Deteção de Code Smells em JavaScript utilizando LLMs em ambiente local

Abstract

Code smells negatively impact software maintainability and evolution. Although LLMs demonstrate potential in static analysis, the reliance on proprietary cloud-based models raises data privacy and intellectual property concerns. This paper investigates the effectiveness of locally executing an LLM (gpt-oss-20b) to detect code smells in JavaScript projects. A study was conducted using a dataset of 840 instances, comparing three prompt engineering strategies: Zero-Shot, Role-Based, and Chain-of-Thought. The results indicate that while the model detects Long Method with high performance, God Class detection is significantly more challenging due to context limitations, and the Chain-of-Thought strategy is crucial for reducing false positives. The findings suggest that local LLMs are a viable and secure alternative for code review, provided that appropriate prompting techniques are applied.