Pi Model Router

A Pi extension that uses Jev to assess a prompt and choose a configured coding model. Routing is off by default.

With your consent, Jev judges the task category, reasoning demand, dependency scope, context-integration demand, and probability of missing critical evidence. The router combines demand scores with your policy and picks the cheapest eligible configured model that meets its required capability score; unclear tasks or insufficient evidence leave the current model unchanged.

Status example: Router: research R0.36 S0.94 I0.61 M36% W0.29 — R reasoning, S dependency scope, I context integration (each 0–2); M probability of missing critical evidence; W weighted, normalized demand (0–1).

Set up

1. Clone this repository. From its root, install dependencies and run the following setup commands:

   npm install
   

2. Add the local package to Pi (use the repository's absolute path):

   pi install /absolute/path/to/prompt-router
   

   For a one-time trial instead, run pi --extension /absolute/path/to/prompt-router/src/index.ts from any directory.
3. Set up Jev credentials: sign in to OpenRouter in Pi with /login openrouter, or set OPENROUTER_API_KEY or TYPESAFE_API_KEY. The router prefers Pi's OpenRouter credentials, then the OpenRouter environment variable, then TypeSafe.
4. Create your routing config (or invoke /skill:model-router-config with exact provider/model IDs and thinking levels to research and prepare one):

   mkdir -p ~/.pi/agent
   cp examples/router.config.json ~/.pi/agent/model-router.json
   

   Edit the options to match models available in your Pi (pi --list-models). Set their provider, model ID, thinking level, cost, and capability scores. The example's scores, costs, and policy are illustrative, not calibrated. Validate a complete config offline with npm run config:validate -- /absolute/path/to/model-router.json; it checks structure and one explicit generalist, not the truth of sourced benchmark or cost claims. You can instead use .pi/model-router.json in a trusted project; see configuration for precedence and format. Make sure Pi can access the coding models you configure.

The package includes the manual-only /skill:model-router-config skill (excluded from Pi's automatic skill prompt). Invoke it explicitly with the candidate models and thinking levels; it researches score/cost evidence, proposes categories, asks which exact option is the generalist if you have not chosen one, seeks approval before replacing a live config, and runs the validator. Run /reload after installation or skill edits to discover it.

Commands in Pi

/model-router-assess off and /model-router-route off also cancel permission. One-shot commands replace auto mode; a manual model or thinking-level change stops auto mode. To resume, run /model-router-auto on and confirm again. A one-shot route changes the active Pi model until you change it again.

Privacy: With consent, Jev receives selected unredacted user/assistant text and the current prompt via OpenRouter or TypeSafe. Auto mode covers future prompts without asking again. Do not use it for sensitive conversations. Details and limits: Pi assessment and routing.

Releases

Pull requests run typecheck, tests, and offline evaluations. To publish, bump the package version, push a matching v<version> tag, and publish a GitHub release for that tag; the release workflow re-runs checks and publishes to npm with provenance. Configure npm trusted publishing for renjithwarrier94/pi-model-router and the Publish to npm workflow before the first release. Publishing directly from pull requests is intentionally disabled.

License

MIT.