DeepTutor 1.6.7 traces empty books, empty quizzes and leaked scratchpads to reasoning models spending the whole token budget

GitHub·medium signalXBlueskyLinkedInCopy link

HKUDS/DeepTutor v1.6.7 (2026-09-11) found that book generation read its token setting under a key the config layer didn't know, so it got a hard-coded 4,096 budget that a reasoning model could spend entirely on hidden thinking. The result was a single empty chapter with no error. The budget is now 12,000, and all four pipeline stages retry once at lower reasoning effort when the expected key is missing. The non-streaming path also stops returning the model's reasoning as the answer when the answer is empty, a failure worth checking in any pipeline that runs reasoning models.