TutorTrace is a dataset and behavioral abstraction pipeline designed for AI‑assisted programming education. By collecting low‑level IDE telemetry (about 180K events) and continuously computing 27 metrics, it transforms learner behavior into 13,633 quantifiable segments. From this foundation we derive a taxonomy that characterizes activity before the first AI query, between consecutive queries, and across the entire session, enabling systems to react not only to what learners ask but also to what they have done beforehand. In a classroom trial, behavior‑aware prompts reduced the proportion of queries made without independent work from 50.0% to 20.7%. As further validation, TutorTrace was applied to two held‑out prediction tasks: forecasting a query within the next 60 seconds (AUROC=0.726) and determining whether an upcoming query reflects guided or dependent help‑seeking (AUROC=0.717), demonstrating downstream utility.
Blogger's Review: This work convincingly shows how fine‑grained IDE actions can be turned into real‑time features, offering a promising route toward scalable, adaptive AI tutoring.