fix(adapters): mcp_tool_interceptor returns raw str for redacted output, breaking langchain-mcp-adapters

[gzak]

Bug

When mcp_check_output returns redacted_data, the mcp_tool_interceptor in AxonFlowLangGraphAdapter returns the redacted string directly:

if output_check.redacted_data is not None:
    return output_check.redacted_data  # returns a raw str

langchain-mcp-adapters expects the interceptor to return an MCPToolCallResult (CallToolResult | ToolMessage | Command). When a plain str is returned instead, _convert_call_tool_result in langchain_mcp_adapters/tools.py attempts to access .content on the string, raising:

AttributeError: 'str' object has no attribute 'content'

Steps to Reproduce

1. Run the langgraph-sample with a query that triggers PII redaction on an MCP tool output (e.g. "tell me everything about alice and bob" — Alice's record contains an SSN).
2. The getUserInfo MCP tool returns "SSN: 123-54-8967. Favorite color: blue.".
3. mcp_check_output redacts the SSN and sets redacted_data.
4. The interceptor returns the redacted str, which langchain-mcp-adapters cannot process.

Expected Behavior

The interceptor should wrap the redacted string in a CallToolResult so the return type is always a valid MCPToolCallResult.

Fix

Wrap redacted_data in a CallToolResult:

from mcp.types import CallToolResult, TextContent

if output_check.redacted_data is not None:
    return CallToolResult(content=[TextContent(type="text", text=output_check.redacted_data)])