Scholarlink

GitHub is one of the most popular collaborative development tools used by many institutions and enterprises. The GitHub issue community serves as the primary tool for project developers to gather program bug information, with GitHub users posting issues based on encountered program bugs. However, a considerable portion of these issues ultimately remain unfixed by developers. We term the issues fixed by developers as ‘dev-fixed’ issues, and conversely, we call them ‘wontfix’ issues. Through manual analysis of 2500 bug-related issues collected from five GitHub projects, we summarize the unresolved issues from four aspects and propose a multi-agent approach based on large language model (LLM) to identify dev-fixed issues. We design the previously summarized four aspects into different agents, independently analyzed by the LLM, and ultimately parallelly provide information to a summary agent for the final determination. Compared to the baseline, our approach shows the best performance in the metric of F1 score, recall rate and precision rate, with the value of 0.852, 0.862 and 0.857 respectively. Ablation experiments demonstrate the critical importance of each agent to our approach.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Multi-agent Collaboration Approach for Identifying Developer-Fixed Issues in GitHub Projects

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Zhengru Han,
  • Bo Jiang,
  • Weihao Xue,
  • Chaoqun Dai,
  • Qiao Huang,
  • Ye Wang

摘要

GitHub is one of the most popular collaborative development tools used by many institutions and enterprises. The GitHub issue community serves as the primary tool for project developers to gather program bug information, with GitHub users posting issues based on encountered program bugs. However, a considerable portion of these issues ultimately remain unfixed by developers. We term the issues fixed by developers as ‘dev-fixed’ issues, and conversely, we call them ‘wontfix’ issues. Through manual analysis of 2500 bug-related issues collected from five GitHub projects, we summarize the unresolved issues from four aspects and propose a multi-agent approach based on large language model (LLM) to identify dev-fixed issues. We design the previously summarized four aspects into different agents, independently analyzed by the LLM, and ultimately parallelly provide information to a summary agent for the final determination. Compared to the baseline, our approach shows the best performance in the metric of F1 score, recall rate and precision rate, with the value of 0.852, 0.862 and 0.857 respectively. Ablation experiments demonstrate the critical importance of each agent to our approach.