copilot agent mode in vscode is doing too much for the wrong reasons leading to needless itterations undoing work #56
Description
Describe your experience
I like using copilot in vs-code to start a basic framework for a project to see if it makes sense and quickly iterate the idea, maybe get some specifics figured out. Then start actually looking into the code to find and fix things that are in need of change to correct common flaws one expects to see from AI code. Copilot often did a lot of side things that i never really asked for, but given the fact there where no real limits, it was fine to babysit and make copilot take out a lot of the things it added needlessly again. However it seems that lately Copilot has begun to more and more make larger assumptions about environments and such, like running on windows, thus creating batch files and setting up environment variables while the code is intended for a simple and standard shared-hosting environment. So you nudge and give hints of the environment hoping to correct and cleanup things, instead Copilot takes off, starts adding even more files and messy pointers trying to now make all the code windows and Linux compatible generating additional shell scripts where it installs Apache, sets up MariaDB and more. Given the target environment, half does not work as a simple user, the other half if it was not a shared host, would break the actual system if applied. though that would mostly not work as it mixes the tool sets for multiple Linux flavors ( things like apt vs dnf or initV and systemd ), so you end up asking agent mode things and then spent about 5x the amount of requests to get most of the mess made out of the code again. While in the past this was more of a free thing that came with "Playing with AI" and the experience was helping Copilot learn and grow, the now introduced limits make this a fairly unfair deal. One would say, then switch to non premium models, while this would be a good option, doing so seems to often end up loosing context and make things worse, or break the few things that work resulting in ping-pong between models to make, break, adjust, fix, bend and eventually getting somewhere, even it that somewhere is back to the initial half mix of environments.
Suggestions for improvement
I am not sure how one could make this better overall, maybe have Copilot ask a few questions about a target environment before agent mode runs off to iterate and make a lot of additional things like init or systemd service files that where never asked for, then at least those would fit the target environment. As it is these extra things take away from the helping hand i think AI should be, It is not bad per definition, though for every bit of helpfulness, it also throws in some noise tipping the balance negatively. The limits on premium requests then make it feel even more like a potential waste. Hitting undo is also not helpful, as now you wasted the request, and have no code either.
Additional context
The feedback is not so much intended as a complaint, I am aware that the computations are not cheap so limits need to be applied, though if those limits then get used up on undoing what is done, it is wasting resources for me in the sense of limited premium requests, and for GitHub in making the initial computations to generate too much, only to then do more generations to undo nearly half of the work done either for more premium requests, or using the non-premium requests a lot more to take small targeted changes in larger quantities as their complexity is often more limited.