Lars Grammel

This commit performs a significant **refactoring** within the **AI SDK's tool handling logic**, extracting the responsibility of filtering active tools into a new, dedicated utility function. The new `filterActiveTools` function encapsulates logic previously embedded within `prepareToolsAndToolChoice` and implicitly used by `generateText` and `streamText`, removing the `activeTools` parameter from these internal calls. This **improves modularity** and prepares for future enhancements by **exposing** this new functionality as `experimental_filterActiveTools` for external use. Comprehensive unit and type tests have been added for `filterActiveTools`, along with updated documentation for its experimental public API.

This commit **adds new documentation** detailing the structure of the **`streamText` pipeline**. A new file, `architecture/stream-text-loop-control.md`, is introduced to provide a comprehensive explanation of its architecture, including an ASCII art diagram. This **documentation enhancement** improves developer understanding of the streaming process's internal organization. The change specifically impacts the project's **architectural documentation**, making it easier to grasp the flow and control mechanisms of the stream.

This commit **introduces and exposes** the `Experimental_ModelCallStreamPart` type within the **`packages/ai` library**, specifically for the `generate-text` subsystem. It **refactors** internal streaming types, renaming `UglyTransformedStreamTextPart` to `ModelCallStreamPart` and standardizing `response-metadata` to `model-call-response-metadata` for improved clarity and consistency. The work also **integrates stream part transformation logic** directly into the `streamModelCall` function and adds comprehensive **tests** for `streamModelCall`, enhancing its robustness. This **feature** provides a more stable and well-defined interface for handling AI model stream events, impacting downstream consumers and updating examples like `ai-functions`.

This commit performs a significant **refactoring** within the **AI stream processing** module, specifically `packages/ai/src/generate-text`. It standardizes the naming of stream parts by renaming `model-call-init` to `model-call-start` and `finish` to `model-call-end` across various transformations and type definitions. Concurrently, it **removes redundant delay logic** for the `finish` (now `model-call-end`) stream part from the `createExecuteToolsTransformation` function, as this delay is already correctly handled by the `streamText` utility. This cleanup improves code consistency, reduces duplication, and enhances the maintainability of the **stream text generation** pipeline.

This commit introduces a **refactoring** to the **AI package's text generation streaming** capabilities by **renaming the internal `stream-start` part type to `model-call-init`**. The `createStreamTextPartTransform` in `packages/ai/src/generate-text/` is updated to convert incoming `stream-start` chunks into the new `model-call-init` type, ensuring compatibility during this transition. Consequently, the `DefaultStreamTextResult`'s `transform` method now checks for the `model-call-init` type, affecting how initial stream events are processed internally. This change is a foundational step towards introducing more dedicated and semantically clear part types for language model call streams.

This commit introduces a **new capability** within the **`@ai` package** by extracting and exposing `experimental_streamModelCall`, a lower-level function for direct **streaming text generation** from language models. This function encapsulates core logic such as prompt conversion, tool preparation, and retry mechanisms, enabling more granular control over LLM interactions. The existing `streamText` function has been **refactored** to utilize this new primitive, and a mechanism to prevent further token emission upon `stopStream` calls has been added, improving stream control. This change lays the groundwork for custom loop controls in `streamText` and provides developers with a powerful new tool for advanced AI streaming scenarios, as demonstrated by a new example.

This commit introduces a significant **refactoring** within the **AI streaming pipeline** by extracting the logic for **tool callback invocation** into a new, dedicated function, `invokeToolCallbacksFromStream`. This change modularizes the `streamText` functionality, making it more maintainable and preparing it for enhanced custom loop control. A **breaking behavior change** is introduced: previously, tool callbacks were invoked before a chunk was forwarded, but now chunks are forwarded *before* callbacks are invoked, affecting the order of operations for consumers of the stream. The `createStreamTextPartTransform` module was also updated to support new chunk types, and comprehensive tests were added for the new `invokeToolCallbacksFromStream` function.

This commit introduces a significant **refactoring** to the **`@ai` package's `streamText` function**, fundamentally altering the timing of **tool execution**. Previously, tools would execute immediately; now, their execution is **delayed** until the model call has fully completed and sent a `finish` chunk, representing a **breaking behavior change**. This is achieved by encapsulating the tool execution logic into a new `createExecuteToolsTransformation` function, which `streamText` now utilizes as a `TransformStream`. The change aims to enable **better support for custom loop control** and provide a clearer separation between model calls and tool execution, which is crucial for external systems like the Workflow SDK.

This commit performs a significant **refactoring** within the **AI package's stream text generation** capabilities, specifically focusing on **type simplification and modularization**. It **replaces the `SingleRequestTextStreamPart` type with a simplified `UglyTransformedStreamTextPart`** and introduces more granular `TextStream*Part` types to better represent individual transformation steps. This work **prepares the `streamText` functionality for enhanced custom loop control** by separating transformation responsibilities, impacting files like `stream-text-result.ts` and `create-stream-text-part-transform.ts`. The architecture documentation (`architecture/stream-text-loop-control.md`) is also updated to reflect these foundational changes.

