AI SDK

Fixed a bug where onInputStart was not called before onInputAvailable during non-streaming tool calls.

Fixed handling of empty HTTP 202 responses while polling videos, and images in Responses API tool results are now preserved.

Both the OpenAI and open-responses providers now throw a clear APICallError when the Responses API returns a 200 status with no output array, replacing the opaque "output is not iterable" TypeError with an actionable message that surfaces upstream error details.

MCP now accepts OAuth metadata that lacks code challenge methods, fixing a compatibility issue with providers that omit this field.

LangChain adapter now allows suppressing outer start and finish UI message chunks, and fixes LangGraph reasoning closing before text starts.

Fixed a security issue where workspace symlink escapes were possible in file tools.

Denied tool output state is now correctly emitted for client-rejected approvals, and onInputStart is called before onInputAvailable during non-streaming tool calls. Updated internal provider utilities and gateway.

Fixed a bug where empty HTTP 202 responses during video polling could cause issues. Includes dependency updates to provider-utils and openai-compatible packages.

Fixed an issue in the LangChain adapter where LangGraph reasoning was not closed before starting text generation.

Fixed a bug where onInputStart was not called before onInputAvailable during non-streaming tool calls, ensuring correct event ordering.

Provider support for end-user identifiers has been added for video generation and editing use cases.

Assistant messages that contain only tool calls now send null content instead of potentially causing issues with providers that expect the field to be absent or empty.

Added the kimi-k3 model and a reasoningEffort provider option for Moonshot AI.

Patch Changes

• 73c6143: Backport: chore(provider/gateway): update gateway model settings files

Added the kimi-k3 model and a new reasoningEffort provider option to the AI SDK Gateway.

xAI provider now supports end-user identifiers for video generation and editing.

Adds the kimi-k3 model and a new reasoningEffort provider option for configuring reasoning behavior.

Patch Changes

• 49a3098: Backport: chore(provider/gateway): update gateway model settings files

Added gateway.experimental_transcription.getToken for minting transcription-bound client secrets.

Structured error data from chat completion SSE streams is now preserved instead of being lost during parsing.