Armin Ronacher: reasoning levels live in the system prompt, and switching them busts your KV cache

mitsuhiko · x · 2026-09-06

Armin Ronacher points out a non-obvious mechanism: reasoning effort is baked into the system prompt for most models, so switching levels mid-session invalidates KV caches. The main exception is Fable 5.1, which allows mid-conversation reasoning switches. His linked post 'What Is Reasoning' explains that reasoning traces are just ordinary text emitted into a scratchpad (visible via GPT-OSS's Harmony analysis/final channels), that early token-budget APIs made effort look like a sampling property, and that his curiosity was sparked by a paper on extracting reasoning traces from closed-weight models.

Related event: Switching reasoning levels invalidates KV cache, developers find(3 posts)→

Original post →

More from coding & agent

coding & agent channel →