Type what a partner said and the choices appear. Pick one, deliver it, and watch the memory count up — partner turn, your words, partner turn — until you reset it. Then show it a photo and the tiles become what you are looking at.
board is one of the runtime's six providers: tiles an app author wrote, with no model involved. Deterministic, which is why it can run in a page. Swap in Gemini, Ollama or the local model and the same engine proposes the words.
Those tiles are real output of the vision pipeline on those exact images, recorded with the runtime's evaluation records. The page replays them; it does not call a model from your browser.
Turn on “use my own symbols” and the artwork changes while the words do not. That is the boundary: you bring the symbols, we supply the conversation.
| On this page | In your code |
|---|---|
| Typing a partner turn | session.hear({text}) |
| Tapping a tile | session.select({choiceId}) |
| Deliver it | session.speak({draftId}) → your speech engine → confirmSpeech({speechId, outcome}) |
| Showing a photo | session.scan({images, mode}) |
| Choices, length, tiles | profile — or setProfile() |
| Reset the conversation | session.reset() |