Ben Kazemi

This commit introduces **new capabilities** for the **Vertex AI `AgentEngine` integration**, enhancing its data model and processing logic. It adds a `memoryId` field to `AgentEngine` memory configurations, enabling more granular memory management, and a `rawEvent` field to `AgentEngine` session events for richer event logging. These changes involve updating the `vertexai/genai` package, specifically modifying `memories.go` and `sessionevents.go` to process these new fields, and extending the core data types in `types/types.go` with `MemoryID` and `RawEvent`. This allows the `AgentEngine` to handle more comprehensive state and event information, improving its overall observability and control.

This commit introduces a **new capability** to the **Vertex AI GenAI client library** by adding a `SessionEvents` field to the `Sessions` struct within `vertexai/genai/sessions.go`. This **feature enhancement** allows the `Sessions` struct to store or manage session-related events, thereby expanding the data model for interactive session handling. The addition enables more comprehensive tracking and management of session lifecycle events within the **Vertex AI module**.

This commit **introduces new capabilities** to the **Vertex AI Go SDK**, primarily enhancing the **`genai` package** by adding session management for agent engines. A new `Sessions` field is added to the `AgentEngines` struct in `vertexai/genai/agentengines.go`, enabling stateful interactions with **Vertex AI agent engines**. Furthermore, the `EventMetadata` type in `vertexai/genai/types/types.go` is extended with `InputTranscription` and `OutputTranscription` fields, alongside new types for agent engine query job configuration and results, providing richer event logging. This work facilitates more robust and detailed interactions with **Vertex AI agent engines** by allowing for better state management and comprehensive event data capture. A minor documentation reformat is also included in `vertexai/README.md`.

This commit performs a **version bump** for the **`vertexai` library**, officially releasing version `v0.17.0`. It updates the library's version in the `.librarian/state.yaml` configuration and sets the internal `Version` constant within `vertexai/internal/version.go` to reflect the new release. This is a **maintenance chore** that prepares the library for its latest iteration, making any new features, bug fixes, or improvements from `v0.17.0` available to users.

This commit performs a significant **refactoring** of the **Vertex AI Go SDK** by extracting shared data types into a **new, dedicated `types` package** within `vertexai/genai`. This change centralizes type definitions, improving code organization and maintainability across the SDK. Consequently, all client-facing functions and internal helper functions within modules like `agentengines`, `memories`, `sandboxes`, `sessionevents`, and `sessions` have been updated to utilize these newly consolidated types. This **maintenance** effort streamlines type management and enhances the overall clarity of the `vertexai/genai` module.

This commit performs a significant **code cleanup and refactoring** within the **Vertex AI Generative AI client library**, specifically targeting internal **Agent Engine** components. It streamlines numerous request parameter conversion functions across files like `agentengines.go`, `memories.go`, `sandboxes.go`, and others. The work involves removing redundant assignments of the `config` field and eliminating unused variable assignments, enhancing code clarity and maintainability. This is an **internal optimization** that improves the codebase without altering any external API behavior or functionality.

This commit **regenerates the documentation** for version 1.56.0, primarily updating the **Generative AI module's API reference**. It includes revised descriptions for `MultiSpeakerVoiceConfig`, `SpeechConfig`, `ThinkingConfig`, and `ThinkingLevel`, along with the addition of new enum values for `ThinkingLevel` and `PartMediaResolutionLevel`. This **maintenance** task also incorporates new definitions for `Icon` and `ToolExecution` objects. The updated documentation ensures that users have access to the most current information regarding available configurations and options within the Generative AI module, improving clarity and usability.

This commit introduces **new capabilities** to the **`google.genai` library** by extending the `ThinkingLevel` enum in `google/genai/types.py`. Specifically, it **adds `MINIMAL` and `MEDIUM` thinking levels**, along with their associated descriptions. This enhancement provides users with more granular control over specifying the desired cognitive complexity for AI model interactions, allowing for finer-tuned request configurations.

This commit introduces a **new capability** by adding `MEDIA_RESOLUTION_ULTRA_HIGH` to the `PartMediaResolutionLevel` enum within the **`google.genai.types` module**. This enhancement allows users to specify an **ultra high resolution** for media content when defining parts, expanding the available options beyond existing resolution levels. It directly impacts how media resolution can be configured for generative AI models, enabling support for higher fidelity inputs.

