Effective coordination of multi-robot swarms in dynamic adversarial environments demands both strategic decision making and adaptive local planning. Human intelligence provides superior high-level reasoning and strategic adaptation, while robotic swarms excel in distributed planning and large-scale execution. We propose a human-in-the-loop framework for multi-robot mission planning that leverages a large language model (LLM) to incorporate human-supervisor guidance as prior constraints, guiding high-level objective adjustments, and conducts local replanning on a selected subset of robots using an evolutionary algorithm solver, minimizing plan perturbations and reducing online computational load. In simulated adversarial scenarios with previously unknown risk zones, our approach improves deadline compliance and adherence to safety, while maintaining high task completion rates and a lower solver effort. This hierarchical integration demonstrates the potential of combining human strategic insight with swarm-level autonomy, offering a promising paradigm for collaborative multi-robot systems operating in contested and high-stakes environments.