Passive aggressive responses

[sarob]

Describe the Bug

Increasing over the past couple of weeks, using auto with anthropic models, the response quality has gone noticeably down. Annoying, but I attempted to adapt with more targeted requests.

Now, the responses are becoming passive aggressive and flat out refusing to do work. Yes, hallucinations happen, where the code does not match the comments. This is different, where instead of writing code, the response removes code and leaves comments like “customer asks cursor to not to do ..” in an empty file.

Sure, this is human behavior. I blame the engineers that are managing the edit model behavior. They accidentally or on purpose coded this behavior into the edit model.

Steps to Reproduce

Passive aggressive behavior is occurring more often. I ignored it at first. Now it is affecting productivity. For example, the edit model refused to fix a bug even after closing chat and restarting cursor. So I quit and started up Windsurf. Problem solved quickly.

Expected Behavior

Hallucinations are expected from time to time. But I expect, the edit model to follow commands and not give me attitude.

Operating System

MacOS

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.2.4 (Universal)
VSCode Version: 1.99.3
Commit: a8e95743c5268be73767c46944a71f4465d05c90
Date: 2025-07-10T16:55:16.443Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin x64 22.6.0

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[mwjt42]

@sarob

using auto with anthropic models

This isn’t a thing.

It’s either Auto, or it’s set to a Model.

We suspect Auto typically uses GPT4.1, it may occasionally route to Sonet 3.5/3.7, but there’s no way to reliably know as the UI doesn’t inform the user.

That said, the Auto model(s) are incredibly passive, to use them I’ve had to add (and regularly remind it of) some very clear proactivity instructions.