Artifact for "Fast First, Flawless Later: Re-evaluating Automated Program Repair for Hot Fixing Time-Critical Bugs"

This repository contains the complete results and artifacts associated with the paper “Can Automated Program Repair Tooling Support Hot Fix? An Empirical Study”.The goal of the study is to evaluate the ability of automated program repair (APR) tools to support time-critical hot fix development, with a focus on effectiveness, efficiency, and patch quality. The results include all run data and per-bug outputs generated by all evaluated tools across the HotBugs.jar benchmark. The dataset includes outcomes for the following APR tools: Search-based: ARJA Template-based: Cardumen Learning-based: RepairLLaMA Agentic / Context-aware: AutoCodeRover (multiple model configurations) and SWE-agent with Gemini-3.1-pro Each tool was executed against the HotBugs.jar dataset according to the experimental protocol described in the associated paper. File Naming Convention Arja, Cardumen, and RepairLlama tool results follow a consistent naming scheme: results_<tool>_<run>.zipRaw tool execution outputs, including logs, generated patches, intermediate build artifacts, and tool-specific diagnostic information.These files are large due to extensive patch generation and are primarily intended for inspection, auditing, or re-analysis. output_<tool>_<run>.zipAggregated, structured results for a specific tool and run.These archives contain files summarizing run status, patch counts, runtime measurements, and classification outcomes used for quantitative analysis. summaries_<tool>_<run>.zipLightweight summaries extracted from outputs.These include per-bug outcomes, success indicators, and metadata used to generate plots and tables in the paper. AutoCodeRover results with both the Llama3 and Llama3:70B settings are stored in autocoderover_results.zip. SWE-agent with Gemini-3.1-pro full results are stored in swe_agent_gemini3.1-pro_fullres.zip. Reruns of timed-out bugs with double budget are stored in swe_agent_gemini-3.1-pro_double_budget_reruns.zip.本仓库收录了与论文《自动化程序修复工具能否支持热修复？一项实证研究》相关的全部实验结果与配套工件。本研究旨在评估自动化程序修复（Automated Program Repair, APR）工具对时效性极强的热修复开发的支持能力，重点关注修复有效性、执行效率与补丁质量三个维度。 本次实验涵盖的全部结果包括所有待测工具在HotBugs.jar基准测试集上运行产生的全流程运行数据与单缺陷修复输出结果。 本数据集包含以下APR工具的实验结果： - 基于搜索的（Search-based）：ARJA - 基于模板的（Template-based）：Cardumen - 基于学习的（Learning-based）：RepairLLaMA - 智能体/上下文感知型（Agentic / Context-aware）：AutoCodeRover（支持多种模型配置） 所有工具均按照配套论文中描述的实验规程，在HotBugs.jar数据集上完成了测试运行。 ### 文件命名规范 ARJA、Cardumen与RepairLLaMA的实验结果遵循统一的命名规则： 1. results_<tool>_<run>.zip：工具原始运行输出包，包含运行日志、生成的补丁、中间构建工件以及工具专属诊断信息。由于该类文件包含大量补丁生成相关内容，体积通常较大，主要用于人工检视、结果审计或复现分析。 2. output_<tool>_<run>.zip：特定工具单次运行的聚合结构化结果包。该归档文件包含用于定量分析的运行状态汇总、补丁数量统计、运行时长测量结果与分类判定结果等内容。 3. summaries_<tool>_<run>.zip：从运行输出中提取的轻量级摘要文件，包含单缺陷修复结果、成功判定标识与元数据，可用于生成论文中的图表与表格。 采用Llama3与Llama3:70B模型配置的AutoCodeRover实验结果，均存储于autocoderover_results.zip中。