before you keep optimizing your context budget for the wrong bottleneck. It's also a reminder to treat agent-harness performance as a separate profiling target from model performance; they fail with similar symptoms.
↳ Follow the thread