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
A thirteenth tool: claude-hook-bench runs every hook of a Claude Code setup with the exact JSON Claude Code sends for its event and reports what each one does: exit code, time, output, whether the output is valid for the event, and whether Claude Code would block, allow, ask, add context or stay silent; before running anything it finds the mistakes Claude Code swallows in silence (a misspelled event, a lowercase matcher, a missing script, a wrong-typed field). Like md-doctor, session replay, chat ferry and permission doctor it only runs on demand, so there is nothing to install or switch on: the setup page lists it with two buttons (list the hooks and check them, run every hook), and status says it runs on demand. The marketplace carries its hook-bench plugin. The end-to-end test gives it a config of its own with four hooks (a deny on rm -rf, an invalid systemMessage, a misspelled event name and a SessionStart hook that prints context): list --json flags the misspelling, run --json calls the deny a deny for rm -rf and silent for npm test, the invalid output a non-blocking error with the field named, and the context hook adds context, and no file under the config folder changes. With a Claude Code CLI at hand it also installs the plugin and runs its skill's command from the installed copy.
Always-on tokens: all eleven plugins together add about 180 by Claude Code's own estimate (hook bench's skill description adds 15), still under the 200 the end-to-end test allows, and the request Claude Code sends the model still mentions none of the user-only skills (now twelve).
Website and README: thirteen tools, eleven plugins, the new tool's card and rows, over 5,800 tests, and the preview images say thirteen.