The agent stops and asks before it acts
A human-in-the-loop checkpoint: a question, some context, and option buttons to decide. While it waits it wears the warning pair that tool calls use for pending, because the two mean the same thing. Once answered, the signal drops to monochrome and the footer echoes the choice.
The question
A title that asks, one line of context, and the choices left to right. Mark the recommended option primary so there is a default to reach for.
The agent wants to write to src/routes/checkout.ts.
With a tool call
The detail slot holds whatever the decision covers. The intended pairing is a pending tool call, so the person can open the arguments before answering.
Review the call, then decide.
{
"path": "src/routes/checkout.ts",
"bytes": 2048
}Deciding
The card is fully controlled: it reports the choice and the consumer renders it back. Pick an option to see the answered state, where the warning drops away and the chosen label is echoed in the footer.
The agent wants to write to src/routes/checkout.ts.
Danger options
A destructive approval takes the danger weight, which renders the destructive button. The card itself keeps the warning pair either way, because the state is still the same: waiting on a person.
This removes the generated fixtures and cannot be undone.