Recoleta
GitHub
Source note

FeatX: Editing Software by Editing Features for Repository-Level Code Evolution

Authors
Xutian Li, Yifeng Zhu, Xianlin Zhao, Yanzhen Zou, Lu Zhang, Bing Xie
Source
arXiv
Published
Collection
Software Intelligence
Software EvolutionCode IntelligenceRepository Level EditingLLM AgentsHuman AI Interaction
Open arXivMarkdown source

Summary

FeatX lets developers change an existing Java repository by editing a feature list, then maps those feature edits to code patches. It reports lower user workload and better function-level localization than ChatGPT, Cursor Agent, and direct LLM baselines on 38 feature-editing commits.

Problem

  • Repository-level feature changes often span several files and functions, so developers must connect product intent to code locations before making edits.
  • Chat and autocomplete tools leave much of the repository context selection to the user, which raises workload in unfamiliar codebases.
  • The paper cites prior work that about 60% of repository maintenance tasks involve feature evolution, making this a common software maintenance case.

Approach

  • FeatX extracts a two-level epic-feature hierarchy from a repository and links each feature to classes, methods, or files.
  • It builds feature groups by combining static dependency signals with semantic similarity from LLM summaries and Sentence-BERT embeddings, then clusters them with the Leiden algorithm.
  • A three-stage Evolution Agent expands context, localizes changed feature intent to code regions, and generates class-wise line-level diffs.
  • The UI has Feature, CodeMap, Agent, and Diff panels so users edit features, inspect mapped code, review the agent plan, and confirm patches.

Results

  • In a 10-person study, mean NASA-TLX workload dropped from 12.5 with ChatGPT to 7.4 with FeatX, a 41% reduction.
  • SUS usability rose from 73 with ChatGPT to 84 with FeatX, a 15% increase; confidence improved with p<0.001.
  • On 38 feature-editing commits across five Java repositories, FeatX reached 41.6% precision, 35.8% recall, and 0.385 F1 for function-level modification localization.
  • Claude-opus-4.5 was the strongest direct LLM baseline with 50.7% precision, 18.4% recall, and 0.270 F1; FeatX reports a 42.6% relative F1 gain over it.
  • FeatX cost $0.07 total in the replay study, compared with $45.05 for direct Claude-opus-4.5 and $1.28 for GPT-4o-mini.

Link