SSE transport: connection succeeds then marked 'Failed' after first event, works fine with raw curl #2783

Describe the bug

Connecting to a remote MCP server using the legacy SSE (Server Sent Events) [deprecated] transport: the initial GET /sse request succeeds (200, correct event: endpoint frame delivered with a valid session_id), but the Inspector UI then marks the server connection status as "Failed" and never proceeds to list tools/resources. The History panel shows the single GET transport request as 200 SSE, with the response body described as "Long-lived stream — 1 event delivered, closed; body not captured".

Driving the exact same server by hand with raw curl (bypassing Inspector entirely) completes the full handshake correctly: GET /sse stays open indefinitely, and POSTing initialize → notifications/initialized → tools/list → tools/call to the returned /messages/?session_id=... endpoint all return correct JSON-RPC responses on the SSE stream, with no server-side errors. This confirms the target server, TLS, and auth header are all fine — the failure is isolated to Inspector's own legacy SSE client path.

Steps to reproduce

1. Start Inspector against a remote MCP server using the Python MCP SDK's SSE transport (uvicorn-based, /sse + /messages/?session_id=... pattern):

   npx @modelcontextprotocol/inspector \
     --transport sse \
     --server-url "https://<your-server>/sse" \
     --header "Authorization: Bearer <redacted>"

2. Open the printed Inspector UI URL, confirm Transport = SSE, Server URL ends in /sse, Authorization header populated.
3. Click Connect.
4. UI shows "Connecting..." then settles on "Failed".
5. In the Requests panel, the single GET transport request shows 200 SSE with 1 event delivered, then the connection is reported closed.

Expected behavior

Inspector should keep the SSE stream open (the same way a direct curl -N connection to the identical endpoint stays open indefinitely) and proceed to auto-initialize / allow listing tools, since the server is fully spec-compliant and responds correctly when driven manually.

Environment

• Inspector version: latest via npx @modelcontextprotocol/inspector (package version 2.12.0 as of testing)
• OS: macOS (darwin)
• Target server: Python MCP SDK, SSE transport, fronted by a cloud load balancer (TLS terminated there)

Additional context

Manually replayed protocol trace (via curl) against the same server, for reference, showing a fully successful initialize → tools/list → tools/call exchange with no errors — confirming the server side is healthy and the bug is isolated to Inspector's legacy SSE client:

event: endpoint
data: /messages/?session_id=<redacted>

event: message
data: {"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2024-11-05","capabilities":{...},"serverInfo":{"name":"<server-name>","version":"<x.y.z>"}}}

event: message
data: {"jsonrpc":"2.0","id":2,"result":{"tools":[{"name":"tool_a", ...}, {"name":"tool_b", ...}, {"name":"tool_c", ...}]}}

event: message
data: {"jsonrpc":"2.0","id":3,"result":{"content":[{"type":"text","text":"<tool output>"}],"isError":false}}

Happy to provide a full HAR/network export or more detail if useful — the server itself cannot be shared since it's a private internal instance.