We forbid declarations after nested rules

Seems potentially problematic that some people may have some garbage followed by lots of declarations, then at some point we add some feature that parses the garbage as a nested rule, and then all the declarations stop applying. Spooky action at distance.

One of the main characteristics of option 3 (later modified with lookahead) is that it allows freely mixing declarations and nested rules. So at this point I don't think it makes sense to restrict that. The proper way to have such restriction would have been choosing option 4 or similar (which BTW seemed better to me).

Right, I argued in previous discussions that forbidding decls after rules is fundamentally problematic. The basic question is "when has a rule occurred?", and this needs to not be dependent on whether a rule is valid or not, as that would mean different browser levels would interpret following properties differently.

So you need to define some notion of when we've found a rule, that allows for invalid rules, and do so in a way that doesn't unduly restrict our future syntax options.

For example, if we ever allow {} in a property, and you write it invalidly so it triggers rule parsing, would that kick the "now there's a rule" switch? Does that mean we actually have to forbid ever using {} in properties?

On the other hand, sticking with the current design that just allows them, and relying on people to not write unreadable stylesheets, suffers from none of these issues.

Not saying it's necessarily a good idea, but playing devil's advocate a bit, we already have that kind of concept for e.g. @import / @namespace / etc, don't we?

Imagine someone wrote this ten years ago:

:has(body) {}

@import "something.css";

(Or something along those lines)

Their @import would stop working on browsers that support :has(), yet that hasn't been ever a concern (and I don't think it should be).

The switch for that is "A valid rule has been parsed", and I think we could do the same, so that random garbage doesn't cause that action at a distance.

I think these are pretty different situations in practice, tho. We add to the set of "things that can go before @import" super rarely -- in fact, we've only done it once, with @layer.

On the other hand, we add new things that would qualify for nesting fairly regularly - we recently added @scope, and just resolved to add @initial-name-tbd. So the set of things that may or may not trigger "no more properties from here on" is regularly changed, presenting new hazards.

Ultimately, the question is still - what are we trying to protect authors from? This sort of interleaving has been allowed in Sass and related tools for many years, with the exact same "pretend all the properties are together at the top" behavior, and never caused notable issues.

For example, if we ever allow {} in a property, and you write it invalidly so it triggers rule parsing, would that kick the "now there's a rule" switch? Does that mean we actually have to forbid ever using {} in properties?

If you mean the property value and I interpret the algorithm correctly, then {} is already allowed by consuming a component value in step 5 of consuming a declaration. And it can't trigger rule parsing because the consumed component value is appended to the declaration's value.

Sebastian

This is in reference to the resolved-on new parsing algo that tries to parse as a decl, then falls back to parsing as a rule if the result wasn't valid. That algo isn't in the Syntax spec quite yet.

That algo isn't in the Syntax spec quite yet.

Should we expect more changes to come (in CSS Syntax 3, at least) after this commit?

color: red; @media { foo: bar } color: green does not seem to parse as I would expect against <declaration-list>, probably because there is no ; after foo: bar. I think } should be provided as a stop token to consume a declaration.

Thanks for spotting that! I hadn't gotten around to updating my own parser to the new text to test it, so I missed that I was mishandling nested constructs.

I believe it's all good now; the three construct consumption algorithms (at-rule, qualified rule, declaration) now all take a "nested" bool, which triggers them to stop early when encountering a top-level unmatched }. This is passed by "parse a block's contents" (which is renamed from "parse a declaration list" since it definitely returns more than just declarations).

<declaration-list> is defined with consume a list of declarations and the procedure to parse the input of CSSStyleDeclaration.cssText is defined with parse a list of declarations. Both procedures are removed.

I do not mind waiting for an update to your parser to validate your updates. This would prevent me from polluting this issue with implementation details.

That said, <declaration-list> could be renamed to <statement-list> and could be parsed with consume a list of statements.

This would allow using <declaration-list> for when a list of declarations (strict) is required (eg. keyframe rule, @font, etc). But I do not know if it would be backward compatible to apply it to existing rules, and your intent may even be to preserve this flexibility to accept declarations/rules within any rule, for future extensibility.

I've done the parser update, but had to stop at end of day before I could get my testing framework updated to the new data structures. Later today I'll have it working. ^_^

But I do not know if it would be backward compatible to apply it to existing rules, and your intent may even be to preserve this flexibility to accept declarations/rules within any rule, for future extensibility.

Yes, all rules already handled at-rules in their blocks even if they only accepted declarations validly, so that's staying, and at that point there's not really any reason to continue having a parsing split. All blocks are parsed the exact same way now in Syntax, accepting all three kinds of constructs (at rules, qualified rules, declarations).

