Reasoning traces are text, not magic
This is a useful demystification of reasoning models: hidden chain-of-thought is still model text routed into a separate channel, and that explains both KV-cache behavior and some leakage paths.
Links: Original source · Shared link
Logged at IST: 2026-08-20 06:30 IST
What it is: Armin Ronacher's short explanation of how reasoning traces work in current reasoning-model APIs.
Gist: The important framing is that reasoning traces are not a mystical second system. They are ordinary model text emitted into a hidden analysis or scratchpad channel before the final answer. The channel markers may be special tokens, but the reasoning content itself is still text generated by the same model.
That makes a few confusing behaviors easier to understand. Reasoning effort is effectively part of the prompt/control context, so changing it changes the prefix and can invalidate KV-cache reuse. Chain-of-thought leaks happen when the learned separation between scratchpad and final channel is tricked, disabled, or routed through odd tool conventions.
Ronacher's useful warning is that “don't think” is often mechanical. Some systems prevent or prefill the tokens that normally open a thinking channel; custom tools or prompts can still create side channels where scratch work appears somewhere it was not meant to go.
Newsletter angle: Good explainer for reasoning-model infrastructure: hidden CoT, KV-cache behavior, prompt-controlled effort levels, and why channel routing is a product/security boundary.