Testing without a phone

@hellohelen-ai/goliath/testing exports fakeModel, an AI SDK language model that reads from a script.

import { createAgent } from "@hellohelen-ai/goliath";
import { fakeModel } from "@hellohelen-ai/goliath/testing";

const model = fakeModel([
  { json: { kind: "tool", tool: "listTasks", brief: "see what is open" } },
  { toolCall: { name: "listTasks", input: {} } },
  { json: { kind: "answer", brief: "reply" } },
  { text: "You have two tasks: buy milk and call mom." },
]);

const agent = createAgent({ model, tools: { listTasks } });
const result = await agent.run("what's on my list");

expect(result.handledBy).toBe("device");
expect(model.remaining()).toBe(0);

Three kinds of reply

Reply	Used for
{ json }	A conductor plan or a worker’s arguments
{ toolCall: { name, input } }	A worker step, when the provider returns a tool call
{ text }	The answer, or the scribe’s brief

Why running out is an error

The script is consumed in order and the model throws when it is empty. A passing test therefore proves the harness sent exactly the prompts you scripted, no more. Assert model.remaining() is zero at the end to prove it sent no fewer.

Asserting on prompts

model.calls holds every call the harness made, in order, with the full prompt. Prompt changes are tested as prompts in this project: if you change what the conductor reads, assert the new shape.

const conductorPrompt = JSON.stringify(model.calls[0].prompt);
expect(conductorPrompt).toContain("Use listTasks before completeTask.");