A teardown of Claude Code's auto-mode safety classifier, looking at the undocumented ruleset that interprets user consent.

Author here, I ran a MITM dump on Claude Code sessions to see what actually gets sent over the wire and what makes the safety classifier tick.