Workarena/Base: Merge init scripts before evaluation.

[conte91]

Fixes #57 and duplicates without the need for downgrading the playwright version.

Playwright's init scripts are not executed sequentially, as per PW's docs:

However, some benchmarks depend on all installed init scripts to be executed in order, which is why the page hangs if they don't (in WorkArena's case, this was because of the function being waited on being undefined - since the script defining it hadn't been evaluated yet).

It shouldn't cause any problem to just concatenate all scripts together, so that we can guarantee FIFO execution order.

[younik]

I think the evaluate lines should be removed

[conte91]

Self-note: this could be an issue in case names defined across scripts shadow each other?

Something like:

add_init_script(`const Math = () => {alert("fake math!");}`)
add_init_script(`Math.random()`)

I think we should be fine though, because even if two scripts' private functions will share the same name, the last one defined will win.

[aldro61]

Hi, thanks for the PR @conte91, @younik. This works for most tasks but currently breaks the chart-related ones. I'm looking into what the problem could be.

[aldro61]

@conte91 Found a regression introduced by this PR — chart tasks (SingleChartValueRetrievalTask, MultiChartValueRetrievalTask, etc.) fail with ReferenceError: findElementInShadowDOM is not defined because merging all scripts into a single string prevents function declarations from being globally accessible when called from outside the init script context (e.g. via iframe.evaluate_handle). Note that window.WORKARENA_LOAD_COMPLETE = true still works because it's an explicit window. assignment — only bare function declarations are affected.

Fix is to add these 3 lines to js_utils.js right after the findElementInShadowDOM function definition (see:

WorkArena/src/browsergym/workarena/tasks/utils/js_utils.js

Lines 37 to 39 in ff7aebd

	// Explicitly expose on window so it is accessible via frame.evaluate_handle()
	// when all init scripts are merged into a single string (see base.py).
	window.findElementInShadowDOM = findElementInShadowDOM;

):

// Explicitly expose on window so it is accessible via frame.evaluate_handle()
// when all init scripts are merged into a single string (see base.py).
window.findElementInShadowDOM = findElementInShadowDOM;

Verified with test_task_general.py (all 165 tests passing, including all chart tasks) and a sample of L2 compositional tasks including DashboardRetrieve* tasks.

Please add these 3 lines and i'll merge as soon as this is included!