-
Notifications
You must be signed in to change notification settings - Fork 851
Description
Disclaimer: I am not a native English speaker. This issue was drafted with AI translation assistance. I apologize for any awkward phrasing.
Bug description
After a successful initialize handshake where the server negotiates down to an older protocol version (e.g. 2025-06-18), the subsequent GET request to open the SSE stream still sends the client's original latest version (2025-11-25) in the MCP-Protocol-Version HTTP header rather than the negotiated version. Servers that validate this header (such as rmcp) reject the request with 400 Bad Request.
Root cause
In both WebClientStreamableHttpTransport and HttpClientStreamableHttpTransport, the MCP-Protocol-Version header is resolved via Reactor Context:
ctx.getOrDefault(McpAsyncClient.NEGOTIATED_PROTOCOL_VERSION, this.latestSupportedProtocolVersion)The negotiated version is written into the Reactor Context by LifecycleInitializer.withInitialization():
.flatMap(res -> operation.apply(res)
.contextWrite(c -> c.put(McpAsyncClient.NEGOTIATED_PROTOCOL_VERSION,
res.initializeResult().protocolVersion())));However, the GET reconnect is triggered as a side effect inside sendMessage() when the transport session is first marked as initialized:
if (transportSession.markInitialized(response.headers()...getFirst(HttpHeaders.MCP_SESSION_ID))) {
reconnect(null).contextWrite(sink.contextView()).subscribe();
}At this point, sink.contextView() does not yet contain NEGOTIATED_PROTOCOL_VERSION because the contextWrite in LifecycleInitializer has not executed yet — it runs after the initialize request's sendMessage() completes. So the GET reconnect falls back to latestSupportedProtocolVersion (2025-11-25).
Timeline:
LifecycleInitializer.doInitialize()→ callsmcpClientSession.sendRequest("initialize", ...)→transport.sendMessage()sendMessage()POSTs to/mcpwith headerMCP-Protocol-Version: 2025-11-25✅ (no negotiation yet, expected)- Server responds with
protocolVersion: "2025-06-18"✅ - Inside
sendMessage(),transportSession.markInitialized(sessionId)returnstrue→ immediately callsreconnect(null)which fires a GET withMCP-Protocol-Version: 2025-11-25❌ - Later,
LifecycleInitializer.withInitialization()runs.contextWrite(c -> c.put(NEGOTIATED_PROTOCOL_VERSION, "2025-06-18"))— too late for the GET in step 4
Environment
- Spring AI: 2.0.0-M3
- MCP Java SDK (
io.modelcontextprotocol.sdk:mcp-core): 1.1.0 - Spring Boot: 4.0.4
- Java: 25
- Transport:
WebClientStreamableHttpTransport(viaspring-ai-starter-mcp-client-webflux) - MCP Server: rmcp 1.2.0 (Rust, Streamable HTTP, supporting protocol version
2025-06-18)
Steps to reproduce
- Set up an MCP server that supports
protocolVersion: "2025-06-18"and validates theMCP-Protocol-VersionHTTP header (rejecting unsupported versions). - Configure a Spring AI MCP client with
spring-ai-starter-mcp-client-webfluxusing default settings (no customsupportedProtocolVersions). - Start the application and observe the HTTP traffic.
Observed:
POST /mcp → MCP-Protocol-Version: 2025-11-25 → 200 OK (negotiated to 2025-06-18)
GET /mcp → MCP-Protocol-Version: 2025-11-25 → 400 Bad Request
Expected behavior
After the server responds with protocolVersion: "2025-06-18", all subsequent requests (including the GET SSE reconnect) should use MCP-Protocol-Version: 2025-06-18 in the HTTP header:
POST /mcp → MCP-Protocol-Version: 2025-11-25 → 200 OK (negotiated to 2025-06-18)
GET /mcp → MCP-Protocol-Version: 2025-06-18 → 200 OK
Workaround
Register a McpClientCustomizer bean to remove 2025-11-25 from the supported versions list, so the fallback value aligns with the server:
@Component
public class McpTransportCustomizer implements McpClientCustomizer<WebClientStreamableHttpTransport.Builder> {
@Override
public void customize(String name, WebClientStreamableHttpTransport.Builder builder) {
builder.supportedProtocolVersions(List.of(
ProtocolVersions.MCP_2024_11_05,
ProtocolVersions.MCP_2025_03_26,
ProtocolVersions.MCP_2025_06_18
));
}
}