Separating Request Persistence from Feedback in Student–LLM Conversations

Repeated requests for code in student–LLM conversations can resemble a feedback loop, but repetition also reflects conversation composition and task continuity. We conduct a descriptive audit of StudyChat using its existing request taxonomy, retaining 14,322 strictly adjacent student-turn pairs. We compare observed transitions with a finite-composition reference that preserves each conversation's length, category counts, and available next-turn positions. A code-writing request is followed by another in 37.17% of eligible cases, versus a source-opportunity-standardized reference of 28.26% (+8.91 percentage points). Question-to-code transitions occur at 9.37% versus 11.01% (-1.64 points). Preserving category counts within conversation thirds reduces these differences to +4.51 and -0.77 points; questions also persist above their reference. Longer code-request prefixes increase raw continuation rates without a monotonic increase in excess persistence. These patterns describe category clustering in the observed corpus, not an identified effect of AI responses. Together with a historical label-dependence audit, they motivate composition-aware, response-specific evaluation before conversational persistence is interpreted as dynamic misalignment.