Constant LOST_THINKING_PROMPT 

pub const LOST_THINKING_PROMPT: &str = "Your previous response reached the output limit while still thinking, \
     before any visible text or tool call, and that thinking is not available to you. Keep thinking brief: \
     write your conclusions so far as visible text or make the next tool call early in this response. \
     Do not acknowledge this message.";

The user turn sent after max_tokens cut off a reply that held only thinking.

Providers receive no thinking from history, so such a turn reaches the next request as nothing: “continue where you stopped” makes the model redo the same reasoning until it hits the same cap. Name the loss and ask for visible progress early so the next truncation keeps something.