@blackbelt-technology/pi-system-one

Typed System-1 decisions (choice / score / noul) for pi extensions and the pi-dashboard server, through one user-controlled registry of backends:

• hosted TypeSafe Jev and any /v1/systemone endpoint (http),
• dashboard-managed local Von / Laya servers (managed, loopback only),
• a pi chat-model role such as @fast as a slow fallback (llm).

Node built-ins only. No runtime dependencies.

Use

import { predict } from "@blackbelt-technology/pi-system-one";

const r = await predict({
  consumer: { id: "my-ext:is_destructive", failurePolicy: "fail-open" },
  state: JSON.stringify({ command: "rm -rf build/" }),
  questions: { destructive: { type: "noul", instructions: "Does this command delete user data?" } },
});

if (!r.ok) {
  // r.reason: no-backend | capability | off-machine | timeout | error
  // apply your declared policy (r.policy)
} else if (r.mode === "enforce" && (r.answers.destructive as { noul: number }).noul >= (r.thresholds.destructive ?? 1)) {
  // act only on calibrated, enforced answers
}

The adapter never allows or blocks anything itself: it reports mode, thresholds and policy, and the consumer decides. Answers are advisory (shadow) until the user measures a backend for your consumer with the dashboard's Test and saves it as enforce.

Configuration

~/.pi/agent/system-one.json, edited from the dashboard settings section "Decision models (System 1)". Nothing leaves the machine unless allowOffMachine is true. Keys come from an env var (e.g. TYPESAFE_API_KEY) or ~/.pi/agent/system-one/auth.json (0600), never from the config file.

Full reference: docs/system-one.md.

License

MIT