ViteHub is still experimental. Expect bugs and breaking changes.

LLM route

Choose one developer-defined route before the main Agent Invocation runs.

llmRoute() adds a pre-invocation model decision that chooses one developer-defined route. It records the chosen route as an Agent Invocation Context Value and does not apply route effects by itself.

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 select exactly one configured choice. It can include recent conversation history, records the decision under a stable id, and exposes the decision as a finish extension.

Configuration

Define stable choice keys with short descriptions. Later callbacks can read the recorded context value and decide how to use the route.

server/agents/support.ts

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

export default defineAgent({
  driver: { model },
  capabilities: [
    llmRoute({
      choices: {
        billing: 'Billing and invoice requests.',
        technical: 'Technical troubleshooting requests.',
      },
    }),
  ],
})

Runtime behavior

llmRoute() runs during the input phase before the main Agent Driver. It resolves a model, renders a decision prompt from the latest user text and configured choices, validates the structured output, and stores the result in invocation context.

The default context id is llm-route. Duplicate writers for the same invocation context value fail early.

Requirements

llmRoute() requires at least one choice. Choice keys must be stable identifiers.

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

Driver support

Agent Driver	Support
Model-backed	Runs the pre-invocation decision and records the route before model execution.
Harness-backed	Runs the pre-invocation decision before harness execution when a model resolver is available.
Custom-run-backed	Records the decision before `driver.run`; custom code decides how to use it.

Inspect and verify

Run one invocation and inspect the context value for llm-route or your custom id. The value should include choice and may include confidence or reason.

Add a test case for an invalid model response if you provide a custom model wrapper. The Capability should reject choices outside the configured map.

Options

Option	Type	Default	Description
`choices`	`Record<string, string \| { label?: string; description?: string }>`	required	Developer-defined route choices.
`history`	`boolean \| number`	`false`	Include recent conversation history in the classifier prompt.
`id`	`string`	`"llm-route"`	Capability id and invocation context key.
`model`	`AgentModelResolver`	Agent model	Model used for the pre-invocation decision.
`prompt`	`string`	generated	Additional classifier prompt text.

Reference

llmGate()
Agent invocations
Source: packages/agent/src/capabilities/llm-route.ts