This commit **refactors** the **AI stream processing logic** within `packages/ai/src/generate-text` by moving stream transformations, specifically the handling of 'finish' parts and tool execution, to earlier stages in the pipeline. This **architectural improvement** separates transformation steps to ensure single responsibility, preparing the system for better custom loop control. Key changes include **renaming** `runToolsTransformation` to `executeToolsTransformation` and updating its usage in `streamText.ts`, along with explicitly handling and enqueuing 'finish' parts in `createStreamTextPartTransform.ts`. The work streamlines the internal handling of stream parts and tool outputs, enhancing the modularity and maintainability of the **stream text generation** feature.

This commit performs a significant **refactoring** within the **`generate-text`** module, specifically by moving the logic for transforming **tool calls, tool approval requests, and tool results** into the `createStreamTextPartTransform` function. This change aims to improve modularity and ensure each transformation has a single responsibility, laying groundwork for better custom loop control. The `runToolsTransformation` function is streamlined by offloading these responsibilities, while `stream-text.ts` is updated to reflect the new API. This enhances the maintainability and clarity of the text streaming pipeline by centralizing tool-related stream part processing.

This commit **fixes a bug** in the **AI package's tool execution pipeline** where **tool input was not included in stream text tool results** for provider-executed dynamic tools. It refactors the `runToolsTransformation` function within `packages/ai/src/generate-text/run-tools-transformation.ts` to unify bookkeeping objects and correctly track tool input using `toolCallsByToolCallId`. This ensures that **all tool results, including those from dynamic tools, now consistently contain their corresponding input**, improving the completeness and accuracy of AI stream outputs. The change is validated by updated test expectations in `stream-text.test.ts`.

This commit **refactors** the **AI text generation stream processing pipeline** by **relocating the transformation logic** for `reasoning-delta`, `file`, and `reasoning-file` stream parts. This **architectural improvement** moves their processing from `runToolsTransformation` to an earlier stage within `createStreamTextPartTransform` in the **`@ai` package**. The change aims to enhance the modularity and single responsibility of stream transformation steps, which is crucial for future **custom loop control** capabilities. Extensive test updates ensure the correct handling and output structure of these stream parts in the new processing order.

This commit **enhances the project's architecture documentation** by introducing two new files that detail core system components. It adds a comprehensive explanation of the **four-level message architecture** in `architecture/message-layers.md`, outlining how messages are structured and processed. Additionally, a new document at `architecture/stream-text-loop-control.md` details the **stream text loop control process**, clarifying its operational flow. This **documentation enhancement** improves clarity and understanding of critical architectural patterns for developers and contributors.

This commit **refactors** the **AI text generation streaming pipeline** within the `@ai` package by introducing a new `createStreamTextPartTransform` function. This utility is responsible for converting `LanguageModelV4StreamPart` objects, specifically `text-delta` chunks, into a standardized `UglyTransformedStreamTextPart` format. The `runToolsTransformation` function is updated to leverage this new transform, centralizing text part conversion and removing its internal `text-delta` handling. This **refactoring** improves the modularity of the `streamText` functionality and is a foundational step towards enabling more flexible custom loop control and advanced stream part transformations.

This commit performs a significant **refactoring** within the **`generate-text` module** by moving the responsibility for `delta` to `text` conversion into an earlier stage of the transformation pipeline. It renames the `delta` property to `text` within the `SingleRequestTextStreamPart` type for `text-delta` and `reasoning-delta` chunks, updating the `runToolsTransformation` function and `stream-text.ts` to process this new property. This **architectural change** enhances the modularity and single responsibility of individual transformation steps, laying groundwork for future custom loop control features. The update includes necessary adjustments to `run-tools-transformation.test.ts` to reflect the new output structure.

This commit implements a **bug fix** within the **OpenAI provider** to address issues arising from incomplete reasoning sequences. Specifically, it modifies the `convertToOpenAIResponsesInput` function to **filter out reasoning parts that lack encrypted content** when the `store` option is set to `false`. This prevents subsequent OpenAI responses requests from breaking due to malformed or unsigned reasoning data, which previously occurred when requests were aborted prematurely. **Tests** have been updated to validate this new filtering logic, ensuring only complete and valid reasoning parts are processed.

This commit **introduces new documentation** outlining the **provider abstraction architecture** within the project. A comprehensive guide is added at `architecture/provider-abstraction.md`, detailing concepts such as **AI functions**, **model specifications**, and **provider implementations**. This **new documentation** serves as a critical resource for coding agents and contributors, enhancing their understanding of the system's design for extending or maintaining AI-related functionalities. Additionally, `AGENTS.md` is updated to include a reference to this newly established architectural document.

feat(open-responses): add Open Responses provider (#11836)

This commit introduces a **new analytical tool** within the **`tools/analyze-downloads`** module to understand AI SDK usage patterns. It adds a TypeScript script, `analyze-ui-frameworks.ts`, which analyzes npm download statistics for various UI framework integration packages. This **developer tooling enhancement** specifically focuses on the last two major releases to ensure data accuracy by excluding older, potentially "tainted" download numbers. The new script provides valuable insights into which UI frameworks are most commonly used with the AI SDK, aiding future development and strategic decisions.