This commit updates the JSHandle class to take a generic representing
the underlying object that it's wrapping. We can then define
`ElementHandle` as a class that extends `JSHandle<Element>` and begin
to get better type inference.
Prior to this commit the following code would have `d` set to `any`:
```
const div: page.$<HTMLDivElement>('div')
const text = await div.evaluate(d => d.innerText)
```
You could work around this in two ways:
```
const text = await div.evaluate<(d: HTMLDivElement) => string>(d => d.innerText)
const text = await div.evaluate((d: HTMLDivElement) => d.innerText)
```
But both of these have two issues:
1. Requires the user to type extra information.
2. There's no type checking: in the code above I could type `d` as
`number` and TS would be happy.
With the change here to `evaluate` the user can now type the original
code:
```
const div: page.$<HTMLDivElement>('div')
const text = await div.evaluate(d => d.innerText)
```
And TypeScript will know that `d` is an `HTMLDivElement`.
This change brings us inline with the approach that @types/puppeteer
takes. If we land this and it works, we can do the same with
`evaluateHandle` to hopefully make a similar improvement there.
BREAKING: because this changes the types, which were previously `any`,
this is technically a breaking change as users using TS could start
getting errors after this change is released.
* chore: enforce pinned dependencies
Because we don't check our `package-lock.json` in, we can end up with
different versions installed locally vs CI, or even two devs having
different versions. Let's pin and enforce we pin every version to
avoid this.
* fix: make `$` and `$$` selectors generic
This means, much like TS's in built `querySelector` type, you can now do:
```ts
const listItems = page.$$<HTMLLIElement>('ul li');
```
And/or:
```ts
const h2 = page.$<HTMLHeadingElement>('h2');
```
And the return value will be of type `ElementHandle<T>|null`, where `T`
is the type you provided. By default `T` is an `Element`, so you don't
have to provide this if you don't care as a consumer about the exact
type you get back.
* chore: fix test assertions
This PR:
1. Makes sure we remove and freshly install Puppeteer before testing our
type defs, to avoid running on stale files.
2. Makes the tests run off `puppeteer.tgz` to avoid having version
numbers in the file name and therefore having to update it when we
bump versions.
* fix: wider compat TS types and CI checks to ensure correct type defs
This PR improves our TS types further to make sure they are usable in a
TS environment where ES Modules are the target output. Our use of
`export =` is problematic this environment as TypeScript does not allow
`export =` to be used and it errors.
The fix for the type issues to avoid `export =` is to instead define the
functions that you gain access to when you import Puppeteer as top level
functions in our `types.d.ts` file. We can do this by declaring them
explicitly in `src/node.ts`. These are then rolled into `lib/types.d.ts`
at build time. The downside to this is that we have to keep those
declarations in sync with the Puppeteer API; should we add a new method
to the `Puppeteer` class, we must add it to the `nodes.ts` declarations.
However, this could easily be automated by a small script that walks the
AST and generates these. I will do that in a follow-up PR, but I
consider this low risk given how rarely the very top level API of
Puppeteer changes. The nice thing about this approach is we no longer
need our script that hacks on changes to `lib/types.d.ts`.
To avoid yet more releases to fix issues in one particular TS
environment, this PR also includes a suite of example setups that we
test on each CI run. Each sample folder contains `good.ts`, which should
have no TS errors, and `bad.ts`, which should have some errors. The test
first packs Puppeteer into a tar, and then installs it from that tar
into each project. This should replicate how the published package
behaves when it is installed. We then check that we get no errors on
`good.ts`, and the expected errors on `bad.ts`.
We have a variety of test projects that cover both TS and JS source
code, and CJS and ESM imports and outputs.
This PR aims to vastly improve our TS types and how we ship them.
Our previous attempt at shipping TypeScript was unfortunately flawed for
many reasons when compared to the @types/puppeteer package:
* It only worked if you needed the default export. If you wanted to
import a type that Puppeteer uses, you'd have to do `import type X
from 'puppeteer/lib/...'`. This is not something we want to encourage
because that means our internal file structure becomes almost public
API.
* It gave absolutely no help to CommonJS users in JS files because it
would warn people they needed to do `const pptr =
require('puppeteer').default, which is not correct.
* I found a bug in the `evaluate` types which mean't you couldn't
override the types to provide more info, and TS would insist the types
were all `unknown`.
The goal of this PR is to support:
1. In a `ts` file, `import puppeteer from 'puppeteer'`
1. In a `ts` file, `import type {ElementHandle} from 'puppeteer'`
1. In a `ts` file, referencing a type as `puppeteer.ElementHandle`
1. In a `ts` file, you can get good type inference when running
`foo.evaluate(x => x.clientHeight)`.
1. In a `js` file using CJS, you can do `const puppeteer =
require('puppeteer')` and get good type help from VSCode.
To test this I created a new empty repository with two test files in,
one `.ts` file with this in:
https://gist.github.com/jackfranklin/22ba2f390f97c7312cd70025a2096fc8,
and a `js` file with this in:
https://gist.github.com/jackfranklin/06bed136fdb22419cb7a8a9a4d4ef32f.
