[benatespina]

Summary

• GitHub evicts a cache nobody has read for 7 days. A pull request can read only the caches saved on main, and a push to main refreshes only the incremental Stryker files of the packages it changed. So after a quiet week, a pull request ran every mutant of the other packages cold. On chore(deps): update non-major dependencies #202, Mutation (core) found no stryker-core- cache (last written 2026-10-10), ran all 6,333 mutants and hit the timeout.
• New workflow stryker-cache.yml: once a day, and on demand, each package restores its stryker-<package>- cache from main. A restore counts as a read, so the cache is not evicted. Each job takes about 3 seconds and needs no checkout.

Why this mechanism

• Restore only, not a scheduled mutation run. Warm caches are already cheap to use. The content stays current because the mutation job rewrites a package's file whenever that package changes on main. Running Stryker on a schedule would cost minutes per package every day just to keep a file alive.
• Daily. A weekly run would read the cache on the same day it is due to be evicted, and GitHub can delay or skip scheduled runs. A daily run tolerates several missed days for about 7 × 3 s of runner time.
• Fails on a miss (fail-on-cache-miss). Reading cannot create a cache that main lacks, so a red run is the signal. The next mutation run of that package on main writes the cache.
• The 10 GB limit is not a concern. Each incremental file is 20 to 165 KB. Older per-commit keys are never matched by the prefix restore, so they age out after 7 days. The repository uses about 2.2 GB, almost all of it pnpm caches. Keeping these caches read also protects them from LRU eviction under the limit.

Validation

workflow_dispatch only works once the file is on main, and pull requests do not trigger a schedule-only workflow. To test it, I pushed the same file with an extra push trigger to a throwaway branch, which can read main's caches, and deleted that branch afterwards:

• Read (cli) restored stryker-cli-a52ee3e… from refs/heads/main. That cache's last_accessed_at moved from 10:44:17Z to 10:50:39Z, which confirms that a restore with no checkout resets the eviction clock.
• The other six failed with "Failed to restore cache entry": today main has no Stryker cache for them, because they were evicted.

After merge

Until main has a cache for every package, the daily run fails for the missing ones. #208 changes ci.yml, so its merge runs mutation for every package on main and writes all seven caches. After that, run this workflow once with workflow_dispatch to confirm that every job is green.

pnpm check exits 0. No changeset: no package changes.