This commit **introduces new `ThinkingLevel` options** within the **`google.genai` library's `types` module**. It **adds `MEDIUM` and `MINIMAL` members** to the `ThinkingLevel` enum, expanding the available configurations for AI model interactions. This **new feature** provides developers with more granular control over the verbosity or depth of AI processing, enabling finer tuning of model behavior. The changes are confined to `google/genai/types.py`, updating the enum definition, docstrings, and descriptions to reflect these new capabilities.

This commit provides a **test fix** by introducing a new utility to standardize dictionary key formats within the **testing infrastructure** for the `google.genai` library. A new helper function, `camel_to_snake_all_keys`, is added to `google/genai/tests/pytest_helper.py` to recursively convert all dictionary keys from camelCase to snake_case. This utility is then applied in `google/genai/tests/live/test_live.py` to ensure that assertions in **live tests** correctly compare data structures, resolving discrepancies in key naming conventions. This **maintenance** task enhances the reliability and accuracy of the **test suite**.

This commit **updates the test suite** for the **`google.genai` library's cached content functionality** to utilize the newer `gemini-2.5-flash` models. It involves **refactoring test parameters** across various cached content creation and retrieval tests, including `test_create.py` and `test_create_custom_url.py`, to specify the updated model identifiers and adjust MIME types. Additionally, test constants in `constants.py` and cached content names in `test_get.py` are updated to align with these new model configurations. This **maintenance** ensures that the caching mechanisms are thoroughly validated against the latest supported generative AI models.

This commit **fixes** a critical issue in the **`google.genai` API client** where **streaming responses** (Server-Sent Events) failed to properly parse and raise API errors embedded within the stream. It introduces robust error detection and handling for both synchronous and asynchronous streaming requests, ensuring that users receive actionable exceptions instead of malformed data. The **error handling subsystem** in `google/genai/errors.py` was refactored to centralize this logic, specifically for methods like `request_streamed` and `async_request`. New test cases were added to verify correct error handling, even with malformed error JSON, significantly improving the reliability and developer experience when interacting with streaming APIs.

This commit **updates the test suite** for the **`google.genai` library**, specifically modifying tests for the `generate_content` functionality. It **migrates these tests** from using the `gemini-1.5-pro` model to the newer `gemini-2.5-flash` model, ensuring that **content generation tests** are run against the latest available models. This **test update** improves the relevance and accuracy of the test coverage for the `generate_content` API. As part of this migration, cached content names within `test_generate_content_cached_content.py` are also updated to align with the new model.

This commit **fixes an omission** by adding crucial configuration fields to the **`google.genai` model types**. It introduces optional parameters such as `temperature`, `top_p`, `top_k`, and `thinking` to the `Model` class and `ModelDict` in `google/genai/types.py`, enabling **enhanced model configuration**. The `_Model_from_mldev` function in `google/genai/models.py` is updated to correctly transfer these new parameters during model conversion. This allows users to specify more granular control over model generation behavior, with related test cases in `google/genai/tests/models/test_get.py` updated to reflect new model versions.

This commit introduces a **new phish filtering feature** within the **`google.genai` library**, specifically for the **`GoogleSearch` and `EnterpriseWebSearch` tools**. It adds a `blocking_confidence` parameter to these search types, allowing users to specify a threshold for blocking phishing content. This **new capability** is implemented by defining a `PhishBlockThreshold` enum in `google/genai/types.py` and integrating validation for the new parameter across internal conversion functions like `_GoogleSearch_to_mldev` in several core modules. The change enhances the content moderation capabilities of these search tools, with new test cases ensuring its proper functionality.

This commit introduces a **new capability** by adding `ADAPTER_SIZE_TWO` to the `AdapterSize` enum within the `google/genai/types.py` module. This enhancement specifically supports **Gemini 2.0 Tuning**, providing an additional configuration option for users to specify a different adapter size during the tuning process. It expands the flexibility and choices available for optimizing Gemini 2.0 models, allowing for more granular control over tuning configurations.