Automated SOLID Violation Detection and Refactoring on Real Life Repositories

Abstract

This paper presents an automated framework for detecting and refactoring SOLID design principle violations in real-world software repositories using Large Language Model (LLM) agents. Recognizing that architectural issues often span multiple files, we implement a dynamic agentic detection system that utilizes Abstract Syntax Tree (AST) summaries and repository navigation tools to identify violations with high accuracy. The framework further attempts to resolve these issues through atomic full-file refactoring in isolated workspaces, validated by native test suites and quantitative code quality metrics. We evaluated our pipeline using the DeepSeek model on nine open-source repositories across Java, Python, and Kotlin. Our experimental results show that while the detection agent achieves up to 73\% accuracy, the refactoring component remains prone to syntactic errors, particularly in compiled languages like Java and Kotlin. Consequently, while the system provides a robust basis for automated design analysis, we conclude that it currently serves best as a semi-automated assistant within a developer-in-the-loop workflow rather than a fully autonomous maintenance tool.