Claude Code Stops Auto Mode From Blocking Itself

Claude Code 2.1.225 fixed a bug that made auto mode count its own safety-filter refusal toward the consecutive-block limit, per the changelog. For anyone running Claude Code unattended, that counting bug was the difference between a task that finished and one that quit after a handful of denials.

The model would deny its own permission check, then count that denial as a block. Enough of them and the run ended early, even though the underlying action stayed denied either way. The changelog doesn't say how many self-denials it took to trip the old limit.

The fix doesn't change what gets approved or blocked. Before, retrying a self-denial burned through the block budget until the run stopped short. Now the model just moves on.

If you've watched a long autonomous session stop cold after a few denials, this is a plausible explanation. Re-run the same workload on 2.1.225 or later before you touch your permission setup.