josh

This commit delivers a **bug fix** for the **Amazon Bedrock integration** by addressing an issue where cryptographic signatures on reasoning blocks were invalidated due to unintended whitespace trimming. Specifically, the `trimIfLast()` function in the **message conversion logic** (`convertToBedrockChatMessages.ts`) was incorrectly modifying signed reasoning text, leading to `ValidationException` errors from Bedrock. The change now ensures that **signed reasoning blocks** are preserved without modification, while unsigned blocks continue to be trimmed as expected, thereby maintaining the integrity of cryptographic signatures and enabling correct multi-turn tool use.

This commit introduces a **new capability** to specify **per-tool timeout overrides** within the `@ai-sdk/ai` package. Users can now define distinct timeout durations for individual tools by adding a `tools` object to the `timeout` configuration, such as `timeout: { tools: { weatherMs: 3000 } }`. This significantly enhances the **tool execution** mechanism by providing more granular control, improving reliability when integrating tools with diverse latency profiles. The change includes **type safety improvements** using template literal types and updates various `CallSettings` types across core functions like `generateText`, `streamText`, and agent-related operations. This allows for a more robust and flexible handling of external tool calls, preventing slow tools from unnecessarily impacting faster ones.

This commit performs a **refactoring** and **bug fix** within the **OpenAI provider** by eliminating a redundant `name` argument from `openai.tools.customTool()`. Previously, custom tools required a separate `name` field that duplicated the SDK tool key, introducing unnecessary complexity for alias mapping. The change simplifies the **custom tool definition API** by automatically deriving the tool name from the SDK tool key (`tool.name`), aligning it with how other provider tools are named. This also removes the `resolveProviderToolName` parameter from the shared `createToolNameMapping` utility in `provider-utils`, streamlining **tool name resolution logic** and reducing maintenance overhead.

This commit introduces **new capabilities** to the **xAI video model integration**, enabling it to properly handle and expose additional data from the xAI API. It now surfaces `costInUsdTicks` within the provider metadata for video generation, offering greater transparency into usage costs. Furthermore, the **`xai` provider package** will now explicitly throw an `XAI_VIDEO_MODERATION_ERROR` when content is blocked by xAI's moderation system, replacing previous generic error messages. This **enhances error handling and data transparency** for users generating videos with xAI, providing more specific feedback and financial insights.

This commit introduces a **new capability** to configure **per-tool execution timeouts** within the **AI SDK**, addressing issues where tools could hang indefinitely. It extends the `TimeoutConfiguration` with a `toolMs` property, allowing users to specify a maximum duration for individual tool calls, which are then aborted if they exceed this limit. This enhancement is integrated into the `generateText` and `streamText` APIs, ensuring that timed-out tools return a `tool-error` to the model, enabling it to respond gracefully or retry. This significantly improves the robustness and reliability of AI applications, particularly in environments with strict execution limits.

This commit implements a **bug fix** for the **xAI provider** to correctly extract **reasoning text** from non-streaming responses. Previously, the SDK would silently drop reasoning because it only read from the `summary` field, while xAI's API often returns this data in the `content` field for non-streaming calls. The `doGenerate` method in `xai-responses-language-model.ts` is now updated to check the `content` field if `summary` is empty, and the `xai-responses-api.ts` schema is extended to support this. This ensures that users of the xAI `generateText` function now receive complete reasoning output as intended.

This commit introduces a **new capability** to the **xAI responses API provider**, enabling users to specify a `reasoningSummary` option in their requests. This feature allows for greater control over how reasoning traces are presented in the API responses, complementing the existing `reasoningEffort` option. The change involves updating the `xai-responses-options` schema, modifying the `doGenerate` method in `xai-responses-language-model` to correctly map this new parameter, and adding comprehensive tests. New examples for `generateText` and `streamText` demonstrate its usage, enhancing the configurability of the xAI integration.

This commit introduces a **breaking change** to the **xAI provider**, making the Responses API the default for direct calls like `xai('model-id')` and `xai.languageModel('model-id')`. This **feature enhancement** aligns the xAI provider's behavior with the OpenAI provider and adheres to xAI's recommendation to use the Responses API over the legacy Chat API. While the **Chat API remains accessible** via `xai.chat('model-id')`, this change impacts existing integrations that relied on the Chat API as the default. **Documentation** and **examples** have been updated to reflect this new default and provide guidance for using the legacy Chat API.

This commit **improves the automated issue labeling rules** for the **Tigent bot** by updating its configuration in `.github/tigent.yml`. This **maintenance chore** addresses previous misclassifications, such as confusing `openai` and `openai-compatible` issues, and incorrectly applying protected labels. It **refines the issue classification logic** by adding a blocklist for sensitive labels, simplifying area labels, making `bug` and `support` mutually exclusive, and introducing new patterns for specific issue types. This change ensures **more accurate and consistent issue categorization** within the **GitHub issue tracking system**, streamlining the triage process for new issues.

This commit introduces **new capabilities** to the **xAI provider** by adding support for `logprobs` and `topLogprobs` parameters in both chat and responses API calls. It extends the provider options for `xai-chat-language-model` and `xai-responses-language-model` to expose these settings, mapping them to the underlying xAI API requests and automatically enabling `logprobs` when `topLogprobs` is specified. This **feature enhancement** allows users to retrieve detailed token probability information, significantly improving the analytical capabilities and control over model outputs within the **xAI integration**. The change includes comprehensive **documentation** updates, new **examples** demonstrating usage with `generateText` and `streamText`, and robust **test coverage** to ensure correct parameter forwarding.

