Kazuhiro Sera

This commit delivers a **bug fix** and **refactoring** for the **AnyLLM responses** processing, specifically addressing issues with replayed input. It updates the `_sanitize_any_llm_responses_input` and `_sanitize_any_llm_responses_value` functions within `src/agents/extensions/models/any_llm_model.py` to ensure proper sanitization occurs *before* validation. The enhanced logic now correctly strips incompatible fields, `FAKE_RESPONSES_ID`, `provider_data`, and `None` values from replayed `AnyLLM` input. This change significantly improves the **robustness and correctness** of `AnyLLM` response handling by preventing validation errors caused by extraneous data. New tests have been added in `tests/models/test_any_llm_model.py` to verify the accurate stripping of these elements.

This commit **refactors** the **automated PR labeling system** to correctly categorize changes related to extensions. It **replaces the deprecated `feature:lite-llm` label with `feature:extensions`** across the project's `.github` workflow. This **maintenance** update modifies the `pr_labels.py` script, specifically the `infer_specific_feature_labels` function, along with the PR label schema and documentation, to ensure accurate mapping. Consequently, all future pull requests concerning **extensions** will be correctly identified and labeled, improving **workflow automation** and **code organization**.

This commit performs **documentation maintenance** by updating the `.changeset/fresh-rabbits-hunt.md` file. The update reflects the prior removal of flaky test coverage within the **`agents-core`** module, specifically for streamable HTTP reconnects. This ensures that the project's **release documentation** accurately records a **test improvement** aimed at enhancing the reliability of the test suite.

This commit **stabilizes** the **test suite** for **concurrent streamable HTTP reconnects** within the `agents-core` package. It **refactors** the `streamableHttpRetry.test.ts` file, specifically the 'coalesces concurrent reconnects' test, by replacing custom promise tracking utilities with `vi.waitFor`. This **test improvement** enhances the reliability and clarity of existing coverage for the HTTP retry mechanism. The change ensures more consistent and accurate validation of how `agents-core` handles concurrent streamable HTTP retries.

This commit provides a **bug fix** for the **`server conversation tracker`** by resolving issues with stale hydrated input IDs. It **removes the problematic reliance on `id(item)`** for tracking sent items within the `hydrate_from_state` function in `src/agents/run_internal/oai_conversation.py`, which previously caused object identity problems after state hydration. This ensures that the conversation tracker correctly identifies inputs regardless of their hydration state, improving the reliability of conversation flow. New **test cases** have been added to `tests/test_server_conversation_tracker.py` to validate that `hydrate_from_state` no longer tracks initial input by object identity for both string and list inputs.

This commit **fixes** an issue by **enabling the OpenAI chat completions input** to correctly accept raw `image_url` content parts. It introduces a **new normalization method** in the `src/agents/models/chatcmpl_converter.py` module, specifically `_normalize_input_content_part_alias`, to properly process these raw inputs. This logic is then integrated into the `_get_messages` function of the `src/agents/models/openai_chatcompletions.py` model, allowing it to directly handle `image_url` content. This **enhancement** improves the flexibility and compatibility of the **chat completions input parsing** for multimodal scenarios.

This commit **improves test stability** within the `agents-core` package by **refactoring** the `waitForBarrier` utility in `packages/agents-core/test/shims/mcp-server/streamableHttpRetry.test.ts`. It introduces a new `trackPromiseState` utility, shifting from `Promise.race` to polling promise states for more reliable asynchronous test synchronization. This **maintenance chore** specifically targets the **test infrastructure** for `shims/mcp-server` functionalities, ensuring more consistent and less flaky test outcomes across the project.

This commit **fixes** an issue within the **`@langchain/openai-agents`** package where the `external_web_access` property for **web search tools** was not correctly forwarded during tool conversion. It **enhances** the `WebSearchTool` type definition and updates the `webSearchTool` and `converTool` functions to properly handle and preserve this setting. This ensures that the `external_web_access` configuration, which dictates whether a web search tool can access external websites, is accurately respected. Consequently, **OpenAI agents** utilizing web search tools will now correctly enforce or allow external web access as intended by the developer.

This commit introduces a **new automated code formatting workflow** for changed TypeScript files within the **`agents-core` and `agents-realtime` packages**. It establishes a **`codex` hook** (`stopRepoTidy.mjs`) that leverages a new `prettier-changed.mjs` script to automatically apply Prettier formatting. This **maintenance chore** significantly enhances code consistency and readability by ensuring that modified files adhere to defined style guidelines. The new `codex` configuration and utility modules (`gitState.mjs`, `hookState.mjs`, `stopTidyPolicy.mjs`) provide the infrastructure for this automated process. This improves the overall developer experience and code quality by integrating style enforcement directly into the development workflow.

