Research
SWE-Chain: First Benchmark for Chained Release-Level Package Upgrades — 155 Transitions, 1,660 Requirements
Introduces a benchmark capturing continuous software maintenance at package-release granularity, where each version transition builds on the agent's prior codebase. Contains 12 upgrade chains across 9 real Python packages with 155 version transitions and 1,660 grounded upgrade requirements synthesized via a divide-and-conquer pipeline aligning release notes with code diffs. Fills a critical gap in coding agent evaluation beyond isolated issue resolution.
Source
↳ Follow the thread