Problem: The test test_openapi_json_schema_accessible was failing because Pydantic couldn't generate a JSON schema for mcp.client.session.ClientSession, which is part of genai_types.ToolListUnion.
Root Cause: The AgentDetails model in src/google/adk/evaluation/app_details.py:37 had tool_declarations: genai_types.ToolListUnion, and ToolListUnion includes mcp.client.session.ClientSession which doesn't have Pydantic core schema support. This model is used in the FastAPI app through EvalCase → Invocation → AppDetails → AgentDetails, causing OpenAPI schema generation to fail.
Solution: Changed the type annotation from genai_types.ToolListUnion to list[Any] in two places:
AgentDetails.tool_declarations
_ToolDeclarations.tool_declarations
This allows Pydantic to generate the OpenAPI schema while maintaining runtime compatibility (the field still accepts the same values).
Co-authored-by: Xiang (Sean) Zhou <seanzhougoogle@google.com>
PiperOrigin-RevId: 863765002
Update existing toolsets to utilize the new toolset authentication framework. Key changes:
- McpToolset: Add _auth_config instance variable, _get_auth_headers()
method to build auth headers from exchanged credentials, and
get_auth_config() override. Auth headers are now included when
creating MCP sessions.
- OpenAPIToolset: Add _auth_config and get_auth_config() to expose
auth configuration to the framework.
- ApplicationIntegrationToolset: Add _auth_config and get_auth_config().
- APIHubToolset: Add _auth_config and get_auth_config().
When ADK resolves toolset auth before calling get_tools(), it populates exchanged_auth_credential on the auth_config. Toolsets can then use this credential when making authenticated requests.
Also update test fixtures in test_apihub_toolset.py to use real auth objects instead of mocks that fail pydantic validation.
Co-authored-by: Xiang (Sean) Zhou <seanzhougoogle@google.com>
PiperOrigin-RevId: 863764941
This change adds framework-level support for resolving toolset authentication before calling get_tools(). Key changes:
- Add _resolve_toolset_auth() method in BaseLlmFlow that iterates
through toolsets, checks for auth config, and resolves credentials
via CredentialManager before tool listing
- Add TOOLSET_AUTH_CREDENTIAL_ID_PREFIX constant for identifying
toolset auth requests
- Add skip logic in auth_preprocessor to not resume function calls
for toolset auth (they do not need it)
- Add get_auth_response() method to CallbackContext for retrieving
auth credentials from session state
- Update CredentialManager to accept CallbackContext instead of
requiring ToolContext
When a toolset needs authentication but credentials are not available, the flow yields an adk_request_credential event and interrupts the invocation, allowing the user to complete the OAuth flow before retrying.
Co-authored-by: Xiang (Sean) Zhou <seanzhougoogle@google.com>
PiperOrigin-RevId: 863543036
This seems a breaking change, but actually credential manager is used internally only and also it won't work if some one call it using callback context
Co-authored-by: Xiang (Sean) Zhou <seanzhougoogle@google.com>
PiperOrigin-RevId: 863534476
The ADK web server's /run_sse endpoint now yields a JSON object like {"error": "..."} when an exception occurs during event generation. The adk_web_server_client is updated to detect this error payload and raise a RuntimeError.
Close#4291
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 863475838
The `_part_has_payload` helper now returns False for parts containing a function response, preventing the addition of fallback user content when a user turn consists solely of tool outputs.
Close#4249
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 863319497
For providers that typically require file IDs (like OpenAI and Azure), if a file URI is an HTTP/HTTPS URL and the MIME type is image, video, or audio, convert it to the corresponding URL-based content type (e.g., "image_url") instead of using the generic "file" type
Close#4112
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 863311703
This change introduces a new function, _validate_agent_import, which attempts to import the user's agent.py file before the deployment process begins. This helps catch common issues such as missing dependencies, incorrect relative imports, or syntax errors in the agent code early, providing more informative error messages to the user.
