Stylable <select> element

[josepharhar]

Edit: I am proposing stylability and other improvements for the <select> element rather than a new element based on the discussion in this issue. The explainer has been updated.

Original description below:

I'd like to propose the new <selectlist> element, as described in this explainer: https://open-ui.org/components/selectlist/

This element still has a lot of open issues as listed in openui, but for now I'd at least like to know if people agree that this is a problem worth solving. The problem is that the existing <select> element is not customizable in terms of appearance or behavior which leads web authors to build their own and miss out of the usability, accessibility, and autofill features of builtin elements like <select>.

[zcorpan]

@whatwg/forms

[annevk]

Thank you for bringing this proposal to the WHATWG Joey! I thought this would be a good opportunity to outline how colleagues and I feel about extending HTML in this area. In particular, we feel that new and existing form controls:

• Should look the same as the operating system controls by default.
• Should be fully styleable by web developers.
• Should generally attempt to follow existing HTML element patterns.
• Should not be redundant with existing HTML form controls.
• Should work on a wide variety of platforms, including those with very small screens, no mouse or touch, etc.
• Should be fully accessible.
• Should not have any l10n or i18n shortcomings.

We understand that the select element can’t address a variety of scenarios due to parser limitations, but the select element could address them in combination with the datalist element. One of our big worries with complete duplication is that we end up not solving the problems with the existing controls and that the duplicated controls will have a variety of shortcomings the older controls did not have.

(Aside: #5791 seems very related to this issue.)

[hsivonen]

I think the explainer doesn't explain the need for a new element in a convincing way, because the multiple examples look addressable in the context of select: flags are addressable by emoji instead of image files, subheadings by the label attribute of optgroup, and color and font-level styling seems feasible without a new element.

(From off-explainer examples, I gather that the ambition level of the feature is much higher than the explainer makes appear.)

[josepharhar]

Thanks for taking a look and discussing with colleagues Anne!

Should look the same as the operating system controls by default.

If we made appearance:none or a new appearance value like appearance:base as discussed here and here have a standardized style, then the default appearance:auto appearance could look the same as OS controls. That would be up to each UA to decide, as it is currently for all other form controls.

Should be fully styleable by web developers.

Agreed! Styleability is the primary driving use case for the current proposal.

Should generally attempt to follow existing HTML element patterns.

Agreed! Domenic left some feedback in this vein and we completely revamped the proposal to use existing HTML element patterns by using new tag names instead of slot and behavior attributes.

Should work on a wide variety of platforms, including those with very small screens, no mouse or touch, etc.

Agreed! Existing elements conform to this by letting the browser/OS provide a native picker which unfortunately can’t be styled or customized by web developers, right? It was recommended that we let the browser use native pickers for selectlist on mobile here. Ideally developers would be enabled to make their own pickers that work on these platforms when they want to as well.

Should be fully accessible.

Agreed! This is also a core requirement for the proposal. I have been working closely with accessibility experts participating in OpenUI to ensure that the accessibility mappings and keyboard behaviors follow the current best practices in ARIA.

Should not have any l10n or i18n shortcomings.

Agreed! The content is fully customizable/replaceable, so it should be localizable and internationalizable, right? The only thing I can think of is supporting right-to-left and vertical writing modes when rendering where the listbox popup goes, which is something we are doing for other form control elements and we can easily do via Anchor Positioning for selectlist. Is there anything else you had in mind here?

Should not be redundant with existing HTML form controls.

We understand that the select element can’t address a variety of scenarios due to parser limitations, but the select element could address them in combination with the datalist element

It sounds like yall would prefer to reuse the existing <select> element instead of creating a new tag name. Am I correct? As you mentioned, we can’t have the new content model exist inside the <select> element due to parser rules, but we could try to do something with how datalist works by adding attributes to <select> which refer to the other elements we want to use, which would perhaps look like <select list=my-listbox button=my-button>. However, I don’t think this would work well. This works for <input list=datalist> because datalists are just used to host text data without doing actual rendering. Since we are actually trying to render an authored <listbox> or <button> element inside the select/selectlist, we don’t have the machinery in the browser to render them from an arbitrary location in the DOM. ShadowDOM is perfect for this and is how existing builtin elements like details and summary work. It is also easier and more idiomatic (like <option> and <summary>) for the new listbox and button elements to know when they should have special behavior for selectlist because they can look at their parent context rather than having to keep track of which select element has an attribute pointing at them. A preview of the State of HTML survey I’ve seen also lists <datalist> as a top developer pain point, which perhaps alludes that the ergonomics of referring to a separate element aren’t great.

