Testing

@vrealtime/testing provides fakeModel(), a RealtimeModel that replays scripted events on a virtual clock and records every command it receives. Tests run fast, offline and deterministically.

import { defineAgent } from '@vrealtime/core';
import { serve } from '@vrealtime/server/node';
import { connect } from '@vrealtime/client';
import { fakeModel } from '@vrealtime/testing';

const pcm = { encoding: 'pcm16', sampleRate: 24000, channels: 1 };

// A model that answers every request by calling set_color 10 ms (virtual time) later.
const model = fakeModel({
  onCommand(command, session) {
    if (command.type !== 'respond') return;
    session.schedule({ type: 'response.start', responseId: 'r1' });
    session.schedule({ type: 'tool.call', responseId: 'r1', callId: 'c1', name: 'set_color', arguments: { color: 'teal' } }, 10);
    session.schedule({ type: 'response.end', responseId: 'r1', status: 'done' }, 11);
  },
});

const server = await serve(defineAgent({ model, instructions: 'test', clientTools: 'accept' }), { port: 0, clock: model.clock, onListening: () => {} });
let color;
const session = await connect(server.url.replace('http', 'ws'), {
  tools: { set_color: { parameters: { type: 'object' }, run: ({ color: c }) => (color = c) } },
});

session.sendText('make it teal');
// ...wait until the model has received the request, then move virtual time forward:
model.clock.advance(20);
// ...wait until color === 'teal'

What the fake model gives you

Member	Meaning
fakeModel({ capabilities, script, onCommand, clock, connectError })	Override capabilities (for example no argument streaming), replay a fixed script, react to commands, share a clock, or fail to connect
model.clock	A VirtualClock: nothing scheduled runs until advance(ms)
model.session	The most recent session; model.sessions lists all of them
session.schedule(event, delayMs) / session.emit(event)	Emit a model event later, or now
session.commandsOf(type)	What the runtime sent the model: respond, cancel, truncate, submitToolResult…

Pass the same clock to the server (serve(agent, { clock: model.clock })) so heartbeats and resume windows run on virtual time too.

Test interruption

The most valuable test of a voice app: interrupt while a tool call is still on its way, and check that it never takes effect.

session.sendText('make it blue');            // the model calls set_color 500 ms later
// ...wait until session.state === 'speaking'
session.interrupt();
model.clock.advance(1000);                   // the late call is due now
// color is unchanged, and no tool.result reports success for it