Ember.js Template Imports: Part 3

Evaluating the tradeoffs of template language designs for tooling.

Assumed audience: Other members of the Ember.js and Glimmer.js communities. (Others are welcome to read and chime in, of course, but this is absolutely an “in-house” discussion!)

In this, the third of a planned five-part series on Ember’s template imports, I am digging into the implications of each of the designs for ecosystem tooling. Previously:

  • Part 1: Introducing the series and walking through the formats.
  • Part 2: Which template imports design has the biggest set of wins for teaching and understanding components?

Recall from those posts that there are four basic formats under discussion:

  • <template> tags with a custom file extension (currently .gjs and .gts)
  • template literals using an hbs literal
  • something like Svelte’s and Vue’s SFC format
  • an imports-only extension of the current format

In those previous posts, I said Part 3 was going to be about Scaling. However, I think it’s more useful to talk about Tooling here. As I have kept working on this series, I’m not actually persuaded that there are particularly meaningful differences between these approaches for scaling codebases which aren’t subsumed in the other topics — especially teaching and testing. So: tooling it is!


There are (at least) five broad categories to consider in evaluating the impact of these formats in terms of tooling:1

  • basic editor integration, e.g. syntax highlighting, code folding, etc.
  • lint tooling, e.g. ESLint and ember-template-lint support
  • formatter support, e.g. Prettier integration
  • language server tooling, including ELS and Glint, and interactions with existing TypeScript support
  • running the components in a server-side context like Fastboot

(Notice that TypeScript support cuts across several of these in various ways, but is most pronounced in the final point.)

Spoilers for this post: there aren’t actually any great outcomes here, my preferred <template> included. All of them have pretty significant downsides. The good news” is that the same thing is true for the formats chosen by Vue and Svelte, and that hasn’t been a serious hindrance to either of them. The problem is tractable for us, too — but it’s probably not tractable unless we pick one format and commit to it, simply because there is a lot of work to be done and we are a fairly small community.


There already exists a basic degree of support for all of these formats in terms of syntax highlighting and and code folding — albeit with some important caveats.

  • The imports-only format works” in all editors I have tried, but without any JS-specific formatting for the imports section. It is simply presented as plain text.

  • SFCs get syntax highlighting out of the box” by VS Code if you use a .hbs extension, which makes sense: the Handlebars syntax highlighting is an extension of basic HTML highlighting, and VS Code’s HTML highlighter has built-in support for embedded langauges in <script> and <style> tags. Other editors — including Vim, Sublime Text, and IntelliJ — generally work similarly here, and for the same reason.

  • Bothhbs and <template> have at least some degree of syntax highlighting support via various editor extensions, e.g. vscode-glimmer for VS Code, which also adds support for treating .gjs and .gts as aliases for the JS and TS syntaxes respectively.

    (Notably, however, for reasons I will cover below, just aliasing to JS and TS is actually not a great move for .gjs and .gts files, for reason I discuss below under Lint tooling. If you try this today, you will see red squiggles everywhere in VS Code and possibly other editors.)

In sum, as far as the most basic editor integration goes, these are all basically a wash.

The same rough mix of support for existing syntax highlighting tooling exists on GitHub, GitLab, and Bitbucket. You can see this in practice by taking a look at how GitHub renders Part 2 of this series. It works surprisingly well already across the board. The hbs and imports-only modes coming out worst in that they just parse as strings and get no highlighting. Both SFCs and <template> highlight reasonably well.2

Lint tooling

When we come to lint tooling — specifically, ESLint and ember-template-lint — the long and short of it is that nothing works particularly well, but SFCs and template literals come out slightly better in one specific way, imports-only basically neutral, and <template> slightly worse.

Out of the box, the existing linting tools simply do not understand template imports. In this regard, it’s a level playing field: all of the approaches on offer will have to implement custom handling. In particular, all of them incorrectly flag anything used only in the template as an unused value. That includes the imports-only mode! For example, in this lightly-modified version of an example from Part 2, the Greet and WeatherSummary components here will both be marked as unused imports, and isBirthday will be flagged as defined but never used.

import { hbs } from '@glimmer/component';
import Greet from './greet.js';
import WeatherSummary from './weather-summary.js';

function isBirthday(dateOfBirth) {
  // ...

export default hbs`
  <Greet @name={{@user.name}} />
  {{#if (isBirthday @user.dob)}}
    <p>Happy birthday!</p>
  <WeatherSummary />

Likewise, in the backing class for this variation on the WeatherSummary component from Part 2, the getCurrentTemp method and the isSet helper function will both be flagged as unused:

import Component, { hbs } from '@glimmer/component';
import { tracked } from '@glimmer/tracking';

const isSet = (val) => val != null;

export default class WeatherSummary extends Component {
  @tracked currentTemp;
  getCurrentTemp = () => {
    this.currentTemp = Math.floor(Math.random() * 100));

  static template = hbs`
      {{on "click" this.getCurrentTemp}}
      Check the weather

    {{#if (isSet this.currentTemp)}}
      <p>The current temperature is {{this.currentTemp}}</p>

We’d see exactly the same warnings about unused code in an SFC format, and for the same basic reasons: we have to inform the JavaScript and template linters about values in the other language.

