Local MCP tool call waits indefinitely after sampling.requested

Start by reproducing the hang with the documented Azure MCP configuration and Copilot CLI command, then trace the flow after the recorded sampling.requested event. Done means the local tool call returns a result or error, or stops after a bounded timeout instead of waiting indefinitely.

Affected version or release

• GitHub Copilot App: 1.1.22
• GitHub Copilot CLI: 1.0.84-5
• Azure MCP: 3.0.0-beta.44 and 2.0.5

Installation context

GitHub Copilot App and CLI on Windows, using a local MCP server over stdio. The behavior was reproduced in fresh, non-interactive CLI sessions.

What happened?

A GitHub Copilot CLI session waits indefinitely after a local MCP server sends a sampling/createMessage request during tool execution.

The MCP server connects and the tool begins execution. Copilot CLI records a sampling.requested event, after which no further tool result, error, timeout, or assistant response is observed.

The MCP server connects:

serverName: azure
status: connected

The tool execution starts:

tool.execution_start
toolName: azure-get_azure_bestpractices
mcpTransport: stdio

The following progress event is emitted:

tool.execution_progress
0.0: Learning about get_azure_bestpractices capabilities...

Copilot CLI then records:

sampling.requested
serverName: azure
mcpRequestId: 1

After this event, none of the following are observed:

• A tool result
• A tool error
• A timeout
• A final assistant response
• Process termination

The session continued waiting until it was terminated using an external 90-second process timeout.

Steps to reproduce

1. Configure Azure MCP as a local MCP server:

   {
     "mcpServers": {
       "azure": {
         "type": "local",
         "command": "npx",
         "args": [
           "-y",
           "@azure/mcp@3.0.0-beta.44",
           "server",
           "start"
         ],
         "tools": ["*"]
       }
     }
   }
   

2. Start a fresh, non-interactive Copilot CLI session:

   copilot `
     --prompt "Call the Azure MCP get_azure_bestpractices tool for resource coding-agent and action all. Return only whether the tool completed." `
     --allow-all-tools `
     --disable-builtin-mcps `
     --log-level debug `
     --output-format json `
     --no-ask-user
   

3. Observe the session output.

Expected behavior

The tool call should complete, return an error, or stop after a bounded timeout. The session should not wait indefinitely without a visible result or error.

Additional context

The same sequence was observed in separate fresh Copilot CLI sessions with:

• @azure/mcp@3.0.0-beta.44
• @azure/mcp@2.0.5

In both tests:

1. The MCP server connected.
2. The tool execution started.
3. A progress notification was emitted.
4. A sampling.requested event was recorded.
5. No subsequent result, error, timeout, or assistant response was observed within 90 seconds.

The active agent turn remains in progress without returning a response. The session must be stopped manually or terminated using an external process timeout.