Versions and compatibility

Package availability

Registry status checked on 17 October 2026:

Use an explicit version when adopting the verified compatibility hotfix:

npm install @yourgpt/[email protected] @yourgpt/[email protected]

Keep the provider peer packages required by your application installed. This command does not install the unreleased features below. Recheck dist-tags before a later upgrade; an unqualified install follows latest, not beta.

The workspace packages are versioned 2.6.0, but that version has not been published to npm as of the check above. The documentation and playground deploy from main separately from npm. See the 2.6.0 upgrade guide for the combined feature set and compatibility changes.

Compatibility fixes in 2.6.0-beta.3

• Structured tool messages: OpenAI-compatible raw-message transports preserve array/object content instead of coercing it to [object Object]. The provider still validates its own content schema. Explicit modelOutput uses its existing provider-specific conversion.
• Anthropic system content: null, non-text values and unsupported blocks are ignored; valid text survives. An explicit prompt repeated as a system message is sent once, while distinct instructions retain their order.
• Responses metadata: saved messages carry reduced continuation state, while full provider output and usage remain on the server-side generation result.
• Malformed tool calls: invalid Responses function calls raise 502 / INVALID_PROVIDER_RESPONSE before execution, so configured recovery can run before any output has been emitted.
• Optional tool calls: generateText handles models that omit toolCalls.

Which system prompt field should I use?

Choose one form for the same instruction. A Runtime systemPrompt is convenient when your server owns the instruction and the message list holds conversation history:

import { createRuntime } from "@yourgpt/llm-sdk";
import { createAnthropic } from "@yourgpt/llm-sdk/anthropic";

const runtime = createRuntime({
  provider: createAnthropic(),
  model: "claude-sonnet-4-5",
});
const systemPrompt = "Classify the email. Do not answer it. Return SYSTEM_OK.";
const emailContext = "Please help me with a duplicate charge.";

const result = await runtime.chat({
  systemPrompt,
  messages: [{ role: "user", content: emailContext }],
});
console.log(result.text);

The message-based equivalent is:

const result = await runtime.chat({
  messages: [
    { role: "system", content: systemPrompt },
    { role: "user", content: emailContext },
  ],
});

For Anthropic, Runtime extracts system text into the provider's top-level system field and removes those messages from the conversation. The nonempty request prompt, or Runtime default, comes first. Distinct system messages follow in order. Duplicate text shared between the explicit prompt and messages is removed; intentional repeated message-only instructions remain.

In 2.6.0-beta.1, the Runtime adapter dropped system-role messages; explicit systemPrompt worked. Beta.2 forwarded system messages but rejected unusual content and could duplicate the explicit prompt. Beta.3 retains forwarding and fixes those regressions. The separate modern Anthropic generation provider already supported system-role messages.

Keep diagnostics out of public responses

Full generation results contain server diagnostics. Return the fields your UI needs, rather than forwarding a complete generateText result. Runtime results still expose top-level billing usage and debug events; remove them at your application boundary after accounting:

const result = await runtime.chat(requestBody);
await recordBilling(result.usage); // Your existing server-side billing function.
const { usage, events, ...clientResult } = result;
return Response.json(clientResult);

Do not remove the reduced message-level continuation metadata needed for Responses replay. Default SSE responses omit usage; includeUsage: true deliberately exposes it. Installing the SDK cannot update your controller's response serialization or dependency pins automatically.

Verification and rollout

The isolated published hotfix passed 371 automated tests on Node 18 and Node 26, 22 live provider checks, 41 packed-consumer regression cases and all 13 public entry-point imports. The downloaded npm archive matched the tested archive. Google, xAI and Fireworks shared transport paths used fixtures, not live keys.

The combined source merged in PR #131 passed 550 automated tests, both application builds and public deployment checks. Feature Lab testing lists live coverage and skips. Direct Jev/Cloudflare calls and physical audio quality still require validation.

For an existing app, pin the intended package versions, retain conversation metadata in custom storage, apply the public-response boundary, and test its actual prompts and tools before rollout. No automatic guardrail policy change is included; see parameters and fallback.