These files included enough code to check that the types were behaving
as I expected.
The fix for our types was to make use of API Extractor, which we already
use for our docs, to "rollup" all the disparate type files that TS
generates into one large `types.d.ts` which contains all the various
types that we define, such as:
```ts
export declare class ElementHandle {...}
export type EvaluateFn ...
```
If we then update our `package.json` `types` field to point to that file
in `lib/types.d.ts`, this then allows a developer to write:
```
import type {ElementHandle} from 'puppeteer'
```
And get the correct type definitions. However, what the `types.d.ts`
file doesn't do out of the box is declare the default export, so
importing Puppeteer's default export to call a method such as `launch`
on it will get you an error.
That's where the `script/add-default-export-to-types.ts` comes in. It
appends the following to the auto-generated `types.d.ts` file:
```ts
declare const puppeteer: PuppeteerNode;
export = puppeteer;
```
This tells TypeScript what the default export is, and by using the
`export =` syntax, we make sure TS understands both in a TS ESM
environment and in a JS CJS environment.
Now the `build` step, which is run by GitHub Actions when we release,
will generate the `.d.ts` file and then extend it with the default
export code.
To ensure that I was generating a valid package, I created a new
repository locally with the two code samples linked in Gists above. I
then ran:
```
npm init -y
npm install --save-dev typescript
npx tsc --init
```
Which gives me a base to test from. In Puppeteer, I ran `npm pack`,
which packs the module into a tar that's almost identical to what would
be published, so I can be confident that the .d.ts files in there are
what would be published.
I then installed it:
```
npm install --save-dev ../../puppeteer/puppeteer-7.0.1-post.tgz
```
And then reloaded VSCode in my dummy project. By deliberately making
typos and hovering over the code, I could confirm that all the goals
listed above were met, and this seems like a vast improvement on our
types.
In `src/common` we now use `fs.promises.X` which we can dynamically
`import`. In a browser environment this code will never run because it's
gated on `isNode` (in a future PR we will add tree-shaking to the bundle
step such that this code is eliminated). By using `import`, we ensure
TypeScript still can track types and give good type information.
In `src/node` we continue to use `util.promisify` but that's not a
concern as that code explicitly is never run in the browser.
* chore: enforce file extensions on imports
To make our output agnostic it should include file extensions in the
output, as per the ESM spec. It's a bit odd for Node packages but makes
it easier to publish a browser build.
* chore: Use devtools-protocol package
Rather than maintain our own protocol we can instead use the devtools-protocol package and pin it to the version of Chromium that Puppeteer is shipping with.
The only changes are naming changes between the bespoke protocol that Puppeteer created and the devtools-protocol one.
I think when the FF version changes this check breaks - let's just make
it check that a version installed rather than the specific version else
this will happen on a continuing basis.
* chore: Don't store revisions in `package.json`
It's quite messy to have to require the `package.json` file in multiple
places purely to find out what revision of a given browser we want to
use. We can also achieve better type safety by placing it in an actual
source file.
This commit makes that change and also tidies up our reliance on
`package.json` within the source code generally; we now only use it to
find the location of the Puppeteer root such that we know where to
install downloaded browsers to.
To avoid using `package.json` to parse the name of the module, we also
now explicitly have an entry point for the Puppeteer module and the
Puppeter Core module. This will make it easier in the future to ship
less code as part of core (e.g. core never needs to download a browser,
so why ship that code?). Core can also then not have any revisions based
info contained in it.
The test install script has also been updated to ensure that
puppeteer-core can be installed correctly too.
Finally, the `install` script has been moved to TypeScript for nicer
typechecking and safety. The functionality of it has not changed.
* chore(agnostic): ship CJS and ESM builds
For our work to enable Puppeteer in other environments (e.g. a browser)
we need to ship an ESM build. This commit changes our config to ship to
`lib/cjs` and `lib/esm` accordingly. The majority of our code stays the
same, with one small fix for the CJS build to ensure that we ship a
version that lets you `require('puppeteer')` rather than have to
`require('puppeteer').default`. We do this with the `cjs-entry.js` which
is what the `main` field in our `package.json` points to.
We also swap to `read-pkg-up` to find the `package.json` file. This is
because the folder structure of `lib/` does not match `src/` now we ship
to `cjs` and `esm`, so you cannot rely on exact paths. This module works
up from the file to find the nearest `package.json` so it will always
find Puppeteer's `package.json`.
Note that we *do not* point any users to the ESM build. We happen to
ship those files so people who know about them can get at them but it's
not expected (nor will we actively support) that people will rely on
them. The CommonJS build is considered our main build.
We may make breaking changes to the structure of the ESM build which we
will do without requiring new major versions. For example the ESM build
currently ships all files that the CJS build does, but given we are
working on the ESM build being able to run in the browser this may
change over time.
Long term once the Node versions catch up we can ditch CJS and ship
exclusively ESM but we are not there yet.
* chore: add test for npm package installing correctly
This command packs up the module and installs it again to check we're
correctly bundling everything we need to allow users to do a fresh
install.
* install realpath
