TensorUI is organized into three main layers:
hal/: platform-facing services such as screen, input, time, memory, strings, and fontsTensorUI/: framework widgets, containers, text input, overlays, and window managementexamples/: current integration surface used to exercise the framework in SDL
Frames are the basic units of rendering and interaction. A widget typically owns one or more frames and attaches render, touch, and lifecycle callbacks to them.
Pointer ownership is tracked through a shared InputSession. This lets gesture intent, click candidates, drag state, and active ownership move through a single framework abstraction instead of scattered global flags.
Framework focus and the active text-input target are owned by shared framework state. Editable controls such as TextField attach themselves to the current text session, and keyboard surfaces consume that session instead of relying on page-specific rebinding on the main path.
VStack and HStack provide the current container baseline. Safe-area helpers in hal/screen are used to adapt content to round and rectangular displays without tuning every screen independently.
The window manager owns a stack of root surfaces. App roots and system overlays are separate surface roles:
- app surfaces represent top-level application roots
- system surfaces represent overlays such as dialogs and toast layers
This distinction keeps system UI responsibilities separate from normal app content.
TensorUI is intentionally being shaped toward embedded OS use rather than a demo-only UI layer. The current priorities are:
- maintain explicit ownership models
- preserve a small and understandable API surface
- keep public documentation concise and durable
- support repeatable verification during continued refactors