Community · E2 · artifact verified
Route coding-agent requests to the right model
A local proxy that sits between a coding agent (Claude Code, Codex) and the model providers: per request, Jev classifies what kind of task it is and how hard it looks, then silently routes to a suitable model from a named three-model collection you configure.
01 · Role in the system
What Jev does here
Every agent request passes through the proxy, which asks Jev two typed questions - task kind and difficulty - and uses the calibrated answers to pick from the model collection: cheap fast models for routine edits, stronger ones for hard reasoning, with the routing decision invisible to the agent. The point is measurement, not magic: the proxy records what routing does to cost, speed and output quality so the model collection can be tuned against real numbers. The router ships as an npm package with the classifier isolated in its own module.
02 · Control boundary
Where Jev sits
Jev as the task classifier inside a model-routing proxy: typed kind-and-difficulty questions per request, calibrated answers select from a configured three-model collection, and logged cost/speed/quality deltas tune the mapping.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Three-model collections only, by design; routing map is user-configured.
- 0 stars; the cost/quality measurements are the author's local runs.
- Classifier misroutes are silent to the agent - quality regressions need the logs to spot.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- kalowery ↗Community · github · public · checked 2026-10-09