An agent that repairs OSS-Fuzz build breakages fixes 62 of 128 reproducible failures, and 29 patches reach maintainers
Synopsis
On the continuous fuzzing platform OSS-Fuzz, some projects stay stuck in build-failure states, so their fuzzers stop producing new targets. FuzzMedic uses collaborating agents to localize root causes across project repositories and OSS-Fuzz configuration, retrieve repair knowledge, record per-round progress and roll back unproductive edits; it repairs 62 of 128 reproducible failures (48.44%), against 27 (21.09%) for the strongest native-budget baseline, CompileAgent. In a PR experiment in oss-fuzz-gen run with Google, 29 patches were reviewed and merged by maintainers across 28 projects.
Interpretation
FuzzMedic turns repair of OSS-Fuzz build failures into an automated loop: starting from the build log, it decides which workspace should be edited, choosing between the project repository and the OSS-Fuzz build configuration, then generates and validates a patch. Such failures were previously handled by manual coordination between platform maintainers, who lack project context, and project maintainers, who often lack fuzzing and infrastructure expertise; generic build-repair tools target Maven dependencies, general C/C++ builds or repository-level compilation and do not cover container, instrumentation and fuzz-target validation constraints. On 128 cases reproduced from real OSS-Fuzz failures, FuzzMedic repairs 62 (48.44%), CompileAgent 27 (21.09%), SWE-Search 17 (13.28%) and Agentless 10 (7.81%); 31 cases are repaired only by FuzzMedic.
The system works round by round: it localizes a root-cause commit, draws on reviewed repair knowledge, compresses each round's build outcome into memory, and rolls the workspace back to an earlier usable version when an edit makes the state worse. Localization uses only the build log and failure timestamp, builds a candidate commit set from the first fatal diagnostic, and confirms a candidate by counterfactual build replay: the parent revision must build, and reapplying the commit must reproduce the same failure stage and diagnostic. The repair knowledge base is distilled from 188 historical failures across 110 projects into 43 error-pattern entries and 13 general guidelines, with all 113 evaluation projects and related incidents excluded; in ablations, removing rollback drops repairs from 62 to 42, removing localization to 46, removing repair knowledge to 52 and removing the memory component to 51.
Perspective
Evaluation is limited to 128 reproducible OSS-Fuzz build failures from 113 projects, of which compilation errors and dependency issues account for 112 cases; success means a patch passes OSS-Fuzz check_build in the reconstructed environment. The real-world deployment is a PR experiment in oss-fuzz-gen run with Google, where patches are human-verified before maintainer review, an experimental integration rather than an unattended production service.
The same 128 cases were used for parameter selection and evaluation, which may yield optimistic estimates; native budgets differ across methods in tokens, time and build attempts, and matched token/time caps leave build and interaction counts unmatched. Passing check_build establishes build restoration, not preservation of harness coverage or semantic equivalence; although the knowledge base excludes evaluation projects, public repairs already present in closed-model pretraining cannot be fully ruled out.
