Like the author of the issue, I had also noticed that we stopped using short names for constants in completion. Which puzzled me a lot given that we explicitly handle that and have unit tests to verify the behaviour.
After reading the specification again, I finally realized the issue.
When we split completion into completion + resolve, there was a mistake in the original PR. The resolve request should never mutate any existing fields of the completion item other than label details and documentation. But instead, we were creating a new completion item from scratch without including all of the original fields.
That means that the insertText we were specifying in completion was getting erased during resolve, which forces the editor to use the label instead, resulting in the behaviour of showing the fully qualified name.
Implementation
We never want to mutate anything in resolve that isn't label details and documentation. We should take the original item as a hash, change only those fields and then return it.
This fixes the issue of using the wrong sorting text and the wrong insertion text, which we're currently seeing in main.
Motivation
Closes #1983
Like the author of the issue, I had also noticed that we stopped using short names for constants in completion. Which puzzled me a lot given that we explicitly handle that and have unit tests to verify the behaviour.
After reading the specification again, I finally realized the issue.
When we split completion into completion + resolve, there was a mistake in the original PR. The resolve request should never mutate any existing fields of the completion item other than label details and documentation. But instead, we were creating a new completion item from scratch without including all of the original fields.
That means that the
insertTextwe were specifying in completion was getting erased during resolve, which forces the editor to use thelabelinstead, resulting in the behaviour of showing the fully qualified name.Implementation
We never want to mutate anything in resolve that isn't label details and documentation. We should take the original item as a hash, change only those fields and then return it.
This fixes the issue of using the wrong sorting text and the wrong insertion text, which we're currently seeing in
main.Automated Tests
Improved the resolve test a bit.