The agent lane

The agent lane renders the same tree for a machine: an agent brain, an MCP client, or a scripted test. It answers the two questions an agent always asks — “what can I do here?” and “how do I do it?” — from the one description.

toolCatalog(bridge)

One MCP-shaped tool per actionable node, each with an inputSchema. An MCP client reads this list and auto-wires the surface with no bespoke code.

import { agent } from 'gamestruct'

agent.toolCatalog(bridge)
// [
//   { name: 'act__ui__start',
//     description: 'Perform "activate" on button "Start game"',
//     inputSchema: { properties: { verb: { enum: ['activate'], default: 'activate' } } },
//     _handle: 'ui:start' },
//   ...
// ]

The _handle field is the stable handle the tool drives — the same handle the human lane walks. That identity is the whole point.

affordances(bridge)

The flat “you can X, Y, Z” list a brain reads before choosing an act.

agent.affordances(bridge)
// [{ handle: 'ui:start', role: 'button', name: 'Start game', actions: ['activate'] }, ...]

toolCall(bridge, name, args)

Dispatch an MCP tools/call onto the bridge. Accepts either the tool name (act__ui__start) or the raw handle (ui:start).

agent.toolCall(bridge, 'act__ui__start', { verb: 'activate' })
// { ok: true, id: 'ui:start', verb: 'activate' }

There is no second code path. The tool call lands on the same act() the human lane uses — which is exactly why the two lanes cannot diverge.