SWE-Together: Evaluating Coding Agents in Interactive User Sessions