Widgets

Send "widgets": true on a turn and the model gets a built-in tool, render_ui, for showing UI: cards, lists, tables, charts, stats, buttons and forms. The widget spec lists every component.

{ "input": "Show revenue by month as a chart", "model": "gpt-oss-120b", "widgets": true }

1. The model calls render_ui with a widget.
2. Cencori checks it against the spec.
3. Valid: it's sent as the tool's output and the model is told the user can see it. Invalid: the errors go back to the model, which fixes the widget and calls again. Nothing invalid reaches the user.
4. The turn continues: usually a short sentence after the widget.

render_ui runs on Cencori, so it never pauses the turn, even with pause_on_tool_calls.

event: tool_call.delta       data: {"action_id":"call_w","tool":"render_ui","delta":"{\"widget\":{\"type\":\"Card\",..."}
event: tool_call.started     data: {"tool":"render_ui","arguments":{},"action_id":"call_w"}
event: tool_call.completed   data: {"tool":"render_ui","action_id":"call_w","output":{"ok":true,"widget_id":"wgt_...","widget":{...}}}

• tool_call.delta carries the widget's JSON as the model writes it, so a chat can draw it early. It's live only (not stored) and isn't sent when an output policy or PII masking needs the full text first. How smoothly it arrives depends on the provider.
• tool_call.completed with output.ok: true carries the checked widget. Draw that one.
• output.ok: false carries errors: the model is fixing it.

The session remembers what was shown: later turns see a note like [Showed the user a Card: Revenue], so "break that down by product" works.

A button or form in a widget sends an action. With Agent UI it becomes the next turn's input:

[The user used the widget ("Break down by product"): action "breakdown" {"by":"product"}]