This work discusses the fundamental limits of Variable-Length Stop-Feedback (VLSF) communication systems, in which transmission continues until an acknowledgment is received. It focuses on developing analytical tools and decoding strategies that approach optimal performance under practical constraints. In one direction, an optimization framework for sparse VLSF codes based on saddlepoint approximations is developed, enabling efficient gradient-based optimization of decoding configurations for common memoryless channels such as AWGN, binary symmetric, and binary erasure channels. A refined decoding rule extending the conventional fixed-threshold strategy is also introduced, yielding tighter achievability bounds. In another direction, VLSF transmission over correlated noncoherent fading channels is studied. Reliability-guaranteed decoding strategies are established by deriving computable finite-blocklength bounds on information density, which enable constructive decoding procedures and provide insight into the impact of channel correlation on system performance.