一句话：提出 IDE 集成工具 TaCoS，用 LLM 从 IDE 与浏览器活动自动生成任务上下文摘要，助开发者中断后快速恢复任务。
软件开发中的频繁打断迫使开发者重建任务的心智与工作上下文，而现有的可视化时间线或手写笔记要么需额外投入，要么难以支持心智上下文重建。本文提出 TaCoS，一个集成于 IDE 的工具，它基于 IDE 与浏览器活动记录自动生成任务上下文摘要，融合推断出的开发意图、实现步骤以及相关工件的链接。作者以 32 名软件开发者开展受控实验，让每人恢复三个编程任务，并分别使用手写笔记、近期编辑时间线和 TaCoS 生成摘要三种线索。结果显示，生成摘要带来的恢复耗时与编辑滞后最短，整体被评为最有帮助的线索；不过时间线条件下的任务成功率最高。参与者认可摘要结构化、低负担地呈现过往活动，但指出其缺少手写笔记中常有的前瞻性信息。研究据此指出，将自动摘要与轻量用户标注结合是支持开发环境任务恢复的可行方向。
Interruptions are a frequent challenge in software development, requiring developers to reconstruct both the mental and working context of their tasks. While prior work has proposed visual timelines or manual notes to aid resumption, these approaches require effort or lack support for mental context reconstruction. Recent advances in large language models (LLMs) offer new opportunities to support task resumption through automatically generated summaries. We present TaCoS, an IDE-integrated tool that generates task context summaries from IDE and browser activity, combining inferred intent, implementation steps, and links to relevant artifacts. We evaluated TaCoS in a controlled lab experiment with 32 software developers, who resumed three programming tasks. Each task was resumed using one of three cues: a manual note, a timeline of recent edits, or a TaCoS-generated summary. The generated summary led to the shortest resumption and edit lag, and was rated as the most helpful cue overall. However, the timeline condition achieved the highest task success rate. Participants valued the summaries’ structured, low-effort presentation of past activity, but noted that they often lacked forward-looking information, which was typically present in the self-authored manual notes. These findings highlight the potential of combining automated summarization with lightweight user annotations to support task resumption in development environments.
@inproceedings{paperbot11871,
title = {TaCoS: Generated Context Summaries for Task Resumption},
author = {Alexander Lill and Valentin Hollenstein and Roy Rutishauser and André N. Meyer and Thomas Fritz},
booktitle = {ICSE},
year = {2026},
doi = {10.1145/3744916.3773226}
}