Jev with Coding Agents

A coding agent fails in two different ways. It can write the wrong patch, and it can run a tool you did not want. Jev is built for the second kind only when you can list the allowed answers.

What the agent still does

The agent reads the repo, edits files, and writes commands. That output is open-ended text. Jev does not produce it. Official limitations still apply: no generation, weak arithmetic, dates as text. See Jev limitations.

What a gate can ask

A hook that runs before a tool call already knows the command, the path, and the latest user message. Those are state. The questions are closed:

• Is this command one the user asked for?
• Does it touch a protected path?
• Is the effect a publish, a force-push, or a recursive delete?

pi-jev-auto-mode implements that split for the Pi agent. Deterministic hard-denies never reach the model. The README says a missing engine, a timeout, or a malformed answer blocks. The measured latency it reports for a judged call is 193–642 ms on eleven ordinary commands. That number is the author's.

jevc starts from the other end. It scans CLAUDE.md, AGENTS.md, and similar files, and compiles rules it classifies as decidable into Jev questions plus a reducer in ordinary code. The README says the offline test corpus needs no API key. The scan is a heuristic. Read the rules it emits before you install the hook.

The use-case writeup is Coding-agent gates.

The official skill

typesafe-ai/skills is TypeSafe's agent skill for the hosted API. It teaches the agent how to call Jev. It does not make Jev the agent. The docs page is Agent skill.

A related cookbook, skill suggestion, uses Jev to pick at most one skill from a catalog. That is a Choice over names you already listed, which is the same shape as agent tool selection.

A pattern that survives the framework

1. Enumerate the legal outcomes in code.
2. Ask Jev only about those outcomes.
3. Put the allow line in code, including what happens when the call fails.
4. Keep a text model, or the coding agent, for anything that has to be written down.

Confidence-gated routing is the threshold half of step 3. Jev as a decision layer is the same split in a browser agent, where the legal outcomes are controls on the page.