Abstract: Large language models (LLMs) have demonstrated outstanding performance in standalone function-level code generation but still struggle with repository-level code generation due to limited ...