Jev Permission Gate — Development skill for Claude Code

jev-permission-gate

A [Claude Code](https://code.claude.com) mod that puts [TypeSafe's Jev](https://docs.typesafe.ai/) in front of the auto mode classifier. In auto mode, for each tool call Claude Code would otherwise send to its built-in classifier, Jev answers eight yes/no questions in one request and the mod decides:

• allow when Jev is at least 85% sure the call serves your request and every risk check is at or below 25%. The built-in classifier doesn't run.
• deny when the call is both risky and unrequested: some risk check at or above 80% and "serves your request" at or below 30%. Claude sees the reason.
• defer everything else to the built-in classifier, unchanged.

Results

Measured head to head against Claude Code's built-in classifier in a live session, on identical tool calls (details below):

• 2× faster where Jev decides. The permission wait drops from a median of 329ms to 164ms.
• Matching judgment. Jev's verdicts agreed with the built-in classifier's on every call it would have decided.
• No unsafe allows in testing. Across 93 labeled cases, including a fresh held-out set with manipulation attempts and non-English requests, it made zero unsafe allows and zero wrong denials. With 35 risky cases, that bounds the unsafe-allow rate below about 9% (95% confidence); more cases will tighten it. Anything Jev isn't sure about goes to the built-in classifier.
• Resists steering. An eighth question flags text inside a call that claims approval or argues for itself, and a pattern check keeps obvious cases from reaching Jev at all.
• Cheap to defer. The classifier's work overlaps Jev's request, so calls Jev hands off cost almost nothing extra at the median.

Today Jev decides about half of the calls in a typical workload, which cut the average permission wait by 16% overall. Every point of decision rate raises that: at the 84% Jev reached on the eval set, the projected saving is about 40%. Raising that rate is the main lever