CSE Roadmap

Core tier · about 6 hours · 14 parts

Part 6 of 14 · 2 min read

What the example measured

The shutdown ladder, the SSE and status-code tables, the header-mismatch and idempotency results, and the transport comparison timings.

Read each table left to right: the situation on the left, what the example observed on the right.

One slow call, three ways. A 40-step job, cancelled partway through, behaves differently by transport:

stdio Streamable HTTP, 2026-07-28 HTTP, 2025-11-25
Progress notification lines on stdout, matched by the progress token SSE events on this request's stream SSE events, each with an event id
Cancel notifications/cancelled close the stream CancelledNotification (disconnecting is not cancelling)
Dropped connection the pipe closes: restart the process stream ends: send the request again GET with Last-Event-ID: the server replays what you missed

What to notice: the cancel signal is a message on two transports and a socket event on the third. That is why section 1's notifications/cancelled and section 4's closed-stream check are different mechanisms for the identical idea.

HTTP status codes in the pages checked:

Situation Status and body
notification accepted 202, no body
missing or mismatched mirrored header 400 and error -32020
unsupported protocol version 400 and an UnsupportedProtocolVersionError
unknown method 404 and a JSON-RPC error -32601
bad Origin 403
GET or DELETE (old shapes) 405

What to notice: two choices are not covered by the specification text: 406 for a missing Accept types list, and 200 with a JSON-RPC error for failures like an unknown tool (-32602). Both are marked as such in the code.

The header-mismatch and idempotency results, already walked through in sections 2 and 3:

Experiment Without the fix With the fix
Spoofed Mcp-Name header HTTP 200, 1 hold created HTTP 400 (-32020), 0 holds created
Dropped connection, retried call 2 attempts, 2 holds (duplicate) 2 attempts, 1 hold (idempotency key)
Dropped connection, origin check hostile page: hold created hostile page: 403, no hold

Transport cost, measured on this machine over loopback with a small call repeated 150 times: starting the stdio server took about 47 to 51 ms, against about 1.5 ms to connect to an already-running HTTP server. A call took about 0.4 ms on stdio against about 1.3 ms over HTTP. The request was 297 bytes on stdio against 470 bytes over HTTP, because every HTTP request carries headers. What to notice: on a real network, with TLS and a proxy, the gap changes, but the shape is the point. stdio is cheaper per call and costs a process; HTTP costs a little per call and costs nothing to start.