Add ToolErrorRecoveryMiddleware as the outermost agent middleware so any
tool-call exception (notably MCP ToolException) is converted into a
ToolMessage with status="error" carrying the raw error text. The agent
can then loop once more and reply to the user in natural language about
what failed, instead of bubbling the exception up through agent.astream
and breaking the SSE response in routes/chat.py.
The recovery layer extracts the inner `text="..."` payload out of the MCP
TextContent repr when present, falling back to str(error) otherwise. It
deliberately re-raises asyncio.CancelledError so task cancellation still
propagates, and sits *outside* ToolMetricsMiddleware so the existing
status=error metric is still emitted before recovery kicks in.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
06974e9744