@happyvertical/smrt-personas

Tenant-owned, context-scoped agent personas for s-m-r-t. A TenantAgent decides whether an agent may run and caps its capabilities; a persona decides how that agent behaves for a tenant and context—its instructions, tools, acting identity, principal, memory scope, and priority.

The package also provides a gated learning loop: feedback reinforces memory, reflection creates pending directive proposals, and a human-authorized approval service activates instruction changes.

Installation

pnpm add @happyvertical/smrt-personas

Add svelte when consuming the optional directive-review UI.

Quick start

import {
  AgentPersonaCollection,
  PersonaResolver,
} from '@happyvertical/smrt-personas';
import { withTenant } from '@happyvertical/smrt-tenancy';

const personas = await AgentPersonaCollection.create({ db: 'app.db' });

await withTenant({ tenantId: 'tenant-1' }, async () => {
  const editor = await personas.create({
    tenantId: 'tenant-1',
    agentClass: '@acme/agents:EditorAgent',
    name: 'newsroom-editor',
    contextType: 'site',
    contextId: 'site-42',
    instructions: 'Prefer local sources and concise summaries.',
    priority: 100,
  });
  editor.setAllowedTools(['facts.search', 'content.update']);
  await editor.save();
});

const resolved = await new PersonaResolver(personas).resolve(
  'tenant-1',
  '@acme/agents:EditorAgent',
  { contextType: 'site', contextId: 'site-42' },
);

Resolution prefers exact context over context type over tenant default, then higher priority. The nearest tenant level with an applicable persona shadows farther ancestors.

Layering model

manifest defaults
    → TenantAgent availability and tool ceiling
        → nearest applicable AgentPersona
            → resolved instructions, tools, identity, and memory scope

Hierarchy resolution intentionally reads ancestor-tenant personas. Run it from a reviewed system/admin path rather than inside a strict tenant interceptor that forbids cross-tenant reads.

Learning and directive approval

1. Feedback records a human or autonomous outcome with a correlation ID.
2. reinforceFromFeedback() updates persona-scoped learning memory.
3. ReflectionRunner turns evidence into a pending DirectiveProposal.
4. DirectiveApprovalService requires personas.activate-directive before an approved rewrite becomes a tenant/persona-scoped prompt override.

Autonomous reflection can propose but cannot activate instructions. Confidence reinforcement is automatic; instruction changes are permission-gated.

Durable agent instances

Personas can also identify independently scheduled instances of a multi-instance agent. schedulePersonaInstance(), buildPersonaInstanceAdmin(), and upgradeSingletonToDefaultPersona() preserve singleton identity while enabling per-persona schedules, memory, and administration.

Public entry points

| Entry point | Purpose | | --- | --- | | @happyvertical/smrt-personas | Models, resolver, learning loop, approval services | | @happyvertical/smrt-personas/svelte | Presentational DirectiveReviewQueue |

Related packages

• smrt-agents owns availability, scheduling, and agent lifecycle.
• smrt-prompts stores activated instruction overrides.
• smrt-users supplies permission resolution and run-as principals.
• smrt-messages supplies the fixed persona messaging tool boundary.

Development

pnpm --filter @happyvertical/smrt-personas test
pnpm --filter @happyvertical/smrt-personas typecheck
pnpm --filter @happyvertical/smrt-personas build

See AGENTS.md for resolver, memory, tenancy, and permission invariants.