screen reader issue since 1.89.0 with open file CTRL+p

[dstpierre]

screen reader issue since 1.89.0 with open file CTRL+p

[TylerLeonhardt]

Looking at NVDA's log, I see this:

d
1 Results  
Search files by name (append : to go to line or @ to go to symbol)  list
abc.go dir1, file results  1 of 1
i
r
slash

with your example:

So we do announce the name, but the second you type another character that gets lost. In 1.88 it would re-say the file.

[TylerLeonhardt]

The change that caused this regression is here:
#208913

cc @joaomoreno

and locally, in main, I reverted that commit and the issue went away.

It looks like we were depending on that insert in order to trigger the a11y event to read things out.

Repro:

• I used NVDA on Windows
• have a workspace with the following structure:
  • dir/
    • abc.go
  • main.go
• Open the file picker (ctrl+P)
• type, slowly, d I r.
• 🐛 file abc.go only gets read out on the d and not the other typed characters

Here's the Quick Pick tree accessibility provider, if it helps:

vscode/src/vs/platform/quickinput/browser/quickInputTree.ts

Lines 284 to 314 in 6373d80

	class QuickInputAccessibilityProvider implements IListAccessibilityProvider<IQuickPickElement> {
	getWidgetAriaLabel(): string {
	return localize('quickInput', "Quick Input");
	}
	getAriaLabel(element: IQuickPickElement): string | null {
	return element.separator?.label
	? `${element.saneAriaLabel}, ${element.separator.label}`
	: element.saneAriaLabel;
	}
	getWidgetRole(): AriaRole {
	return 'listbox';
	}
	getRole(element: IQuickPickElement) {
	return element.hasCheckbox ? 'checkbox' : 'option';
	}
	isChecked(element: IQuickPickElement): IValueWithChangeEvent<boolean> | undefined {
	if (!element.hasCheckbox || !(element instanceof QuickPickItemElement)) {
	return undefined;
	}
	return {
	get value() { return element.checked; },
	onDidChange: e => element.onChecked(() => e()),
	};
	}
	}

And where we setChildren:

vscode/src/vs/platform/quickinput/browser/quickInputTree.ts

Line 1116 in 6373d80

this._tree.setChildren(null, elements);

I'm surprised by this behavior, esp since we don't use an identity provider anymore.

[joaomoreno]

The tree seems to be doing the right thing for performance sake: it is not modifying the DOM. Before #208913, the list row was thrown away and rendered again for no good reason. That had the nice side effect of A11Y being read as we desired.

Trying to find a way to kick NVDA to still read it, even though nothing has changed.