一句话：提出基于 LLM 智能体的 PL-IaC 程序自动修复框架，结合语义图重建与约束引导诊断，修复成功率最高达 61.8%。
基础设施即代码（IaC）已成为工业云交付的常规环节，PL-IaC 框架让工程师能用熟悉的编程抽象构建基础设施，但修复有缺陷的 PL-IaC 程序在实践上仍代价高昂且困难。现有工作多聚焦缺陷分析与静态检查，缺乏面向 PL-IaC 的端到端自动修复框架；而通用软件的程序自动修复技术既缺少 PL-IaC 领域知识，也缺乏对拓扑依赖的有效建模与引导，难以修复拓扑相关缺陷。本文提出 InfraAPR，一个基于智能体的 PL-IaC 自动修复框架，包含三个协同模块：面向阶段特化的 LLM 推理 Agent 模块、重建图式 PL-IaC 语义的语义映射模块，以及基于结构化诊断识别约束违反并生成修复程序的约束引导修复模块。在工业基准上的实验表明，InfraAPR 持续优于各基线，修复成功率最高达 61.8%，相对最强基线平均提升 15.4%，显示出对工业 PL-IaC 工作流的实际价值与可靠性提升。
Infrastructure as Code (IaC) is now a routine part of industrial cloud delivery, and Programming Languages IaC (PL-IaC) frameworks further enable engineers to build infrastructure using familiar programming abstractions. However, repairing buggy PL-IaC programs in practice remains costly and highly challenging. Existing quality assurance efforts for PL-IaC have largely concentrated on defect analysis and static checking, while end-to-end automated repair frameworks for PL-IaC programs remain largely absent. In addition, automated program repair techniques designed for general-purpose software often lack PL-IaC-specific domain knowledge, as well as effective modeling and guidance for topology dependencies, making them ineffective for repairing topology-dependent bugs in PL-IaC settings. This paper presents InfraAPR, an agent-based automated repair framework for PL-IaC programs. To enable end-to-end automated repair, InfraAPR integrates three coordinated modules: (1) an InfraAPR Agent Module for phase-specialized LLM reasoning, (2) a Semantic Mapping Module for reconstructing graph-based PL-IaC semantics, and (3) a Constraint-Guided Repair Module that diagnoses constraint violations and generates repaired programs based on structured diagnostics. We evaluate InfraAPR on an industrial benchmark. Experimental results show that InfraAPR consistently outperforms the baselines, achieving a repair success rate of up to 61.8% and improving the strongest baseline by 15.4% on average. These results demonstrate its substantial practical value and improved reliability for industrial PL-IaC workflows.
@inproceedings{paperbot12523,
title = {InfraAPR: An Agent-Based Repair Framework for Cloud Infrastructure-as-Code Programs},
author = {Quan Chen and Zhao Liu and Chunfang Li and Yifan Zhang and Ning Hu and Qing Liao},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834526}
}