change the client to send the elicitation/create result immediately on SSE receipt, independent of the tools/call HTTP lifecycle

[Agent-Hellboy]

I guess currently it depends on HTTP lifecycle.
but how?

here is the log message form my implementation

  INFO:     incoming POST /mcp/http
  2026-01-05 12:08:38,117 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:290 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [CLIENT->SERVER] request_received session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb client_host=10.163.2.82 method=tools/call id=5 phase=receive
  2026-01-05 12:08:38,117 DEBUG pymcp.core.state.cancellation:create_token:33 [DATA][CANCELLATION][TOKEN] created token=5 phase=create
  2026-01-05 12:08:38,118 DEBUG pymcp.core.state.elicitation:request_elicitation:51 [DATA][ELICITATION][SEND] queued_elicitation rpc_id=3ed6e74d-
  a9b1-444d-bd2d-ed08786ef48b session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb payload={"jsonrpc":"2.0","id":"3ed6e74d-a9b1-444d-bd2d-
  ed08786ef48b","method":"elicitation/create","params":{"mode":"form","message":"Provide environment and confirm deployment.","requestedSchema":
  {"type":"object","properties":{"environment":{"type":"string","title":"Environment","enum":["development","staging","production"]},"confirm":
  {"type":"boolean","title":"Confirm deploy?"}},"required":["environment","confirm"]}}}
  2026-01-05 12:08:38,118 DEBUG pymcp.transports.mcp_http_streamable:sse_stream:205 [TRANSPORT][MCP_HTTP_STREAMABLE][SERVER->CLIENT] sse_send
  session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb phase=stream_send
  2026-01-05 12:08:58,118 DEBUG pymcp.core.state.elicitation:request_elicitation:61 [DATA][ELICITATION][TIMEOUT] elicitation_timeout
  rpc_id=3ed6e74d-a9b1-444d-bd2d-ed08786ef48b session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb
  2026-01-05 12:08:58,118 DEBUG pymcp.core.state.cancellation:clear:101 [DATA][CANCELLATION][TOKEN] cleared token=5 phase=clear
  2026-01-05 12:08:58,119 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:395 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [SERVER->CLIENT] response_json session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb method=tools/call id=5 status=200 phase=respond
  INFO:     handled POST /mcp/http status=200 duration_ms=20002.8
  INFO:     incoming POST /mcp/http
  2026-01-05 12:08:58,180 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:290 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [CLIENT->SERVER] request_received session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb client_host=10.163.2.82 method=None id=3ed6e74d-a9b1-444d-
  bd2d-ed08786ef48b phase=receive
  2026-01-05 12:08:58,180 DEBUG pymcp.core.state.session:resolve_elicitation_response:470 [SESSION][ELICITATION][LATE] no_pending_future
  session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb rpc_id=3ed6e74d-a9b1-444d-bd2d-ed08786ef48b keys=['jsonrpc', 'id', 'result']
  payload={"jsonrpc":"2.0","id":"3ed6e74d-a9b1-444d-bd2d-ed08786ef48b","result":{"action":"accept","content":
  {"environment":"development","confirm":true}}}
  2026-01-05 12:08:58,180 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:357 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [CLIENT->SERVER] response_ignored session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb id=3ed6e74d-a9b1-444d-bd2d-ed08786ef48b body_keys=['jsonrpc',
  'id', 'result'] body={"jsonrpc":"2.0","id":"3ed6e74d-a9b1-444d-bd2d-ed08786ef48b","result":{"action":"accept","content":
  {"environment":"development","confirm":true}}} phase=ignore_response
  2026-01-05 12:08:58,180 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:366 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [CLIENT->SERVER] pending_elicitation_ids session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb pending=[]
  INFO:     handled POST /mcp/http status=202 duration_ms=1.6

What I did here is, from a server, I am waiting for Future to be resolved, but it never got resolved(it timed out) because the client didn't promptly respond back to my elicitation request sent over sse stream, it sent the response back after tools/call got over

[evalstate]

This is a common MCP Client issue -- over HTTP if the elicitation request is not answered within the timeout window the connection aborts. You can adjust the HTTP timeout settings to help overcome this.

[Agent-Hellboy]

This is a common MCP Client issue -- over HTTP if the elicitation request is not answered within the timeout window the connection aborts. You can adjust the HTTP timeout settings to help overcome this.

No, the client is waiting for tools/call HTTP round-trip. I guess it should promptly respond back to the see event.

[Agent-Hellboy]

This is a common MCP Client issue -- over HTTP if the elicitation request is not answered within the timeout window the connection aborts. You can adjust the HTTP timeout settings to help overcome this.

I don't know whether it's a coincidence or what, but this is the exact response I got from an llm when I shared the issue with LLM.

[evalstate]

LOL 100% coincidence. I guess weeks of peering at llm outputs does that to a person. (this wasn't llm generated either).

So the scenario is

1) client -> tool/call -> server (POST with SSE response)
2) server -> elicitation/request -> client
3) human -> doesn't respond quickly
4) POST request from Step 1 times out

yes?

[Agent-Hellboy]

3. human -> doesn't respond quickly

No, human has responded, but the server is still waiting for a response from the client, because the fast-MCP client is not sending back a response promptly, it is waiting for tools/call to get over.

