For AI agents

This page covers both sides of SINA's agent story: how to integrate an agent so its output goes through the gate, and the machine-readable resources an agent (or coding assistant) can consume directly.

A prompt or style guide is advisory: the model can misread it, forget it, or be steered around it. The gate is code that runs on your server between the model's request and your render, so following the rules stops being the model's responsibility.

What the agent sends

You give the model a menu of intent verbs. fintechIntentManifest() from @sina-design-system/fintech returns one entry per verb, each carrying a propsSchema (JSON Schema derived from the exact Zod object the gate runs) you can drop into an LLM tool definition:

[
  { intent: "wire_transfer", kind: "governed", summary: "Move money by wire", propsSchema: { /* … */ } },
  { intent: "list_transactions", kind: "display", summary: "Show recent activity", propsSchema: { /* … */ } },
  // … 50 verbs in total
]

The model picks a verb and supplies props. It describes a goal, not a screen.

What the governed integration withholds

• Component names. The manifest lists verbs. SINA maps a verb to a component after the gate runs, so the model's output has no field in which to request a specific component.
• Real data for reads. When the model asks to show a balance or a transaction list, your server supplies the values. The model picks what to show, not what the numbers are.

Applied to the wire example: the model cannot select the confirm dialog (it has no component vocabulary), cannot approve the transfer in the same payload (the approver must differ from the initiator, and the initiator id is set by your server, not the payload), and cannot get a small approval attached to a larger amount (the approval binds to a hash of the exact terms, recomputed server-side). Those are properties of the rule code, and they hold only where requests actually go through the gate.

Try it

This demo runs the real gate on a server when you press Run. The opening scenario streams a fabricated "Confirm" button; strict parsing rejects the extra field before any value is read. Swap scenarios to see where the boundary falls.

The emulator has three modes: mock (a recorded request), direct (hand-written JSON), and live (a real model). The decision is identical in all three because the gate does not know where a request came from. The full walkthrough is the $60k wire.

Docs your agent can read

Point a coding assistant at these and it can consume the docs without scraping:

The JSON catalog follows the draft Design System Documentation Schema (DSDS 0.15.2). One manifest links an entity file for every primitive (props included), every design token (values in a W3C Design Tokens file), every fintech intent, and every guide. Pattern entities carry the same propsSchema as the manifest above, so a tool can load the verb menu straight into its own definitions.