一句话：提出 VSCode 扩展 FlightRecorder，细粒度追踪 AI 助手与开发者的代码演化数据。
AI 编码助手正在改变软件开发，但开发者与 AI 代理协作时，代码如何在细粒度交互层面演化仍不清楚。现有遥测无法区分代理生成的修改、行内补全与人工编辑，也无法精确追踪其发生的时间与位置。为此，作者提出 FlightRecorder：一个轻量级 VSCode 扩展，在不干扰开发者工作流的前提下收集细粒度代码演化数据。该扩展集成三类数据源：代理模式编辑快照、行内补全和聊天历史，并全部加上时间戳以实现精确归因。研究者可导出结构化数据，并支持匿名化与基本可视化。通过提供复制包与演示视频，FlightRecorder 使研究者能够探究 AI 辅助开发中的基本问题。
AI coding assistants like GitHub Copilot are transforming software development. Yet fundamental questions remain about how code evolves as developers collaborate with AI agents, particularly at fine-grained levels of interaction -- a gap that persists partly because existing telemetry cannot distinguish agent-generated changes from inline completions and human edits, nor precisely track when and where they occur. To address this gap, we present FlightRecorder, a lightweight VSCode extension that collects fine-grained code evolution data without disrupting the workflows of developers. Our extension integrates three data sources: snapshots of agent-mode edits, inline completions, and chat history, all timestamped for precise attribution. Researchers can export these data in structured formats alongside anonymization and basic visualization functionality. By making such fine-grained evolution data accessible, FlightRecorder enables researchers to investigate fundamental questions about AI-assisted development. We provide a replication package, including the code for FlightRecorder and the data used throughout our paper, at https://doi.org/10.5281/zenodo.20134365. We further provide a screencast demo of our tool at https://youtu.be/bZR6aPROVMU.
@inproceedings{paperbot12206,
title = {FlightRecorder: Tracing Fine-Grained Development with AI Assistants},
author = {Tobias Dick and Benedict Bliem and Kallistos Weis and Alisa Welter and Sven Apel},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834607}
}