My plan on reshuffling the productions is just to define one generic production, and then probably some sub-productions that automatically imply certain restrictions on what's valid so you don't have to say it in prose. But they won't change the parsing behavior.

Unfortunately, it appears that Chrome shipped before we could decide on this, and now authors are already blogging about the "gotcha".

Hopefully it's not too late to change this.
I think we should really try to avoid any rewriting that changes order of declarations.

Since we’re already resolved that nested MQs will wrap their contents with an implicit & {}, perhaps we can do the same thing with any declarations after the first nested rule, as @emilio proposes in the OP?

I still feel very strongly that we should not change this, and the current behavior is the best. Again, this is the exact behavior that Sass (and I suspect other preprocessors) have had for a decade+ already, and it has not been a problem there. (Largely because people just don't write that code - they put their declarations first, then their nested rules.) I don't think we should try to add more behavior for something that has proven itself to not be a problem in practice.

MQs wrap all naked properties in an & {}; style rules obviously cannot do this, so the behavior would be inconsistent either way. As the spec is currently written, style rules and MQs are each internally consistent with a single behavior for all naked properties. I think it would be a (probably minor) bad thing for style rules to have two behaviors, depending on relative ordering of rules and declarations.

While I would like to see a world where the order would stay as written, I think I agree with Tab on that this is the behavior that was implemented in all preprocessors: less, sass and stylus, at least some CSS-in-JS solutions (tested in styled-components, couldn't find a good playground to share), and nested PostCSS plugin (can be tested here), lightningCSS (but it polyfills the current spec).

Given this was done literally everywhere, I'd say it is not worth it to change this.

The CSS Working Group just discussed [css-cascade] [css-nesting] Figure out whether we're fine with "shifting up" bare declarations after rules.

I thought it was very odd that the WebKit poll seems to be favoring Option 1 (shifting up), so I posted a couple more polls:

• Twitter/X
• Mastodon

which so far seem to be showing a very different picture with more than 3 out of 4 expecting no shifting. There are even people incorrectly explaining what happens (1 2)

For the minority that expects the current behavior, it often seems to be due to misconceptions around how specificity works (1 2 3 4 5). Others mentioned that while not shifting feels more natural, shifting has been beaten into them by existing implementations (e.g. 1 2 3 ).

There is also this older poll but it's phrased as a quiz, which influences the data, as people expect the result to be weird. Even so, it shows an even split between shifting and no shifting.

I think the current behavior is extremely unintuitive, and 10 years down the line we will regret opting for consistency with today’s preprocessors over predictable, natural behavior. In fact, we even have a TAG principle advising against this exact thing: Prioritize usability over compatibility with third party tools.
I'd even vote for entirely disallowing declarations after rules over the current behavior, as it would give us more time to figure this out.

I’m quite concerned that wording seems to influence the result so much, as it indicates we don't yet have a good picture of author expectations. Perhaps we need more data here. Maybe an MDN short survey would help?

Agenda+ to resolve to temporarily disallow declarations after rules while we try to get better data here, so that we don't get stuck in a situation where we can't change the behavior due to web compat.

FWIW we (WebKit) were also very surprised by our poll results (it actually changed over the weekend, the first days were Option 2 winning 70/30).

Disallowing declarations after rules seems the "worst" solution because it means we have to determine the "a rule have been parsed" trigger which could cause some compatibility issue if we extend what correctly parse as a rule in the future. However, I'm not sure what would be the risk if we allow declarations after rules like now and just follow the cascade so last one wins?

Might be good to resurface/link the CSSOM aspects of this as I don't see any mention of those in this issue.
If I recall correctly it was an issue for CSSOM if declarations and rules could be interleaved.

(maybe there is a clever way to resolve those?)

Disallowing declarations after rules seems the "worst" solution because it means we have to determine the "a rule have been parsed" trigger which could cause some compatibility issue if we extend what correctly parse as a rule in the future.

Could you please elaborate on that? Not sure I follow what you mean by "a rule has been parsed trigger".
FWIW I don't think disallowing declarations after rules is a good long-term solution; I'm only proposing it so that we don't back ourselves into a corner while we deliberate and time passes.

However, I'm not sure what would be the risk if we allow declarations after rules like now and just follow the cascade so last one wins?

I think the argument against that is inconsistency with preprocessors (and perhaps the existing Nesting implementations? Though usage of these in the wild right now is effectively nil, and even smaller where this changes things). Not sure if there's any other counterargument.

Might be good to resurface/link the CSSOM aspects of this as I don't see any mention of those in this issue. If I recall correctly it was an issue for CSSOM if declarations and rules could be interleaved.

(maybe there is a clever way to resolve those?)

From what I remember, we resolved that by resolving that bare declarations can be wrapped with & { ... }, though I can't find that right now.