One of our big worries with complete duplication is that we end up not solving the problems with the existing controls and that the duplicated controls will have a variety of shortcomings the older controls did not have.

I agree that duplication should be avoided. Selectlist is not a complete duplicate of select. <select> can draw outside the bounds of the browser window, and <selectlist> has a completely different rendering model that allows full customization. If it were possible to shoehorn into <select> that would be nice, but it looks infeasible.

[josepharhar]

I think the explainer doesn't explain the need for a new element in a convincing way, because the multiple examples look addressable in the context of select: flags are addressable by emoji instead of image files, subheadings by the label attribute of optgroup, and color and font-level styling seems feasible without a new element.

Thanks for the feedback! I am improving the examples here: openui/open-ui#918

[lukewarlow]

flags are addressable by emoji instead of image files

I just wanted to point out Windows doesn't provide flag emojis so this specific example is a good use case for needing images not emoji.

[annevk]

Existing elements conform to this by letting the browser/OS provide a native picker which unfortunately can’t be styled or customized by web developers, right?

If you use appearance:none they can usually be styled. And we should do more work here to add something akin to appearance:base to make that more ergonomic for web developers. I don't think the solution should be adding new elements that don't have platform theming by default as that would make the HTML language inconsistent.

ShadowDOM is perfect for this and is how existing builtin elements like details and summary work.

For <datalist id=options> we'd have to clone the options into the select element's shadow tree for rendering purposes, I think, much in the same way we probably have to clone one of the options as the selected option. As long as this happens in the internal shadow tree it's largely not observable and should be okay.

It would also have the benefit of having a consistent design with <input type=text list=options>, so we'd tackle combo boxes at the same time (or at least make it possible to reuse the design for them). Since the <datalist> continues to mainly serve as a source of data I don't think there's any need to track what element is pointing to it.

(There are some challenges here with how to match these options and their elements from CSS, but that seems like a surmountable problem.)

I haven't looked at survey in detail, but I'd assume that if <datalist> is mentioned as a pain point it would mainly be around not being able to style it, which is something we'd have to tackle for combo boxes and select alike.

[josepharhar]

I’ve been discussing with @annevk and @pxlcoder about how to move forward here, and we have reached agreement on some things. Please correct if I misrepresent anything!

• Reusing the <select> element will work
• We can change the parser for <select> to allow particular new child tags like <button> and <datalist>
• We can work incrementally, first by making these parser changes in the spec etc.
• Using <datalist> as a child of <select> will work to replace the listbox with custom content

[lukewarlow]

I'm concerned by the implications of the multiple attribute. How does it work with the new parsing?

That being said I am excited by the progressive enhancement possibilities.

[gregwhitworth]

If you use appearance:none they can usually be styled. And we should do more work here to add something akin to appearance:base to make that more ergonomic for web developers.

@annevk The majority you cannot and it isn't solely about styling but also extending them. While select may be able to be re-used; one of the benefits of introducing the new elements that are interoperable by default and not OS platform styled does make the platform consistent rather than trying to retro-fit all the build-in UI controls. I think we may want to fork this into a more generic issue specific to the introduction of new elements as I'm curious how your assertion holds up for input type=switch that not only allows styling but extensibility without a new element; type=file, type=progress, etc?

Using <datalist> as a child of <select> will work to replace the listbox with custom content
This makes sense to me outside of my position that we should introduce a new element, not a new CSS attribute to solve this problem as it's not just about styling

@josepharhar @annevk so will this likewise have the child of <selectedoption> and ability to modify the initial shadow DOM?

[josepharhar]

@josepharhar @annevk so will this likewise have the child of <selectedoption> and ability to modify the initial shadow DOM?

