[Resolved, not a bug] mcp_session() timeouts were caused by duplicate registration, not Positron #238

Update (2026-10-07): not a Positron bug. In my repro, btw_mcp_session() was called by hand and from ~/.Rprofile, so the R process held two mcptools sockets. With a single registration, the Positron console handles repeated tool calls fine. The original report is kept below for reference.

Summary

An R session in the Positron console registers with the MCP server (it appears in
list_r_sessions), but every tool call sent to it times out after 120 seconds. The same
setup works correctly from plain R started in a terminal.

Environment

• macOS Tahoe 26.7.1 (aarch64)
• R 4.6.1
• Positron 1.130.0 (from the app's Info.plist)
• btw 1.5.0, mcptools 1.0.3, later 1.4.8, nanonext 1.10.3
• MCP client: Claude Code 2.1.293 (VS Code extension running inside Positron)
• Server config: Rscript -e "btw::btw_mcp_server()"

Steps to reproduce

1. In the Positron R console, run btw::btw_mcp_session() (also run from ~/.Rprofile).
2. From the MCP client, call list_r_sessions. It returns 1: my-project (Positron).
3. Call select_r_session(1). It succeeds.
4. Call any tool, e.g. btw_tool_run_r with ls(), while the console is idle at >.

Expected: the code runs in the Positron session and returns its objects.
Actual: Timed out waiting for a response from the selected R session after 120 seconds.

What I tried

• Restarting R, reinstalling btw/mcptools, and reconnecting the MCP server: same result.
• Plain R in a terminal (same machine, same project, same .Rprofile): works.
  interactive() is TRUE, and objects created in that console are visible to the tools.

Possibly related

• When no session is selected or registered, tools quietly run in a fresh non-interactive
  R process (interactive() == FALSE, empty environment). This made it look as if the
  Positron session was working when it wasn't. A warning or a visible flag when falling
  back would help.
• btw_tool_env_describe_environment on an empty global environment returns a result with
  text: null. The client rejects it with an MCP schema validation error.

Debugging details

• btw tools run fine when called directly in the Positron console
  (btw_tool_sessioninfo_platform() takes 0.24 s).
• later, promises, and nanonext::recv_aio() → as.promise() callbacks all fire
  normally in the Positron console when idle.
• The session answers the first list_r_sessions probe. After that, a forwarded
  tools/call never arrives: tracing mcptools:::mac_open in the session shows no call.
  nanonext::unresolved(mcptools:::the$raio) is TRUE, and there is no
  [mcptools] session error message. The session then stops answering
  list_r_sessions probes as well.
• The server log shows TO SESSION: {"method":["tools/call"], ...} for each failed request.
• Plain terminal R on the same machine, with the same packages and .Rprofile, works reliably.
• Closing a second Claude Code instance (a second btw_mcp_server()) made no difference.