input_required not being handled correctly #1073

Activity

1. cliffhall commented on Feb 6, 2026

   

   cliffhall

   on Feb 6, 2026

   Last edited by cliffhall

   Member

   More actions

   Hi @Randgalt

   I think there's a disconnect. Looks like you're running 0.19.0. The PR we're testing is on 0.20.0.

   Maybe you're not running the PR branch I linked to.

   If you are, then I think perhaps your tool should have execution.taskSupport === "forbidden" because the tool itself isn't going to be run as a task, it is going to start a task on the client.

   The presence of the "Run as Task" button in your screenshot is a giveaway that either you're not on my branch OR your tool is not configured correctly.

   The Everything server has a tool that starts a client side task called trigger-sampling-request-async. Let's test with it.

   Try this

   1. Run the Everything server npx @modelcontextprotocol/server-everything@latest streamableHttp
   2. Run the Inspector from this branch
   3. Click Connect
   4. Click List Tools
   5. Scroll down list and click "trigger-sampling-request-async" tool in list
   6. Enter "Hello" in prompt field
   7. Note "Run as Task" checkbox is not present
   8. Click "Run Tool"
   9. See you're on the Elicitation screen
   10. Enter "goodbye" in the text field of the elicitation form
   11. Click Submit
   12. See you're back on the Tool screen and the tool completed
   13. Look at the Server History pane, it is newest on top
   14. There is a tool/call with success at the top
   15. Next there is a notifications/tasks/status with completed
   16. Next there is a notifications/tasks/status with input_required

   What you should see

   Screen.Recording.2026-02-06.at.6.06.43.PM.mov
2. Randgalt commented on Feb 6, 2026

   

   Randgalt

   on Feb 6, 2026

   Last edited by Randgalt

   Author

   More actions

   Let me see if I understand correctly. We have a tool, called "task" that wishes to run as a task on our server. So we set the execution.taskSupport to required. Am I not understanding how this works?

   {
         "name": "task",
         "description": "Test task with elicitation",
         "inputSchema": {
           "type": "object",
           "properties": {},
           "required": [],
           "$schema": "https://json-schema.org/draft/2020-12/schema",
           "description": "Test task with elicitation"
         },
         "annotations": {},
         "execution": {
           "taskSupport": "required"
         }
       }

   Note this is what we return for the tools/call result:

   {
     "content": [],
     "task": {
       "createdAt": "2026-02-06T23:30:47.576226Z",
       "lastUpdatedAt": "2026-02-06T23:30:47.576226Z",
       "pollInterval": 100,
       "status": "working",
       "taskId": "e14b50b5fe9c4072ba00f97c9629ceaa",
       "ttl": 3600000
     }
   }
3. cliffhall commented on Feb 7, 2026

   

   cliffhall

   on Feb 7, 2026

   Member

   More actions

   Ok, gotcha. We have an Everything server tool with a mode that is meant to test this use case, but it seems not to send the actual elicitation request. We'll get that sorted so we can more properly test the scenario. But are you certain that your server has sent the elicitation request? You said it sent a task status of input_required but didn't mention if elicitation request was sent.
4. Randgalt commented on Feb 7, 2026

   

   Randgalt

   on Feb 7, 2026

   Last edited by Randgalt

   Author

   More actions

   You said it sent a task status of input_required but didn't mention if elicitation request was sent.

   As I understand the protocol the ElicitRequest is returned when the client calls task/result which it is meant to do when the status becomes input_required. I’ve tested this with the pending Java client implementation and it works.

   With the Inspector we never get the task/result call.
5. cliffhall commented on Feb 7, 2026

   

   cliffhall

   on Feb 7, 2026

   Last edited by cliffhall

   Member

   More actions

   @Randgalt try with STDIO.
   In trying the Everything Server's "Simulate Research Query" tool which tests elicitation request generated from a server task, I was seeing:

   

   Getting into the code I found the note @olaservo had left about the gap. The Typescript SDK doesn't support this for anything but STDIO at the moment.

   

   Tried it with STDIO and it works fine.

   stdio-elicitation.mov
6. Randgalt commented on Feb 7, 2026

   

   Randgalt

   on Feb 7, 2026

   Author

   More actions

   Ok. Our server is HTTP/SSE only. Will wait for that PR

   👍React with 👍1Reacted by Cliff Hall
7. cliffhall commented on Feb 7, 2026

   

   cliffhall

   on Feb 7, 2026

   Member

   More actions

   Ok. Our server is HTTP/SSE only. Will wait for that PR

   I've pinged the TS SDK team about it. The PR has been outstanding since December. They're in the process of creating a V2 SDK and so V1 isn't getting quite as much attention as it was before. But this is a spec compliance matter, so I think it'll get tended.

   👍React with 👍1Reacted by Jordan Zimmerman
8. galagaevdc commented on Apr 3, 2026

   

   galagaevdc

   on Apr 3, 2026

   Contributor

   More actions

   I've prepared the fix for this issue in PR #1174
9. cliffhall

   closed this as completedin #1174on Jun 7, 2026