Bug 1 — Agent questionnaire / prompt countdown
During an Agent conversation, the UI sometimes shows a questionnaire or structured prompt from the Agent, with what appears to be a countdown. If I do not submit an answer before the countdown finishes—even though the questionnaire remains visible on screen—the Agent proceeds as if I skipped the questionnaire instead of still being able to respond.
Bug 2 — Rollback in one chat affects another chat’s applied edits
When using multiple chats in parallel and both conversations produce or apply code changes, performing a rollback / undo of applied changes in one chat can also undo edits that were already applied from a different chat.
Steps to Reproduce
Bug 1
Start an Agent conversation where the Agent presents a questionnaire or multi-choice prompt.
Leave the questionnaire on screen without selecting/submitting until the countdown completes (or appears to expire).
Observe whether the Agent treats the interaction as skipped even though the UI still showed the questionnaire.
Bug 2
Open Chat A and apply code changes via the Agent (accept/apply edits as you normally would).
Open Chat B (another conversation) and also apply code changes, possibly touching overlapping files.
In Chat A, use rollback / undo for that conversation’s applied changes.
Bug 1 (countdown timeout): This is a known issue. The agent can sometimes proceed after a prompt or approval times out, even though the UI still shows it. Our team is actively investigating this behavior. Updating to the latest Cursor version may help, as some improvements in this area have shipped since your build (April 28). You can update via Help > Check for Updates or the Command Palette (Attempt Update).
Bug 2 (cross-chat rollback): This is a known limitation of how checkpoints currently work. For now, the safest approach is to avoid overlapping file changes across parallel chats, or to commit/stash changes from one chat before reverting another.
Could you also try updating to the latest version and let us know if Bug 1 still occurs? That would help us narrow down whether the specific fix that shipped addresses your case.
Confirming Bug 1 (questionnaire timeout treated as a skip) still reproduces on a much newer build than the original report.
Environment
Cursor 3.7.19 (arm64)
macOS 26.5.1 (build 25F80)
@mohitjain you asked the OP to retest after the post–April-28 improvements — I can confirm it is still happening on 3.7.19. In an Agent session today the AskQuestion prompt resolved with the synthetic “Questions skipped by the user, continue with the information you already have” string while I was still deciding how to answer. The questionnaire was still visible on screen, but the agent had already received the skip and would have proceeded on an uninformed default path had I not intervened.
This matches the root cause confirmed in the related (now-closed) thread AskQuestion tool can return synthetic skip string with highly variable, unpredictable delay — the orphaned gRPC promise resolving to an empty result that is byte-identical to a deliberate skip. The most damaging part is that the agent cannot tell a timeout from a real skip, so for any consequential prompt it silently treats the timeout as user consent.