<p>Session-based recommendation aims to predict the next item from short anonymous interaction sequences. Although existing methods have achieved notable progress in single-behavior sequence recommendation, multi-behavior scenarios still present three key challenges. First, heterogeneous behaviors such as click, add-to-cart, and purchase are often modeled as homogeneous transitions, making it difficult to distinguish behavior-specific intentions. Second, within-session transition dependencies and cross-session multi-behavior collaborative relations are rarely modeled in a unified framework. Third, time-domain encoders may smooth abrupt intent shifts and rhythmic behavioral patterns, leaving frequency-domain structures insufficiently explored. To address these issues, we propose frequency-enhanced dual-graph modeling for multi-behavior session-based recommendation (MF-DGRec). MF-DGRec jointly constructs a local session graph and a global multi-behavior graph, and further introduces a time–frequency dual-path encoder to capture both contextual dependencies and spectral patterns. On top of this, behavior-aware intent disentanglement and frequency-aware dual-graph routing are designed to adaptively fuse local and global evidence for different behavioral intentions. Experiments on three real-world datasets show that MF-DGRec consistently achieves the best performance. Compared with the best available baseline, MF-DGRec improves HR@10 by 10.48%, 2.38%, and 7.06%, and improves MRR@20 by 9.92%, 5.37%, and 5.62% on the three datasets, respectively. These results verify that jointly modeling heterogeneous behaviors, dual-graph collaborative structure, and time–frequency signals leads to more discriminative session representations and more accurate next-item prediction.</p>