issue triage

Prior work lacks empirical characterization of problem-solving processes in software development. Method: Integrating grounded theory coding, sequential pattern mining, and multidimensional statistical analysis on 356 Mozilla Firefox issue reports, this study extracts fine-grained, reusable problem-solving process patterns from collaborative textual artifacts. Contribution/Results: We identify 47 empirically grounded process patterns—challenging the traditional linear assumption by revealing pervasive nonlinearity: 73% of fixes involve iterative backtracking or parallel activities. The resulting process landscape and pattern catalog systematically characterize distributional regularities across issue types, defect categories, and repair durations. This advances understanding of real-world engineering complexity and provides an evidence-based foundation for process optimization, collaborative tool design, and developer support.

To address the high cost and low efficiency of manual triage for bug reports in large-scale software projects, this paper proposes a GitHub-integrated AI assistant that, for the first time, unifies multi-task deep learning across the full bug-report management pipeline—including duplicate detection, severity prediction, and fix-file recommendation. The method integrates a BERT-based semantic model, a graph neural network (GNN) for code localization, and a lightweight GitHub App architecture to deliver end-to-end interpretable recommendations. Evaluated on multiple benchmark datasets, the approach achieves F1 scores of 0.82–0.89. A real-world user study with professional developers demonstrates an average 47% reduction in triage time and a 76% recommendation adoption rate. These results significantly advance the automation level and practical utility of defect management in industrial settings.

This work addresses the limited trust developers place in AI-generated bug reports due to their frequent lack of actionability and reproducibility. The authors propose a novel approach that integrates code coverage analysis with large language models (LLMs) to automatically detect defects in uncovered code regions and generate structured bug reports containing severity ratings, reproduction steps, and repair suggestions. A key innovation is an LLM-driven prioritization mechanism that substantially outperforms traditional rule-based methods. Evaluated on 13 Python projects, the method produced 10,467 reports; manual assessment of the top 130 revealed an 84.6% validity rate. Compared to CoverUp, it achieves higher defect validity (81.0% vs. 76.2%), a 50% improvement in P@3, and a 41% gain in mean reciprocal rank (MRR).

Continuous Integration (CI) practices suffer from severe monitoring deficiencies: developers largely neglect critical metrics such as “build health” and “time-to-fix failed builds,” while mainstream CI services offer only weak native monitoring capabilities, forcing reliance on fragmented and often redundant third-party tools. Method: We conducted a triangulated investigation—including documentation analysis, developer surveys, functional audits of CI platforms, and case studies of open-source projects—to systematically identify cognitive gaps and practical monitoring needs. Contribution/Results: Our study provides the first empirical evidence that although over 80% of developers track test coverage, only a minority monitor build health or timeliness; further, all major CI services lack built-in multidimensional monitoring support. These findings establish an evidence-based foundation for designing next-generation CI monitoring frameworks and prioritizing tooling enhancements.

This paper addresses the challenge of low reproducibility in software bug reports caused by ambiguous or incomplete reproduction steps (S2Rs). To tackle this, we propose a cross-modal quality assessment method that synergistically integrates large language model (LLM)-based semantic understanding with dynamic UI analysis. Our approach constructs a program state model and enables fine-grained semantic alignment between S2R text descriptions and corresponding GUI interaction actions, thereby bridging the lexical diversity and programmatic semantic gap inherent in prior methods. We introduce the first LLM-driven framework that deeply unifies natural language parsing with GUI state modeling. Evaluated on standard benchmarks, our method achieves a 25.2% improvement in F1-score for S2R quality labeling and a 71.4% gain in F1-score for missing-step completion over state-of-the-art approaches—significantly enhancing both reproducibility and debugging efficiency.

This work addresses the challenge that bug reports in open-source projects often lack reproducible tests, hindering effective repair. The authors propose a multi-stage agent framework that decomposes test reproduction into four phases: defect localization, root cause analysis, test planning, and test generation. For the first time, this approach integrates tool-augmented mechanisms and task decomposition strategies, leveraging code-text graph retrieval, runtime environment interaction, and large language models to achieve repository-level code understanding and flexible test construction. Evaluated on SWT-bench-lite and SWT-bench-verified, the method achieves reproduction success rates of 58.43% and 70.30%, respectively—substantially outperforming existing approaches—while maintaining a low average cost of only $0.14 per instance and demonstrably enhancing downstream repair performance.

This work addresses the limitation of existing bug reports—often lacking critical information—which hinders the effectiveness of automated program repair (APR). The authors propose TrajSpec, a novel approach that introduces trajectory-guided reasoning and hierarchical evidence representation. By leveraging execution trajectories collected via proxy runs, TrajSpec extracts multi-level evidence to refine bug reports and further enhances them through integration with repository context. Combining trajectory-guided inference, hierarchical evidence modeling, and large language model–driven report generation, TrajSpec significantly improves the Pass@1 performance of multiple APR systems on SWE-Bench Lite, raising it from 47.00% to as high as 72.00%. Ablation studies confirm the contribution of each component to the overall effectiveness.