OpenAI Responses

Responses first shipped in LLM SDK 2.6.0-beta.2; use the compatibility-fixed 2.6.0-beta.3 with Copilot SDK 2.5.1-beta.3. There is no separate runtime.response() method.

Use api: "responses" to select OpenAI Responses, including calls without tools. The installed openai peer must expose responses.create; an older client fails with an actionable error. A compatible Chat endpoint does not automatically support Responses. See release availability before upgrading an existing application.

Stateless conversation replay

import { generateText, streamText } from "@yourgpt/llm-sdk";
import { openai } from "@yourgpt/llm-sdk/openai";

const model = openai("gpt-6-luna", {
  api: "responses",
  responses: { store: false },
});

const first = await generateText({
  model,
  prompt: "Remember the project code ORCHID-742.",
});
const second = await streamText({
  model,
  messages: [
    ...first.response.messages,
    { role: "user", content: "What is the project code?" },
  ],
});

console.log(await second.text);
const history = (await second.response).messages;

Preserve complete returned messages, including providerMetadata, modelOutput and tool call IDs. Assistant metadata holds ordered continuation items, encrypted reasoning where supplied, and assistant phase. It excludes the raw response envelope, native usage and diagnostic extensions. With store: false, the SDK requests encrypted reasoning automatically. It only replays completed items.

generateText().response.messages includes the final assistant turn and tool turns. Stream readers and final promises share one execution; reading the result does not issue a second request or execute tools twice.

Copilot preserves this state in JSON/SSE continuation and serialized UI messages. Custom Runtime storage must retain StorageMessage.metadata.providerMetadata and .modelOutput, then restore them onto each message. The existing YourGPT cloud storage adapter does not yet guarantee lossless Responses replay after a remote session reload.

Stored continuation

With store: true, read the response ID from the server-side result at result.providerMetadata.openai.response.id. Pass it on the next generation as openai: { responses: { previousResponseId: id } }, with only new input and the system instructions that still apply. Do not also resend the same full history. The SDK does not persist or advance response IDs for you. Omitting store leaves the provider default unchanged. OpenAI conversation state describes the provider's retention and continuation behavior.

Configuration and caching

api accepts "auto", "chat", or "responses". Responses options also select Responses; combining them with api: "chat" fails locally. Set defaults on openai, createOpenAI or OpenAIAdapter. Per-call overrides go in generateText({ openai: { api, responses }, ... }) or runtime.chat({ config: { openai: { api, responses } }, ... }).

Cache eligibility depends on the provider, model and prompt. These options do not guarantee a hit, savings or a latency improvement. Inspect server-side usage to measure results; a zero cached-token count is a valid response. Reasoning effort remains a separate setting; see parameters.

Public responses and errors

The generation result's top-level providerMetadata contains full provider output and usage for server diagnostics. Do not serialize that complete result to public clients. Runtime still exposes billing usage and debug events; omit them after server-side accounting. Default SSE helpers omit usage; includeUsage: true is an explicit opt-in. The upgrade guide shows the response boundary.

Malformed function calls fail before execution with 502 / INVALID_PROVIDER_RESPONSE, allowing the configured fallback policy to recover before output starts. Failed, cancelled or incomplete streams do not execute partial function calls or return a successful final result.

Native image tool results have their own guide. Background jobs and native MCP approval flows are outside this integration. Use the Feature Lab to check streaming and non-streaming replay.