Existing programming knowledge tracking studies have improved model performance by introducing code representation techniques and modeling the relationships between exercises and skills. But the impact of two critical factors, code quality and difficulty, has not been sufficiently explored. This limitation restricts the performance of models in distinguishing learners’ programming ability and accurately modeling their knowledge states. Thus, this paper proposes a Code Quality and Difficulty Aware Programming Knowledge Tracing (CQD-PKT) framework based on attention mechanisms, which aims to explore the impact of three key features of code quality, exercise difficulty and skill difficulty on modeling learners’ programming ability. Specifically, the framework first incorporates difficulty features using a self-attention mechanism to enhance the representation of both exercises and skills. Second, to avoid the subjectivity of manual scoring, the deep pretrained model DeepSeek is employed to achieve automated and objective code quality score. Finally, the framework employs a cross-attention mechanism to evaluate the impact of code quality on programming ability and to capture learners’ application of programming skills with varying difficulty during the coding process. The experimental results on real-world datasets demonstrate that CQD-PKT outperforms the best baseline model in predicting learners’ future performance. The AUC and ACC are respectively improved by approximately 2% and 1%.