Walter Korman

This commit introduces a **new capability** to the **AI Gateway provider**, enabling users to retrieve detailed information about specific AI generations after they are created. It adds a `getGenerationInfo()` method to fetch comprehensive data, including cost, token usage, and latency, by using a `generationId` now exposed in `providerMetadata` for `generateText` and `streamText` responses. This feature significantly enhances observability and allows for better programmatic analysis and cost management of AI Gateway interactions.

This commit introduces **spend reporting support** to the **AI Gateway provider**, enabling developers to programmatically query usage metrics, costs, and token consumption for their AI applications. A new `getSpendReport()` method is added to the `GatewayProvider` and `GatewaySpendReport` classes, allowing data to be filtered and aggregated by various dimensions such as model, user, or tags. This **new feature** significantly enhances the `@ai-sdk/gateway` package by providing critical cost visibility and usage analysis capabilities. The change includes comprehensive test coverage, updated **documentation** in `content/providers/01-ai-sdk-providers/00-ai-gateway.mdx`, and new example scripts in `examples/ai-functions/src/gateway/` to demonstrate its usage.

This commit introduces a **new capability** by adding support for passing `metadata.userId` to **Anthropic models** within the AI SDK. This feature allows developers to include an external identifier for end-users in API requests, enhancing tracking and analytics. The change affects the direct **Anthropic provider**, as well as Anthropic models accessed via **Amazon Bedrock** and **Google Vertex**, ensuring consistent `userId` propagation. It involves updates to the `anthropic` package, relevant documentation, and includes new test cases and examples to demonstrate its usage.

This commit introduces **new capabilities** to the **KlingAI provider**, adding comprehensive support for the **`kling-v3.0-motion-control` model**. It enables enhanced facial consistency through `elementList` support for motion control, previously exclusive to I2V, and makes the `watermarkEnabled` option universally available across all video generation modes. This **feature enhancement** allows users to leverage the latest KlingAI motion control advancements, with updated documentation and examples reflecting these new functionalities.

This commit introduces **support for OpenAI custom tools** within the **OpenAI provider**, enabling users to define and utilize their own tools with grammar formats. It also **resolves critical bugs** related to **aliased tool names**, preventing runtime errors like `AI_APICallError` or `AI_NoSuchToolError` that previously occurred when provider tool names differed from SDK keys. The changes ensure correct end-to-end mapping of tool choices, parsing, and streaming, allowing for robust execution of both forced and unforced custom tool calls. This **new capability** and **bug fix** significantly enhances the flexibility and reliability of **tool usage** in the **OpenAI integration**, providing a more seamless developer experience.

This commit introduces a **new feature** to the **`gateway-provider`** module, enabling it to **pass the Vercel project ID** for enhanced observability. It implements logic to read the `VERCEL_PROJECT_ID` environment variable and include it as an `ai-o11y-project-id` header in outgoing requests. This change significantly improves the ability to track and identify requests by their originating Vercel project when using the AI SDK gateway in Vercel deployments, ensuring better insights into usage patterns.

This commit introduces **Server-Sent Events (SSE) support** for **video generation** within the **`gateway` provider**, a **new capability** designed to prevent HTTP timeouts during long-running video processing. By utilizing **heartbeat keep-alive messages**, the connection is maintained, significantly improving the reliability of video generation requests. The implementation in `packages/gateway/src/gateway-video-model.ts` now sets the `accept: 'text/event-stream'` header, parses SSE events using `parseJsonEventStream`, and properly handles structured SSE error events. **Tests** in `gateway-video-model.test.ts` were updated to validate SSE stream chunks, error handling, and heartbeat comments, ensuring the robustness of this feature.

This commit introduces a **new capability** by adding the **`@ai-sdk/bytedance` provider package** to the AI SDK. This integration enables support for **ByteDance's Seedance video generation models**, allowing users to perform both text-to-video and image-to-video generation. The new provider expands the SDK's multimedia generation capabilities, offering configurable options for watermarks, audio, and polling behavior. It includes comprehensive documentation and multiple examples within the `AI functions demo` to showcase various generation scenarios.

This commit **adds new capabilities** to the **Kling AI provider** by integrating support for **Kling AI v3.0 models**, specifically `kling-v3.0-t2v` and `kling-v3.0-i2v`. It enables advanced features such as **multi-shot video generation**, **voice control**, and **element control** for more sophisticated video creation within the AI SDK. The implementation in `packages/klingai/src/klingai-video-model.ts` and `klingai-video-settings.ts` allows users to leverage these new options. This **feature addition** significantly expands the video generation options, with updated documentation and new example scripts demonstrating multi-shot workflows.

This commit introduces **comprehensive video generation support** for the **xAI provider** within the AI SDK, integrating the `grok-imagine-video` model. This **new feature** enables users to perform **text-to-video, image-to-video, and video editing** operations directly through the SDK. It includes the implementation of the `XaiVideoModel` class, new provider methods, and extensive examples demonstrating basic generation, image-to-video from URLs or base64, and advanced video editing with chaining and concurrency. The update significantly expands the AI SDK's multimedia capabilities by adding a powerful new video provider.

