Describe the bug

While experimenting with developing MCP servers, I killed the STDIO MCP server process spawned by crush.

Once the tool was called, it printed an ERROR transport error: failed to write request: write |1: broken pipe.

The right sidebar still showed a green dot next to the MCP server and the agent continued to try to use the tool. No recovery occurred. Work here might dovetail with (#739) and extend to restart policy?

It is still using context with a dead tool description and doesn't display the server state.

Setup

MacOS, Claude, screentime_mcp

To Reproduce

1. Install a STDIO MCP server and configure crush to use
2. Run Crush and load the MCP server
3. observe success
4. ps -elf | grep <mcp_server and kill it, etc
5. try some tool use and see what happens

Source Code

N/A

Expected behavior

I would expect that upon an irrecoverable transport failure, that the MCP Server would have a red dot and would not be included in further tool use.