Where did we fail? -- Reproducing build failures in embedded open source software

Continuous integration (CI) builds in embedded systems frequently fail due to complex cross-compilation, board configurations, and toolchain constraints. Build logs are short-lived and difficult to reuse across heterogeneous runners and log formats, making debugging and failure analysis a persistent challenge. To address this, researchers developed PhantomRun—a unified abstraction layer and publicly reusable dataset that standardizes how CI build logs and metadata are retrieved, stored, and reproduced.

In an empirical evaluation with 4,628 failing CI runs from embedded open-source projects, PhantomRun reconstructed 91.8% of the original builds and preserved execution outcomes in 98% of cases. The reproduced builds closely matched their originals, differing only in timestamps or minor nondeterministic reordering. PhantomRun exposes all build artifacts in a uniform, machine-readable format, enabling large-scale historical CI reconstruction and longitudinal studies of failure patterns. This work was presented at EASE 2026 and is available on arXiv.