MCP reload reuses startup workspace config after .github/mcp.json changes

[zoherghadyali]

Describe the bug

Copilot CLI keeps using the MCP configuration snapshot loaded when the session started. If a workspace MCP server fails to initialize and .github/mcp.json is corrected while the session remains open, reloading/restarting MCP servers retries the old command instead of reading the updated file.

The on-disk file contains the corrected command, but reload diagnostics report that the configuration signature is unchanged and launch the original command again.

Affected version

• Copilot CLI: 1.0.81-7
• Platform: Windows
• MCP transport: stdio

Steps to reproduce

1. Create .github/mcp.json in a workspace with a deliberately observable failing server:

   {
     "mcpServers": {
       "config-probe": {
         "type": "stdio",
         "command": "python",
         "args": [
           "-c",
           "import sys; print('OLD_CONFIG', file=sys.stderr)"
         ],
         "tools": ["*"]
       }
     }
   }

2. Start Copilot CLI in that workspace. The server exits during initialization and its stderr contains OLD_CONFIG.

3. Without exiting Copilot CLI, change only OLD_CONFIG to NEW_CONFIG in .github/mcp.json.

4. Open /mcp and reload/restart the MCP servers.

5. Inspect the MCP status and session log.

Expected behavior

Reload should reread .github/mcp.json, compute a new configuration signature, and launch the command containing NEW_CONFIG.

Actual behavior

Reload keeps the startup signature and launches the original command containing OLD_CONFIG. In the observed session, the equivalent diagnostics were:

mcp discover_and_start_root {
  "signature_matches": true,
  "previous_signature": "...old command...",
  "signature": "...old command..."
}
mcp unchanged reload reuses live graph
MCP server stderr: ...old command output...
Recorded failure for server config-probe

The file on disk already contained the new command when these messages were recorded.

This also occurs when replacing a nonexistent relative launcher path with a valid absolute path and adding an environment override: subsequent reloads continue retrying the original relative path.

Impact

A user cannot repair a failed workspace MCP server without terminating and relaunching Copilot CLI. The reload/restart action appears to run, but it retries stale settings, which makes configuration troubleshooting confusing.

Workaround

Exit Copilot CLI completely and start a new process after editing the workspace MCP configuration.

Related issues

• .mcp.json workspace config not loaded at startup — requires /mcp reload #2198 covered workspace MCP configuration not loading at startup; that issue was fixed by making reload discover the config. This report is the inverse: startup discovers the config, but reload does not observe later edits.
• Workspace .mcp.json detected by 'mcp list'/'mcp get' but not connected in actual agent session (interactive/-i/-p) #4542 covers workspace configuration being detected but not wired into newly started sessions. This report concerns in-session configuration refresh after the server was already discovered.