Large Language Models (LLMs) are prone to producing bugs when generating code. For instance, LLMs often misuse APIs from libraries they are not familiar with. Incorrect API usage can severely impact the functionality of the generated code. However, existing error detection and program repair methods lack the targeted handling of API hallucinations and are often unsuitable when faced with data scarcity, such as private libraries. Therefore, a lightweight and transferable method for detecting and repairing API misuses is needed. This paper introduces the StepwiseGraphFix framework, which detects API misuses in LLM-generated code by retrieving and comparing relevant subgraphs from the code library’s knowledge graph. It then applies a stepwise refinement strategy to complete the repair. We constructed the UnifiedFixEval dataset for program repair in unseen libraries by integrating five existing code generation datasets, and validated our method against several other baselines on this dataset. Experimental results show that our approach outperforms current state-of-the-art methods, increasing the proportion of successful repair tasks by 8.1%, with average improvements of 4.0% in test pass rate. Ablation studies further confirm each component’s contribution to overall performance, and the multi-round repair mechanism improves repair quality.