2.6.0 upgrade guide
Promote the tested beta features to stable SDK versions and understand the compatibility changes
The isolated 2.6.0 production candidate in PR #130 carries the feature set tested in LLM SDK 2.6.0-beta.3 and Copilot SDK 2.5.1-beta.3, including the compatibility hotfix. That PR’s runtime source matches those betas. The combined workspace and playground also contain the unreleased next-beta features described below; they are not part of that baseline comparison. npm publication is a separate step; merging the production PR does not publish packages.
Installation
After the stable versions are published:
npm install @yourgpt/llm-sdk@2.6.0 @yourgpt/copilot-sdk@2.6.0Use 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/llm-sdk@2.6.0-beta.3 and @yourgpt/copilot-sdk@2.5.1-beta.3.
Included in this release
| 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 |
The release 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. Feature guides mention the beta versions where the APIs first appeared; those APIs are included in stable 2.6.0.
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. The backend's TRAINING_CONFLICT schema and failure policy still need application-level evaluation.
Next beta
Native Decisions for Jev/TypeSafe, OpenAI, OpenRouter, and Cloudflare, experimental realtime voice, and image generation/editing are implemented in the workspace source and new Feature Lab, alongside later UI/PR adaptations. They remain unpublished on npm and were outside the isolated PR #130 baseline. Use the availability guide to distinguish workspace APIs from installed package exports. Remaining provider/caching and persistence work stays on the roadmap.
For historical beta versions and validation evidence, see the beta release record.