Question 1

Is the 90% skill-pick number an independent evaluation?

Accepted Answer

Not fully. It comes from 181 labelled turns, of which 127 are real sessions pulled from the author's own Claude Code history and the rest are synthetic. Against a keyword baseline the router picks the right skill 90.0% of the time versus 59.4%, and under-provisions the model on 3.3% of turns versus 19.9%. Useful signal, but the labels and the labelled turns share an author with the implementation.

Question 2

What happens if the Jev call is too slow?

Accepted Answer

There is a deadline, and exceeding it is not an error: the router answers from its offline regex heuristic instead, and route.source reports "cache", "shortcut" or "fallback" so you can tell a keyword answer from a Jev one. Bare acknowledgements and your command prefix are even cheaper, routing in 0ms with no call at all. Because the shipped deadline came from someone else's machine, calibrate before trusting it.

Question 3

Does this call the model for me as well?

Accepted Answer

No. The router is the only component that needs a TypeSafe key; executing the routed turn stays your harness's job on whichever provider it already uses. In the clone, src/catalog/provider.ts holds the provider config, and only npm run chat consults it. That file covers Anthropic natively and treats Groq, Together, OpenRouter, DeepSeek, Mistral, vLLM, LM Studio and Ollama as the same chat-completions shape.

Question 4

How do you keep the router from enabling a dangerous tool?

Accepted Answer

Thresholds live in your catalogue, not in the model's answer. Each tool declares a risk of read, write or execute, which sets the bar it must clear, and you can override that per tool. The reasoning is spelled out in the README: the message driving the decision is untrusted input, and Jev does not treat state as hostile by default, so the floor has to be enforced in code.

Question 5

How do I calibrate the router's deadline?

Accepted Answer

Run npm run calibrate from a clone; it times the round trip from where you are and writes HARNESS_ROUTER_DEADLINE_MS into .env, because the shipped 600 ms was fitted from one machine in Buenos Aires.

Question 6

Which routing decisions does jev-harness-router make in one call?

Accepted Answer

One Jev call answers the turn's tools through one Noul per tool plus a ranking Choice, its skill through a Choice gated by four request-shape Nouls, and a Score from which model tier and effort budget are derived in code.

Question 7

What risk thresholds gate a tool in jev-harness-router?

Accepted Answer

Risk sets the bar a tool must clear: read at 0.35, write at 0.6 and execute at 0.8, overridable per tool with threshold, so the safety floor lives in code rather than in the model's answer.

Jev Harness Router

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community