Switching reasoning levels invalidates KV cache, developers find

Armin Ronacher explained that most models encode reasoning effort in the system prompt, so switching reasoning levels mid-conversation invalidates the KV cache. A developer confirmed the issue on pidotdev 0.85.0, where changing thinking levels rewrites reasoning.effort and refreshes the cache.

2026-09-06 ~ 2026-09-06 · 3 related posts