Could you please elaborate on that? Not sure I follow what you mean by "a rule has been parsed trigger".

Yeah, I've elaborated on this in the past when Emilio suggested disallowing properties after rules.

So, this is a parsing switch. In theory parsing switches are doable; we explored a few of these earlier when working thru Nesting options. But the switch needs to be reliable - the earlier discussion about @nest; being the switch worked, because we know exactly what to look for and don't expect that to change in the future.

But if the parsing switch is "a rule of some kind is seen", that's hard. The most obvious interpretation of that is "a valid rule of some kind is seen" - that's well-defined, but it means that authors can see unexpected differences in parsing behavior if the rule in question is supported in some browsers but not others, as older browsers will throw out the rule and continue allowing declarations, while newer browsers will see it and disallow declarations. (And consider: an invalid selector makes the style rule invalid, and we do new selectors all the time.)

So ideally we define invalid rules as also triggering this. But then we open up the full syntax space of what an "invalid rule" actually is. Is foo bar:baz {...}; an invalid rule? Or is it a new property syntax? We can define something for this, but it means restricting our future evolution capabilities, to a larger extent than what we've already done for Nesting.

FWIW we (WebKit) were also very surprised by our poll results (it actually changed over the weekend, the first days were Option 2 winning 70/30).

Have the results been restricted? I remember seing opposite numbers yesterday.

I watched the results of the poll very carefully in the first couple days. It was consistently 37-39% for Option 1, and 61-63% for Option 2.

It's very telling when you hit enough votes (like 50 or 100) and the results stabilize. As more and more votes came in, the results did not change.

Sept 28 at 2:50pm ET.

Sept 29 at 12:50pm ET.

Then over the weekend, another 1800 votes came in, with an overwhelming preference for Option 1. Almost three times as many votes came in over the weekend? Long after we stopped promoting the article on social media? With a radically different result?

That's a bot.

So we decided to close the survey and post the last known results from before the traffic pattern became highly suspect.

You can see a similar preference for Option 2 in Lea's survey: https://mastodon.social/@[email protected]/111177156433448874

@emilio Either CSSGroupingRule.cssRules returns a list of CSSRule + CSSStyleDeclaration or we use a new CSSGroupingRule.elements or something else.

It's really the same idea, except that we use CSSStyleDeclaration instead of a potential @nest. But it has the advantage of not being explicitly writeable by a user (you can't write distinct consecutive CSSStyleDeclaration blocks in CSS - but you can if you really want with CSSOM).

(For the example above, merging the two CSSStyleDeclaration would makes more sense indeed)

Sure they could, but we don’t design APIs based on the 1% of cases. Far better to have reasonable behavior in the 99% of cases and magic in the 1% than the other way around, even if the magic case in the former seems more magic.

Basically my opinion is that the more magic we have for this, the worse it is. There's a minimum amount of magic we need for usability's sake (automatically wrapping the declaration in this new rule when parsing), but beyond that I strongly prefer for everything to be as normal and unsurprising as possible. Every bit of magic is another sharp corner we have to be aware of and design around, forever, vs something predictable and boring.

Importantly, pay attention to the use-case you're talking about - when does an author serialize a CSS rule and manually examine the results? That seems... super rare to me. It's the sort of thing I, personally, have only ever done to debug things. Putting aside the question of whether it's better to serialize the implicit rule or not, does this case occur often enough that it's worth doing anything custom at all? In my opinion, definitely not.

[talk about using CSSStyleDeclaration]

Same point, really - is it worth doing anything more complicated than just "here's a new type of at-rule, it acts like you'd expect"?

I don't understand why you think it's acceptable to magically add @nest, but not magically remove it. They seem like two sides of the same coin to me, and if anything doing one without the other is inconsistent. Having @nest magically appear in my code when I try to roundtrip it seems way weirder than seeing a manually authored @nest disappear in the same case. And it's still unclear to me what is the value of allowing authors to write @nest at all. If there is none, then the case of an author writing @nest and seeing it disappear if they try to roundtrip will be exceedingly rare.

Importantly, pay attention to the use-case you're talking about - when does an author serialize a CSS rule and manually examine the results?

I believe certain types of dev tools do this to at least some degree.

There's a minimum amount of magic we need for usability's sake (automatically wrapping the declaration in this new rule when parsing), but beyond that I strongly prefer for everything to be as normal and unsurprising as possible.

Are you saying that the expected frequency with which something is encountered doesn't factor in the API design choices you make?

Every bit of magic is another sharp corner we have to be aware of and design around, forever, vs something predictable and boring.

Wait, I thought we were talking about author ergonomics, when did it become about our convenience?

In the interests of resolving this 90-comment issue, would it be OK to open a new issue on the details of @nest, or do we need to figure it all out to make any progress?

