pi-jev-router-pro

Single-surface TypeSafe Jev model router for the Pi coding agent.

One virtual model — jev/auto by default — routes every request to a physical {model, thinkingLevel} pair chosen by a Jev typed judgment, an ordered candidate chain, and a fail-open eligibility filter (parking, quota, budget). It replaces the legacy pi-jev-model-router + pi-jev-model-router-profile-switcher duo with one extension.

jev/auto:high  →  zai-coding-cn/glm-5.3-flash:medium      (dispatch recorded per assistant message)

Why

• Judgment routing. Every user prompt is classified by TypeSafe Jev (task kind, complexity, capability deserved, deep-reasoning need) and dispatched to the matching tier: quick | standard | high | premium | xpremium. Kind specialists (kindModels) are tried first when the judged tier is at or above their minTier.
• Mid-turn transient fallback. On a retried request with a transient error (429 / 5xx / 529 / rate-limit / timeout / network), the router advances the chain past the failed target — mid-turn, mid-chain-of-thought — using pi's virtual-model retry hook. Chain exhausted → the router throws, and the run ends with the original error visible.
• Permanent failure parking. 401/403, model_not_found, and decommissioned/retired/ withdrawn models (identified by name in the error text) are parked for recovery.parkMs (default 1 h), persisted to disk, and skipped pre-flight on later turns.
• Profiles are first-class data. Reads your existing ~/.pi/agent/pi-jev-model-router-profiles.json unchanged; switching is in-memory plus one state file — no config rewrites, no reloads.
• Cache-aware moves. An avoidable tier switch inside the demand deadband (default 0.25) is suppressed so the provider's prompt cache stays warm.

Install

pi install npm:pi-jev-router-pro

Requires pi ≥ 1.4 (virtual models) and a TypeSafe API key in TYPESAFE_API_KEY (~/.pi/.env works). Development loop: pi --extension ./src/index.ts.

Migration from the legacy duo

1. Install this package. On first run the legacy pi-jev-model-router.json is offered as an implicit default profile (read-only) — run /jev profile save default-import to persist it.
2. Your profiles file keeps working untouched. The active profile is stored in ~/.pi/agent/pi-jev-router-pro-state.json; a legacy config that exactly matches a known profile is detected automatically.
3. Keep both installed while verifying. pi warns once per session about duplicate routing surfaces. While the legacy router is installed, the pro extension skips its jev_route tool registration (pi fails extension loads on tool-name conflicts) — the tool appears automatically once the legacy duo is removed. Remove the legacy duo (pi uninstall) only after the smoke checklist is green.
4. Legacy commands keep working: /jev-profile, /jev-profiles, /jev-profile-add, … are aliases of the /jev profile … surface.

Commands

Dormancy

Selecting a physical model manually parks the router (one warning, no routing, no entries); /jev auto or selecting jev/auto re-activates it. While dormant the extension is fully inert.

Consent modes

mode is preserved from the legacy router: auto switches silently; confirm prompts before switches (tier jumps and fallbacks) per confirm.tiers/timeout policy; notify never switches — it dispatches the current model and reports what it would have done. Env knobs JEV_ROUTER_MODE and JEV_ROUTER_OFF are honored.

Files

The legacy pi-jev-model-router.json is read-only to this extension.

Fault injection (offline fallback verification)

// ~/.pi/agent/pi-jev-router-pro.json
{
  "faultInjection": { "firstTransient": "429", "onRequest": 1, "mode": "throw" }
}

The first onRequest routed dispatches fail with the synthetic transient error, so pi's real retry path runs and the fallback becomes visible against your real second candidate. "mode": "advance" skips the candidate without a real failure (purely cosmetic simulation). The fallback decision entry is labeled fault injection. Remove the block when done.

Config compatibility

Config shapes (routes, kindModels, tiers quick|standard|high|premium|xpremium, RouteTarget {provider, model, thinkingLevel?, minTier?, priority?, minQuota?}, free, budget, cache, confirm, recovery) are compatible with the legacy router's schema. With useDefaultModels: false, built-ins are gone entirely. New knobs are optional and default to legacy-equivalent behavior: budget soft/hard 0.7/0.9, cache deadband 0.25, free policy prefer, recovery parkMs 3 600 000.

Development

pnpm install
pnpm typecheck   # tsc --noEmit
pnpm test        # vitest — offline decision-table tests, no network

License

MIT — see LICENSE.