[BUG] MCP servers connect successfully then get torn down with no reconnect on resumed sessions

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Bug: MCP servers connect successfully then get torn down with no reconnect, specific to resumed sessions

Environment

• Claude Code VS Code extension v2.1.220 (win32-x64)
• Windows 11 Pro 10.0.26200
• MCP servers affected: glsuite, mssql (both custom stdio Python servers defined in project .mcp.json) — but the pattern looks server-agnostic; it happened to both simultaneously.

Summary

On a resumed chat session, MCP servers connect successfully and report tools available, then are deliberately torn down by the extension ~3 seconds later. The log says it clears the connection cache "for reconnection," but no reconnection attempt is ever logged again for the rest of the session — leaving all MCP-provided tools permanently unavailable (ToolSearch reports "none found" for any MCP tool name) until the user starts a brand-new (non-resumed) chat.

Suspected mechanism

On startup, the extension spawns two Claude child processes: one throwaway instance used only to warm the config cache ("Loading config cache by launching Claude (no channel)...", resume: undefined), and one for the actual resumed session (resume: <session-id>). Both independently start and successfully connect to the same MCP servers. Shortly after, the throwaway instance's cleanup path issues "Terminating MCP server process tree" — but this appears to tear down (or otherwise invalidate) the connection state needed by the real resumed session too, and the subsequent "reconnection" it logs intending to do never fires.

A brand-new (non-resumed) chat does not exhibit the dual-spawn + teardown pattern and connects normally, which points at the resume code path specifically (likely the interaction between the config-cache warm-up spawn and the resumed-session spawn) rather than anything server- or config-specific.

Impact

Every MCP-provided tool (in this case glsuite, mssql, plus presumably github, sharepoint, browser, chrome-devtools, playwright per the same .mcp.json) is unavailable for the entire duration of a resumed session after this happens, with no user-facing error explaining why — it just reports tools as not found. Confirmed workaround: starting a new chat session (no resume) restores normal MCP connectivity immediately.

What Should Happen?

Resuming should not terminate MCP servers

Error Messages/Logs

### Log evidence
From `Anthropic.claude-code.Claude VSCode.log` (timestamps same session, `window1/exthost/Anthropic.claude-code/`):


15:03:24.674  MCP Server running on port 38966 (localhost only)
15:03:26.874  Loading config cache by launching Claude (no channel)...
15:03:33.117  Spawning Claude with SDK query function ... permission mode: default,    resume: undefined
15:03:33.128  Spawning Claude with SDK query function ... permission mode: acceptEdits, resume: <session-id>   <- this is the resumed chat
...
15:03:34.125  MCP server "glsuite": Starting connection with timeout of 30000ms
15:03:34.166  MCP server "mssql":   Starting connection with timeout of 30000ms
15:03:34.170  MCP server "glsuite": Starting connection with timeout of 30000ms   <- second instance (dup spawn)
15:03:34.216  MCP server "mssql":   Starting connection with timeout of 30000ms   <- second instance (dup spawn)
...
15:03:37.640  MCP server "mssql":   Successfully connected (transport: stdio) in 3476ms
15:03:37.640  MCP server "mssql":   Connection established with capabilities: {"hasTools":true, ...}
15:03:37.696  MCP server "mssql":   Successfully connected (transport: stdio) in 3484ms
15:03:37.697  MCP server "mssql":   Connection established with capabilities: {"hasTools":true, ...}
15:03:37.758  MCP server "glsuite": Successfully connected (transport: stdio) in 3640ms
15:03:37.758  MCP server "glsuite": Connection established with capabilities: {"hasTools":true, ...}
15:03:37.764  MCP server "glsuite": Successfully connected (transport: stdio) in 3601ms
15:03:37.785  MCP server "glsuite": Connection established with capabilities: {"hasTools":true, ...}
...
15:03:40.578  MCP server "glsuite": Terminating MCP server process tree
15:03:40.590  MCP server "mssql":   Terminating MCP server process tree
...
15:03:41.166  MCP server "mssql":   UNKNOWN connection closed after 3s (cleanly)
15:03:41.167  MCP server "mssql":   Cleared connection cache for reconnection
15:03:41.179  MCP server "glsuite": UNKNOWN connection closed after 3s (cleanly)
15:03:41.180  MCP server "glsuite": Cleared connection cache for reconnection


No further line containing `glsuite`, `mssql`, or `reconnect` appears anywhere later in the same log file, including at:


15:04:40.772  ToolSearchTool: select failed — none found: glsuite_export_task, glsuite_get_task_attachments, glsuite_get_task_spec, mssql_connect


— a minute after the teardown, in the same (resumed) session, with the reconnection that was "cleared for" never actually attempted.

Steps to Reproduce

Repro

1. Have an existing chat session with MCP servers configured in .mcp.json.
2. Fully quit VS Code (all windows) and relaunch, resuming the prior session.
3. Observe: MCP tools are unavailable in the resumed session (ToolSearch finds nothing), even though nothing about the servers, config, or environment changed.
4. Workaround: start a new chat instead of resuming — MCP tools connect and stay available normally.

Claude Model

Sonnet (default)

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.220

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

_No response_