Previously, when some iframe started/finished a new navigation, we
could have removed and then re-addded load/domcontentloaded on the
main frame.
Drive-by: unflake wheel test in Firefox.
It could be that iframe was blocking some event, like load or networkidle,
and we never updated the lifecycle after the iframe was detached. This
lead to goto and other navigation commands to never resolve.
As of July 14, 2022, Safari Technoly Preview returns the following:
```
Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/16.0 Safari/605.1.15
```
Fixes#15474.
Notes:
* page-level requests that are also handled by a SW's fetch handler, should not be interceptable at the page-level
* `Network.requestWillBeSent` does not provide enough metadata for Playwright to fire the `request` event at that time, so it does it as soon as it gets to the end of the request lifecycle
* `clip` option is always passed from the client code
* with this change, we can no longer capture screenshot of a blinking
caret; the browser-side API doesn't have this capability.
The following options now work across languages:
- `recordHar`
- `serviceWorkers`
In addition, object properties are now sorted alphabetically.
Drive-by: fixed `--target` help message to include all available targets.
Resolves#14255.
PNGs should be preferred as they are deterministic, but increasing this
limit should be fine for JPEG users. (Looking through jpeg-js source code, the actual memory allocation is based on the size of the image—so unless a user is hitting this limit already—this should not impact the memory consumption of Playwright.
- Never use open shadow root for highlight. This messes up
our selectors that accidentally match internal preview elements.
- Remove failing electron test that we do not care about.
- Skip `channels.spec.ts` in non-default mode.
Previously, we only validated/converted on the way to the server,
but not from the server.
Validating both ways catches issues earlier, and allows us to
perform automatic conversions, for example only converting
buffers to base64 when sending over wire.
Potential fixes to avoid startup races:
- Wait for "juggler listening" message.
- Make sure `transport.onclose` is called when connecting to the transport after the actual pipe closure.
- Do not resolve raw headers upon `loadingFinished`, since they may still come later in
`responseReceivedExtraInfo`.
- Introduce separate promises for `encodedBodySize`, `transferSize` and `responseHeadersSize`.
- Make sure we resolve each of them either with data available
from the browser, or a fallback calculation.
- Set raw response headers for redirects on WebKit.
- Do not stall on cached responses in Chromium, they have erroneously set `hasExtraInfo` flag.
- Use `transferSize` that is available in Firefox protocol.
fix(network): make allHeaders wait until all header are available
Before, calling `allHeaders()` from `page.on('request')` would yield
provisional headers instead.
With these changes:
- In Firefox, all headers are available immediately.
- In Chromium, all headers are available upon requestWillBeSentExtraInfo.
- In WebKit, all headers are available upon responseReceived.
- In all browsers, intercepted requests use "provisional" headers
as all headers, since there is no network stack to change the headers.
Drive-by: migrated Chromium to `hasExtraInfo` flags that simplifies
the logic quite a bit.
When target element is inside a non-main frame, there could be an
overlay in some of the parent frames that intercepts pointer events.
However, we never detected this case.
Currently, `loadstate` and `load` are two separate events in the protocol,
and are fired in this order. As a result, `waitForLoadState()` sometimes
resolves before the `'load'` event is fired, which is unexpected.
Also fixes a flaky test that assumed `load` event comes after `domcontentloaded`
for the empty page, which is not always a case in Chromium.
- `har` option is now an object `{ path, fallback }`.
- Allows falling back to `abort()`, `continue()` or throwing.
- Matches based on url + method.
- Follows redirects in the HAR file.
- Nice error/stack when throwing.
- Tests.
Adds cross-browser support for easily allowing/blocking Service Workers via a Context option.
Includes plumbing for Playwright Test's `use`.
Resolves#14522.
Relates #1090.
Supercedes #14321.
`npx playwright open --save-har=file.har` produces a HAR file that
can be later inspected or used for mocks.
`--save-har-glob="**/*.{png,woff2}"` option can be used to
produce a smaller har that only contains entries matching the glob.
Language override behavior changed upstream in WebKit/WebKit@039ebd9
New logic is closer to the actual behavior of WebKit on macOS, meaning that when the user changes system language the actual locale changes according to some weird OS rules:
ru-RU => navigator.language === 'ru'
fr-CH => navigator.language === 'fr-FR'
es-MX => navigator.language === 'es-MX'
Our locale emulation is aligned with that, so setting locale to fr-CH will result in fr-FR etc.
fix: sanitize URLs with vbscript:
The vbscript: protocols can be used to run scripts in much the same way as the javascript: protocol. This PR adds in validation for those aforementioned protocols in snapshotterInjected.ts and snapshotRenderer.ts.
This restores the old hit target check, in addition to the new
hit target interceptor.