The idea (as I see it) is to still have a <selectedoption> element that can be a child of the <button> element we use to replace the <select>'s default button.

The UA shadowroot will certainly adapt if and when needed to support slotting in and replacing the default button and listbox.

Does that answer your question? I'm not sure if I understood it well.

[josepharhar]

Adding agenda+ to move this to stage 1

[josepharhar]

I updated the explainer to be <select> instead of <selectlist>: https://open-ui.org/components/selectlist/

[annevk]

To me that does not seem like the right approach. That's still the line in the sand approach I referred to above. Showing options as a "popover" should purely be the result of appearance:bikeshed and apply to the <select> element as it stands today. The same goes for styling the currently-selected value. appearance:auto might not allow for that, but appearance:bikeshed does.

I'm not sure what's lacking about keyboard focus, and a11y, but it seems like that should also be a universal improvement.

[brechtDR]

@annevk said:

One of our big worries with complete duplication is that we end up not solving the problems with the existing controls and that the duplicated controls will have a variety of shortcomings the older controls did not have.

I just want to note that if this was one of the big worries, that it should be out of the way with an attribute as well.

It does feel as @mfreed7 mentioned that an attribute is likely the more natural pattern and that children deciding parser behaviour feels like an anti-pattern (declaring the intent twice, based on children). Parsing based on children could also lead to more author errors.

[annevk]

Why would it be out of the way with an attribute? An attribute as you all have envisioned creates the exact kind of mode switch we'd want to avoid.

I also don't follow what you're saying about the HTML parser. Perhaps you can describe the dilemma you see there in more concrete terms? The entire parser is structured around branching on tag names which is exactly what we'd be doing here. It wouldn't affect the parent one bit. And as I mentioned above it seems worthwhile investigating if we can allow things like <select><option><img> Some text</select> as well. No reason to require a <datalist> wrapper for a simple enhancement like that if we can avoid it.

[brechtDR]

Perhaps you can describe the dilemma you see there in more concrete terms?

It wouldn't affect the parent one bit.

Even if it doesn't really affect the parent in a technical way. It feels like I'd be telling the browser: I want a select, and then telling again in the children: I still want a select, but I want it styleable.

This is a complete different way than saying it inside of an attribute, that is what authors are used to. Example: I want an input and I want it to be a (type)date. aka: do those things for me based on the attribute.

This would be the equivalent that HTML should be like this:

<input type="date" />
vs
<input><datalist>datepicker things</datalist></input>

<select bikeshed="bikeshed" />
vs
<select><datalist>Now it's styleable</datalist></input>

Note that a select without any options in, still looks like a select. So in that case, children don't matter. Except it won't show a listbox.

(Exaggerated example, but wanted to make a point of how it feels)

I'm sorry if I can't give you any more technical reasons as I'm not a browser engineer, just an interested front-end developer (from what I hearing, both ideas are technical possible). I think I made my point as it is about consistency and general tutoring and feeling (yes, I believe that last one can be important). I would be utterly astonished if I would be the only developer thinking that way. Consider this "user feedback" from me. Maybe it could be asked around in developer communities a bit more.

[annevk]

You're still thinking of a mode switch. I'm saying that the select element we know today should be styleable as well in the exact same way. There's no before and after. The parser changes will enable developers to include more types of descendants, but that's it. Everything else should apply universally.

[mfreed7]

You're still thinking of a mode switch. I'm saying that the select element we know today should be styleable as well in the exact same way. There's no before and after. The parser changes will enable developers to include more types of descendants, but that's it. Everything else should apply universally.

Given the very high usage of the existing select, with its current behaviors and idiosyncrasies, I would have serious doubts that we could just change all of them in a way that would be web compatible. There will need to be an opt-in.

[gregwhitworth]

Given the very high usage of the existing select, with its current behaviors and idiosyncrasies, I would have serious doubts that we could just change all of them in a way that would be web compatible. There will need to be an opt-in.

Agreed, there are bound to be sites that have complex content that currently we're throwing away for a variety of reasons and would not be backwards compatible.

An additional thing to consider is that this has to be contained within the viewport for security reasons. So there has to be some opt-in to get this new behavior.

I want a select, and then telling again in the children: I still want a select, but I want it styleable.

