This paper investigates dialogue game agents using the LM Playschool Challenge with a 2B open-weight model, identifying that failures stem not only from broad knowledge gaps but also from local decision errors such as repeated guesses, malformed actions, and ignored feedback. The authors propose a diagnosis-guided post-training recipe—acquire, repair, and preserve—targeting these specific failure modes. The approach addresses the challenge of maintaining state across turns and selecting valid actions under evolving constraints in interactive dialogue settings.