0.48 uses many steps, have to keep responding in chat for it to actually completely apply changes

Bug Reports
1

In Cursor 0.45 the AI usually got most requests right in one go. But in 0.48 it keeps asking if I want to continue applying changes, and I have to keep answering ‘yes’ and keep including context for it to continue actually applying the changes.

This often results in the chat just hallucinating stuff and destroying solutions that it already correctly implemented in a previous step.

This happens in Cursor 0.48 in chat ‘ask’ mode.

Cursor 0.45 worked, but Cursor 0.48 like this often it takes more time than just manually doing the work.

How to reproduce: Compare Cursor chat ‘ask’ in Cursor 0.45 with Cursor 0.48 ‘ask’ mode

Sometimes I have to respond multiple times in the chat for it to eventually, actually apply the changes.

Cursor version details:

Version: 0.48.9 (Universal)
VSCode Version: 1.96.2
Commit: 61e99179e4080fecf9d8b92c6e2e3e00fbfb53f0
Date: 2025-04-12T19:19:47.994Z
Electron: 34.3.4
Chromium: 132.0.6834.210
Node.js: 20.18.3
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.3.0

2

For anyone facing a similar issue: I solved it by asking for an investigation of the issue first, and then asking for a solution (with hints of what my preferred solution is).

At the same time, asking the chat “give me code I can APPLY”, tends to simplify things as well. Besides giving it requests with a smaller scope.