Skip to main content

A2UI artifacts

Boundary statement​

A2UI renders a validated component model; it does not execute model-authored HTML or JavaScript. Unknown component types and unsupported artifact shapes fail closed.

Chat can display declarative output and collect structured user input inside the conversation. The current artifact bridge supports confirmation, single-choice, text input, and JSON-form responses, plus display artifacts rendered through the shared React surface renderer.

Interaction lifecycle​

  1. A run emits an artifact with an ID, type, title, content, metadata, and status.
  2. The client parses an object-shaped payload and maps supported legacy input types into the approved component catalog.
  3. The renderer owns local form state and disables action submission while a response is pending or resolved.
  4. A valid action posts a structured response to /api/uar/runs/{run_id}/artifact-response.
  5. The block displays captured, pending, invalid, or failed state. A malformed JSON form remains local and is not submitted.

The protocol catalog and UAR entity-extension catalog are described in Events, AG-UI, and A2UI. The package-level React, Lit, and Svelte renderers have different ownership roles; only @prometheus-ags/a2ui-uar is the first-party React product renderer.

Profile limits​

  • server-full packages the React artifact block and HTTP response path.
  • embedded-mobile may render the same declarative semantics through a host-owned surface, but it does not inherit the browser component or transport evidence.
  • minimal can expose protocol/runtime paths but carries no packaged renderer claim.

See A2UI testing for the development-only trigger surface and Chat for the customer workflow.