Unsupervised domain adaptation for object detection (UDA) has achieved remarkable progress under well-lit conditions. However, significant domain discrepancies between daytime and nighttime scenes pose persistent challenges for nighttime object detection. To mitigate performance degradation caused by day-to-night domain shifts, we propose CAP-Shift, a novel unsupervised domain adaptation framework specifically designed for nighttime object detection. CAP-Shift incorporates a Light-Aware Nighttime Augmentation (LANA) module, which enhances adaptability to nighttime scenarios by decoupling luminance information for separate modeling of nighttime degradation and detail enhancement. Building upon this, we further introduce a Dynamic Class-aware Threshold Adaptation (DCTA) strategy that dynamically adjusts confidence thresholds for each class during training to effectively filter noisy pseudo-labels and suppress false negative predictions. Extensive experiments on two public datasets demonstrate that our method can effectively reduce domain discrepancies and improve the quality of pseudo-labels.