MetaMorphic

Mylonite — Change the order. Reveal the fault. Mylonite is an AI-assisted concurrency testing system that discovers, reproduces, and verifies race conditions in software services that share state through Redis or Valkey. A service can pass every unit test and still lose data when two requests execute at the same time. For example, two requests might read the same balance, independently update it, and overwrite each other's changes. Traditional memory race detectors can miss these failures because the shared state lives outside the application's memory. Mylonite combines Meta's Muse Spark with deterministic execution control. Muse analyzes the service's source code and operation catalog, then proposes structured concurrency experiments: which operations to execute together, what setup they require, and which command interleavings are worth testing. A custom delay-only proxy controls when commands reach Redis or Valkey without modifying them. An independent, code-based contract oracle checks correctness. When a violation is found, Mylonite reproduces it 20 times and records a verifiable execution trace. The model proposes experiments; it never decides whether a bug exists. We also reproduced a known lost-update defect in the OpenTelemetry Demo's shopping-cart service. We tested the approach against conventional methods. On two held-out lost-update defects created after our method freeze, Muse-guided testing found the bug in a median of one execution per target. The strongest non-AI systematic method required 764 and 97 executions respectively, under the same evaluation budgets. The result is deliberately nuanced: Muse was significantly more efficient in executions, but conventional systematic testing was faster and cheaper overall because each execution was very inexpensive. Across 200 negative-control trials involving five methods, there were no reported false positives. Mylonite demonstrates a research-backed direction for AI-guided software testing: use model reasoning to select promising experiments, while deterministic tools retain control over execution, verification, and reproduction. Built during the hackathon, with open-source code, reproducibility instructions, recorded results, and explicit limitations.