1min.

1MIN · AI & AGENTS

Claude Code adds a human approval requirement for tool-check hooks

Claude Code 2.1.304 lets a tool-check hook mark an approval request with decider: "person". Only the permission dialog or a host callback can approve it; when neither can be asked, the request is refused with the hook’s explanation. The release also fixes plugins that could keep noninteractive runs alive by spawning or resuming agents after a stop or turn limit.

Anthropic · Oct 19, 2026

Why it matters

Agent extensions need explicit approval behavior and stop conditions when the same workflow runs unattended.

More context

This is a released CLI update. Extension authors can use the new field in a mod’s tool.check response and inspect e.isReadOnly on the event. Try a disposable workflow that requests approval with and without an available person; verify refusal and termination before relying on it for automation.

Keep in mind

These are documented release behaviors, not an independent security evaluation. Human-only approval applies to requests marked by the hook; it does not automatically review every tool call.

Original read

More engineering ideas, one minute at a time.

Explore more stories