This commit introduces a **new automated code quality enforcement mechanism** by integrating a **Codex Stop hook** for targeted Ruff tidying. It adds configuration files (`.codex/config.toml`, `.codex/hooks.json`) to enable Codex hooks and define a `Stop` hook that executes a new Python script, `stop_repo_tidy.py`. This script automatically formats and fixes linting issues in **modified Python files** using **Ruff**, ensuring consistent code style and reducing manual effort in the developer workflow. This **developer tooling enhancement** improves overall code quality and maintainability.

This commit performs a **documentation update** to align code examples with recent changes, specifically addressing issue #2792. It **refactors** Python code snippets within the **models documentation** by renaming the `provider` variable to `client`. This consistency change affects `docs/models/index.md` across **English, Japanese, Korean, and Chinese language versions**, ensuring that users viewing these examples receive accurate and up-to-date information. This is a **maintenance** task focused solely on improving the clarity and correctness of the **documentation**.

This commit **fixes** an issue by **refactoring the project's docstring parsing mechanism to directly depend on `griffelib`**, replacing the previous reliance on `griffe`. This change primarily impacts the **`agents` module**, specifically `src/agents/function_schema.py`, where import statements are updated to reflect the new dependency. The `uv.lock` file is also updated to manage this dependency switch and refresh related documentation tools like `mkdocstrings`. This ensures more robust and direct docstring processing, particularly for generating accurate function schemas.

This commit **improves the reliability of a flaky test** within the **`realtime` module's testing suite**, specifically for **OpenAI integration** and **handshake timeout configuration**. It **refactors an existing test** in `tests/realtime/test_openai_realtime.py` to use `websockets.serve` for better stability and **adds a new test** to further validate handshake timeout behavior. This **maintenance work** ensures the continuous integration pipeline accurately reflects the module's functionality without intermittent test failures.

This commit **updates the project documentation** by refining the `README.md` file. It specifically **enhances the acknowledgements** for open-source projects and introduces new sections detailing **optional dependencies** and **project management tools**. This **documentation update** improves the clarity and completeness of the project's public-facing information, making it easier for users and contributors to understand its ecosystem and dependencies.

This commit **fixes** an issue where private tool metadata was inadvertently persisted within session items, addressing #2776. It introduces robust **data sanitization** across the **agent's session management** and **input item processing** to ensure internal SDK-specific tool call metadata is stripped. New helper functions like `strip_internal_input_item_metadata` and `_normalize_compaction_session_items` are now applied during session item addition, compaction, saving, and key generation within `src/agents/memory/openai_responses_compaction_session.py` and `src/agents/run_internal/session_persistence.py`. This prevents SDK-only details from being stored in **persisted session data** or exposed to external APIs, improving data cleanliness and preventing potential integrity issues.

This commit delivers a **bug fix** to ensure the `reasoning_effort` parameter is handled consistently and portably across different **LiteLLM providers**. It **refactors** the `src/agents/extensions/models/litellm_model.py` module by extracting the resolution logic for `reasoning_effort` into a new private method, `_get_reasoning_effort`, which is then used by `get_response`. Additionally, the **testing suite** is enhanced with a parameterized test to verify that `reasoning.summary` is correctly ignored and a warning is logged when `reasoning_effort` is passed. This improves the robustness of **AI agent interactions** by standardizing how reasoning effort is communicated to various underlying language models, preventing unexpected behavior.

This commit performs a **maintenance** update by adjusting the version constraint for the `litellm` dependency. It specifically targets the project's **dependency lock file**, `uv.lock`, to ensure a consistent and potentially newer version of this library is used. This update helps maintain the project's dependencies and ensures compatibility or access to the latest features/fixes from `litellm`.

This commit **fixes a test flakiness issue** within the **`agents-core` package's test suite**. It specifically addresses **false positives in the reconnect barrier check** within the `streamableHttpRetry.test.ts` for the `mcp-server` shim. The **test is refactored** to remove a deferred promise and now utilizes `vi.waitFor`, which prevents spurious failures. This change significantly improves the **reliability and accuracy of the `streamableHttpRetry` test**, ensuring more trustworthy continuous integration results.

This commit **hardens** the **streamable HTTP reconnect tests** within the **`agents-core` package** by introducing new helper functions. Specifically, `formatError`, `withTimeout`, and `waitForBarrier` are added to `streamableHttpRetry.test.ts` to improve test robustness. These additions incorporate crucial timeout and early settlement checks, which **fix** previous test flakiness and ensure reliable validation of the HTTP retry mechanism. The overall impact is a more stable and dependable test suite for this critical networking component.

This commit **upgrades the pnpm version** from `10.32.1` to `10.33.0` across all **GitHub Actions CI/CD workflows**. This **maintenance chore** ensures that automated processes like `changeset` generation, `docs` building, `release` publishing, and `test` execution utilize the **latest stable version of pnpm**. The update affects the `.github/workflows/` configurations, providing potential stability and performance improvements for the project's continuous integration and deployment pipelines by leveraging the updated package manager.