When /plugin marketplace add <owner>/<repo> fails during the initial clone (transient network error, DNS blip, interrupted download, partial tar extraction), Claude Code leaves a broken clone cached under ~/.claude/plugins/marketplaces/<owner>-<repo>/. Every subsequent /plugin marketplace add call short-circuits on the cache entry and fails with the same error, even after the upstream repo is fixed and pushed. The only recovery is a manual rm -rf of the cache directory — there is no user-visible indicator that the cache is the source of the failure. 2. Clean up cache on failure — if the initial clone / extraction fails, delete the partial cache entry before returning the error so the next retry starts from a clean slate.

TL;DR

Implement a cache invalidation mechanism to handle failed plugin marketplace additions, either by re-fetching on failed parse or cleaning up the cache on failure.

Guidance

• Consider implementing option 1: Re-fetch on failed parse, where if the cached marketplace manifest fails to parse on a subsequent add call, the cache is invalidated and re-cloned from the remote.
• Implement option 2: Clean up cache on failure, where if the initial clone/extraction fails, delete the partial cache entry before returning the error, allowing the next retry to start from a clean slate.
• To verify the fix, test the /plugin marketplace add command with a flaky network connection, then retry after fixing the network issue and observe if the command succeeds without manual cache removal.
• Document the manual recovery recipe in the README, as a temporary workaround until the fix is implemented.

Example

No code snippet is provided as the issue does not imply a specific code change, but rather a design decision on how to handle cache invalidation.

Notes

The fix should be implemented in the Claude Code CLI to handle cache invalidation, and the chosen solution should be tested on both macOS and Linux environments to ensure compatibility.

Recommendation

Apply workaround: Document the manual recovery recipe in the README and implement option 2: Clean up cache on failure, as it is simpler and matches the user's mental model, until a more robust solution can be implemented.