Felix Arntz

This commit introduces a **new capability** to the **AI SDK Core** by adding the `isLoopFinished()` stop condition helper. This helper allows `ToolLoopAgent` instances to run indefinitely, or until the underlying model naturally ceases making tool calls, providing a flexible alternative to fixed `stepCountIs` limits. The change primarily affects **agent loop control** within the `ai` package, enabling more unconstrained agent execution for users who prefer natural termination. Extensive **documentation updates** across various agent and reference guides have been included to reflect this new option and streamline existing examples.

This commit primarily provides a **bug fix** for the **`notify-released` script**, which was generating incorrect **NPM package version URLs**. It **removes unnecessary encoding** of the package name within `.github/scripts/notify-released/index.mjs`, ensuring that release notifications now link directly to the intended package versions on npmjs.com. Additionally, a **maintenance chore** updates the bundle size limit in `packages/ai/scripts/check-bundle-size.ts` from 590KB to 600KB. This improves the accuracy of release communication and adjusts build constraints for the `ai` package.

This commit introduces a **new capability** to the **Google Generative AI provider**, enabling users to specify a `serviceTier` for their model requests. This allows for explicit selection between standard, flex, or priority processing tiers, directly impacting the cost and performance characteristics of interactions with Google Gemini API models. The change updates the `GoogleLanguageModelOptions` schema and the core `google-generative-ai-language-model.ts` to accept and process this new parameter, which is fully **backward compatible** as it remains optional. **Documentation** has been updated, and new **examples** for `generateText` and `streamText` have been added to demonstrate how to leverage this feature.

This commit resolves a **build artifact synchronization issue** affecting the **`codemod` package** by regenerating its `README.md` file. Specifically, it adjusts the column alignment within the codemod tables in `packages/codemod/README.md` to enhance readability. This **maintenance chore** ensures that running `pnpm build` no longer produces unexpected diffs, thereby improving the **developer experience** and build consistency across the project.

This commit **refactors** the **AI Gateway** module by renaming its primary configuration type from `GatewayLanguageModelOptions` back to `GatewayProviderOptions`. This **maintenance** change corrects a previous misnomer, as the gateway's options are truly provider-wide (e.g., routing, fallbacks, tags) rather than specific to a language model. All relevant **documentation**, **examples**, and internal type definitions have been updated to consistently use the more accurate `GatewayProviderOptions` type. To ensure backward compatibility, `GatewayLanguageModelOptions` is retained as a **deprecated alias**. This improves the semantic clarity and accuracy of the **AI Gateway**'s configuration API.

This commit performs a **maintenance chore** by **increasing the allowed bundle size limit** for the **AI package**. It updates the `packages/ai/scripts/check-bundle-size.ts` script, raising the bundle size threshold from 580KB to 590KB. This adjustment prevents recent failures in the `check-bundle-size` CI job on `main`, accommodating minor and expected increases in the **AI package**'s bundle size without blocking continuous integration.

This commit performs **maintenance** on the **Google provider** by **removing an obsolete image generation model** from its supported types. It updates the `GoogleGenerativeAIModelId` type union within the `google` package to no longer include `gemini-2.0-flash-exp-image-generation`, which Google has marked as deprecated. This ensures the provider's model definitions remain current with Google's API, preventing users from attempting to utilize a non-existent or unsupported model. The change primarily affects the **Google Generative AI services** integration, keeping the model list accurate and up-to-date.

This commit introduces **support for configuring reasoning effort** within the **Mistral provider** for the AI SDK. It implements the mapping of the AI SDK's top-level `reasoning` parameter to Mistral's `reasoning_effort` (e.g., `"high"` or `"none"`) for models like `mistral-small-latest`, while also adding a direct `reasoningEffort` provider option. This **new capability** enhances the **Mistral chat language model** by allowing users to control the model's reasoning process, including handling compatibility warnings for non-exact matches. The change also updates the list of supported Mistral model IDs and includes new **documentation** and **examples** to guide usage.

This commit performs a significant **refactoring** within the **`ai/core` package** by **removing the `timeout` property from the `CallSettings` type** and making `CallSettings` non-generic. This change addresses an unnecessary generic parameter that previously propagated through many files, improving **type clarity and reducing complexity**. Functions like `generateText`, `streamText`, and `ToolLoopAgentSettings` now explicitly define `timeout` as a standalone property in their settings, preserving their public API while centralizing the `TimeoutConfiguration` generic where it's truly needed. Additionally, dead code for `timeout` handling in `getBaseTelemetryAttributes` has been removed, further **simplifying the codebase**. This is considered a **low-severity breaking change** for direct consumers of `CallSettings` but aims to enhance the maintainability of the core AI types.

This commit introduces **warning mechanisms** for the **`perplexity`**, **`mistral`**, and **`prodia`** AI SDK providers. It ensures that these providers now emit an explicit warning when a custom `reasoning` parameter, which they do not natively support, is passed to `generateText` or `streamText`, preventing silent failures. This **enhances user feedback** and clarifies behavior for unsupported configurations. Additionally, the **`architecture/provider-abstraction.md` documentation** has been updated to provide comprehensive guidelines for all providers on how to properly handle the new `reasoning` parameter.