@brechtDR just a reminder that this solution we're discussing is solely for changing the DOM structure and its defined behaviors on that structure. The opt-in to make it base styles so that you can interoperabley style it will come from the CSS property.

[annevk]

The opt-in for different styling (within viewport) is appearance:bikeshed as now oft-repeated. And I think you have to be more concrete on the other roadblocks you see as it seems feasible to me and a much better model for web developers (and end users) overall.

[lukewarlow]

Isn't the viewport security model relevant to new parsing too? If we're allowing arbitrary content isn't that bad enough without full styling?

[annevk]

User agents get to decide what they propagate. I would imagine we would not let anything through but images and text. The richer data structure ultimately needs to be able to flatten to text options for a11y and cross-platform purposes anyway. That happens to also work well for native popup windows.

[gregwhitworth]

User agents get to decide what they propagate. I would imagine we would not let anything through but images and text. The richer data structure ultimately needs to be able to flatten to text options for a11y and cross-platform purposes anyway. That happens to also work well for native popup windows.

I feel like we're going in circles and possibly conflating the issues since you can theoretically add support for images on their own in select outside of aligning on a standardized DOM and behavior structure that would require it to be rendered within the viewport. So if we want to have that conversation maybe we should open a separate issue @annevk?

Regarding the below comment @annevk: Apologies I am still confused on the "line-in-the sand". We'll need to fix up the DOM in some manner that is backwards compatible. The Open UI CG decided on attribute, am I understanding your position correctly that you only want a CSS solution for that?

To me that does not seem like the right approach. That's still the line in the sand approach I referred to above. Showing options as a "popover" should purely be the result of appearance:bikeshed and apply to the element as it stands today. The same goes for styling the currently-selected value. appearance:auto might not allow for that, but appearance:bikeshed does.

[annevk]

@gregwhitworth I'm not really sure what you're saying. The DOM can already be whatever (using append() and friends or XML). However, not through HTML syntax. That part requires HTML parser changes. Making those changes doesn't require an attribute.

[gregwhitworth]

@annevk It seems we're speaking past each other at this point. My goal was to try and get back to @mfreed7 question and your answer seemed to state that no changes are needed to the HTML specification and we only need a CSS proposal.

[annevk]

No? We need HTML parser changes. And if we're going to allow more elements we also need various semantic changes and possibly API changes. I just don't see the need for a mode switch / line in the end / whatever on the HTML side. There's no need for an old and a new select element.

[mfreed7]

@gregwhitworth I'm not really sure what you're saying. The DOM can already be whatever (using append() and friends or XML). However, not through HTML syntax. That part requires HTML parser changes. Making those changes doesn't require an attribute.

One important thing, for compat, is that doing this with append() currently doesn't render properly on any browser I checked. E.g.

  <select id=s></select>
  <script>
    const div = document.createElement('div');
    div.innerHTML = '<option>Hello<option>there';
    s.appendChild(div);
  </script>

With your current proposal, that would start rendering the options, I believe. Because the parser would now allow HTML with the equivalent wrapper <div>.

[annevk]

I don't think that's true. We will need to define the semantics of the select element and when and how option elements have meaning. We might well decide that we ignore select element children other than option/optgroup/hr/datalist (and maybe button?). (I also don't think I suggested we should change the parser behavior for <select><div>.)

Let's try what @domenic suggested in Chat and explore the different pieces to the select element and how they all need to be thought through carefully:

1. Parsing: how do we enable richer data structures given that currently the "in select" insertion mode drops most start tags on the floor. I think here we should look into allowing the "datalist" start tag (and maybe "button"?). "img" as a child of "option" would also make sense to me. "datalist" (and maybe "button") would have to trigger some insertion mode that allows more arbitrary children.
2. Semantics: how do we extract options and option text out of these new structures. We'll need this for the existing API, for "native appearance", etc.
3. Styling: a new "base appearance" has to be opt-in and has to deal with the existing data structures and the new richer data structures. An internal shadow tree is also relevant here. I don't think an internal shadow tree is necessarily relevant for the "native appearance", as we'd just extract text options (possibly with images and some slight formatting?) for that. But an internal shadow tree will also be relevant for appearance:none. It might well be possible that this internal shadow tree is always constructed. Requires some more exploration.
4. Keyboard and a11y behavior: these should work well for the existing and new data structures. Just like we don't want to regress on either for new controls, we also don't want to leave older controls behind.

