Jevengineeringgate — AI skill for Claude Code

jevengineeringgate — a calibrated decision layer and risk gate for AI coding agents

[](https://github.com/RavenRepo/jevengineeringgate/actions/workflows/ci.yml) [](LICENSE) [](https://nodejs.org) [](#measured-cost)

**An LLM creates the work. Something else should decide what happens next.**

Every AI agent has the same leak: a frontier model sits in a loop answering yes-or-no, picking the next worker, and scoring relevance. Those calls never needed generation. This repository moves them to [Jev](https://typesafe.ai), the System One decision model from TypeSafe AI — and then **enforces the result deterministically**, as host hooks rather than as a rule an agent has to remember.

LLM generates  →  Jev decides  →  code executes  →  humans handle uncertainty

• ~292 input tokens, ~0.98s, ~$0.0000123 per decision (measured, not quoted)
• Thresholds are fitted, not guessed — a sweep over 26 labeled cases
• 0 wrong, 0 unsafe, 25/26 primary-label hits at the current operating point
• 48 offline tests, no API key required to run them

---

Why a prose rule is not a gate

Most agent setups put the guardrail in an instruction file:

*"Before any risky, destructive, or irreversible action, run the risk gate."*

That cannot work. It asks the agent to **already know the action is risky** — which is the exact judgment the gate exists to make. The gate only fires when it was not needed.

A [`PreToolUse` hook](docs/jev.md) does not need to be remembered. It runs on every tool call, before execution, and returns `deny`, `ask`, or nothing.

Install

git clone https://github.com/RavenRepo/jevengineeringgate.git
cd jevengineeringgate
npm