* fix(core): catch all MCP tool execution errors instead of re-raising
MCP tools are external user-configured servers - any failure during
tool execution is expected and should be returned as (error_msg, False)
to the agent, not raised as an exception that hits Datadog as a 500.
Previously:
- base_client.py only caught McpError/ToolError, re-raised everything else
- fastmcp_client.py (both SSE and StreamableHTTP) always re-raised
Now all three execute_tool() methods catch all exceptions and return
the error message to the agent conversation. The agent handles tool
failures via the error message naturally.
This silences ~15 Datadog issue types including:
- fastmcp.exceptions.ToolError (validation, permissions)
- mcp.shared.exceptions.McpError (connection closed, credentials)
- httpx.HTTPStatusError (503 from Zapier, etc.)
- httpx.ConnectError, ReadTimeout, RemoteProtocolError
- requests.exceptions.ConnectionError
- builtins.ConnectionError
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
* fix(core): log unexpected MCP errors at warning level with traceback
Expected MCP errors (ToolError, McpError, httpx.*, ConnectionError, etc.)
log at info level. Anything else (e.g. TypeError, AttributeError from
our own code) logs at warning with exc_info=True so it still surfaces
in Datadog without crashing the request.
🐾 Generated with [Letta Code](https://letta.com)
Co-Authored-By: Letta <noreply@letta.com>
---------
Co-authored-by: Letta <noreply@letta.com>
ddcfeb26b1