@volter/twin-merge

The Merge Gateway twin (Protocol 3), serving what Open Autonomy's model rail calls (and through it RH2's, Volter Harness's and Substrate's workers) and what Volter Editor's Pi calls in Substrate (journeys/demand.json, journeys/decisions.json) at api-gateway.merge.dev:

• POST /v1/chat/completions (OpenAI's chat completions, streamed and not) on the kernel's openaiWire, its usage carrying cost and the answer service_tier;
• POST /v1/responses in Gateway's native Responses shape (an output message with finish_reason, usage with cost; streamed as response.stream snapshots and a response.done, with no [DONE]);
• POST /v1/messages (Anthropic's Messages, streamed and not) on the kernel's anthropicWire, its usage with no cost;
• GET /v1/models, the catalogue filtered by provider and vendor, paged by cursor/next_cursor, each model with its access_required.

With include_routing_metadata (or its header), routing names the model and vendor that served and carries cost_usd and Merge's 5% merge_fee_usd. Every call is priced from the twin's catalogue (Open Autonomy's default model at the per-token prices it measured against Gateway, spec/SOURCE.md) and drawn from the organization's credit; a used-up credit is 402 budget_exceeded before any turn. The catalogue also holds Claude Haiku 4.5 at Anthropic's list prices; a gated provider (Anthropic, for an organization made on or after 2026-09-04 21:00 UTC) is refused 403 vendor_access_required until support grants it. Every other operation is unmodeled.

The twin runs no model. A turn is the World's scenario's when it scripts one (handlers/merge.json, one vocabulary for the three surfaces), else the wires' labeled deterministic stub.

The World's doors: POST /_twin/organizations ({ name?, credit_usd?, access? }: a Pro organization and its key, mg_…, made at the World's clock) and POST /_twin/access ({ organization_id, provider }: support's grant of a gated provider). The credential door answers MODEL_GATEWAY_API_KEY and MERGE_GATEWAY_API_KEY.

Gateway's core job is routing API calls, so this pack has no screens. Its dashboard-only organization and access acts are World doors. It serves native Responses without an OpenAI wire-format or Codex originator header; stored Responses, routing policies, Management API resources, media and the other operations in unmodeled answer the gap.

A native Responses request uses an array of input messages:

const response = await fetch("https://api-gateway.merge.dev/v1/responses", {
  method: "POST",
  headers: { Authorization: `Bearer ${process.env.MERGE_GATEWAY_API_KEY}`, "Content-Type": "application/json" },
  body: JSON.stringify({ model: "zai/glm-5.3-flash", input: [{ role: "user", content: "hello" }] }),
});
console.log(await response.json());

A World scenario can script a tool call in handlers/merge.json:

{
  "handlers": [{
    "id": "weather-tool",
    "on": { "hasTool": "weather", "lastMessageIsToolResult": false },
    "respond": { "toolCalls": [{ "name": "weather", "arguments": { "city": "Boston" } }] }
  }]
}