Mistral’s docs say to replay a reasoning model’s thinking on every turn, and warn that stripping it significantly degrades output quality. LiteLLM strips it. So does Qwen Code when it talks to Mistral directly. Here is how to tell whether your gateway is doing this to you, a third runtime patch that fixes it for Mistral without rebuilding anything, and what happened to a coding-agent run that had been going round in circles.
Le Chonk Forgets: Replaying Mistral’s Thinking Through LiteLLM
