Re-running hook-approved calls through the permission classifier sits behind a flag whose fallback is off.
When a hook approves a tool call that would run in auto mode, Claude Code can now stop treating that as final and run the call through the permission classifier anyway, with a note that the hook vouched for it. The branch sits behind a flag whose built-in fallback is off, so without a remote value hook approvals still bypass the classifier as before.
- Gate:
tengu_virtual_knuth, defaulting to off in the shipped code. - Logs
tengu_auto_mode_hook_allow_funneledwith tool name, whether it is MCP, and the hook's source bucketed to settings, plugin, skill or other. - Auto-mode decision events gained a field recording the hook vouch.
- Debug output says the hook approved the tool use but auto mode requires classifier adjudication.
tengu_auto_mode_hook_allow_funneled
Strings lifted out of the shipped bundle, so the claim above can be checked against them.
tengu_virtual_knuth Not enough to sayNothing here resolved what this flag was doing on this version, so nothing here should be read as on or off.
This account: no value returned · anonymous baseline: no value returned · compiled default in v2.1.246: off
Read once, for one account on one subscription tier, against v2.1.246. It isn't a statement about your account. What a flag value here can and cannot tell you
New in this build: tengu_virtual_knuth