Repository navigation
Description
What happened?
POST SSE streams open with a priming event (id, plus retry when retryInterval is set). The standalone GET stream doesn't, so until the server sends something on it the client has no event id to resume from. If closeStandaloneSSEStream() is called in that window, StreamableHTTPClientTransport reconnects on schedule but without Last-Event-ID, and anything sent while it was disconnected sits in the event store and is never delivered.
SDK client against an in-process server (retryInterval: 300), one notification sent right after the close and one after the reconnect:
[19ms] client GET, Last-Event-ID=(none)
[221ms] server closeStandaloneSSEStream()
[528ms] client GET, Last-Event-ID=(none)
received: [ 'after-reconnect' ] <- 'during-gap' was stored as e2, never replayed
writePrimingEvent() is only called from the POST path; handleGetRequest() and replayEvents() never prime. v1.x (1.32.2) has the same code in webStandardStreamableHttp.ts.
What did you expect?
The GET stream primed on open and on resume, same as POST, so the client reconnects with Last-Event-ID and gets what it missed.
The 2025-11-25 transports spec, for GET streams: "If the server closes the connection without terminating the stream, it SHOULD follow the same polling behavior as described for POST requests: sending a retry field and allowing the client to reconnect." The POST rules it points to only allow closing after an event ID has been sent.
Two things a fix has to handle: the priming event has to be stored before the stream is registered, otherwise a concurrent send() can land ahead of it, and replay should skip stored priming events because _GET_stream is reused across connections. The Python SDK primes resumed GETs but not fresh ones, so it has the same gap on first connect. I have this on a branch with tests and can open a PR if the approach works for you.
Related: #2499
Code to reproduce
// Shows the missing priming event. Output:
// POST: "id: 30847839-..._0\nretry: 1000\ndata: \n\n"
// GET: ""
import { McpServer, WebStandardStreamableHTTPServerTransport } from '@modelcontextprotocol/server';
import type { EventStore } from '@modelcontextprotocol/server';
const events = new Map<string, unknown>();
const eventStore: EventStore = {
async storeEvent(streamId, message) {
const id = `${streamId}_${events.size}`;
events.set(id, message);
return id;
},
async replayEventsAfter() {
return '';
}
};
const transport = new WebStandardStreamableHTTPServerTransport({
sessionIdGenerator: () => crypto.randomUUID(),
eventStore,
retryInterval: 1000
});
await new McpServer({ name: 'repro', version: '0' }).connect(transport);
const init = await transport.handleRequest(
new Request('http://localhost/mcp', {
method: 'POST',
headers: { 'content-type': 'application/json', accept: 'application/json, text/event-stream' },
body: JSON.stringify({
jsonrpc: '2.0',
id: 1,
method: 'initialize',
params: { protocolVersion: '2025-11-25', capabilities: {}, clientInfo: { name: 'c', version: '0' } }
})
})
);
const sessionId = init.headers.get('mcp-session-id')!;
const initReader = init.body!.getReader();
console.log('POST:', JSON.stringify(new TextDecoder().decode((await initReader.read()).value)));
await initReader.cancel();
const get = await transport.handleRequest(
new Request('http://localhost/mcp', {
headers: { accept: 'text/event-stream', 'mcp-session-id': sessionId, 'mcp-protocol-version': '2025-11-25' }
})
);
setTimeout(() => transport.closeStandaloneSSEStream(), 100);
console.log('GET: ', JSON.stringify(await get.text()));
await transport.close();SDK version
@modelcontextprotocol/server 2.4.0 (main @ 5ee54ee)
Area
Transports
Activity
A7mdMe commentedon Nov 21, 2026
Standalone GET Priming, Reconnection Semantics & Gapless Replay Considerations
This is an acute issue in production MCP deployments—especially where gateways, load balancers, or Cloudflare enforce connection rotation (e.g. 60–100s idle limits) and the server relies on
closeStandaloneSSEStream()to gracefully cycle long-lived connections.Root Cause & Protocol Mechanics
Under the MCP 2025-11-25 / 2024-11-05 transport specification, clients rely on SSE
id:frames to track the stream cursor. When opening a standalone GET stream, if the server emits nothing until an application event occurs, the client'sLast-Event-IDremainsnull. WhencloseStandaloneSSEStream()terminates the connection:retry: <interval>, but sends noLast-Event-ID.EventStoreduring the disconnect-to-reconnect gap are permanently bypassed and never replayed.Two Critical Implementation Details for the Fix
When implementing stream priming on the GET path, two concurrency and replay details are essential:
The priming event (
id: ${streamId}_1\nretry: ${retryInterval}\ndata: \n\n) must be persisted in theEventStorebefore the active stream listener is registered. If registered prior to priming, a concurrenttransport.send()can acquire sequence1while the priming frame arrives with sequence2out-of-order.Because
EventStoreretains the initial priming record, any subsequent resumption usingLast-Event-IDfrom an earlier point (or across stream resets) must filter out empty priming frames (isPriming: trueor empty payload) so clients are not spammed with empty heartbeats.Ensuring headers include
X-Accel-Buffering: noandCache-Control: no-cache, no-transformprevents intermediate proxies (Nginx, reverse proxies) from holding the priming event in transit.Disclosure: I maintain MCP Production Kit, an open-source hardening suite for MCP servers. We implemented and verified this standalone SSE stream priming, monotonic sequence tracking, and gapless replay pattern in our test suite (Module 30).