This way, we got some coverage for iframes and other quirky cases,
but keep the bullet-proof hit target check in place.
Fixes focus and blur management when `page.locator(…).focus()` and `page.locator(…).type(…)` are used which was regressed by 7a5b070 (#13510).
#13510 relied on an implicit assumption that this (conditional) [`blur`](7a5b070e95/packages/playwright-core/src/server/injected/injectedScript.ts (L672)) call would always be followed by a call that resulted in a newly focused element via this [`focus`](7a5b070e95/packages/playwright-core/src/server/injected/injectedScript.ts (L674)) call.
However, some elements are [not focusable](https://html.spec.whatwg.org/multipage/interaction.html#focusable-area), so we were blurring incorrectly, and losing focus that we should have maintained.
Two regression tests were added that pass on the commit prior to 7a5b070e95 (and match manual testing/expectations):
* `page.locator(…).focus()`: _keeps focus on element when attempting to focus a non-focusable element_
* `page.locator(…).type(…)`: _should type repeatedly in input in shadow dom_
Additionally, a third test (_should type repeatedly in input in shadow dom_) was added to check the invariant from #13510 that states:
> This affects [contenteditable] elements, but not input elements.
and allows us to introduce the targeted fix (contenteditble check before blur) without breaking FF again.
And _should type repeatedly in contenteditable in shadow dom with nested elements_ was added to ensure the above fix works with nest contenteditble detection.
Fixes#14254.
References #14037.
We used to query and save all non-zero `scrollTop`/`scrollLeft` to restore the page perfectly. However, this became very slow with Chromium v100 regression (see issue).
This change only queries and stores `scrollTop`/`scrollLeft` for elements that are ancestors of the action target. The resulting snapshot does not perfectly recreate the original screen, but should be good enough to inspect the action.
fix(chromium): work around about:blank issue on Chromium
We don't receive the `loaderId` which translates to `newDocumentId`,
so we expect the same-document navigation. Instead, we can wait
for any new-document navigation as a workaround, only for `about:blank`.
This also reverts commit f0f65fa247.
We should not fast-return if someone tries to run
`npx playwright install-deps` on Debian.
Instead, we should warn users, fallback to Ubuntu 20.04 dependencies,
and proceed.
Currently, when I choose "Java" in the sources list and then
click on the page to generate the "click" action, sources reset
to "JavaScript". This is very inconvenient.
This changes the logic to only forcefully change files if either
old or new file is a user file, not a generated one.
Use cases considered:
- run `codegen`, click around, choose different language, click more;
- run script with inspector, pause, click "Record" and record an action;
- same as above, but then continue and see that user source is revealed.
This fixes a few issues:
- strict mode was producing false negatives if multiple query paths
lead to the same element being picked;
- in some cases the number of intermediate items in the list was
exponential and crashed quickly.
What changed:
- `visible` engine is a real engine now;
- `capture` selectors are transformed to `has=` selectors for
easier implementation;
- chained querying switched from a list to a set to avoid
exponential size.
This allows us to use the full retry logic of rimraf in the `onexit` handler.
Note this is already covered by failing on Windows test
`should remove temp dir on process.exit`.
We skip waiting for "visible" state that enforces non-zero size.
Other invisible conditions like "display:none" fail during the
actual "scrolling" step and will retry.
feat: support experimental doc entries
- Params/options/members are marked as experimental in the docs.
- `experimental.d.ts` is generated that contains all types and
includes experimental features.
- `experimental.d.ts` is references in our tests so that we
can test experimental features.
- `fonts` option is restored as experimental.
Reason: turns out Debian Buster requires just one source list to
install `ttf-ubuntu-font-family` font.
All other dependencies are satisfied.
Fixes#13530
Firefox has a bug: calling `node.focus()` does make the node focused,
but some internal "current contenteditable node" is not changed.
Blurring the previous one and focusing the new one helps.
This was introduced in #7500 to fight `ERROR: The process "4436" not found.`
messages when killing a process that did already exit.
Since then, we no longer inherit stdout/stderr, so the error message
should not appear anymore.
This is a speculative fix to leftover tmp directories.
When users issues SIGINT twice, we enter `gracefullyClose()`
twice, and shortcut the second time. It turns out, we do
not wait for directories removal.
Note: it is unknown how often we reach this codepath in practice.
- This supports `role=button[name=Hello]` similarly to CSS selectors.
- Does not change `_react` or `_vue` behavior that insist on quoting the string.
- Uses CSS notion of "identifier" characters.
This patch will check if browser channel is already installed.