In the interests of resolving this 90-comment issue, would it be OK to open a new issue on the details of @nest, or do we need to figure it all out to make any progress?

I think we can probably resolve on the general proposal, then discuss serialization in a separate issue.

Same point, really - is it worth doing anything more complicated than just "here's a new type of at-rule, it acts like you'd expect"?

More complicated for browser engine indeed, but less complicated for CSS users (no need to learn about a @nest rule which will maybe create some confusion with the nesting selector & ..etc)

More complicated for browser engine indeed, but less complicated for CSS users

Not sure I agree. This requires to expand the API surface of CSSOM quite a bit, or changing APIs in ways that are confusing / inconsistent (like not returning rules from .cssRules), or unclear...

If you're not inspecting stylesheets via script, it doesn't matter either way. If you do, then IMO the @nest rule is easier to deal with than something new / more complicated. It just works like a style rule.

Given that, and that this is somewhat of a time bomb which we probably want to fix sooner rather than later, I tend to think that the simplest / less magical solution is best.

• It's implementable in a very straight-forward / performant way in all engines.
• It raises less open / follow-up questions.
• It's (IMO) not worse for authors than the current state of things or alternatives.

@tabatkins regarding the proposal, I'm assuming:

• @nest doesn't need to be a grouping rule (though I guess it doesn't matter much either way).
• With "style rules" in your proposal you mean style declarations, presumably, right?

Fully agree with @emilio.

I initially thought that we should serialize back to bare declarations for outer cssText since it's shorter, nicer, and basically "why not". But the round-trip issues and other edge cases convinced me the other way.

@LeaVerou You said that we should consider another name for @nest, would that mitigate the situation? If the name better explains what it does/is, maybe its presence will feel less offensive? @this? @self? @wart?

changing APIs in ways that are confusing / inconsistent (like not returning rules from .cssRules)

Yes, let's avoid this, please. Getting creative about the OM structure is even more intrusive than getting creative about its serialization.

The CSS Working Group just discussed [css-cascade] [css-nesting] Figure out whether we're fine with "shifting up" bare declarations after rules, and agreed to the following:

• RESOLVED: Tab specs this by representing @nestin the CSSOM, with an inline issue about@nest as a syntactic construct that authors could write being something we might want to get rid of

To reiterate my points a bit from the call:

• Fixing this problem in any reasonable way will affect (positively) approximately all CSS authors (by making properties in MQs/etc work better), and unblock CSS from evolving in certain ways that we want it to (adding mixins).
• The issues brought up in the latter half of this thread are exclusively about the CSSOM, specifically reading and modifying it. I, personally, have never once read or modified the CSSOM in my entire webdev career; I've only ever done so to write WPTs.
• So, to a first approximation, these issues affect nobody. ^_^ More precisely, the only audience I know of that actually reads/modifies the CSSOM in non-trivial ways (more than just, say, serializing an entire stylesheet and putting it somewhere else) are CSS tool authors. I'd estimate there to be a few hundred people in this category, across the entire world; a thousand on the extreme end of the possible range. This audience is also going to be significantly more experienced with coding in general.
• None of the issues listed here are about a use-case being difficult or impossible, which is usually what we can use to justify solving a use-case for such a small audience. They are convenience features.
• And the convenience itself is arguable in most cases. They're mostly about making the OM more consistent in some ways, but every single one comes at the cost of making it less consistent in other ways.

So in summary, past the necessary complexity/magic that's at the core of the issue (grouping interleaved declarations in some way so they don't merge with the initial set of decls), any further complexity we add is, as far as I can tell, solely to add minor convenience to some use-cases for a miniscule audience of already-experienced developers. Thus, none of them justify the spec and impl complexity cost they would impose.

I'm more than happy to be shown wrong on this regard - if you have an example of some change to my base proposal that would either give the "CSS tool authors" audience a large benefit (especially if it unblocks something that is currently hard or impossible), or that would give the "CSS general audience" even a mild benefit, or that would give some third audience I'm not aware of a benefit that is appropriately weighted against its size, then I'd joyfully consider it.

But let’s not make this a permathread to go over all the alternatives and options. This issue was specifically about changing the “shifting up” behavior, which we have resolved to undo (to great author benefit). Please open new issues on specific problems or alternative solutions so that we can have more focused discussion about them.

@mdubet could you open an issue on the idea of using multiple CSSStyleDeclarations instead of a new @nest rule?

@LeaVerou would it make sense to open a new issue to explore how far we could go in making things browser-internal and not exposed to developers?

@kbabbitt if you think it is possible to solve the setProperty() problem while keeping the non-shifting behavior, please open a new issue

(and anything else I have missed, please open a new issue)

@LeaVerou It looks like @nest has been removed from the specification again right? This history I think needs an article detailing it.