You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
mcp: tool calls abort and restart the server with a large (229-tool) MCP server #59037
Calling any tool from the REA MCP server (229 tools) aborts the invocation and restarts the server. A trivial call (get_navigation_context({})) fails the same way as a large one (binary_session({})), so this is not response size.
The server log shows the connection succeed, then the run ends with no error and no shutdown line, and ~10-20s later a new run starts with a new run= id. Four distinct runs in 8 minutes, each right after a tool call:
Reproduced 3/3 attempts. The MCP server itself is healthy: rea doctor reports "healthy": true and a direct stdio handshake against the same command returns 229 tools. Registration is reported aligned. No Request timed out precedes any abort, and there is no error in the log - the process simply disappears.
Note: the registration carries an environment.PATH override because this Windows session's inherited PATH predates the Node.js install; without it the server cannot launch at all. It is not the cause here - the connection succeeds.
I couldn't trigger this on Linux. With OpenCode 2.0.30 and the same server ([email protected] as a local MCP server, rea (229 tools) in the catalog), calling get_navigation_context({}) through Code Mode worked normally. It returned REA's own "No app is open" result. With codemode: false, a direct binary_session({}) call also worked. The OpenCode service stayed up, and nothing was interrupted. So the MCP call path doesn't seem to break with a server this size in general.
In your log, the process that restarts is OpenCode's own background service (cli starting with a new run= id), not the REA server. That looks like the Windows desktop app restarting its background service, which is what these issues describe (ordered by likelihood):
Activity
opencode-agent commentedon Oct 25, 2026
Thanks for the detailed report and the run IDs.
I couldn't trigger this on Linux. With OpenCode 2.0.30 and the same server (
[email protected]as a local MCP server,rea (229 tools)in the catalog), callingget_navigation_context({})through Code Mode worked normally. It returned REA's own "No app is open" result. Withcodemode: false, a directbinary_session({})call also worked. The OpenCode service stayed up, and nothing was interrupted. So the MCP call path doesn't seem to break with a server this size in general.In your log, the process that restarts is OpenCode's own background service (
cli startingwith a newrun=id), not the REA server. That looks like the Windows desktop app restarting its background service, which is what these issues describe (ordered by likelihood):This needs a maintainer with Windows and the desktop app to reproduce. Two things would help:
Background service is unresponsiveorEvent stream stalled.opencodefrom a terminal on that machine, outside the desktop app.