Claude Code Fixes Blocking Hooks That Failed Open

Claude Code hooks could exit code 2, report a block, and still let the tool run anyway if the hook's JSON failed schema validation, per the changelog.

That's the guardrail working backwards. You write a hook to stop something. The hook does its job and reports failure. The tool runs anyway, because a formatting problem in its output never reached you.

Version 2.1.214 fixes it. The same release also fixes a separate bug from v2.1.212, where continue:false halts got dropped if a tool failed or completed mid-stream. Two different mechanisms for stopping execution, both quietly broken before this fix.

If you've built hooks as a safety layer, a lint check, a policy gate, a cost limiter, test them against a deliberately malformed payload. Clean JSON alone won't tell you whether the block actually holds. A hook you've never watched fail is a hook you don't actually know works.

A related item in the same release notes covers a separate permission-bypass sweep, including a directory wildcard rule that auto-approved writes across the whole tree. Worth checking whether any of your own hook or permission configs assumed enforcement that was, until this release, optional.