Proceedings of the International scientific and practical conference ―New York Global Science Conference 2026‖ (May 18-20, 2026) / Publisher website: www.naukainfo.com. – New York, USA, 2026. - 322 p.

128 As software architectures become increasingly distributed, the complexity of diagnosing automated test failures grows exponentially. In a unified E2E testing pipeline where a single scenario interacts with desktop applications via Appium and simultaneously verifies state persistence through REST APIs, a failure can originate from multiple points of failure. Traditional test reporting tools typically only capture the final assertion failure, such as a missing UI element or an unexpected HTTP status code. Consequently, automation engineers are forced to manually sift through massive execution logs, network captures, and server-side traces to perform Root Cause Analysis (RCA). This manual debugging process creates a massive bottleneck, severely diminishing the return on investment of automated test suites within rapid release cycles. To mitigate this challenge, the integration of Large Language Models (LLMs) into the diagnostic layer offers a paradigm shift from manual inspection to automated intelligence. The core technical workflow begins immediately after a test failure occurs within the C# execution framework. Instead of merely logging a standard exception, a custom-built diagnostic collector aggregates a comprehensive contextual "snapshot" of the system state. This snapshot includes the exact Appium/WinAppDriver [1] stack trace, the raw JSON request and response bodies from recent RestSharp [2] or HttpClient [3] calls, the current page source XML, and any relevant console logs. This unstructured, heterogenous data is then formatted into a structured prompt and transmitted to an LLM provider via private enterprise endpoints. The role of the LLM agent is to act as an expert system capable of semantic correlation. Modern models, such as Claude or GPT, possess deep contextual understanding of programming logic, network protocols, and testing patterns. By ingesting the failure snapshot, the model can reason through the chronological sequence of events. For instance, if the desktop UI failed to display a user profile, but the preceding API validation log indicates that the backend returned an HTTP 500 Internal Server Error due to a database deadlock, the LLM can instantly correlate