The PsyDefDetect shared task focuses on classifying nine psychological defense mechanisms in multi-turn dialogues, a problem complicated by severe label imbalance and the implicit, context-dependent nature of defenses. In this work, we investigate several approaches for dialogue-level defense detection, including supervised baselines and large language model (LLM)-based pipelines. Our primary system is a retrieval-augmented LLM framework with hierarchical prediction and lightweight heuristics for decision calibration. Experiments on the PSYDEFCONV dataset show that LLM-based methods improve overall performance compared to supervised baselines, but still struggle with fine-grained distinctions, especially for minority labels. These findings highlight the challenges of modeling implicit psychological constructs in dialogue.