Enhancing Adaptive Test Healing with Graph Neural Networks for Dependency-Aware Decision Making

Flaky tests are a major obstacle in modern CI/CD pipelines, leading to unreliable feedback, increased reruns, and developer frustration. Our previously published adaptive healing framework combined Large Language Models (LLMs) and Reinforcement Learning (RL) to automate flaky test recovery, but it assumed test independence and failed to account for structural dependencies between tests. In this paper, we introduce a significant extension to that baseline: a Graph Neural Network (GNN)-based Test Dependency Mapping layer that models intertest relationships. By integrating GNN embeddings with LLM-classified failures, the RL agent becomes dependency-aware, enabling more precise and efficient healing decisions. We evaluate the enhanced framework on a real-world industrial platform, a social lifestyle application actively used by thousands of users for health, nutrition, and coaching. Results show a 90% reduction in flaky test-related costs and faster, autonomous resolution of dependency-induced failures.