@geminixiang/pi-jev

Ask Jev, TypeSafe's calibrated decision model, typed boolean/choice/score questions from any Pi conversation. Jev never generates text; it returns probabilities against options you define, so a jev call is a cheap, calibrated alternative to judging something by eye.

Backed by @geminixiang/jev, a pi-ai-shaped SDK covering TypeSafe, OpenRouter, Vercel AI Gateway, and Cloudflare Workers AI as interchangeable backends.

Install

pi install npm:@geminixiang/pi-jev

Requires Pi 0.82.1 or newer (native provider registration and provider-level auth resolution). After installing, restart Pi or /reload, then configure a backend with Pi's interactive /login:

TypeSafe prompts for a plain API key, not OAuth. Pi owns credential persistence and /logout typesafe removes the saved key. The registration has no chat models: TypeSafe appears in /login, not /model; use the jev tool to call it. OpenRouter's browser login is also usable: it mints an OpenRouter API key billed from your credits. No Jev-specific copy of Pi's auth file is needed.

Alternatively, keep using Jev's existing environment variables:

The tool checks backends in this order: TypeSafe → OpenRouter → Vercel → Cloudflare. Within each backend, compatible Pi-resolved auth takes precedence; otherwise it falls back to that backend's Jev environment configuration. An earlier backend's environment key still wins over a later backend's Pi key. Pass provider in a call to restrict selection to one Jev backend ID from the table above.

Credentials are resolved from the executing tool's ctx.modelRegistry on every call, so Pi handles saved keys, configured key expressions, and runtime overrides. The active chat model does not select the Jev backend, and unrelated subscription credentials (such as Codex or Claude) are never reused. Cloudflare reuse is limited to Workers AI, not Cloudflare AI Gateway; the resolved Workers account ID accompanies its bearer key. Provider-specific environment values are not copied wholesale.

For safety, Pi credential reuse is limited to the standard service endpoints and a resolved API key with no custom headers. Redirected providers, custom auth headers, header-only auth, and incomplete Workers account configuration are skipped in favor of Jev's environment fallback. Chat base URLs are never passed to the decision API. Pi auth-resolution errors stop the call with a sanitized login hint; evaluation failures are not retried on another backend. Configure a Jev environment key if your Pi proxy configuration is incompatible. /logout does not disable environment keys. The account/key must also have access to Jev on the selected service; configuration alone does not verify permissions or billing.

Skill

Installing the package also installs the jev skill, which tells the agent when to reach for the tool (classification, detection, scoring, ranking, routing, extraction over known candidates, verification) and the rules that keep its use sound: ask narrow atomic questions, batch everything about one state into one call, rank with one score question per item rather than a choice over orderings, and read confidence before acting on an uncertain answer.

Tool

jev({ label, state, questions: { <id>: { type, instructions, criteria? } }, provider? })

A direct pass-through of the Jev decisions API: state is text or a JSON object/array, questions is any number of boolean / choice / score questions evaluated against that state in one request. See the skill for question shapes and an example, and @geminixiang/jev for the full API this wraps.

Development

npm test --workspace @geminixiang/pi-jev
npm run check --workspace @geminixiang/pi-jev

License

MIT