Runtime projection
Agent UI starts from typed runtime events, durable snapshots, artifact facts, and evidence facts.
Project agent facts into controllable conversation, process, task, artifact, evidence, and session surfaces.
| Contract | What it answers |
|---|---|
| Projection model | How do runtime facts become user-visible UI state without becoming authority? |
| Message parts | Which text, reasoning, tool, action, artifact, and evidence parts should be rendered separately? |
| Standard surfaces | Which surface answers conversation, process, task, artifact, evidence, and session questions? |
| User actions | Which buttons, approvals, interrupts, edits, exports, and queue controls are controlled writes? |
| Hydration | What should appear immediately, and what can load progressively? |
| Acceptance | Which behavior-level scenarios prove the UI is compatible? |
UI owns the projection layer. Runtime owns execution facts, Artifact systems own deliverables, Evidence owns trust records, and Knowledge owns source-grounded context.