一句话：提出对话式 AI 智能体工具 Path-Guider，在 IDE 中集成路径/切片、符号执行与差分测试，辅助 COBOL 程序理解与调试。
以 COBOL 为代表的大型机应用仍在支撑关键企业系统，但工具支持有限、与开发现代环境集成薄弱，使其调试、维护与变更影响分析十分困难。程序切片、符号执行、差分测试等分析技术虽在当代语言中已有充分研究，却难以直接适用于大型机语言，也缺乏面向开发者的自然交互入口。本文提出 Path-Guider：一个基于聊天交互的 AI 智能体，用于 COBOL 程序理解与调试。其核心是一个任务特定的智能体模式，统一编排多项程序分析能力，包括执行路径与切片检索、基于符号执行的测试输入生成、支持部分输入的模拟执行（mocked execution），以及用于变更影响分析的差分测试。开发者可在现代 IDE 内以自然语言查询直接调用这些分析。论文以公开演示视频的形式展示该工具的原型与交互流程。
Mainframe applications written in languages such as COBOL continue to support critical enterprise systems, yet their debugging, maintenance, and impact analysis remain challenging due to limited tool support and poor integration with modern development environments. While program analysis techniques such as program slicing, symbolic execution, and differential testing are well studied for contemporary languages, their applicability to mainframe languages and accessibility through natural developer interactions are still limited. In this paper, we present Path-Guider, a chat-based AI agent for COBOL program understanding and debugging. Path-Guider introduces a task-specific agent mode that integrates multiple program analysis capabilities, including execution path and slice retrieval, symbolic-execution-based test input generation, mocked execution with partial inputs, and differential testing for change impact analysis. Developers interact with these analyses through natural language queries directly within a modern IDE. The demo video of Path-Guider is available publicly 1
@inproceedings{PimparkhedeKSYS26,
title = {Path-Guider: An IDE-based Agentic AI Tool for Path-based COBOL Program Analysis},
author = {Sameer Pimparkhede and Pankaj Kumar Kalita and Diptikalyan Saha and Toshiaki Yasue and Fumiko Satoh},
booktitle = {FSE},
year = {2026},
doi = {10.1145/3803437.3806422}
}