In Multitask chat, questions get lost

Support Bug Reports
, ,
1

Where does the bug appear (feature/product)?

Cursor CLI

Describe the Bug

According to Opus 5.5:
What happened: question cards came back to me marked “interrupted” with no selections. That happened at 3:01, at 4:36, and for the AI-kick/sledge pair earlier. According to the side chat, an open card is discarded, answers included, whenever a new turn lands in the chat. That new turn can be a message you send or a background agent finishing. Answers typed into the chat box do arrive.

According to Grok from the side chat:
It’s a Cursor question-card bug, and this chat’s multitasking makes it show up constantly. The orchestration is reacting to a missing answer, not dropping one it already had.

The card and the chat box cancel each other. On Monday you hit this directly: you answered the first question on the card, typed the second in the chat box, and the card came back cancelled with nothing attached. An open question card is thrown away, answers included, the moment a chat message is sent. Text typed in the box but not sent never reaches the agent either.

Today’s pair was the same failure, stretched out. The AI-kick and sledge questions sat on a card for about 22 minutes, then the card was interrupted and returned no selections. The agent only sees that interrupt, so it correctly said both were still open. Your “I answered them both” message is what finally delivered the yeses, as ordinary chat text. That message was also delivered twice, back to back.

Multitask mode keeps tripping it. Background agents finish and post into the same chat while a card is still up. That new turn interrupts the card the same way a sent chat message does. The agent then re-asks, because it never received the click.

Answers typed straight into the chat, including quoting the question, do arrive. The card is the part that gets discarded.

Steps to Reproduce

Have a multitask chat monitoring several subagents. Those subagents then ask questions. Even fully answering them, the chat won’t recognize they have been answered and will insist I canceled. Although the diagnosis listed said they were all interrupted, the last one was completed. I pasted a screenshot of the answers and the agent picked it up then from the screenshot.

Expected Behavior

You need a way for the orchestrating agent to manage and present the questions in a non-interrupting manner. Simply typing in the chat shouldn’t dismiss them. Simply having another question arrive shouldn’t dismiss the ones I’m answering. And when you don’t answer, either after time or interruption, it shouldn’t just assume you don’t care.

Operating System

Windows 10/11

Version Information

This chat: the main door-system chat, agent ID 6aeef096-983c-44d9-b7ec-6b3d7a318d57, running Claude Opus 5.5.
The side chat next to it: agent 0bd83235-1447-4f75-b110-0d1f95cb3c8a, running Grok 4.7.
Cursor: version 3.21.18, build d5c0e77a0214208f36b56d42e8e787de88d02ea4, on Windows 10.0.26200. That’s from Help > About Cursor, as the side chat reported.

For AI issues: add Request ID with privacy disabled

4ce641c7-56aa-48d8-a01c-8be9e7fc0db5

Does this stop you from using Cursor

No - Cursor works, but with this issue

7

Hey @jseghers-hwg, thanks for the report!

Two things are going on here. First, sending a message in the chat while a question card is open currently skips that card, and anything already selected on it is not sent. So for now, either answer everything on the card and submit it before typing, or put all of your answers in the chat message itself (those always arrive, as you noticed).

Second, a question card raised by a subagent can get dropped when a different background agent finishes and posts into the same chat. That one is a known issue we’re already tracking, and I’ve added your report to it.

In the meantime, telling the orchestrating agent to ask its questions as plain text in the chat, instead of with question cards is one possibility you could consider if you are repeatedly hitting this and this should avoid losing answers while several subagents are running.

8

The orchestrator agent came up with that suggestion too. It will do as a workaround.

The restriction to not be able to ask questions in chat while figuring out how to answer a question is something I think should be fixed. Too many things in the cursor chat UI are fragile. I often lose the side chat just by opening a file. The modal nature of question cards and losing answers just because I accidentally typed an answer into the chat edit box instead of the other-answer edit box.

Maybe have a question stack as a margin tool that shows all the questions for different subagents. That would let you answer them in the order that makes sense to your own mental model of what’s important.

And the agents taking a dismissal as “do it your own way” has sent the subagents building completely wrong things–and costing a lot of tokens doing it.