When you guys gonna get your shit together with VSCode/Copilot?

[BlackShadow]

It's just utter mess and i really started to doubt you people even working on this... People is not gonna come here and give detailed issues everytime. we don't have to. We're not gonna spoon feed you everytime, just fix your garbage for once Microsoft, we are paying money to this crap!

1. LLM models performing worse. Same LLM's performing much better in other AI agent IDE's like Cursor. VScode just creating utter garbage code, hallucinations everytime, constant hangs, it attempts to edit code and does nothing on return.

2. Shitty models in Agent mode. Why we still have fcking Sonnet 3.5? Google just released Gemini 06-05 few days ago and we still have 2.0 fcking Flash... Wake up! They released 2.5 Flash and Pro 06-05

3. Every agent model is terrible except Sonnet 4 and it will work fine IF you're lucky. Other IDE's like Cursor have Opus 4 Agent from day one. I love in agent mode Gemini Pro just goes ahead and deletes 700 lines of code and it asks why it's not working anymore.

4. Getting rate limited because your shitty LLM's can't do shit i have to prompt them every f*cking time. And it says stuff like "Ah, youre right let me fix the code" and proceeds to fuck it up even more...

5. Where's AI features like Memory that helps you properly utilizing codebase, learns from it's mistakes. Proper Thinking modes. Obviously VSCode doesn't have them.

[whtmst]

I also want to add that the models behave disgustingly and do not obey rules not globally defined in the profile folder, nor those within the project.

The only solution for models to strictly follow the rules is to set up an MCP memory server, feed the model all your rules (global and project-specific), tell it that it has extracted everything important and stored it in memory, and then for each task tell it something like:
Check the memory bank and perform ...the task...
In this case, it doesn't start getting lost in the project, and I'm talking about a simple, minimal personal website here :)

I won't say anything about competitors, but as soon as the trial period ends, and if the errors aren't fixed, I'll probably look towards your competitors.

[utapyngo]

I also want to add that the models behave disgustingly and do not obey rules not globally defined in the profile folder, nor those within the project.

I also noticed that it sometimes ignores some of the instructions. It looks like the instructions are summarized before sending to LLM. If it is the case, there should be an option to disable this behavior.

[whtmst]

I also noticed that it sometimes ignores some of the instructions. It looks like the instructions are summarized before sending to LLM. If it is the case, there should be an option to disable this behavior. @utapyngo

Oh, don't even get me started on summarization 😂 Sometimes it'll be working on a task, and halfway through, this summarization kicks in. After some waiting, 70% of the time the model either starts the task over, or continues but does something completely irrelevant.

The idea of summarization itself is great, but it probably needs some fixing 😏

And I agree with the previous speaker - right now, there's nothing to work with besides Claude 4. The other models seem to intentionally act dumb and do the exact opposite of what you ask, especially Gemini 2.5 Pro Preview.

[amunger]

Thanks for the feedback - We understand that it takes significant effort to write up quality bug reports or feature requests, and are grateful to those who file them. As you noted, there are plenty of areas for improvement in the current product and we are hard at work trying to make things better.