SSE transport hangs indefinitely on long-running requests (Mixpanel segmentation queries)

[TweedBeetle]

Summary

When using mcp-remote to connect to Mixpanel's hosted MCP server (mcp.mixpanel.com/mcp), tool calls frequently hang for 5-55+ minutes before the user cancels them. The exact same queries return in 1-3 seconds when immediately retried, suggesting the SSE transport drops or stalls rather than the upstream server being slow.

Environment

• Client: Claude Code (CLI)
• Transport: npx mcp-remote https://mcp.mixpanel.com/mcp --static-oauth-client-metadata '{"scope":"..."}' --allow-http
• Platform: macOS

Reproduction

1. Configure Mixpanel MCP via mcp-remote as shown above
2. Issue a run_segmentation_query tool call with a group-by parameter (e.g., segment chat_started by properties["is_returning_user"])
3. Observe the call hangs indefinitely (no response, no error)
4. Cancel and retry the identical call — it completes in 1-3 seconds

Observed behavior

Over a single session (~65 Mixpanel MCP calls), 6 calls hung indefinitely:

Query	Hang duration	Retry time
Segmentation with group-by	46 min	1.4s
Same + filter	1.8 min	~1s
get_events (list all)	12.3 min	5.8s
Segmentation, 2-month range, group-by	4.4 min	1.9s
Segmentation, group-by error_type	4.4 min	~1s
Two parallel segmentation queries with type: "unique"	55.8 min each	~1s

The remaining ~59 calls completed in 1-5 seconds normally.

Pattern

• Calls with group-by clauses and parallel execution seem more susceptible
• When 2-3 calls are issued simultaneously, one completes while others hang
• No error is ever returned — the call simply never resolves
• Instant success on retry rules out upstream API slowness

Possible relation to #107

Issue #107 describes SSE connections being terminated after 5 minutes of inactivity, with undici's idle-body timeout being the culprit. This may be related — perhaps the response stream stalls briefly and the connection enters a broken state without propagating an error, causing the call to hang forever instead of timing out.

Expected behavior

Requests should either complete or fail with a timeout error within a reasonable window (30-60 seconds), not hang indefinitely.