Merge https://github.com/google/adk-python/pull/4025
**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):**
- Closes:
- #3950
- #3731
- #3708
**2. Or, if no issue exists, describe the change:**
**Problem:**
- `ClientSession` of https://github.com/modelcontextprotocol/python-sdk uses AnyIO for async task management.
- AnyIO TaskGroup requires its start and close must happen in a same task.
- Since `McpSessionManager` does not create task per client, the client might be closed by different task, cause the error: `Attempted to exit cancel scope in a different task than it was entered in`.
**Solution:**
I Suggest 2 changes:
Handling the `ClientSession` in a single task
- To start and close `ClientSession` by the same task, we need to wrap the whole lifecycle of `ClientSession` to a single task.
- `SessionContext` wraps the initialization and disposal of `ClientSession` to a single task, ensures that the `ClientSession` will be handled only in a dedicated task.
Add timeout for `ClientSession`
- Since now we are using task per `ClientSession`, task should never be leaked.
- But `McpSessionManager` does not deliver timeout directly to `ClientSession` when the type is not STDIO.
- There is only timeout for `httpx` client when MCP type is SSE or StreamableHTTP.
- But the timeout applys only to `httpx` client, so if there is an issue in MCP client itself(e.g. https://github.com/modelcontextprotocol/python-sdk/issues/262), a tool call waits the result **FOREVER**!
- To overcome this issue, I propagated the `sse_read_timeout` to `ClientSession`.
- `timeout` is too short for timeout for tool call, since its default value is only 5s.
- `sse_read_timeout` is originally made for read timeout of SSE(default value of 5m or 300s), but actually most of SSE implementations from server (e.g. FastAPI, etc.) sends ping periodically(about 15s I assume), so in a normal circumstances this timeout is quite useless.
- If the server does not send ping, the timeout is equal to tool call timeout. Therefore, it would be appropriate to use `sse_read_timeout` as tool call timeout.
- Most of tool calls should finish within 5 minutes, and sse timeout is adjustable if not.
- If this change is not acceptable, we could make a dedicate parameter for tool call timeout(e.g. `tool_call_timeout`).
### Testing Plan
- Although this does not change the interface itself, it changes its own session management logics, some existing tests are no longer valid.
- I made changes to those tests, especially those of which validate session states(e.g. checking whether `initialize()` called).
- Since now session is encapsulated with `SessionContext`, we cannot validate the initialized state of the session in `TestMcpSessionManager`, should validate it at `TestSessionContext`.
- Added a simple test for reproducing the issue(`test_create_and_close_session_in_different_tasks`).
- Also made a test for the new component: `SessionContext`.
**Unit Tests:**
- [x] I have added or updated unit tests for my change.
- [x] All unit tests pass locally.
```plaintext
=================================================================================== 3689 passed, 1 skipped, 2205 warnings in 63.39s (0:01:03) ===================================================================================
```
**Manual End-to-End (E2E) Tests:**
_Please provide instructions on how to manually test your changes, including any
necessary setup or configuration. Please provide logs or screenshots to help
reviewers better understand the fix._
### 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.
- [ ] ~~Any dependent changes have been merged and published in downstream modules.~~ `no deps has been changed`
### Additional context
This PR is related to https://github.com/modelcontextprotocol/python-sdk/pull/1817 since it also fixes endless tool call awaiting.
Co-authored-by: Kathy Wu <wukathy@google.com>
COPYBARA_INTEGRATE_REVIEW=https://github.com/google/adk-python/pull/4025 from challenger71498:feat/task-based-mcp-session-manager f7f7cd0c9c96840361c30499d08c33a189f57d86
PiperOrigin-RevId: 856438147
There was an extra test_mcp_toolset in the tools/ directory with only one test; I moved it into the main file.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 856419611
The retry_on_errors decorator in mcp_session_manager.py now catches asyncio.CancelledError and re-raises it immediately, ensuring that cancellation requests are not suppressed or retried
Close#4009
Co-authored-by: George Weale <gweale@google.com>
PiperOrigin-RevId: 852468382
Retrying only on closed_resource error is not enough to be reliable for production environments due to the other network errors that may occur -- remote protocol error, read timeout, etc. We will update this to retry on all errors. Since it is only a one-time retry, it should not affect latency significantly. Fixes https://github.com/google/adk-python/issues/2561.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 834514264
Currently logic to check for a disconnected session only checks for certain headers but doesn't detect all cases, leading to situations where it tries to connect to a session that is down. This adds logic so that we ping the server to check if it is disconnected. Fixes https://github.com/google/adk-python/issues/3321.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 833487825
Currently logic to check for a disconnected session only checks for certain headers but doesn't detect all cases, leading to situations where it tries to connect to a session that is down. This adds logic so that we ping the server to check if it is disconnected. Fixes https://github.com/google/adk-python/issues/3321.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 832460068
Retrying only on closed_resource error is not enough to be reliable for production environments due to the other network errors that may occur -- remote protocol error, read timeout, etc. We will update this to retry on all errors. Since it is only a one-time retry, it should not affect latency significantly. Fixes https://github.com/google/adk-python/issues/2561.
PiperOrigin-RevId: 831153803
Retrying only on closed_resource error is not enough to be reliable for production environments due to the other network errors that may occur -- remote protocol error, read timeout, etc. We will update this to retry on all errors. Since it is only a one-time retry, it should not affect latency significantly. Fixes https://github.com/google/adk-python/issues/2561.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 831123085
Currently ADK web hangs and logs "AGSI callable returned without completing response" when the server is unreachable. To fix, set timeouts for connecting to server.
Co-authored-by: Kathy Wu <wukathy@google.com>
PiperOrigin-RevId: 828648928
The `BaseTool` expects the run_async to return a json-serializable object. By model_dump the McpTool result explicitly can allow what ADK runtime sees is identical to what is persisted in the session event list.
Before the change, runtime sees CallToolResult instance and Session persists its serialized dict.
https://github.com/modelcontextprotocol/python-sdk/blob/main/src/mcp/types.py#L916-L922
PiperOrigin-RevId: 822465432
Add a header_provider param which is a callable[ReadonlyContext, Dict[str, Any]] for users to build headers in MCPToolset
fix: https://github.com/google/adk-python/issues/3156
PiperOrigin-RevId: 820412372