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 863310033
This CL enhances asyncio event loop management and test isolation.
- **BigQuery Analytics Plugin:** Ensure the asyncio event loop is consistently closed within the BigQuery analytics plugin. This prevents potential resource leaks. Add checks to handle potential deadlocks in Python 3.13+ when creating loops during interpreter shutdown.
- **Test Thread Pool Cleanup:** Introduce a pytest fixture (`cleanup_thread_pools`) to automatically shut down and clear all tool-related thread pools after each test run in `test_functions_thread_pool.py`. This improves test isolation and prevents order-dependent test failures.
- **Streaming Test Loop Restoration:** Refactor event loop handling in `test_streaming.py`. A new `_run_with_loop` method is introduced in the custom test runners to create a temporary event loop for each test execution, run the coroutine, and crucially, restore the original event loop afterwards. This prevents tests from interfering with each other's loop state.
- **Resource Closure:** Ensure services are closed properly in tests by adding `await service.close()` in `test_service_factory.py` and using `async with session_service` in `test_session_service.py`.
PiperOrigin-RevId: 863305565
Improves test reliability by guaranteeing the BigQueryAgentAnalyticsPlugin is shut down after each test execution. This is achieved by wrapping test logic in try/finally blocks, using a yield fixture, or employing a new asynccontextmanager to ensure the plugin.shutdown() method is always called, preventing potential resource leaks between tests.
PiperOrigin-RevId: 862951816
Since we will save MCP resource contents to the artifact service, no need to do post-processing based on mime type. Also fixed the implementation; the correct method to call is session.read_resource(uri).
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 862885513
Merge https://github.com/google/adk-python/pull/4276
This adds 2 mypy GitHub Actions checks:
* `mypy.yml` runs mypy on every PR. This will list all errors, so it shouldn't block until errors are resolved.
* `mypy-new-errors.yml` compares the PR branch to `main` to ensure no new mypy errors are introduced. This check can be made blocking now and can be removed once all errors are resolved in favor of `mypy.yml`.
### Checklist
- [x] I have read the [CONTRIBUTING.md](https://github.com/google/adk-python/blob/main/CONTRIBUTING.md) document.
- [x] I have performed a self-review of my own code.
- [x] I have commented my code, particularly in hard-to-understand areas.
- [x] I have added tests that prove my fix is effective or that my feature works.
- [x] New and existing unit tests pass locally with my changes.
- [x] I have manually tested my changes end-to-end.
- [x] Any dependent changes have been merged and published in downstream modules.
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/4276 from sararob:mypy-gh-actions 63130773c4862e90e352d4cb79221d78a6b9f5fd
PiperOrigin-RevId: 862226296
This allows users to access MCP resources on their own within agent logic / using custom tools. I plan on also later adding it to the agent state.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 861910520
* The main fix is in eval_metrics.py
* Removed some deprecated paths in the FastAPI server
* Added a unit test to catch future breakages
* Bump fastapi version to be 0.124.1 to capture the fix in https://github.com/fastapi/fastapi/pull/14482
* Removed the upper-bound restriction on fastapi version which was used to temporarily fix the issue
Fixes#3173
Co-authored-by: Liang Wu <wuliang@google.com>
PiperOrigin-RevId: 861867708
Details:
* The Sampler allows ADK agent optimizers to request evals and receive detailed eval results that can be used to guide agent optimization.
* This interface allows developers to run custom evaluations if needed for their agent. An implementation to interface with the inbuilt LocalEvalService shall be published shortly in a follow-up PR.
* The AgentOptimizer interface describes the general framework for an ADK agent optimizer that uses the Sampler interface to optimize an ADK agent.
* data_types.py contains generic types and base classes to allow intercommunication between the AgentOptimizer, Sampler, and developer code.
Co-authored-by: Keyur Joshi <keyurj@google.com>
PiperOrigin-RevId: 861849691
