I just thought of something. This whole thing is going to break de-duplication that we do for completions we get back from the jupyter kernel.

Right now, since the jupyter extension creates its own pylance, we can get the completions from pylance and remove those from jupyter that are the same thing.

After this change, the pylance server won't be started by jupyter and we won't be able to ask it directly for completions.

We need some way to find the specific completion provider (or the language client). So I think we'd need yet another API to pass the language client through to jupyter.

We could do that after this though.

This is the line where we ask pylance directly for its completions:
https://github.com/microsoft/vscode-jupyter/blob/8f7f38e2bc664836af236e8e2ad9f1b85130c3c9/src/intellisense/pythonKernelCompletionProvider.node.ts#L222

We'd need some way to get to this language client from jupyter still (well unless VS code has an API to return a specific completion provider but I don't think it does).

Originally posted by @rchiodo in #19087 (comment)