Conflict-Core Extraction for Explainable Package Version Planning with Language Models

Abstract

Modern software development relies heavily on open-source ecosystems, necessitating robust package management systems to handle complex dependency graphs. As applications scale, dependency networks often become deeply nested, leading to version conflicts that are computationally difficult to resolve and conceptually challenging for human developers to understand. Traditional package managers utilize Boolean satisfiability solvers to guarantee correct version planning; however, when constraints are unsatisfiable, these solvers typically output cryptic error messages. This lack of explainability creates a significant bottleneck in software engineering workflows. This paper proposes a novel framework for explainable package version planning by combining Conflict-Core Extraction with Large Language Models. By isolating the Minimal Unsatisfiable Subset of dependency constraints, the proposed system distills massive dependency graphs into a precise conflict core. This core is then translated into natural language prompts and processed by language models to generate highly contextual, human-readable explanations and actionable resolution strategies. Through extensive empirical analysis across multiple open-source ecosystems, we demonstrate that this methodology significantly reduces the cognitive load required to debug dependency conflicts. The results indicate a substantial improvement in diagnostic accuracy and a marked decrease in resolution time compared to traditional package management diagnostic outputs