LLM gate

llmGate() adds a pre-invocation model decision that classifies a request into allow or reject categories. It can stop the Agent Invocation before the main Agent Driver runs.

Installation

Import the Capability factory from -hub/agent/capabilities and add it to defineAgent({ capabilities }). Use the configuration example below as the starting point, then tighten modes, policies, stores, and providers for the Agent boundary.

What it adds

The Capability asks a model to choose one configured allow or reject category. It records the decision as an Agent Invocation Context Value, exposes it as a finish extension, and throws LlmGateRejectedError when the selected category is rejected.

Configuration

Define allow and reject categories with stable keys. Use the message option when the host should return a product-specific rejection message.

import { defineAgent } from '@vite-hub/agent'
import { llmGate } from '@vite-hub/agent/capabilities'

export default defineAgent({
  driver: { model },
  capabilities: [
    llmGate({
      allow: {
        support: 'Support request the Agent can answer.',
      },
      reject: {
        unrelated: 'Request unrelated to support.',
      },
    }),
  ],
})

Runtime behavior

llmGate() runs during the input phase. It resolves a model, renders a classifier prompt from the latest user text and configured categories, validates the structured output, and stores the decision in invocation context.

When the decision rejects the request, ViteHub throws LlmGateRejectedError with status code 403.

Requirements

llmGate() requires at least one allow category and one reject category. Category keys must be stable identifiers.

The Capability requires either an explicit model option or an Agent model resolver available to Capabilities.

Driver support

Inspect and verify

Run one allowed and one rejected invocation. Inspect the context value for llm-gate or your custom id, then verify rejected requests stop with LlmGateRejectedError.

Check that the gate does not attach, remove, or grant Capabilities. It records a decision; later behavior must read that decision explicitly.

Options

Reference

• llmRoute()
• rateLimit()
• Source: packages/agent/src/capabilities/llm-gate.ts