jev-router (drmaas)

A way to route coding-agent behavior through Jev, summed up by the README as Jev decides, LLM generates. It ships an OpenCode v2 plugin, Cursor hooks and a skill that teaches an agent when to reach for Jev instead of the LLM.

Turn steering through session.hook('context'), tool-call steering with redirects via tool.hook('execute.before'), clarify-versus-act and permission gating in permission.hook('evaluate'), a post-failure advisor, subagent routing, code review with reviewTrigger stop, edits or both, and on-demand decisions through a jev_ask tool.

A beforeShellExecution policy gate and a stop verifier, written in stdlib-only Python, per the README; the skill file skills/jev-router/SKILL.md teaches an agent when to use Jev versus the LLM.

Jev answers typed questions in milliseconds at $0.042 per million input tokens with output free, and the plugin aims to spend fractions of a cent per turn to avoid mistakes that cost full LLM turns.

From TypeSafe cookbooks on a 488-request, 182-skill roster: wrong skill loads drop from 16.8% to 7.3% and needless loads from 9.8% to 4.0% with the skill_suggestion pattern, a two-call rank and rerank. The README notes its own roster is smaller.

The README names skill_suggestion, confidence-routing and intent-routing, and points at docs.typesafe.ai/llms.txt as the upstream API source of truth.

The captured excerpt cuts off mid-link in the reference-results section, so install and configuration steps, key setup and the full plugin options are not visible; the repo shows 0 stars and no license.