The Jev-enator

A set of Claude Code hooks that use Jev to make cheap, calibrated judgment calls inside the agent loop, where a full LLM call would be too slow and too expensive. The README's six hooks cover blocking dangerous commands, noticing quiet failures and judging turn completion.

The danger gate and failure notice are enforcing by default. The completion check is log-only until you decide from your own data whether to enforce it, the scope check is log-only in all cases, the reads ranker runs in shadow, and the clear advisor only notes when a prompt starts a new task in a big context.

The README argues the surface pattern and the actual harm routinely disagree in both directions, and says every row of its comparison table is a fixture in tests/test_jev_gate.py. The captured excerpt begins that table but not its rows.

The project's badges state about 350 ms latency and roughly $0.00004 per check, which the README frames as cheap enough to run on every tool call and every turn. A chat model in the same permission hook would cost seconds and cents per call.

It records a verdict per turn on the Stop event and, by default, blocks nothing. The README says its accuracy is an open question because it depends on intent that is not visible in the transcript, which is where cheap classification is weakest, so it collects evidence first.

Yes — the hooks judge by calling Jev. The repository itself has no dependencies beyond Python 3.10+, and the badges show CI running the 86/86 fixtures. The README's WHY.md link is billed as the two-minute explanation of what breaks without the hooks.

Jake Reardon (jakenbear on GitHub) built it as a community project, MIT licensed. This entry is catalogued from the README captured on 2026-10-15; the excerpt is truncated, so installation details are not reproduced here.