DART-SD addresses limitations in training multi-turn tool-calling agents by modeling task execution as an Interaction-State Transition Graph that captures the combinatorial solution space created by order-independent subtasks. Rather than forcing diverse valid solution paths into a single training trajectory, the framework identifies ‘Critical Topological Breakpoints’ during rollouts and retrieves success-based recovery references for those points. It then applies breakpoint-guided localized supervision that computes training loss only on the recovery steps, preserving otherwise-valid reasoning prefixes from gradient damage.