pi-local-classifier

A Pi coding-agent extension that turns localhost typed-decision backends into native Pi classifiers. Run a small decision model beside Pi, point this extension at it, and every Pi extension can call it through ctx.modelRegistry.classify() — approvals, routing, ranking, and judging with no cloud round trip, no API key, and no per-call cost.

This package is backend-agnostic: it works with any server speaking the wire format below. It neither bundles nor requires any specific model.

Install

pi install npm:pi-local-classifier

Requires Pi 1.0 or later (uses the native classifier registry).

Configure

List backends as id=url pairs before starting Pi:

export LOCAL_CLASSIFIERS="decide=http://127.0.0.1:8133,extra=http://127.0.0.1:8000"

Each entry registers as local-classifier/<id>. Malformed entries are skipped with a warning; with nothing configured the provider stays unregistered and classifier consumers fall back gracefully. Check registration and backend health anytime with /classifier-status.

Use from an extension

const clf =
  ctx.modelRegistry.findOfType("classifier", "local-classifier", "decide") ??
  ctx.modelRegistry.getModelsOfType("classifier")[0];
if (!clf) return fallback();
const result = await ctx.modelRegistry.classify(
  clf,
  {
    state: { command },
    questions: {
      destructive: {
        type: "bool",
        instructions: "Is this operation hard to reverse?",
        criteria: { true: "data loss that is hard to recover", false: "easily reversible" },
      },
    },
  },
  { signal: ctx.signal },
);
if (result.stopReason !== "stop") return fallback();
const p = result.answers.destructive; // { type: "bool", probability: 0.97 }

Notes for consumer authors:

• Prefer a pinned model id when your logic depends on absolute thresholds — calibration does not transfer across backends. Relative uses (ranking, argmax routing) can take the first available classifier.
• classify never rejects: failures arrive as stopReason: "error" (or "aborted"). Check it before reading answers.
• Keep a non-classifier fallback (heuristics, recency, fail-open/closed per your policy) so the extension works with no backends configured.

Compatibility

Backends must implement SystemOne-style input/output: POST <base>/v1/systemone with { model, state, questions }, where questions are bool (sent as the wire type noul, with true/false criteria folded into the instructions), choice, or score; responses carry answers keyed by question id plus usage.input_tokens. Any server implementing that schema — open or proprietary, local or remote — works; the examples above use localhost because private, free, ~100ms judgments are the point.

Development

npm test   # node --test, no dependencies to install

Layout: extensions/local-classifier.ts is the whole extension; tests/local-classifier/ holds the node --test suite (registration, env parsing, question/answer mapping, per-model routing, abort behavior).

License

MIT