0x7067/claude-jev

README preview

# claude-jev

Stop paying frontier-model prices for small judgments.

A coding agent makes dozens of quick calls every turn. What kind of prompt is this? Does it need tools? Does this edit break a rule in `CLAUDE.md`? Which turns still matter when the context fills up? This plugin sends those questions to [TypeSafe's Jev](https://docs.typesafe.ai/introduction). Jev is a System One model: it returns typed answers, not text, and it doesn't write code.

What you get:

- **Rules that hold.** Edits that break your instruction files get blocked with a file:line citation. Re-scoring the same 247 accepted edits at v0.23.0 blocks 23 of them (9.3%); the v0.21.0 run blocked 22 (8.9%), 17 of them under a single repo's own comment-ban rule.
- **Compaction in about a second instead of a minute or two.** Jev keeps the exact rows that matter instead of writing a summary. Planted user constraints survived 100% of the time.
- **Right-sized subagents.** Each spawn gets a model tier. A brief that changes files but leaves out paths, acceptance criteria, verification, or commit policy is sent back once.
- **Routing hints.** Each prompt gets a one-line hint such as "one search" or "focused edit, narrow verification."

Every hook fails open. An error, a missing key, or a timeout produces no output and never blocks a prompt. Compaction falls back to Claude Code's own summary.

## Install

### Claude Code (Python)

```bash

View full repository on GitHub →