2.6.0 upgrade guide
Features, package availability, and compatibility changes in the 2.6.0 source
This guide covers the combined SDK features and compatibility fixes in the 2.6.0
source on main, including Decisions, Responses, image tools, audio and realtime
voice. The documentation and playground use this source. Check
package availability before upgrading an npm
installation; merging source and deploying the website do not publish packages.
Installation
After the stable versions are published:
npm install @yourgpt/[email protected] @yourgpt/[email protected]Use matching stable versions when validating the upgrade. The beta tag is a separate release channel for subsequent work. Provider clients such as openai and @anthropic-ai/sdk remain optional dependencies for the corresponding chat adapters.
Until stable publication, use the verified versions @yourgpt/[email protected] and @yourgpt/[email protected].
Features in 2.6.0 source
| Feature or fix | Guide |
|---|---|
| OpenAI reasoning-model parameters and selective compatibility fallback | Parameters, fallback |
| Anthropic Runtime system-role preservation | System prompts |
| OpenAI Responses transport, stored/stateless continuation, and tool replay | Responses |
| Native image tool results on OpenAI Responses and Anthropic | Image results |
| Buffered OpenAI text-to-speech and recorded-file speech-to-text | Speech, transcription |
| Browser microphone recording, cancellation, and transcript review | Voice input |
| Structured output translation and token usage fixes across existing adapters | Structured output, token tracking |
| Native Decisions for OpenAI, Jev/TypeSafe, OpenRouter and Cloudflare | Decisions |
| OpenAI image generation and editing | Image generation |
| Experimental browser voice calls with Copilot tools | Realtime voice |
| Modern OpenRouter reasoning controls and replay | Reasoning controls |
The source also retains the tested fixes for mixed server/client tool turns, tool-loop iteration limits, cancellation, stream reader reuse, interleaved tool arguments, screenshot attachments, trailing usage, and public CJS/ESM declarations. The version guide lists which capabilities are available in each published package.
Compatibility checks
- Anthropic instructions: Runtime uses
systemPrompt; modern generation usessystem. System-role messages reach Anthropic. The effective explicit/default prompt comes first, followed by distinct system instructions in order. Text repeated between that prompt and a system message is sent once. Unsupported non-text system content is ignored; valid text is retained. - OpenAI parameters: reasoning requests omit unsupported temperature values with a warning. Use supported
reasoningEffort: 'none'explicitly when sampling is needed. Custom compatible servers can select passthrough behavior. - Continuation: preserve reduced continuation metadata, native tool output, and matching tool-call IDs in saved history. Raw provider diagnostics are not replay state. At application response boundaries, remove top-level Runtime billing
usageand debugeventsafter accounting. A transcript containing only visible text cannot restore every Responses conversation. - Tools: raw array/object tool results are preserved rather than converted to
[object Object]; provider content schemas still apply. Malformed Responses calls can trigger configured fallback before output or tool execution begins. Native image tool output requires a supported provider/endpoint. A screenshot with native output does not add a duplicate user-image turn. - Audio: speech and transcription are separate APIs from chat generation. Browser recording returns a transcript for review; it does not submit a chat message automatically.
Known limits
The hosted YourGPT storage adapter does not yet preserve all Responses/image replay metadata. Use a lossless custom store and test save/reload/continue for those workflows. Physical microphones, browser permission flows, and recorded-container compatibility still require validation on your target devices.
OpenAI is the built-in TTS/STT provider. Live media checks covered inline images and WAV/MP3; other supported formats have fixture coverage. Cache option forwarding is available for OpenAI Responses, while unified cache usage/cost accounting remains pending. See storage and prompt caching.
Prompt delivery checks do not establish strict JSON output or application guardrail accuracy. Evaluate your application's schemas and failure policy with representative inputs.
Trying the features
Use the Feature Lab to try the SDK with your provider keys. Direct Jev and Cloudflare account tests and physical microphone validation remain pending. Provider caching and persistence follow-ups are tracked on the roadmap.
For published versions and validation evidence, see the release history.