You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Nightly fuzz job discards its corpus and uploads artifacts from the wrong path #24
Two defects in .github/workflows/fuzz.yml make the nightly fuzz run close to worthless. Both are small and independent of everything else in #22.
The corpus is deleted every run
# Start fresh each run
rm -rf "corpus/${TARGET}" "artifacts/${TARGET}" "coverage/${TARGET}"
No seed corpus is committed either. So three hours on every available core, every night, re-deriving the same shallow coverage from zero. Coverage-guided fuzzing is only worth the machine time if the corpus compounds.
Fix: drop the rm -rf of corpus/, and persist the corpus across runs with actions/cache keyed on the target name. Optionally commit a small seed corpus.
Artifact uploads point at a path that does not exist
The job does cd fuzz (the crate is at the repo root) and writes to fuzz/coverage/... and fuzz/artifacts/.... Both upload steps read from dev/fuzz/coverage/... and dev/fuzz/artifacts/...:
With if-no-files-found: warn this fails silently. On a crash we would get a red job and lose the reproducing input — recoverable from the libfuzzer log output, but painfully.
Fix: correct both paths to fuzz/.... Consider if-no-files-found: error on the coverage upload so a future path regression is loud.
Acceptance
A second nightly run starts from the previous run's corpus.
Coverage and artifact uploads produce non-empty artifacts.
A deliberately introduced crash uploads a usable reproducer.
Activity