Github Copilot + GPT issues when working on UE5

Development Getting Started & Setup
1

Hi

I am experiencing an issue when working in configuration VS2022+GitHub copilot Agent+GPT-5.6+UE5 project. Everything is updated.
Github enterprise account with blocked function “Suggestions matching public code”, and disabled: “Copilot Memory”, “Copilot CLI”, “Copilot in GitHub Desktop”, “GitHub Copilot app”.

“Copilot Chat in the IDE” is enabled, and this is a feature I use (thru built-in into VS2022 Github Copilot Chat tab).

When starting new chat first one-two answers are ok, but subsequent is nearly always cut to first sentence, or brake in a middle without any message.
I suspect it may be caused by context run out, but tasks are no very complex, but possibly LLM goes deeply into UE5 source code and bloats it own context.

My project is moderate in size (about 60 source code files + some blueprints), I removed UE5 hooks from .sln, but it not helps.

It is still a hypothesis, because there is no messages, as well there is no signs of context compression etc. LLM just starts to return with single sentences. Also it uses minimal amount of tokens for such minimal responses.

I don’t know how to control an context size or how to debug this issue.

When prompting “please continue” etc. in most cases is still return short, incomplete answer, but from time to time it works well…

It do not happens when I am using Claude LLM, so I switched LLM as a workaround now, but I would like GPT as I think it is better and cheaper.

Any ideas how to fix it?
Are you experience similar issues in similar configuration?

2

Hello there @HellFireHE!

Since everything works fine with Claude, using the same setup and project, I think we can rule out UE and your project as the cause. The problem is more likely in how Copilot handles the conversation with that specific GPT model.

As for your hypothesis, it would make sense, as large batches of code will pile up eventually. Considering that, the first thing I would test, is to start a fresh chat right after the issue takes place. Open a new chat, and send the same request. If it answers normally, the problem is the context building up during the conversation.

Alternatively, try switching to Ask mode in Copilot. That mode won’t search through your code on its own, so it will use less context. If the answers come out complete there, that points to Agent mode getting overwhelmed by the ammount of code.

You can also check Copilot’s logs, inside VS, at View > Output, pick GitHub Copilot from the “Show output from” dropdown. If something is cutting the responses short, it could show up there, even when the chat shows nothing. If the fresh chat test fixes it consistently, I could be worth reporting the scenario directly to GitHub, or using the Copilot feedback option in VS.