Jev Model Routing

A host-agnostic routing layer that sits between a coding harness and its models, choosing capacity per round from trajectory evidence like tool calls, failures and context pressure, running as a user-scoped daemon and local OpenAI-compatible proxy.

An OpenCode plugin that keeps the session on a cheap sticky parent model so the cache stays warm, and asks Jev whether a turn is hard before escalating through jev_escalate to temporary stronger subagents.

A Claude Code plugin that picks the model tier and thinking effort for every turn with Jev and weighs each switch against what it costs, since moving a long conversation to another model rewrites its prompt cache and can cost more than it saves.

An interactive experiment for model routing: build a pool of OpenRouter models, hand Jev a task, and watch it pick a model with a full probability distribution — then compare that pick with the answer you judge best and export the record as JSON.

Experimental deterministic model routing for coding agents with Jev in the loop: assign capabilities to models and copy one workflow prompt into Codex Desktop. The author's own Pokédex pilots did not show a cost win and he recommends against adopting it.

Pi extension that has Jev score each user request on a continuous 0-2 scale and cuts that range into five bands choosing one of three model gears (light, standard, heavy) plus a thinking level. A plain "continue" can inherit the current gear without a new Jev call.