If it is, it'll abort installation with the following error:
```
aslushnikov:~/prog/playwright$ npx playwright install msedge
Failed to install browsers
Error:
╔═════════════════════════════════════════════════════════════════╗
║ ATTENTION: "msedge" is already installed on the system! ║
║ ║
║ "msedge" installation is not hermetic; installing newer version ║
║ requires *removal* of a current installation first. ║
║ ║
║ To *uninstall* current version and re-install latest "msedge": ║
║ ║
║ - Close all running instances of "msedge", if any ║
║ - Use "--force" to install browser: ║
║ ║
║ npx playwright install --force msedge ║
║ ║
║ <3 Playwright Team ║
╚═════════════════════════════════════════════════════════════════╝
```
To re-install browser channel, use `--force`.
Fixes https://github.com/microsoft/playwright/issues/13061
It turns out that "non stalling evaluate" can stall in Chromium
in some weird conditions, like `document.open` after some weird
`iframe.src` value.
We now only hide highlight in those frames where we did install
highlight in the first place.
- Added docs to `selectors.md`.
- `[pressed]` and `[checked]` do not match `"mixed"` states.
- Disallow `[name]` shorthand without a value.
- Renamed `includeHidden` to `include-hidden`.
Previously, any unpaired quote in the text selector "escaped"
everything till the end of the selector string, and so any
subsequent chained selectors, including ">>" separator were ignored.
An example of misbehaving selector: `text=19" >> nth=1`.
Now, when text selector contains a non-leading quote, selector parser
does not assume it should escape ">>" separator and correctly
tokenizes all selectors from the chain.
Note that this behavior is a workaround for the fact that our
text selectors is somewhat poorly defined in this area. That said,
this workaround seems to be safe enough. It still does not work for
unpaired leading quotes like this: `text="19 >> nth=1`.
This has two values:
- `"hide"` to hide input caret for taking screenshot
- `"initial"` to keep caret behavior unchanged
Defaults to `"hide"`.
Fixes#12643
This introduces `role=button[name="Click me"][pressed]` attribute-style
role selector. It is only available under `env.PLAYWRIGHT_EXPERIMENTAL_FEATURES`.
Supported attributes:
- `role` is required, for example `role=button`;
- `name` is accessible name, supports matching operators and regular expressions:
`role=button[name=/Click(me)?/]`;
- `checked` boolean/mixed, for example `role=checkbox[checked=false]`;
- `selected` boolean, for example `role=option[selected]`;
- `expanded` boolean, for example `role=button[expanded=true]`;
- `disabled` boolean, for example `role=button[disabled]`;
- `level` number, for example `role=heading[level=3]`;
- `pressed` boolean/mixed, for example `role=button[pressed="mixed"]`;
- `includeHidden` - by default, only non-hidden elements are considered.
Passing `role=button[includeHidden]` matches hidden elements as well.
This patch prints a friendly instructions in case Docker image version
mismatches Playwright version and there are missing browser
dependencies.
With this patch, Playwright will yield the following error:
```
root@f0774d2b2097:~# node a.mjs
node:internal/process/promises:279
triggerUncaughtException(err, true /* fromPromise */);
^
browserType.launch:
╔════════════════════════════════════════════════════════════════════════════════════════════╗
║ Host system is missing dependencies to run browsers. ║
║ This is most likely due to docker image version not matching Playwright version: ║
║ - Playwright: 1.22.0 ║
║ - Docker: 1.21.0 ║
║ ║
║ Either: ║
║ - (recommended) use docker image "mcr.microsoft.com/playwright:v1.22.0-focal" ║
║ - (alternative 1) run the following command inside docker to install missing dependencies: ║
║ ║
║ npx playwright install-deps ║
║ ║
║ - (alternative 2) use Aptitude inside docker: ║
║ ║
║ apt-get install libgbm1 ║
║ ║
║ <3 Playwright Team ║
╚════════════════════════════════════════════════════════════════════════════════════════════╝
at file:///root/a.mjs:3:10 {
name: 'Error'
}```
Fixes#12796
Co-authored-by: Dmitry Gozman <dgozman@gmail.com>
This patch:
- Adds 3 new host platform types:
* `generic-linux` and `generic-linux-arm64` for the unsupported
linux distributions
* `<unknown>` for non-supported OS versions
- Prints a warning when downloading Ubuntu browser builds on
unsupported Linux distribution
- Makes sure launch doctor prints all missing shared libraries
on unknown Linux distributions
- Also prints an `apt` command as an alternative to Playwright CLI
dependency installation.
Supports inline regex in addition to string: `_react=BookItem[author = /Ann?a/i]`.
This is similar to `text=` selector, but applies to `_react` and `_vue`
selectors. In the future, will also apply to `role=` selector.
Previously, we preserved input/textarea values by providing
`value` attribute or text child. This produces DOM that does not
actually match the original page.
This change starts using special attributes to modify values
directly when rendering.
Same treatment is also applied to options in `select` and
`checked` property of checkboxes and radio buttons.
