Changelog

Spec version 0.2.0 (draft)

Generated from the HyperStudy source. Machine-readable: agent-spec.json.

0.2.0 (2026-10-21)

Voice engines are shown what the participant sees. A [context] user item can now carry input_image content (a data: URL) after its text: each image on a screen is attached to that screen's item once per session, and the item's images list says what each one is (attached, attached earlier, could not be loaded). Questions attach their choice, item and dimension images the same way.
New context items: participants (the other participants by name), timer (a screen's time limit), experimenter_message, revealed (another participant's answer shown to this one), and paused now names who the study is waiting for in waitingFor. A screen item's text is the screen as read (rich-text tags removed) and is omitted when empty.
Conformance suite version 6 adds image_context: an engine must accept an image item and name the colour of a solid red image. Reports under earlier suite versions no longer pass preflight.

First published draft. It covers the model endpoint contract for text agents, the HyperStudy Realtime Profile for voice engines, the study tool catalogue and silent answering of on-screen questions, pausing turn detection while a participant cannot be heard, and the conformance suite at suite version 5. Engine audio times restart when turn detection is switched back on.