This commit performs extensive **documentation maintenance** by **fixing numerous stale internal links and correcting various typos** across the project's guides and reference materials. It primarily updates paths from `/docs/guides` to `/cookbook/guides` in many **Cookbook guides** and **Getting Started documentation**, ensuring correct navigation for topics like streaming, RAG, and multi-modal chatbots. Additionally, it addresses specific link issues in **AI SDK Core**, **AI SDK UI**, and **AI SDK RSC** sections, alongside general textual typos in **provider documentation** for Bedrock, Google Generative AI, and observability tools. This work significantly enhances the **accuracy and usability of the documentation**, improving the overall developer experience by preventing broken links and clarifying content.

This commit **adds new example files** to demonstrate the usage of the recently integrated `gpt-5.3-codex` OpenAI model. It provides specific examples within the **`ai-functions` module** for both **`generateText` and `streamText`** functionalities. This **new capability** ensures comprehensive example coverage for the new model, making it easier for developers to understand and adopt `gpt-5.3-codex` in their applications. These examples are crucial for showcasing how to interact with the model for both direct text generation and streaming responses, enhancing the overall developer experience.

This commit provides a **documentation update** for the **AI Functions examples** within the `AGENTS.md` file. It **clarifies the new folder layout** for these examples, detailing the expected pattern `examples/ai-functions/src/<function>/<provider>/` and identifying `basic.ts` as the standard provider entry file. Additionally, the commit updates the example command for running these functions and explicitly advises against creating flat top-level provider files. This **maintenance** work improves the clarity and accuracy of the **agent documentation**, ensuring developers understand the current structure and best practices for contributing to AI function examples.

This commit performs **maintenance** on the **`packages/openai`** module's test infrastructure by **restoring stable web-search fixtures and snapshots** that were unintentionally altered in a previous change. Specifically, it reverts `openai-web-search-tool.1.json` and `openai-responses-language-model.test.ts.snap` to their pre-`#12706` state, reducing test noise and improving the signal for future changes to web-search action mapping. Additionally, it updates the OpenAI model used in the `examples/ai-functions/src/generate-text/openai-web-search-tool.ts` and `examples/ai-functions/src/stream-text/openai-web-search-tool.ts` examples from `gpt-5.1` to `gpt-5-mini`. This ensures test stability and clarity while keeping the runtime fix from the prior commit.

This commit performs a **maintenance chore** by **updating the Tigent bot's labeling rules** within `.github/tigent.yml` to improve issue classification. It refines the criteria for the `ai/core` area label, explicitly referencing `Output` and noting `generateObject`/`streamObject` as deprecated. Additionally, the rules for `reproduction provided` are made stricter, requiring actual code snippets, and the bot is now prevented from automatically applying `pull request welcome` and `good first issue` labels. This ensures **more accurate and consistent issue categorization** by reserving certain labels for human judgment and preventing conflicting labels like `bug` and `support` from being applied simultaneously.

This commit **introduces a new configuration file**, `.github/tigent.yml`, to define the operational parameters for the **`tigent` bot**. This **configuration setup** specifies the bot's prompt and crucial labeling instructions, establishing how it will interact with and categorize project items. The addition of this file is a **maintenance** step that lays the groundwork for the `tigent` bot's automated workflows within the project's GitHub environment.

This commit **fixes invalid MDX heading ID syntax** within the **documentation** for the `custom-memory-tool` guide. Specifically, it removes an explicit, incorrectly formatted heading ID from the 'Appendix: Helper Functions' section in `content/cookbook/00-guides/07-custom-memory-tool.mdx`. This **documentation fix** ensures proper rendering and parsing of the guide, preventing potential issues with heading anchors or build processes. It is a follow-up to a previous related change, improving the overall quality and consistency of the project's documentation.

This commit **refactors** the **Rev.ai provider's transcription model tests** to improve their maintainability and realism. It migrates tests from using inline JSON generation to a **fixture-based pattern**, introducing three new JSON fixtures that accurately represent real Rev.ai API responses for job submission, status, and transcripts. The tests now utilize `prepareJsonFixtureResponse` to load these fixtures and employ `toMatchSnapshot()` for robust response assertions, making them more resilient to API changes and easier to understand. This change significantly enhances the reliability and clarity of the **Rev.ai integration tests**.

This commit **refactors** the **Fal provider's transcription tests** to adopt a more robust and maintainable **fixture-based pattern**. It replaces the previous inline JSON response generation with actual API response data stored in `fal-transcription-queue.json` and `fal-transcription-result.json` fixtures. This change improves the realism and reliability of the **`fal` transcription model tests** by using `prepareJsonFixtureResponse` and `toMatchSnapshot()` assertions. The tests within `fal-transcription-model.test.ts` are now better organized into descriptive blocks, enhancing clarity and maintainability. This **refactoring** ensures that the integration tests more accurately reflect real API interactions, making future test development and debugging more efficient.

This commit **refactors** the **ElevenLabs provider's transcription tests** by migrating them to a **fixture pattern**. It replaces inline JSON response preparation with `prepareJsonFixtureResponse`, which now reads from a newly recorded `elevenlabs-transcription.json` fixture, enhancing test realism and maintainability. The tests are also reorganized into `describe` blocks and updated to use `toMatchSnapshot()` for assertions, improving clarity and robustness for the **`elevenlabs-transcription-model.test.ts`** file. This **test maintenance** effort improves the reliability of the **ElevenLabs transcription model** by using more accurate API response simulations.