The implementation for SIGIR 2026: Learning to Retrieve from Agent Trajectories.