AuraTracer智迹闻
中文

EVENT DOSSIER

RefactorPlatform: An Open-Source Harness for Controlled Evaluation of Repository-Scale Refactoring Agents

2026-09-07 12:00 Science 🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
4mentions
SummaryAI generated

On September 7, 2026, arXiv released an open-source evaluation tool called RefactorPlatform, designed to provide a isolated testing environment for code agents in cross-file refactoring tasks. This platform supports isolated runtime of workspaces, real-time terminal stream monitoring, AST validation, and telemetry data export, ensuring the reproducibility and auditability of evaluations. In the 100 multi-file task tests included in RefactorBench, the block-based strategy based on the abstract syntax tree (AST) achieved 25% to 30% better performance than the naive block-based approach; while the retrieval enhancement mode increased token overhead, it kept the unit refactoring cost unchanged; moreover, lightweight single-agent configurations were more performant than sub-agent configurations.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
GitHub Copilot CLIOpenRouterRefactorBenchRefactorPlatform

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
GitHub Copilot CLI × Op…1GitHub Copilot CLI × Re…1GitHub Copilot CLI × Re…1OpenRouter × RefactorBe…1OpenRouter × RefactorPl…1RefactorBench × Refacto…1

SignalsSIGNALS

Keyword heat
  • RefactorPlatform1
  • OpenRouter1
  • GitHub Copilot CLI1
  • RefactorBench1

All reports (1)SOURCES

A arXiv cs.AI en 2026-09-07 12:00

RefactorPlatform: An Open-Source Harness for Controlled Evaluation of Repository-Scale Refactoring Agents

arXiv:2609.04898v1 发布开源评估工具 RefactorPlatform,用于隔离并测试代码代理在跨文件重构任务中的设计选择。该平台固定环境并显式控制模型架构(OpenRouter、GitHub Copilot CLI)、执行模式(基线、检索增强、多智能体)及提示词特异性,支持隔离工作区运行、实时终端流、AST 验证及可导出遥测数据。在 RefactorBench 100 个多文件任务上测试显示,基于 AST 的分块比朴素分块提升 25-30%,检索增强虽增加 token 开销但保持单位重构成本不变;且轻量级单智能体(86%)优于子智能体配置(66%)。RefactorPlatform 旨在使重构代理评估具备可复现性与可审计性。