FakeLocalAiHost class

A scriptable LocalAiHost that records what reached it and lets a test push events back the way a real platform would.

Published rather than hidden in this package's own tests: an app built on flutter_local_ai cannot otherwise exercise its AI paths in a unit test, because there is no OS model behind flutter test.

final host = FakeLocalAiHost()..response = 'hello';
debugLocalAiHost = host;
addTearDown(() async {
  debugLocalAiHost = null;
  await host.dispose();
});

expect((await FlutterLocalAi().generateText(prompt: 'hi')).text, 'hello');
expect(host.sessions.single.transcript.toString(), 'hi');

Streaming needs the test to play the platform's part, because nothing generates tokens on its own:

final chunks = <String>[];
final done = session.getResponseAsync().listen(chunks.add).asFuture<void>();
await pumpEventQueue();
host.emitToken(1, 'partial');
host.emitDone(1);
await done;

So does a tool call: the fake plays the model asking for one, through the same registry the native host dispatches with, so a tool loop can be tested without a device.

final result = await host.invokeTool(1, 'weather', {'city': 'Rome'});
expect(result, '{"temp":21}');
expect(host.toolCalls.single.toolName, 'weather');