一句话：提出 Reprodgen 分阶段框架，从论坛与 Issue 的非正式 bug 描述重建可执行的 buggy-patched 程序对，并分别评估可执行性与行为忠实度。
讨论论坛和 GitHub Issue 中常描述真实 bug 及其修复，但往往缺少完整代码、数据、依赖与执行上下文，导致难以复现。本文提出 Reprodgen，一个分阶段框架：先从非正式问题描述中抽取并精炼语义引导，再生成带有 mock 数据与推断依赖的完整程序，最后分别评估生成程序的可执行性与行为忠实度。作者在 203 个 Stack Overflow 帖子上评估，最强配置可执行 60.0% 的 buggy 程序与 52.2% 的 patched 程序，行为成功率分别为 58.1% 与 51.7%；在 50 个 GitHub Issue 上，可执行率分别为 96.0% 与 56.0%，行为成功率为 62.0% 与 46.0%。消融实验表明，结构化引导能提升 buggy 代码的可执行性，但不能提升其行为忠实度，说明二者应分别评估。
Discussion forums often describe real bugs and fixes but omit complete code, data, dependencies, and execution context. We present Reprodgen, a staged framework for reconstructing executable buggy-patched program pairs from such informal problem descriptions. Reprodgen extracts and refines semantic guidance, generates complete programs with mock data and inferred dependencies, and evaluates both executability and behavioral faithfulness. On 203 Stack Overflow posts, the strongest configuration executes 60.0% of buggy and 52.2% of patched programs, with behavioral success rates of 58.1% and 51.7%, respectively. On 50 GitHub Issues, the strongest configuration executes 96.0% of buggy and 56.0% of patched programs, with behavioral success rates of 62.0% and 46.0%, respectively. An ablation on GitHub Issues shows that structured guidance improves the executability of buggy code but not its behavioral faithfulness, reinforcing the need to evaluate these properties separately.