一句话：探索用本地部署的 LLM 自动生成面向调试的代码导览，以降低新开发者上手与资深开发者带教的负担。
新开发者入职会带来可观的额外开销，往往需要资深开发者投入带教与撰写文档。代码导览（code tour）——对关键代码片段的结构化指引——能够简化这一过程，但人工编写仍然费时费力。本文提出一项研究计划，探讨如何利用本地可运行的大语言模型自动生成以调试为导向的代码导览；相比云端模型，本地模型在保密性、稳定性与可复现性上更具优势。作者计划应对三个关键挑战：其一，选取相关的代码片段；其二，生成清晰且具备上下文感知能力的解释；其三，自动评估所生成导览的质量。最终目标是自动产出持久、可验证的文档产物，帮助初级开发者浏览并理解代码库，从而减轻资深开发者的入职培训负担。
Onboarding new developers is a considerable overhead, requiring senior developers to provide mentorship and documentation. Code tours—structured guides to key code sections—can streamline this process, yet their manual creation still requires time and effort. This research investigates how to automate the generation of debugging-focused code tours using locally runnable Large Language Models (LLMs), which provide more confidentiality, stability, and reproducibility than cloud-based models. We plan to address three key challenges: (1) selecting relevant code segments, (2) generating clear and context-aware explanations, and (3) automatically evaluating the quality of generated tours. Ultimately, our goal is to reduce the onboarding burden on senior developers by automatically generating persistent, verifiable documentation artifacts to help junior developers navigate and understand a codebase.