Equipping Agents with "Perspective Glasses": Mitigating Blind Exploration via Dynamic Observability for Issue Resolution

DAIRA: Dynamic Analysis-enhanced Issue Resolution Agent DAIRA is an automated repair framework that deeply integrates dynamic analysis into the agent decision loop. By adopting a Test Tracing-Driven workflow and equipping agents with "Perspective Glasses" (Dynamic Observability), DAIRA mitigates the issue of blind exploration common in existing static analysis-based approaches. 📂 Repository Structure 🏗️ Project Overview The directory structure is organized to align strictly with the Research Questions (RQs) and experimental setup described in the manuscript. DAIRA/ ├── Motivating Example/ # 🔎 Case Studies with Gemini 3 Flash Preview ("Perspective Glasses" in action) │ ├── sympy_17630_traj/ # Full execution trajectory for SymPy-17630 │ └── matplotlib-22719_traj/ # Full execution trajectory for Matplotlib-22719 ├── prompt/ # 🧠 Prompts ├── swebench result/ # 📊 Evaluation Results │ ├── DAIRA/ # Ours Results │ └── SWE-agent/ # Baseline Results ├── RQ1-Effectiveness/ # 🏆 RQ1: Effectiveness Analysis (vs. SOTA) ├── RQ2-Model_Impact/ # 🤖 RQ2: Model Generalizability Analysis ├── RQ3-Robustness/ # 🛡️ RQ3: Difficulty Robustness Analysis ├── RQ4-Ablation_Study/ # 🧪 RQ4: Ablation Study (Component Contribution) └── result_image/ # 🖼️ Paper Experimental Results Figures └──traj/ #Full Execution Trajectory For All 1. Qualitative Analysis Motivating Example/: Detailed case studies demonstrating the "Perspective Glasses" effect. sympy_17630_traj/ & matplotlib-22719_traj/: Contains full execution trajectories (.traj), and patches (.patch). These examples illustrate how DAIRA successfully resolves complex defects (e.g., polymorphic control flows, implicit type degradation) where baselines failed due to a lack of intermediate execution state observation. 2. Core Components prompt/: Contains the System Prompts and Analyse Prompts designed to guide the agent in utilizing dynamic analysis tools and receiving structured analysis report. swebench result/: The raw evaluation logs (JSON format) generated from running DAIRA and baseline methods on the SWE-bench Verified dataset. Includes results for: DAIRA (Ours) SWE-agent (Baseline) Across models: DeepSeek V3.2, Gemini 3 Flash Preview, Qwen 3 Coder Flash. 3. Experimental Results (RQ1 - RQ4) This section corresponds to the quantitative evaluation of the framework: RQ1-Effectiveness/: Data and scripts for Effectiveness Analysis. SOTA_Methods.xlsx :Instance-level resolution data for DAIRA and SOTA baselines (e.g., SWE-agent, AutoCodeRover). weien_plot.py : performance comparison figures. RQ2-Model_Impact/: Data for Model Generalizability Analysis. Different_model.xlsx:Evaluates the framework's performance across different LLM backbones (DeepSeek, Gemini, Qwen) to verify that DAIRA's improvements are model-agnostic. different_model_performence.py:Code for generating performance analysis charts across different models. RQ3-Robustness/: Data for Robustness Analysis. different_method_different_level_reslovedrate.py: Code for plotting resolution rate maps of DAIRA and other state-of-the-art methods across tasks of varying difficulty. The data for the tables in this section is based on the Different_model.xlsx data sheet. RQ4-Ablation_Study/: Detailed Data for Ablation Study. Verifies the contribution of individual components by removing them: without Dynamic Analysis Tool without Test Tracing-Driven Workflows without Trace Log Semantic Analysis DAIRA(ours) SWE-agent(baseline)