Testing

Copy Markdown View Source

Exercise an agent module end-to-end without a phone, against Guava's v1/test-agent endpoint. Two entry points on Guava.Testing.

Scripted: session/3

Runs the agent and hands your function a Guava.Testing.Session, stopping it when the function returns.

Guava.Testing.session(MyAgent, fn session ->
  alias Guava.Testing.Session

  Session.say(session, "Hi, what are your hours?")
  Session.wait_for_turn(session)
  IO.puts(Session.get_transcript(session))

  Session.evaluate(session,
    ["The agent stated its business hours."],   # pass criteria
    ["The agent transferred the call."]         # fail criteria
  )
end)

Guava.Testing.Session functions: say/2, wait_for_turn/1, recv/2, get_transcript/1, evaluate/3, stop/1. Seed variables with Guava.Testing.session(MyAgent, fun, variables: %{"tier" => "gold"}).

Automated: roleplay/3

An LLM plays the caller against your agent. Returns the (still-running) session so you can evaluate then stop it.

session = Guava.Testing.roleplay(MyAgent, "You are a frustrated customer trying to cancel.")
Guava.Testing.Session.evaluate(session, ["The agent attempted to retain the customer."], [])
Guava.Testing.Session.stop(session)

In ExUnit

test "answers an hours question" do
  Guava.Testing.session(MyAgent, fn session ->
    Guava.Testing.Session.say(session, "What time do you open?")
    Guava.Testing.Session.wait_for_turn(session)
    Guava.Testing.Session.evaluate(session, ["The agent gave opening hours."], [])
  end)
end

These hit the live API, so gate them behind a tag and provide GUAVA_API_KEY. This repo excludes @tag :live by default; run with mix test --include live.

Unit-testing your callbacks without the network

Callbacks are plain functions on your module, so you can call them directly and assert on what they do — no phone, no LLM, no live session. Guava.Testing.MockCall gives you a Guava.Call handle that records the commands a handler emits and serves field/variable reads from an in-memory store.

test "assigns the intake task on call start" do
  mock = Guava.Testing.MockCall.new()

  {:noreply, _state} = MyAgent.handle_call_started(mock.call, MyAgent.initial_state())

  assert [%Guava.Commands.SetTask{task_id: "intake"}] =
           Guava.Testing.MockCall.commands(mock)
end

Pre-seed fields and variables a handler reads back, then assert on its result:

test "rejects an invalid email" do
  mock = Guava.Testing.MockCall.new(fields: %{"email" => "nope"})

  assert {:reply, {:error, _reason}, _state} =
           MyAgent.handle_validate("email", mock.call, "nope", MyAgent.initial_state())
end

MockCall functions: new/1 (:id, :call_info, :fields, :variables), set_field/3, put_variable/3, commands/1 (structs), command_maps/1 (wire maps), and clear/1 (drop recorded commands between simulated turns). Pass a %Guava.CallInfo.WebRTC{} as :call_info to exercise WebRTC-only guards.

A mock exercises a handler in isolation; it does not run the call runtime, so runtime orchestration (e.g. validators firing automatically when a task completes) belongs in a live session/3 test. For a full call loop with recorded commands, drive the internal Guava.Call.Runtime with an injected :emit function (see the SDK's own test/guava/agent_test.exs).

Next: Deployment.