Enable MCP graceful error handling by default and retrieve background session context task exceptions. This prevents unhandled AnyIO TaskGroup transport failures from bubbling through the event loop and abruptly terminating live streaming sessions with 1006 abnormal closure.
Change-Id: Ib4669fb57cddfc6ce61f9a3991b80db3f2b3e9ab
In Gemini 3.1 live, LlmResponse packets sometimes contain only grounding metadata - don't skip these responses
Change-Id: I9e2c1080395588d61aefffd1b6570102930e1fe0
- Claude Opus 4.7 rejects `thinking.type: "enabled"` with a 400 error
and requires `thinking.type: "adaptive"`; adaptive is also the
recommended mode for Opus 4.6 / Sonnet 4.6 where `"enabled"` is
deprecated.
- The genai `ThinkingConfig.thinking_budget = -1` (AUTOMATIC) sentinel
is the natural mapping for adaptive: the model picks the depth
itself. Callers opt in with `-1`; positive values keep the existing
`"enabled"` path so older models are unchanged.
- Bump `anthropic>=0.78` in `pyproject.toml`; 0.78 introduced
`ThinkingConfigAdaptiveParam`. The previous floor (`>=0.43`) let
Kokoro install `anthropic==0.75.0`, which crashes the new tests
with `AttributeError: module 'anthropic.types' has no attribute
'ThinkingConfigAdaptiveParam'`.
PiperOrigin-RevId: 917449729
* Implements live inference in evaluation_generator.py using Runner.run_live().
* Updates base_eval_service.py, local_eval_service.py to support live mode data structures and connection handling.
Testing: Added unit tests:
* test_generates_inferences_with_user_simulator_live
* test_live_session_manually_triggers_callbacks
* test_live_session_manually_triggers_callbacks_with_tools
* test_perform_inference_with_use_live
* test_perform_inference_single_eval_item_live
* test_perform_inference_single_eval_item_non_live
PiperOrigin-RevId: 916231029
Adds @functools.lru_cache to find_context_parameter so the inspect.signature
+ typing.get_type_hints lookup runs once per function, not on every MCP
confirmation callback or declaration build. No public surface change.
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 916204929
AgentTool.run_async only extracted text parts from the inner agent's
response, silently dropping code_execution_result.output and
executable_code.code. Outer agents using an inner agent with a code
executor saw nothing.
Close#5481
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 916196604
Previously, an empty `candidates` list without `prompt_feedback` resulted in an `UNKNOWN_ERROR`. This change updates the logic to handle such cases as a successful completion with no generated content, which is valid for certain model interactions like tool-driven turns.
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 916115022
`part_to_message_block` iterated `content` char-by-char when a tool
returned it as a plain string (e.g. `LoadSkillResourceTool`'s
`{"content": <file text>}`), producing `"H\ne\nl\nl\no"` instead of
`"Hello"`. Guard the list branch with `isinstance(..., list)` and add
a sibling branch that passes a scalar string through directly, matching
Anthropic's `content: str | list[ContentBlockParam]` shape.
Close#5358
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 916109239
This change introduces logic to identify events containing requests for tool confirmation or auth credentials. The compaction process will now stop before any such "Human-in-the-Loop" (HITL) events, ensuring that the full context of the interaction is preserved and not summarized away. This applies to both sliding window and token threshold compaction strategies.
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 916108771
When judge_model_config is None, LlmRequest raises a ValidationError
because it requires a config. We now construct a default GenerateContentConfig
if one isn't provided.
Close#5677
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 916087055
This CL implements the class in the integrations folder, used specifically for the Skill Registry API.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 915627057
This change enables the Google Cloud Telemetry exporter to use mTLS endpoints. It checks for the availability of client certificates and respects the GOOGLE_API_USE_CLIENT_CERTIFICATE environment variables to determine whether to use the mTLS-specific endpoint and configure the session accordingly.
PiperOrigin-RevId: 915541335
This change removes support for the deprecated `--session_db_url`, `--artifact_storage_uri`, and `--verbosity` flags from the ADK CLI. It also simplifies the service URI handling in `cli_deploy.py` by always using the new `--session_service_uri`, `--artifact_service_uri`, and `--memory_service_uri` flags, regardless of the ADK version.
The deprecated flags has been more than 1 year
Co-authored-by: Shangjie Chen <deanchen@google.com>
PiperOrigin-RevId: 915100881
This CL introduces logic to handle scenarios where a non-ADK agent transitions to the `TaskState.input_required` or `TaskState.auth_required` states. It intercepts these events and converts them into a synthetic ADK `FunctionCall` event.
PiperOrigin-RevId: 914877123
When resuming a session with Anthropic (Claude) models, tool_use IDs were
getting stripped during content replay, causing BadRequestError from the
Anthropic API. Two fixes: (1) preserve function call IDs in the contents
flow for AnthropicLlm models, (2) sanitize invalid/empty tool_use IDs
to match Anthropic's required pattern (^[a-zA-Z0-9_-]+$) instead of
passing empty strings.
Fixes#5074
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 914550905
Merge https://github.com/google/adk-python/pull/5597
**Please ensure you have read the [contribution guide](https://github.com/google/adk-python/blob/main/CONTRIBUTING.md) before creating a pull request.**
### Link to Issue or Description of Change
**1. Link to an existing issue (if applicable):**
- Related: https://github.com/google/adk-python/issues/4971
**2. Or, if no issue exists, describe the change:**
_If applicable, please follow the issue templates to provide as much detail as
possible._
**Problem:**
A2A support lacked pluggable or persistent task store backends, forcing a strict default to `InMemoryTaskStore`.
**Solution:**
Extended `ServiceRegistry` and `ServiceFactory` to support URI-driven configuration for A2A task stores. Added built-in support for `memory://`, `postgresql://`, `mysql://`, and `sqlite://` schemes. Plumbed the options down into `to_a2a()` and `get_fast_api_app()`, ensuring connection strings are securely redacted from application logs.
### Testing Plan
**Unit Tests:**
- [x] I have added or updated unit tests for my change.
- [x] All unit tests pass locally.
Summary: Ran `pytest` against `test_agent_to_a2a.py`, `test_fast_api.py`, `test_service_registry.py`, and `test_service_factory.py`. All 131 unit tests passed successfully.
**Manual End-to-End (E2E) Tests:**
Compiled the package distribution wheel via `uv build` and successfully validated its installation inside a clean isolated sandbox environment via `uv pip install dist/google_adk-1.32.0-py3-none-any.whl[a2a]`. Verified that the application instantiates default in-memory stores and custom persistent URI stores
### 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/5597 from allen-stephen:feat/a2a-task-store-support f53ba0fa35fa3c28e4923b6850a5873fd5293525
PiperOrigin-RevId: 914496800
Liang Wu