EINTR errors in file watcher cause MCP initialization timeout

[ludocarceles]

Summary

MCP server initialization times out due to unhandled EINTR (interrupted system call) errors in the bundled Node.js CLI's file watcher. This causes sessions to crash and enter a timeout/recovery loop.

Environment

• OS: macOS 15.2 (Darwin 25.2.0)
• Python: 3.13
• SDK version: Latest (installed via pip)
• MCP tools registered: 68 custom tools

Error logs

Exception: Control request timeout: initialize

This error originated either by throwing inside of an async function without a catch block, 
or by rejecting a promise which was not handled with .catch(). The promise rejected with the reason:
Error: EINTR: interrupted system call, watch '/var/folders/h4/.../T/...-IBTOOLD-IN.QnhTSAefXh'
    at watch (unknown)
    at new FSWatcher (node:fs:29:31)
    at watch (node:fs:296:23)

Impact

• Every session eventually hits the 60-second initialization timeout
• Sessions enter an infinite timeout → recovery → timeout loop
• The application becomes unusable for multi-turn conversations

Root cause analysis

The bundled CLI subprocess uses file watchers for IPC. When a system signal interrupts the watch() syscall, Node.js throws an EINTR error. This error is not caught/retried, causing the initialization to hang until it times out.

Expected behavior

EINTR is a standard Unix signal indicating a syscall was interrupted and should be retried. The file watcher should:

1. Catch EINTR errors
2. Retry the watch operation
3. Not propagate the error up as a fatal failure

Workaround

Setting CLAUDE_CODE_STREAM_CLOSE_TIMEOUT=180000 (3 minutes) gives more time for initialization to eventually succeed, but doesn't fix the underlying issue.

Reproduction

1. Create an MCP server with ~50+ tools
2. Use ClaudeSDKClient in a long-running server process
3. Send multiple messages over time
4. Observe Control request timeout: initialize errors in logs