I'm having a hard time reviewing this because, despite the bold text in the beginning, it seems to mix up the enumerated attribute handling into the same PR as the non-enumerated attribute work.

Apologies I think it got changed by a build task after I updated it. If you look purely at the first commit it should be easy enough to review. I think I'll revert the enumerated stuff and move it to a separate tag on PR.

I've dropped the second commit from this PR so it should be reviewable now.

Would it be possible to start with [ReflectURL] and take them in turns? I don't really know what @domenic would prefer so wait for him to say something, but I would find it easier if we did them incrementally.

To clarify you mean add a PR that does just [ReflectURL] and then a follow up that does just [Reflect]?

I'm happy to split this however reviewers think is best.

This specification text direction, which builds on top of and preserves the existing prose-based infrastructure, is different than what I was expecting. I was picturing that we would rebuild the specification to be centered around the extended attributes, getting rid of concepts like "reflected target", "reflected IDL attribute", "reflected content attribute", "treated as a URL", "limited to only positive numbers", etc. Instead we would do something like:

IDL attributes annotated with the [Reflect] content attribute have the following getter steps: [...] and the following setter steps: [...]

and then everything there would be phrased in terms of the IDL concepts, like the extended attribute arguments, the IDL attribute type, this, or the presence of other auxiliary extended attributes.

I think my vision is cleaner, as it has fewer layers of abstraction. It also prevents people from attempting to do reflection in prose, which seems like a good idea. And it's closer to how one might implement these. However, it's certainly less incremental. What do you think?

So my thought process was to build this in a way that was as incremental as possible. Which is why it's currently in the form I've used.

I don't know that we'll ever be able to fully get rid of spec prose for reflection, but it might be that the end situation has few enough special cases that their prose should just be rewritten to not mention reflection and then we could drop many of these existing concepts.

My other hesitation along the same lines is I don't know what other specs use these algorithms. I didn't want to accidentally break URL reflection from another spec by removing the prose concept (though I'll admit it's rather vague at the moment so it's ideal anyway).

I think my preference would be to write them in terms of the existing prose initially and then once we've moved everything across (including looking at other specs). We can possibly rewrite how they're defined to drop the old approach. It would be purely editorial (I guess normative because we're removing existing algorithms?).

That way we don't break existing stuff and avoid duplication at the same time?

One other question is are we aiming to go for 100% coverage in IDL. Or is it okay if we end up with some still defined manually.

There's a number of odd cases that aren't currently possible to do in the IDL proposal.

• Document.dir reflects the content attribute of the html element

• form.action and form.formAction's missing and empty value default is to return the document's URL instead (Uses ReflectSetter to handle the simple setter case)

• .autocomplete returns IDL-exposed autofill value internal state (Uses ReflectSetter to handle the simple setter case)

• canvas.width and canvas.height are limited to non-negative integers per attribute parsing but the current spec prose and a quick test in chrome shows that setter steps are not equivalent to [ReflectNonNegative,ReflectDefault=X]. (it doesn't throw when you set an invalid value, unlike the existing reflection)

• .tabIndex tabIndex has a defaultValue that's dependent on element itself. It's currently defined on a mixin so this wouldn't work. It's even based on where the summary element is in the DOM so I don't think this could be handled by IDL. (Potentially we could use [Reflect] and then just override the defaultValue using spec prose, but that would require we keep the default value concept in prose.) (Uses ReflectSetter to handle the simple setter case)

• Various document IDL attributes reflect content attributes from the body element (or use an empty string if it's a frameset) so again that couldn't be defined in IDL.

The above is just some examples of what I've found, there's also the element internals cases that I haven't looked into yet and might need extra IDL or might not be doable.

Hopefully some of these changes improve the clarity of the definitions. I've also added a [ReflectSetter] as it turns out WebKit already has it and it cuts down on a number of basic setters.

I don't know that we'll ever be able to fully get rid of spec prose for reflection, but it might be that the end situation has few enough special cases that their prose should just be rewritten to not mention reflection and then we could drop many of these existing concepts.

I was more hopeful that we'd be able to get rid of the spec prose. After your change, the only normative references I see to "reflect" are enumerated attributes, and the ones you list as not possible with IDL. But, see below for how keeping it might be the best way to make some of the existing cases more rigorous.

My other hesitation along the same lines is I don't know what other specs use these algorithms. I didn't want to accidentally break URL reflection from another spec by removing the prose concept (though I'll admit it's rather vague at the moment so it's ideal anyway).

Personally, I think we should not be hesitant about breaking other such specs. The only other specs that should be using these concepts are:

• MathML and SVG. (But from what I recall, these are not very good about defining reflection rigorously.)
• Monkeypatch specs that are intended to one day go into HTML. These need to deal with upstream breakage all the time.

I think my preference would be to write them in terms of the existing prose initially and then once we've moved everything across (including looking at other specs). We can possibly rewrite how they're defined to drop the old approach. It would be purely editorial (I guess normative because we're removing existing algorithms?).

I think this is fair.

Even if we keep some prose concepts, I think it'd be nicer to clean up as much of the "public API" as possible. For example, getting rid of definitions like "default value" or "as a URL" or "limited to only non-negative numbers" and say that if you want those behaviors, you have to just use the extended attributes. (This could be done by having the algorithm steps inspect the reflected IDL attribute, instead of taking arguments.)

One other question is are we aiming to go for 100% coverage in IDL. Or is it okay if we end up with some still defined manually.

In my opinion, it's definitely fine to end up with some defined manually. I think the way they currently abuse "reflect" is not rigorous, and rewriting them to just be manual might be better. Or maybe the best road here is to figure out ways to define them rigorously using the prose concepts that underlie the extended attributes. (E.g., defining custom "get the element" algorithms for Document might be enough to rigorize those attributes.)

Ah yeah the lowercasing was something I'd thought of too forgot to comment about it, it seems webkits does that automatically at the very least. I think I lean towards doing it automatically just because it cuts down on noise and then the string case as you say can be for stuff where it's genuinely a different name (think encoding reflecting enctype or I think there's some hyphenated ones dotted about)

@annevk or @domenic am I okay to mark this as supportive from a browser implementor perspective or would you rather I file standards positions?