Flaky tests, characterized by inconsistent results across repeated executions, present significant challenges in software testing, especially during regression testing. Recently, there has been emerging research interest in non-idempotent-outcome (NIO) flaky tests—tests that pass on the initial run but fail on subsequent executions within the same environment. Despite progress in utilizing Large Language Models (LLMs) to address flaky tests, existing methods have not tackled NIO flaky tests. The limited context window of LLMs restricts their ability to incorporate relevant source code beyond the test method itself, often overlooking crucial information needed to address state pollution, which is the root cause of NIO flakiness.

This paper introduces NIODebugger, the first framework to utilize an LLM-based agent for fixing flaky tests. NIODebugger features a three-phase design: detection, exploration, and fixing. In the detection phase, dynamic analysis provides critical information (such as stack traces and custom test execution logs) from multiple test runs, which helps in understanding accumulative state pollution. During the exploration phase, the LLM-based agent identifies and provides instructions for extracting relevant source code associated with test flakiness. In the fixing phase, NIODebugger repairs the tests using the information gathered from the previous phases. NIODebugger can be integrated with multiple LLMs, achieving patching success rates ranging from 11.63% to 58.72%. Its best-performing variant, NIODebugger-GPT-4, successfully generated correct patches for 101 out of 172 previously unknown NIO tests across 20 large-scale open-source projects. We submitted pull requests for all generated patches; 58 have been merged, only 1 was rejected, and the remaining 42 are pending. The implementation of NIODebugger is provided as a Maven plugin accessible at https://github.com/NIOTester/NIODebugger.