This commit **fixes** an issue in the **`gateway` package** by adding support for `unsupported` and `compatibility` warning types in the **video model's response parsing**. Previously, these specific warnings from video providers were not correctly handled, leading to incomplete information. The **`gateway-video-model.ts`** schema was updated to use a discriminated union for proper type validation, and new tests were added in `gateway-video-model.test.ts` to verify this **bug fix**. This ensures that all relevant warning types are now accurately processed and displayed, improving the robustness of **video response handling**.

This commit provides a **bug fix** for the **Google Vertex AI provider**, specifically addressing issues with **image-to-video generation**. Previously, the `mimeType` was not being passed along with base64-encoded image data, causing the Vertex AI API to fail. The fix ensures that the image's media type is correctly included when calling the `generate` function in `google-vertex-video-model.ts`, enabling successful video creation from images. This impacts the **Google Vertex** integration, allowing users to reliably generate videos from base64-encoded images.

This commit **adds comprehensive support for text-to-video and image-to-video generation** to the **Kling AI provider** within the AI SDK. It introduces new model IDs, implements API integration for these modes in `klingai-video-model.ts`, and extends provider options in `klingai-video-settings.ts`. This **new capability** significantly expands the provider's functionality beyond motion control, offering users a complete suite of Kling AI video generation options. Example implementations in `examples/ai-functions/src/generate-video/` and updated documentation are also included to guide usage.

This commit introduces **video generation support** to the **Alibaba provider** within the AI SDK, enabling users to create videos via text-to-video, image-to-video, and reference-to-video methods. It adds a new `.video()` factory method to the `AlibabaProvider` and implements the `AlibabaVideoModel` for asynchronous task creation and polling. This **new capability** significantly expands the Alibaba integration, providing comprehensive provider options for customization. The change includes new unit tests, updated documentation in `content/providers/01-ai-sdk-providers/32-alibaba.mdx`, and several new examples in `examples/ai-functions/src/generate-video/` to demonstrate its usage.

This commit **enhances the XAI provider integration** by adding support for new `response.reasoning_text.delta` and `response.reasoning_text.done` chunk types. This **maintenance update** ensures that **streaming reasoning text** from the XAI API is correctly processed and converted into `reasoning-delta` events within the SDK's standardized output. The **XAI response schema** (`xai-responses-api.ts`) and **language model processing logic** (`xai-responses-language-model.ts`) are updated to handle these new formats. This **fixes** the handling of evolving API responses, preventing data loss for reasoning streams, and includes a new example demonstrating streaming reasoning from the `grok-code-fast-1` model.

This commit introduces a **new feature** to the **`examples/ai-functions`** module by adding **duration tracking to the `Spinner` class**. The spinner now displays elapsed time in milliseconds, seconds, or minutes, providing users with real-time feedback on the duration of ongoing operations. This enhancement primarily affects the `spinner.ts` file, where methods like `start`, `succeed`, `fail`, and a new `formatDuration` method were updated or added. Additionally, the `presentVideos` function in `present-video.ts` received a minor **refactoring** for improved code clarity.

This commit **enhances the Google Vertex AI provider** by **updating its video generation capabilities** to support the latest `veo-3.1-generate-001` models. It **adds comprehensive documentation** for Google Vertex video generation, detailing usage, provider options, and a model capabilities table that highlights audio generation support. Specifically, the `GoogleVertexVideoModelId` type in `packages/google-vertex/src/google-vertex-video-settings.ts` is updated, and the documentation in `content/providers/01-ai-sdk-providers/16-google-vertex.mdx` and `content/docs/03-ai-sdk-core/38-video-generation.mdx` is expanded. This **feature enhancement** ensures users can leverage the most current video generation models and understand their features, including audio generation support.

This commit **fixes** an issue in the **MoonshotAI provider** by **enabling usage information tracking** during streaming operations. Previously, token usage statistics were not reported for streaming responses, making it difficult to monitor resource consumption. The change involves setting the `includeUsage` option to `true` within the `createChatModel` configuration in `packages/moonshotai/src/moonshotai-provider.ts`. Additionally, **new example files** (`moonshot-cache.ts`) have been added to demonstrate both regular and streaming text generation with large contexts, verifying the correct reporting of token usage. This **enhances the reliability of cost tracking** and resource management for users of the **MoonshotAI provider**.

This commit introduces a **new feature** by integrating the **KlingAI provider** into the AI SDK, enabling **motion control video generation** capabilities. A new **`@ai-sdk/klingai` package** is added, implementing authentication via JWT, integrating with the KlingAI video generation API, and supporting both standard and professional quality modes with customizable watermarking and audio options. This allows users to generate videos by transferring motion from a reference video to a character in an image, complete with comprehensive error handling and polling for asynchronous tasks. The change includes new **documentation** and **example usage** within the AI functions demo, significantly expanding the SDK's video generation offerings.

This commit introduces a **new capability** to the **AI SDK's model resolution system**, enabling **video models to be resolved from string identifiers** using the global default provider. It updates the `resolveVideoModel` function and extends the `customProvider` to support video models, bringing consistency with how other model types are handled. This **enhancement** also adds `videoModel` methods to various **provider implementations** such as `FalProvider`, `GoogleVertexProvider`, `GoogleGenerativeAIProvider`, and `ReplicateProvider`. Furthermore, it **improves error handling** by providing clearer guidance when the default provider does not support video models. This change simplifies how users specify video models and streamlines their integration across the SDK.