pi-jev-router: a minimal Pareto-optimal OpenRouter model router for pi, based on Jev

Jev-classified model routing for pi.

Model rules

The router ranks the whole OpenRouter catalog for each task. It filters the catalog to models that fit the task, computes a Pareto frontier over quality, cost and latency, then picks the knee point. The knee is the frontier member farthest from the chord that joins the cheapest and dearest models. It needs no weights, so the pick follows the catalog and the recorded evidence on every task. A frontier too small or too flat for a knee falls back to a weighted value function.

Quality blends recorded runs from eval/results with the Artificial Analysis index. Writing quality uses the EQ-Bench Creative Writing v3 Elo in src/writing-prior.ts instead, because a general intelligence index says nothing about prose. Cost is a per-task estimate from catalogue prices, replaced by recorded mean cost as runs accumulate. Latency uses recorded means when they exist.

There is one selection policy. src/selector.ts holds a small fallback table for the case where the catalog cannot be fetched and no frontier exists.

Writing tasks get STE and Humanizer rules. Code tasks run unit tests. Planning tasks get a structure check.

Commands

/router           show status
/router shadow    recommend only (default)
/router auto      switch models per task
/router off       stop routing
/router frontier  show the frontier for the last decision
/router pin <id>  force a model
/router pin off   release the pin
/router budget <usd>
/router test      classify a sample task and show the pick

Benchmarks

npm run bench              # fixed baseline vs frontier router
npm run bench -- --all     # adds the pinned role arm as a third control

Five tasks, three code and two non-code, each in an isolated workspace with an independent verifier. Last recorded run (eval/results/, 2026-09-20):

Arm	Pass	Total cost	Cost per success
fixed_frontier (Sonnet 5)	5/5	$0.5722	$0.1144
router_role (pinned routes)	5/5	$1.5732	$0.3146
router_frontier (knee)	5/5	$0.1993	$0.0399

The frontier arm sent all three code tasks to z-ai/glm-5.3-flash, planning to anthropic/claude-sonnet-5 and writing to openai/gpt-5.4-mini. Classification overhead is excluded and five tasks do not establish a general saving rate.

Setup

npm install

Set OPENROUTER_API_KEY, or reuse the key stored in ~/.pi/agent/auth.json. Copy config/models.openrouter.json into ~/.pi/agent/models.json.

Auto-load globally:

mkdir -p ~/.pi/agent/extensions
ln -sf ~/Documents/GitHub/pi-jev-router ~/.pi/agent/extensions/pi-jev-router

Then start pi normally (no -e required).

Tests

npm test

Worktree isolation

Workers dispatched via dispatch_task execute in isolated Git worktrees:

• Worker edits code on a separate task branch (task/<id>).
• If verifierCommand passes (exit 0), changes merge cleanly into the repository.
• If verification fails or aborts, the worktree is cleaned up without leaving dirty changes.

Files

• src/classifier.ts — Jev call. One request, five questions.
• src/catalog.ts — fetch, cache and normalise the OpenRouter catalog.
• src/evidence.ts — per-model, per-work-kind statistics from eval/results.
• src/frontier.ts — Pareto dominance, the knee point and the weighted fallback.
• src/selector.ts — scores models and selects one. Pure functions. No model calls.
• src/board.ts — SQLite task state.
• src/dispatch.ts — worker spawn and handshake.
• extensions/router.ts — pi hooks and commands.
• eval/ — benchmark tasks, verifiers, runner.