[mfreed7]

1. Parsing: how do we enable richer data structures given that currently the "in select" insertion mode drops most start tags on the floor. I think here we should look into allowing the "datalist" start tag (and maybe "button"?). "img" as a child of "option" would also make sense to me. "datalist" (and maybe "button") would have to trigger some insertion mode that allows more arbitrary children.

Ok, that might alleviate some of the compat concerns. We're going to try to measure how often existing <select>s contain <button> and <datalist> to gauge the impact.

However, the primary #1 use case is support for arbitrary content. It might be ok to restrict the direct children of the <select> to a few things (like option, button, datalist), but it would absolutely break the use case to add restrictions to what can live inside the <datalist> or <button>, including within the contained <option>s.

2. Semantics: how do we extract options and option text out of these new structures. We'll need this for the existing API, for "native appearance", etc.

Agreed. Likely just textContent but there are nuances for sure.

3. Styling: a new "base appearance" has to be opt-in and has to deal with the existing data structures and the new richer data structures. An internal shadow tree is also relevant here. I don't think an internal shadow tree is necessarily relevant for the "native appearance", as we'd just extract text options (possibly with images and some slight formatting?) for that. But an internal shadow tree will also be relevant for appearance:none. It might well be possible that this internal shadow tree is always constructed. Requires some more exploration.

Agreed. There was a concern raised by Apple - that conditioning the shadow tree based on CSS property values might be difficult to implement. Is that not the case any longer? (We are going to try to prototype it in Chromium to see what breaks, but we haven't done that yet.)

4. Keyboard and a11y behavior: these should work well for the existing and new data structures. Just like we don't want to regress on either for new controls, we also don't want to leave older controls behind.

Here, doing this is likely web-incompatible, since the behavior changes will be quite observable. Did you read through the various behaviors? 1, 2, 3, 4, 5, 6, 7, 8. Many of those differ from existing browser behavior, for at least some browser/platform permutations. My guess is that many of those changes would need to be opt-in. I'd be happy to be proven wrong here.

[annevk]

1 — I don't think I said otherwise?

3 — I'm not sure the concern was raised by Apple? The minutes suggest Emilio brought it up. In any event, we went over that already here: w3c/csswg-drafts#5998 (comment). I'm not sure why you keep coming back to a 2021 set of minutes when there's been ample discussion since.

4 — Those will need further discussion looking at those issues. In particular for the "native appearance" we would not want to deviate from platform conventions and at least thus far we haven't been that prescriptive on specific end user interactions.

[mfreed7]

1 — I don't think I said otherwise?

Oh good! I think I misunderstood this comment of yours: '"img" as a child of "option" would also make sense to me'. I took that to mean only img would be allowed as a child. I suppose based on this comment you meant only when the <option> was outside of a <datalist>. Good, thanks.

3 — I'm not sure the concern was raised by Apple? The minutes suggest Emilio brought it up. In any event, we went over that already here: w3c/csswg-drafts#5998 (comment). I'm not sure why you keep coming back to a 2021 set of minutes when there's been ample discussion since.

The comment you made immediately after the one linked above says "WebKit at least hasn't fully ruled out that option", which isn't the same as "WebKit supports this option". You also said "I think the details of the CSS-based solution are still up for debate." I didn't see any more conclusive comments - please link them if I missed it! If you read my comment, I'm just trying to get some directional clarity before we invest further in changing the behavior of the prototype again.

Keyboard and a11y behavior: these should work well for the existing and new data structures. Just like we don't want to regress on either for new controls, we also don't want to leave older controls behind.

4 — Those will need further discussion looking at those issues. In particular for the "native appearance" we would not want to deviate from platform conventions and at least thus far we haven't been that prescriptive on specific end user interactions.

I think perhaps I misunderstood your initial comment as saying you wanted the keyboard and a11y behavior to work the same in "native" and "interoperable" mode. If not, great.