test: Move from CDP to BiDi

[KKoukiou]

See cockpit-project/cockpit#20832

Drop chrome-remote-interface NPM dependency.

Explicitly add a "glob" devDependency to follow suit with cockpit-project/cockpit@680decc155a
It was previously used implicitly through a transient dependency of something else, but our esbuild po-plugin uses it explicitly.

[KKoukiou]

@martinpitt i have two tests that check exception handling fail with this bidi convertion.

both do something like: 253 b.eval_js("window.setTimeout(function() { myNonExistingFunction()}, 0);")

And because the UI handles unhandled JS errors with window.onerror I expect to see the error dialog.

With bidi this does not propagate properly. Any ideas what's missing?

Otherwise I will disable the two tests.

[martinpitt]

Some notes:

• The old CDP API didn't have reliable exception reporting, it was async and delayed. That was the reason why you had to do these loops:

  b.eval_js("window.setTimeout(function() { myNonExistingFunction()}, 0);"
  # Some round trips, one of which should update the deferred exception
  for _ in range(0, 5):
      b.wait_in_text("#critical-error-bz-report-modal-details", "myNonExistingFunction is not defined")
      time.sleep(2)

With BiDi, this will happen immediately and reliably if the command or returned promise rejects. However, the called command does not fail, as it just starts a timer and does not return a promise. So you can drop the loop (as it's pointless), and drop the with self.assertRaises(RuntimeError): as the eval_js() itself will reliably not fail. You only get the console error:

error: ReferenceError: myNonExistingFunction is not defined

from the timer when it fires, but as nothing pays attention to that, this doesn't cause any crash.

• Calling the timeout crash fn did not trigger Cockpit's general "Oops" handler with CDP either. This happens in CDP context, not page context. Supposedly Anaconda's error handler uses a different approach than cockpit's Oops handler.

We test that in cockpit in https://github.com/cockpit-project/cockpit/blob/main/pkg/playground/exception.js , and the test clicks that button to provoke a crash. Taking that idea, I was trying to do this not with a timer, but with setting an event handler inside of the actual page:

b.eval_js("window.document.getElementById('toolbar').addEventListener('click', () => myNonExistingFunction())")
b.click("#toolbar")

Doing that in the JS console/inspector and clicking the toolbar works fine, but not through eval_js. Apparently this is running in a "shielded" context.

@KKoukiou Sorry, I'm afraid this is a hard problem. My suggestion would be to add that crasher into the actual page somehwere, in a hidden place (making it conditional on running as test or so?), and then activate it with a click.

[KKoukiou]

@martinpitt I just realized that the test correctly triggers the exception. I see the exception handling UI in the screenshots. The problem is that RuntimeError was never raised, aka - removing with self.assertRaises(RuntimeError): fixes the issue. Thanks for taking a look.

[marusak]

Thanks!

[M4rtinK]

Looks good to me. :)

[martinpitt]

@KKoukiou this looks great now, thanks! I'm glad it was something easy after all.