Claude Code Jev — Development skill for Claude Code

Claude Code × Jev

Auto-mode classifies every action before Claude Code runs it: `allow`, `block`, or `ask`. That gate fires dozens of times an hour, in the hot path, ahead of every tool call. It is a System 1 job, a fast reflex, but it runs on a System 2 model, a frontier chat model built for reasoning.

This repository runs that gate on a System 1 model instead. Jev is a purpose-built classifier that returns a typed decision with calibrated confidence. It installs as a `PreToolUse` hook and reaches Jev through OpenRouter's typed decisions endpoint.

Anthropic's own auto-mode design already leans this way: the first stage emits a single-token decision and only calls reasoning for flagged actions. The point here is to make that first stage an actual reflex model, and keep the reasoning model for the flagged minority where it earns its cost.

Reflex on a reflex model

A System 2 model reasons; it is flexible and slow. A System 1 model recognizes; it is narrow and fast. Permission gating wants recognition. Below is what you get when a reflex task runs on a reflex model, measured over five passes of the 18-case fixture (90 live decisions, `typesafe/jev-1.13`).

For a ~540-token classification call:

	System 2 (chat model as classifier)	System 1 (Jev)	Win
Latency	~4 s (LLM-judge baseline)	264 ms	~93% lower
Cost per call	~$0.0022 (frontier, same context)	$0.0000227	~99% lower
Accuracy on the typed task	reference	80-81% match, 0 dangerous allowed	held

Read those honestly:

• Latency ~93% lower is against a published LLM-judge baseline, not a live measurement of Claude Code's own classifier. Against a bare single-token frontier call the gap is smaller, but you still skip frontier queueing and time-to-first-token.
• Cost ~99% lower (about 100×) is the durable win. A purpose-built classifier is not priced like a frontier model, however few tokens you make that model emit.
• **Accura