Jev Claude Router

A Claude Code plugin that picks the model and effort for every turn with Jev, TypeSafe's decision model, and weighs each switch against what it costs. It is an extended fork of satviksinha/jev-model-router with cost-aware routing, compaction by Jev, session persistence and other additions.

At turn start, Jev answers two questions in one request: which tier should handle the prompt and how hard that model should think. The plugin checks the answer against the conversation's window, confidence, price and a ceiling, then rewrites each model request in the turn to the result, showing a status line with the tier, effort, Jev confidence and latency.

Because Claude's prompt cache is per model: moving a long conversation to another model rewrites the whole context into that model's cache, which at a few hundred thousand tokens costs dollars. A router that follows every pick can cost more than it saves, so this one only switches when the switch is worth it — and says so when it is not.

haiku for trivial work like a lookup, a rename or a yes or no; sonnet for straightforward and minor tasks with no real decision to make; opus for plain implementation carrying some complexity; and fable for planning, brainstorming, architecture and systematic debugging. What Jev is told each tier is for lives in TIER_CRITERIA in hooks/policy.ts.

The README lists it: each prompt you type, up to its first 12,000 characters, to pick the tier and effort; a subagent's task as the model wrote it when it is spawned; a finished task's one-line summary in certain configurations, never the task's result; and the conversation at a compaction. Everything goes to the provider you configured — TypeSafe or the Vercel AI Gateway — and nowhere else.

In Claude Code's plugin store, the router keeps each session's routing history, including the first 400 characters of each prompt, for the last 20 sessions. A prompt is also cut at the first thing shaped like a task notification before being graded or kept, wherever that appears.

Compaction sends the conversation to Jev, building on tamaratran/fast-jev-compaction; the README documents it as a feature this fork adds, and running /jev compact off stops it. The quick-start section was truncated in the digest, so the exact setup flow is in the repository's README.