From Seed to Scope: Reasoning to Identify Change Impact Sets

Change impact analysis (IA), which identifies the set of co-changed program elements (i.e., impact set), is critical for several software engineering tasks. However, existing IA approaches struggle with a trade-off between precision (correctly detecting impacted elements) and recall (detecting all relevant ones). More importantly, they are limited in intent-aware settings, where co-changing elements are determined by a given change intent. In this work, we propose Ripple, an intent-aware IA approach that leverages large language models (LLMs) to capture change dependencies by linking intent to program elements and estimating their co-change relationships. To address the precision-recall tradeoff, we adopt a two-phase design:

(1) a seed-to-scope expansion strategy that expands the impact set using evolutionary and dependence coupling to improve recall, and (2) a plan-then-predict strategy where an LLM-generated change plan refines impact estimation for higher precision. We evaluate Ripple on real-world commits from Apache projects, achieving a 39.7%-380.8% improvement in F1-score over existing IA approaches. In addition, Ripple introduces flexibility, allowing users to prioritize higher precision or recall based on their preferences.