When auto mode blocks your agent, what should happen next?

Auto mode is the default in Claude Code now. A classifier checks every tool call, runs the safe ones and blocks the risky ones. Most days that is exactly what you want.

The part I keep coming back to is what happens after a block.

Anthropic's post says a block redirects Claude "to take a different approach". The message the model gets after a denial says it may try other tools that do the same job. So it goes looking. Each attempt is another turn and more tokens, and sometimes a path you like less than the command it started with.

Claude Code only stops to ask you after 3 blocks in a row, or 20 in a session. That ask is a prompt in a terminal you might not be sitting in front of. Anthropic also says auto mode will occasionally block benign actions. Those cost the most, because the agent works around something you would have said yes to in a second.

What we built in Pushary: on the first block, the agent waits. Your phone or Mac notch shows the blocked command. Yes lets Claude retry that exact call once. No keeps it blocked. Only a signed-in person can say yes, never an API key the agent can read. Everything auto mode allows still runs without a ping.

Limits, so nobody is surprised:

- It waits up to 10 minutes, then Claude gets the normal block.

- Codex is covered for shell commands through its auto-reviewer, not for file patches.

A question for anyone running auto mode all day: when it blocks something, do you want the agent to try another way, stop, or ask you? And how often is the block wrong?