jev-use
A Claude Code, Codex and pi plugin that routes agent steps needing no text output to TypeSafe's Jev judgment model, reporting roughly 230-millisecond decisions at about $0.02 per thousand judgments with typed escalation back to the LLM.
jev-use hands tasks that need no written output to Jev, a judgment model, so the coding agent's own LLM only handles steps that require writing text or code, while fast yes-or-no or multiple-choice decisions run through Jev instead. Recorded demos include a directions task where Jev makes ten click decisions while the LLM writes the two prose steps, gating dangerous shell commands with a typed reason in about 230 milliseconds and zero LLM tokens, and a context-compaction pass that judges 200 messages in seven calls.
It installs with one command that wires itself into whichever of Claude Code, Codex or pi it finds, and can run keyless in a mock mode for testing. Configuration accepts a TypeSafe, OpenRouter or Vercel AI Gateway API key. Internally, a protocol module defines the check, pick and rate question types and escalation reasons, a dispatch module decides what never needs to reach Jev at all, and a redaction module strips credentials from a gated action before it is sent.
It is written in JavaScript and installable as an npm library with zero runtime dependencies on the judgment path.