How is this different from classifying every message?

Only first tasks and /route new reclassify while continuations keep the model, with exact in-memory caching, no fuzzy matching, and exact provider/model IDs inside the Pi available and --models range.

What exactly leaves my machine when Jev is on?

Task text, optional briefing, image-presence flag, tool needs, floor, plus new-task and failure flags go to pinned jev-1.13.0; history body, tool output, system prompt, image content, catalog, and prices never do.

Will a cheaper model actually save money on long context?

Not necessarily. The router estimates price with the recent cache-read ratio, treats unknown prices as unknown rather than free, and checks context and output space before switching.

What if routing fails mid-session?

Auto mode with no qualified model does not submit input, shadow only suggests, the breaker persists in health.json with recovery leases, and Pi auto-retry stays independent with no automatic replay.

How do I install and initialize it?

Install into Pi with pi install git:github.com/win4r/pi-jev-router, restart any running Pi session, then run /route init, edit the generated model config, and finish with /route reload and /route doctor. It needs Node.js 22.19+ and Pi 0.84.2 to 0.86.x.

What are the model tiers and how is the config found?

Tiers are quick, general and frontier, each entry a full provider/model ID with a tools boolean. Config resolution is an explicit --jev-config path, else a trusted project .pi/jev-router.json, else $PI_CODING_AGENT_DIR/jev-router.json. The file is replaced whole, not merged, and unknown fields or duplicate IDs error.

What did the first synthetic evaluation show?

Over 16 synthetic cases it made 13 real Jev requests and 3 local-rule cases, and 15 of 16 matched the preset raw tier or rule label, using 9,282 input and 1,100 output tokens. The author is explicit that this is not downstream programming-task success.

CommunityJev Agents & MCP

Pi Jev Router — win4r

Pi Coding Agent extension that routes only at task boundaries, using Jev judgments plus declared capabilities, availability, price estimates, and failure state.

See highlights

Open github.com

Stars: 11
Language: TypeScript
License: MIT
Published by: Community
Author: win4r
Last verified: Oct 2026

win4r/pi-jev-router

User repository on GitHub

View on GitHub

Task-boundary model routing for Pi Coding Agent, powered by TypeSafe Jev. Conservative policies, exact caching, and observable failover.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is Pi Jev Router — philippdubach.

Tagscommunity jevbest catalog-import

Works forJev for Classification Jev for Routing Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

First tasks and /route new reclassify; continuations keep the model so follow-ups are not downgraded as new tasks.
Briefings stay in-process via /route brief up to 6000 bytes; history text, tool output, and image content are never sent.
Unknown, low-confidence, or thin-context judgments keep the qualified current model instead of forcing a switch.
Exact in-memory cache, anti-flap limits, persisted breaker with recovery leases, plus preview, explain, doctor, and stats.

Watch out

Confidence reflects distribution concentration rather than correctness, costs and cache ratios are estimates, and Windows plus long-term project traffic are untested. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

duet-agent

Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.