v0.2.2

Fixed

• drf-mcp tools are now actually executed in-process. DrfMcpToolset
  extended ExternalToolset, whose tools are kind="external" — Pydantic-AI
  defers those: it yields the call to the client and ends the run, never
  invoking the toolset's call_tool. So drf-mcp tool calls were handed off and
  silently dropped (no TOOL_CALL_RESULT, the model never continued, and an
  AG-UI client's pending indicator would hang). The toolset now advertises its
  tools as kind="function", so the run loop runs them via the per-user
  MCPCallContext and streams a real TOOL_CALL_RESULT. Regression test drives
  a full agent run, not just a direct call_tool.