Repository navigation
API cache is not write-atomic. #252
Description
The cache's store method is not atomic. It is protected by a mutex, but this mutex only prevents concurrent in-process writes. If multiple invocations of the gh CLI are run in parallel then two concurrent writes to the same cache file can occur, corrupting the cache.
Similarly if a process dies mid-write then a partially written file can end up in the cache.
A simple fix for this would be to make the cache write atomic by writing to a temporary file and then doing an atomic move to overwrite the cache entry.
Activity
hyrious commentedon Aug 7, 2026
I just hit this issue with
gh 2.97.0on macOS while running:The command failed consistently with:
With
GH_DEBUG=api, the failure occurred while reading the cachedPullRequest_fieldsGraphQL introspection response. The cached body was visibly corrupted and contained a duplicated/interleaved JSON prefix:The request completed in about 800 µs and returned an older
Dateheader, confirming that it was served from the local cache rather than the network.Clearing the cache fixed it immediately:
After that, the same
gh pr mergecommand completed normally. This looks like a concrete instance of the non-atomic cache write described here.fix(api): make HTTP response cache writes atomic