  2026-01-05 12:08:38,118 DEBUG pymcp.core.state.elicitation:request_elicitation:51 [DATA][ELICITATION][SEND] queued_elicitation rpc_id=3ed6e74d-
  a9b1-444d-bd2d-ed08786ef48b session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb payload={"jsonrpc":"2.0","id":"3ed6e74d-a9b1-444d-bd2d-
  ed08786ef48b","method":"elicitation/create","params":{"mode":"form","message":"Provide environment and confirm deployment.","requestedSchema":
  {"type":"object","properties":{"environment":{"type":"string","title":"Environment","enum":["development","staging","production"]},"confirm":
  {"type":"boolean","title":"Confirm deploy?"}},"required":["environment","confirm"]}}}

  2026-01-05 12:08:38,118 DEBUG pymcp.transports.mcp_http_streamable:sse_stream:205 [TRANSPORT][MCP_HTTP_STREAMABLE][SERVER->CLIENT] sse_send
  session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb phase=stream_send

here the server is sending the request

  2026-01-05 12:08:58,118 DEBUG pymcp.core.state.elicitation:request_elicitation:61 [DATA][ELICITATION][TIMEOUT] elicitation_timeout
  rpc_id=3ed6e74d-a9b1-444d-bd2d-ed08786ef48b session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb

Elicitation timed out, as I have created a future for elicitations that doesn't get resolved in the specified timeout window.

  2026-01-05 12:08:58,119 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:395 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [SERVER->CLIENT] response_json session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb method=tools/call id=5 status=200 phase=respond
  INFO:     handled POST /mcp/http status=200 duration_ms=20002.8

which has made tools/call post got completed

  INFO:     incoming POST /mcp/http
  2026-01-05 12:08:58,180 DEBUG pymcp.transports.mcp_http_streamable:mcp_http_streamable_post:290 [TRANSPORT][MCP_HTTP_STREAMABLE]
  [CLIENT->SERVER] request_received session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb client_host=10.163.2.82 method=None id=3ed6e74d-a9b1-444d-
  bd2d-ed08786ef48b phase=receive
  2026-01-05 12:08:58,180 DEBUG pymcp.core.state.session:resolve_elicitation_response:470 [SESSION][ELICITATION][LATE] no_pending_future
  session_id=d4826a09-e53a-4e4f-b1f5-227d914abdcb rpc_id=3ed6e74d-a9b1-444d-bd2d-ed08786ef48b keys=['jsonrpc', 'id', 'result']
  payload={"jsonrpc":"2.0","id":"3ed6e74d-a9b1-444d-bd2d-ed08786ef48b","result":{"action":"accept","content":
  {"environment":"development","confirm":true}}}

Here, I have received the request, which will arrive only after completion of tools/call. which is of no use since tools/call completes.

should client maintain a double channel?

Sorry for bugging you if my implementation has an issue.

[evalstate]

OK - here is fast-agent 0.4.22 running against the MCP reference "everything" server.

Recording_2026-01-05_190759_small.mp4

One common issue is that the Elicitation Request may not have been sent via the POST channel (this is easy to miss in the SDKs). https://x.com/evalstate/status/1978832801010000369?s=20

You can check that in fast-agent here:

If you can send over the reference server happy to take a look and improve error handling/robustness as needed.

[Agent-Hellboy]

One common issue is that the Elicitation Request may not have been sent via the POST channel

I am sending it via GET stream as it's server->client.
Do you mean the server should send an Elicitation Request as a post request?

https://github.com/modelcontextprotocol/python-sdk/blob/main/src/mcp/client/streamable_http.py#L524

I guess this send response back to the server on a new socket and doesn't wait for the in-flight tools/call waiting on another socket.

I guess in my scenario, my server is not allowing a new connection before tools/call gets over.

#602
Forcing a new connection is making it work.

If you can send over the reference server happy to take a look and improve error handling/robustness as needed.

It's proprietary.

I have verified the client with python-sdk server , I don't think client has any issue

[Agent-Hellboy]

The interesting part is that it's working fine with the cursor as the client, but not with fast-agent.

[evalstate]

Can you share the MCP Server (or a cut down version which demonstrates this case)?

Is the Elicitation unsolicited (it is NOT make in context of a Client->Server Request in which case the GET is correct)? I want to make sure that the JSON-RPC RequestID matching is 100%

[Agent-Hellboy]

Can you share the MCP Server (or a cut down version which demonstrates this case)?

I will try to test with a server created from go SDK, maybe some other language as well. A server created with Python SDK worked fine. But as I told you, my server(my own implementation of spec) worked fine with the cursor

Is the Elicitation unsolicited (it is NOT make in context of a Client->Server Request in which case the GET is correct)?

There is no issue with the flow; I successfully tested it with the cursor.

I want to make sure that the JSON-RPC RequestID matching is 100%

yes I can validate it.

The problem is that FastAgent does not make a post request before an in-flight tools/call is completed. Please check #602

anyways I will try to share a script which works with the cursor and not with Fast Agent.

[evalstate]

Hey -- I've spent a fair bit of time on this, and added this test server to the PR, but am unable to reproduce the scenario. I'm really keen to do that before modifying the code / improving the error handling. Can you see what VSCode does with your MCP Server too?

https://github.com/Agent-Hellboy/fast-agent/blob/4575634c2326d1ab746dac32e49e93bc001410bd/examples/mcp/elicitations/blocking_elicitation_server.py

[Agent-Hellboy]

Can you see what VSCode does with your MCP Server too?

worked fine with VSCode.

but am unable to reproduce the scenario

I can understand. I will try to share a minimal script.

@evalstate please wait a few days , no rush, let me debug a bit.

[evalstate]

Cool - thanks for that and let me know if you need any help. I can see from the logs that something is queuing up. You can also share code in a private repo if that helps.

[Amigol]

Can you see what VSCode does with your MCP Server too?

worked fine with VSCode.

but am unable to reproduce the scenario

I can understand. I will try to share a minimal script.

@evalstate please wait a few days , no rush, let me debug a bit.

Hi. Have you figured out the problem?

[evalstate]

Hi. Have you figured out the problem?

I can't reproduce with the latest version of fast-agent (as per comment on PR).