standardharness

standardharness is a small, model-agnostic TypeScript boundary for durable, owner-scoped agent turns. It came out of familiar's Gemini memory companion, where conversation may be agentic but identity and consent cannot be.

It provides three deliberately narrow primitives:

• StandardHarness enforces the turn order: load bounded context, persist the accepted user message, call your agent, then persist only a non-empty reply;
• authorizeOwnerBoundTurn rejects the wrong channel or actor before context loading or persistence; and
• buildUntrustedContextPrefix bounds recent messages and labels them as data, never instructions.

It does not select a model, provide tools, store long-term memories, or make authorization decisions for you.

Install

npm install standardharness

Example

import {
  authorizeOwnerBoundTurn,
  buildUntrustedContextPrefix,
  StandardHarness,
} from "standardharness";

const scopeId = authorizeOwnerBoundTurn({
  ownerId: account.id,
  actorId: request.user.id,
  channel: request.channel,
});

const harness = new StandardHarness({
  store: {
    loadContext: (scope) => conversations.load(scope),
    appendMessage: (scope, message) => conversations.append(scope, message),
  },
  assistantModel: "your-model-name",
});

const reply = await harness.run({
  scopeId,
  kind: "text",
  userMessage: request.text,
  respond: async (context) =>
    model.generate({
      system: "Follow application policy.",
      prompt: buildUntrustedContextPrefix({
        messages: context.recentMessages,
      }) + `\n\n[CURRENT USER MESSAGE]\n${request.text}`,
    }),
});

The caller still owns tool definitions, model prompting, storage retention, and access policy. See familiar's evaluation page for the concrete Gemini + Durable Object use case and checked-in synthetic tests.