I encountered a critical data loss bug in Cursor. While it appears to be intermittent, the impact is severe—all unsaved instructions are permanently destroyed.
Problem Summary
Under certain conditions, a newly created Agent conversation spontaneously closes without any warning, and the entire conversation—along with all typed but unsubmitted instructions—is permanently deleted. There is no draft state, no recovery mechanism, and no way to retrieve the lost work.
Steps to Reproduce
Open Cursor
Create a new Agent conversation (Cmd/Ctrl+I)
Start typing detailed instructions in the input field
Without submitting the message, open a documentation file or switch focus to another panel/document
[Bug Occurs] The Agent conversation panel unexpectedly closes
Check the conversation list—the new conversation is completely gone with no draft state
Expected Behavior
The conversation should remain open regardless of focus changes, or
Unsaved typing should be automatically saved as a draft, or
A save/discard warning should appear before the conversation closes
I discovered something significant that changes how we should understand this bug.
What Happened
After the initial conversation disappeared, I created a second Agent conversation and:
Typed instructions in it
Started the conversation
Clicked Stop before completion
Closed that conversation
Then: The original “lost” conversation reappeared as an Agent conversation window. It was still treated as a new conversation and did not return to the conversation list, but all of the instructions I had previously entered were still there.
What This Suggests
This appears to be a conversation state management issue rather than actual data loss. The conversations may be entering an unintended hidden state rather than being deleted.
Areas That May Need Review
Conversation visibility — Conversations disappearing from the list without being deleted
State transitions — Recovery happening unexpectedly when other conversations are created/closed
UI feedback — No indication that conversations exist but are hidden
Persistence logic — Unclear conditions that cause conversations to reappear
Potential Impact
Users experience concern over lost work (though data may be retained)
Recovery is unpredictable and accidental
The underlying behavior needs clarification from the team
I’d appreciate the development team’s insight into whether this is expected behavior or an unintended state transition.
This bug has persisted since version 3.x and remains unresolved, particularly within the agent window. If you click “Stop” immediately after sending a message, there is a chance the entire conversation will simply vanish. Additionally, if you encounter the occasional “cyber” block associated with Claude or GPT models while using the tool, a sent message might be “crashed” back into the input field. If you have already engaged in a multi-turn conversation, having a previously sent message reappear in the input field causes severe confusion in the agent’s context for subsequent exchanges.
It’s tied to the Preview / Markdown toggle on a .md file. If your cursor is still in the Agent input when you click Preview, the chat tab gets replaced instead of the markdown file. Nothing is deleted, which matches your conversation coming back later with the typed text intact.
Workaround: click inside the .md file first, then click Preview.
We’ve let the team know and this is an issue we’re tracking. I’ll post here when there’s an update.
@jojolionss - the Stop-button behavior you’re describing looks like something separate. Could you open a new thread for it with your version and exact steps so it gets tracked properly?