Now, all of the options other than the imports-only format work more or less correctly for all of the parts of any given module which aren’t related to templates. (The imports-only format simply doesn’t connect the two layers at all at present, so there are no false positives… but I don’t think we can call that a win.) Perhaps surprisingly, <template> actually appears to do slightly better than the others in terms of recognizing that we are actually using values in module scope — but this is because it’s attempting to parse <template> and the contents of it as JSX. This means that editors which use the TypeScript Language Service (including for their JavaScript support) get very confused and report syntax errors everywhere, because Glimmer templates and JSX aren’t compatible.

The net here is that we have to implement a custom parsing layer for any of these formats to have usable linting integration.

Formatter tooling

The story is similar for existing formatter tooling with Prettier: none of the formats work well across the board.

  • Out of the box, Prettier basically works for the template side of SFCs, but doesn’t work at all for the JS side.

  • Exactly the inverse is true for template literals: Prettier works for the JavaScript side but not at all for the content between hbs, which it treats as a string (no surprise there: remember from Part 2 that that’s exactly what it is semantically!).

  • For the <template> proposal, Prettier simply fails to parse, and so formatting does not work at all.

  • With the imports-only proposal, like SFCs, Prettier works for the template side, but doesn’t format the imports section at all.

In sum, as with the lint tooling, we actually need to implement custom language support to make any of these work correctly. However, it’s worth acknowledging that the template literals and SFC proposals are halfway there, whereas (in very different ways) the <template> and imports-only proposals are much worse off.

Language server tooling

Finally, we come to language server tooling and integration. Most of the JavaScript ecosystem uses the TypeScript Language Service to support features like documentation-on-hover, go-to-definition, and refactoring. That includes React, since TS has built-in support for JSX; Vue, Angular, and Svelte via custom language server integrations; and Ember/Glimmer, via the various experimental ELS implementations and Glint (which it itself used by some of the other language servers). With any of the proposed formats, we would need to create a language server which understood the format and could connect it to the TS LS.

Per Dan Freeman and James Davis, who built and maintain Glint, there is very little difference in effort in supporting hbs vs. <template>, and because these all compile to the same primitives, even SFCs are tractable. The main challenge there is handling the same custom scoping semantics with the default export as I described as odd in Part 2. However, that is the same basic issue as supporting Glimmer components in Ember apps today: something Glint and the experimental ELSs already do.

Notably, Glint also supports GlimmerX, which uses the same syntax as the template literals proposal. This means that we get the integration for free” (really, for Dan’s and James’ hard work in 2019 – 2020). To get the same support for <template>, we would need to update the implementation of the Babel transform for <template> to provide some data about the original string, so that we can map invocations, error messages, and so on. We would have to build something similar for SFCs (albeit from scratch, since no implementation exists whatsoever yet for them).

However, there’s a problem here that’s easy to miss: because we’re giving new semantics to template literal strings, we have to override existing TypeScript’s existing understanding of what JS and TS files mean. In all cases, this is work.

  • For the <template> proposal, this is somewhat tractable and there are a variety of ways to approach it: the custom language integration means we can potentially leave normal” TS files alone and just add information to TypeScript via something like Glint. Doing it that way requires doing a build pass to provide the info, though. The alternative is to disable the TS LS in favor of something like Glint.

  • For SFCs, the story is very similar to that with <template>, though with a qualification: we could do similar to what Vue’s Vetur language server does and provide a blanket type definition for TypeScript, roughly like this:

    import Component from '@glimmer/component';
    declare module '*.glimmer' {
        export default class extends Component {}

    That would make the TypeScript side type-check — though not particularly helpfully — so a tool like Glint would then do its own pass over those as well. You will end up with multiple layers of feedback in your editor — one from TSServer and one from Glint — but that may be fine (and if we went that direction, we could provide a language server plugin to make that experience nicer). All of these proposals require Glint running over top of tsc/TSServer; the difference here is taht it means that you don’t have to disable the original TSServer to make your editor work.