This commit introduces a **new capability** by **migrating seven AI providers** within the SDK to support the recently added top-level `reasoning` parameter. This enables users to control the AI model's thinking process by translating the standardized `reasoning` input into each provider's specific configuration, such as `reasoning_effort` for Groq and OpenAI-compatible models, or `enable_thinking` and `thinking_budget` for Alibaba. The **core logic** of providers like Alibaba, Cohere, DeepSeek, Groq, OpenAI-compatible, Open-Responses, and xAI has been updated, along with the addition of new **examples** and **test cases** to ensure correct functionality. This **feature enhancement** significantly expands the configurability of AI models across the SDK, completing a major integration effort for reasoning parameters.

This commit introduces a **new top-level `reasoning` parameter** to the `generateText` and `streamText` functions, providing a standardized and portable way to configure AI model thinking behavior. This **feature** abstracts away provider-specific `providerOptions` for common reasoning settings, improving **code portability** across different AI providers. It updates the **AI SDK Core** `LanguageModelV4` spec and includes new utility functions in `provider-utils` like `mapReasoningToProviderEffort` to translate the generic `reasoning` enum into provider-specific configurations for **OpenAI**, **Anthropic**, **Google**, and **Amazon Bedrock**. Existing `providerOptions` retain precedence for backward compatibility and granular control, but this change significantly **simplifies the developer experience** for configuring reasoning, with extensive documentation and example updates.

This commit introduces a **new `reasoning-file` type** within the **`ai` core package** and **`provider` v4 specification**, allowing files used for internal model reasoning to be distinctly separated from general content files. This **enhancement** significantly improves the structure of AI model outputs, particularly for `generateText` and `streamText` results, by preventing duplicate file entries in `GenerateTextResult` and `StepResult`. The **`google` language model** and its examples are updated to leverage this new type, ensuring more accurate and semantically correct handling of file-based reasoning in AI applications. This change resolves a **bug** where reasoning files were erroneously included in top-level `files` arrays, leading to redundant data.

This commit **fixes** a type dependency issue within the **`provider` package's v4 language model specification**. Previously, various v4 spec types, such as `LanguageModelV4CallOptions` and `LanguageModelV4GenerateResult`, were incorrectly importing and utilizing shared v3 types like `SharedV3ProviderOptions` and `SharedV3Headers`. This **bug fix** updates all affected `language-model/v4` files to correctly import and use their dedicated `SharedV4*` equivalents from `shared/v4`. The change ensures the **v4 language model specification** is self-contained and independent of v3 types, improving API clarity and preventing potential type conflicts or unexpected behavior.

This commit **fixes test instability** within the **Azure OpenAI provider** by increasing the timeout for a specific test. Specifically, the `should stream image generation tool results include` test in `azure-openai-provider.test.ts` now has a 6-second timeout. This change addresses intermittent failures caused by the slow parsing of large base64 image data in test fixtures, which previously exceeded the default 5-second timeout. The adjustment ensures more **reliable CI/CD pipelines** by preventing unrelated test failures due to performance bottlenecks during fixture processing.

This commit performs a **refactoring** of the **`revai` provider**, updating its internal type definitions to use **V4 types** from the shared `provider` package. Specifically, it renames all V3 type references to their V4 equivalents and updates the `specificationVersion` to `'v4'` within `packages/revai/src/revai-provider.ts` and `packages/revai/src/revai-transcription-model.ts`. This **maintenance** work affects the `RevaiProvider` interface, `createRevai` function, and `RevaiTranscriptionModel` class, ensuring the provider aligns with the ongoing spec version upgrade across the project. Although the V3 and V4 types are currently identical, this change establishes future compatibility and consistency for the **`revai` integration**.

This commit **refactors** the **Replicate provider** to align with the new **V4 type definitions** from the core `provider` package. It updates all internal type references from V3 to V4 and sets the `specificationVersion` to `'v4'` within the `replicate-image-model.ts`, `replicate-video-model.ts`, and `replicate-provider.ts` files. This **maintenance** task is a pure type migration, preparing the **Replicate provider** for future V4-specific features without introducing immediate functional changes, as V3 and V4 types are currently identical. The change ensures consistency across providers as part of a broader spec version upgrade.

This commit performs a **type migration** for the **`prodia` provider**, updating its internal type definitions and `specificationVersion` from `'v3'` to `'v4'`. Specifically, the `ProdiaImageModel` and `createProdia` function now implement `ImageModelV4` and `ProviderV4` respectively, aligning with the **ongoing spec version upgrade** across all providers. This is a **refactoring** change that renames V3 type references to their V4 equivalents. As V3 and V4 types are currently identical, there is **no functional impact** but it prepares the provider for future V4-specific features. This **maintenance** task ensures the `prodia` provider adheres to the latest specification.

This commit **refactors** the **Perplexity provider** by **migrating** its internal type definitions to use the **V4 specification** from the shared `provider` package. It updates type references in functions like `convertPerplexityUsage`, `convertToPerplexityMessages`, and within the `PerplexityLanguageModel` class, along with setting the `specificationVersion` to `'v4'`. This **maintenance** task aligns the provider with an ongoing spec version upgrade across the project. While a pure type rename with no immediate functional changes, it prepares the `perplexity` module for future evolutions of the language model specification.

This commit **refactors the `open-responses` provider** to align with the project's ongoing specification upgrade by **migrating all type references from V3 to V4**. It updates type imports, interface definitions, and the `specificationVersion` to `'v4'` across the `open-responses` package, including its core provider, conversion functions, and tests. This **maintenance task** prepares the **`open-responses` provider** for future V4 specification enhancements, ensuring forward compatibility. While V3 and V4 types are currently identical, this change is crucial for future development and introduces no immediate functional impact.