Relay Target

allow された prepared execution には、OpenAI Compatible clientでそのまま使えるrelay targetが含まれます。providerがOpenRouter、OpenAI、Anthropic、OpenAI Compatible、Local Connectorのどれであっても、アプリはこの抽象を優先します。

field	用途
openAiCompatibleModel	実際に呼ぶmodel id
openAiCompatibleClientOptions	{ apiKey, baseURL }。OpenAI compatible clientへ渡す
openAiCompatibleChatCompletionsUrl	raw fetch用の /chat/completions URL
openAiCompatibleRequestHeaders	raw fetch用のAuthorization / content-type headers

OpenAI compatible client options

declare const prepared: import("@vel4ai/sdk").PreparedExecutionAllowed;
declare function createOpenAIClient(input: {
  apiKey: string;
  baseURL: string;
}): { chat: { completions: { create(input: unknown): Promise<unknown> } } };

export async function callViaClient(messages: unknown[]) {
  const client = createOpenAIClient(prepared.openAiCompatibleClientOptions);

  return client.chat.completions.create({
    model: prepared.openAiCompatibleModel,
    messages,
  });
}

raw fetch

declare const prepared: import("@vel4ai/sdk").PreparedExecutionAllowed;
declare const messages: Array<{ role: "user"; content: string }>;

export async function callViaRawFetch() {
  const response = await fetch(prepared.openAiCompatibleChatCompletionsUrl, {
    method: "POST",
    headers: prepared.openAiCompatibleRequestHeaders,
    body: JSON.stringify({
      model: prepared.openAiCompatibleModel,
      messages,
    }),
  });

  return response.json() as Promise<unknown>;
}

provider差分をアプリに漏らさない

Anthropic backend もVela側でOpenAI Compatible shapeへ寄せます。アプリが直接Anthropic SDKへ分岐すると、委任・監査・失効の境界が崩れるため、prepared executionのrelay targetを優先してください。