Do not force instructions on the user nor change user settings

[aeschli]

See microsoft/vscode#280665
We get complaints about unwanted instructions being loaded on every request.
Copying instructions in the user data folder is bad practice. Same as changing user settings.

There's the 'chatInstructions' contribution point that allows extension to provide instructions.

When doing that do not specify 'applyTo:**' as this will add to the context of every request, causing cost and vslow down.
Instead do not specify 'applyTo' and just define a description. The LLM gets a list of all descriptions and URIs and the agents can pick up instructions on demand..

[microsoft-github-policy-service]

Thank you for contacting us! Any issue or feedback from you is quite important to us. We will do our best to fully respond to your issue as soon as possible. Sometimes additional investigations may be needed, we will usually get back to you within 2 days by adding comments to this issue. Please stay tuned.

[MuyangAmigo]

@aeschli Thank you for reporting this, we will look into it. While we will figure out a way for not using applyTo:**, I want to share that in AI Toolkit we have setting that controls whether to append to instructions:

[Burhan-Q]

Funny story, the default for this setting once the extension is installed is "on" (meaning it's an opt-out setting). I disabled this in both User AND Workspace settings (since it was auto-enabled in both), and the behavior persists despite disabling and restarting VS Code.

[Burhan-Q]

This setting has led to some strange and undesirable LLM behavior (at least outside of using the AI Toolkit extension); to be clear the behavior coincided with the instructions being injected into the chat and does not occur in a workspace without the extension installed. It's likely that the extension will get uninstalled globally if this setting persists in this way, despite my appreciation for the extension otherwise.