  • In the case of imports-only, we have to disable the TS LS, because we have to stitch the script and template files together to create the correct context. Otherwise, the backing classes will always and unavoidably report that there is no usage of anything which is only used in templates.

  • For the template literals proposal, it might initially seem like we could just integrate with the TS LS plugin tooling. After all, the docs say that one of the intended uses for plugins is:

    Enable new errors or completions in string literals for a custom templating language

    Unfortunately, they also specify that one of the things language plugins cannot do is:

    Customize the type system to change what is or isn’t an error when running tsc

    The net of this is that we can add errors in a standard TS LS plugin, but we cannot remove them. We’re stuck with all those warnings about unused values! What’s more, plugins aren’t loaded during normal commandline typechecking or emitting.” This is why Glint works the way it does today: as a custom language server and CLI. To get consistent behavior between our build/CI environments and our editors, we still need to run everything through a custom pipeline. That would leave us with the downside of having two separate paths for type-checking vs. editor support… or with the alternative of disabling existing TypeScript support for those files.

Net, the hbs implementation initially appears to have a very small edge on language server implementation, because it already exists in support of GlimmerX — but we should not take this as a particularly important constraint: it’s basically just slight variations on the same underlying sets of tradeoffs. (Once again, everyone who isn’t using JSX really wishes that the syntax extensions part of TS were pluggable. Alas.)


One common reason a few people have suggested we ought to prefer hbs is that they think it makes it easier to support running Glimmer templates in server-side environments — that is, that it would make it viable to use hbs as an actual import which works without needing any compilation step.

Unfortunately, while that sounds appealing, it isn’t actually true, at least today! The problem is the mismatch in semantics discussed in Part 2:

  1. The scoping semantics are wrong, because hbs literals aren’t actually string literals.3 Remember: if you write this code…

    const Greeting = hbs`<p>Hello, {{@name}}!</p>`
    const Summary = hbs`
      <Greeting @name={{user.name}}>
      {{! ... }}

    …it does not work without a transform. Greeting isn’t in scope” for the Summary component! While we could work around this by introducing some other invocation form for hbs where it also takes the scope as an argument, that makes the ergonomics much worse — close, in fact, to the original precompileTemplate invocation that is the compile target.

  2. Additionally, remember that the scoping semantics are also wrong when you switch to a component with a backing class, because of the mismatch between a static class field and the semantics of component templates. If you want the this value to work correctly, you have to introduce some degree of processing. At a minimum, we would need to rewrite the internals of getComponentTemplate to go look up that static field — not an impossible hurdle, by any means, but a real and significant change to the current design. (As to whether it’s otherwise well-motivated, I refer you to the rest of the series!)

Moreover, I’ll go further here and say that I don’t think there’s any particular value to being able to run a Glimmer component without any build step. Having a build pipeline is extremely normal for both client- and server-side code — and it can even be done fairly transparently and on demand for server-side code, e.g. with @babel/node or ts-node. If someone wants to run Glimmer component code natively in a Node runtime, they can precompile it using our standard build tools or they can simply use @babel/node to integrate the transform automatically.

Net, I take this to be something of a non-issue for the design choice here, as it requires some non-zero degree of extra work compared to today’s baseline regardless and there are straightforward options for this regardless of the design chosen.


In this particular comparison, the template literals proposal clearly comes out with a small edge. In most categories, it’s the same or slightly better than the other options, as we can see in this table:

Consideration <template> Template literals SFCs Imports-only
JS working working working not working
templates working working working working
JS partial partial partial no
templates no no no yes
needs custom parser yes yes yes yes
JS no partial partial no
templates no partial partial no
needs custom parser yes yes yes yes
LS effort small none medium small
Server-side requires compilation yes yes yes yes

These differences are very small, though. Accordingly, I still believe <template> is the best choice — because the small deltas here are fairly straightforward to tackle, and because I think the issues around Teaching described in Part 2 and around Testing as I will describe in Part 4 profoundly outweigh these small tooling differences.


  1. I provide examples here in terms of things like ESLint and Prettier, but it’s important to recognize that these are categorical costs. If we choose at some point to switch our linting over to something like RSLint for the sake of its speed, we would have to pay any costs associated with a given format there as well. ↩︎

  2. That <template> more or less works surprised me; it appears to be a function of the overloading of <template> discussed as a downside in the last post. Supporting something like <Template> or <Glimmer> would require more work: it highlights more or less reasonably (though not necessarily correctly”) until the closing </Glimmer>, which does not highlight correctly — but content after it highlights correctly again. ↩︎

  3. Dan Freeman pointed this out on the Ember Discord in response to the update where I added this section. I knew this, and in my COVID-recovery-induced mental haze, I totally forgot it. Thanks, Dan! ↩︎