Describe the request
Body:
I’m a paying Cursor user, and I’m genuinely angry about a recent regression in how Agent applies code changes.
What used to work (and why I chose Cursor):
When the Agent modified code for a specific task, I would see a clear before/after comparison — red/green inline diffs — and a simple Apply (or Accept) control for each change / each location. I could review this feature’s edits, accept what I wanted, and reject what I didn’t — file by file, chunk by chunk. That workflow was Cursor’s killer advantage: great UX + visual diff + granular control. It felt safe to let the Agent work because I was always one click away from approving or discarding a proposed change.
What I get now:
Changes often seem to land in files before I’ve consciously approved them.
The chat shows a session-level “Review +1181 -413” button — that’s a bulk summary of the entire conversation, not a list of “here are the 5 places I changed for this feature — Apply or reject each one.”
Diff View in the editor shows highlighting, but that’s not the same as the old Apply flow. Highlighting without a clear per-change Apply is just “show me damage after the fact.”
I searched settings for “Inline Diffs” — nothing useful. Auto-Run Mode is confusing and doesn’t restore the old mental model.
Ctrl+Shift+J doesn’t even open settings on my machine. The product has become harder to configure, not easier.
Why this matters:
I’m not asking for “no review.” I’m asking for the review model that made Cursor better than alternatives — review at the granularity of the Agent’s proposals, not only at the end of a long session with hundreds of lines in one Review panel.
Without that, I lose trust in every Agent run. I can’t tell whether logic is broken until I’ve already absorbed a huge diff or dig through Git. That’s the opposite of “AI pair programming with guardrails.”
What I want back (minimum):
Per-proposal or per-turn Apply/Reject — when the Agent edits code for a task, list the affected files/hunks and let me Apply or Reject before (or without) silently committing everything.
Clear separation between “preview / pending” and “applied to disk” — Diff View should not be a substitute for explicit acceptance.
Settings that are discoverable — if behavior moved to Auto-Run / Review / something else, name it consistently and document it. Don’t hide the old “always accept” behavior behind a popup and then remove the way to turn it off.
Honest threat (not a bluff):
Cursor’s edge was UX + visual diff + control. If you keep shipping Agent workflows that auto-apply and only offer session-level Review, you’re giving up the reason many of us didn’t stay on plain VS Code or jump to other tools. I’m already looking at Codex and other options. I don’t want to leave — I want the product I paid for back — but I won’t keep paying for an Agent that writes first and asks questions later.
Please treat this as high-priority product feedback, not a niche power-user request. A lot of us build production code; we need incremental, visual, reversible approval — not a single “Review 1000 lines” button at the bottom of a chat.
Thanks for reading. Happy to share version info, screenshots, and repro steps if needed.
— A frustrated long-time user
Operating System (if it applies)
Windows 10/11