Downloading PDF files freeze the agent

[n-sviridenko]

When the agent clicks a PDF file link to download it, it opens PDF in a new tab (due to Chromium's default PDF viewer capability), this makes the agent losing the initial page (where there are other links to click) and not coming back. And then it just hangs on that PDF page forever.

[n-sviridenko]

In order to reproduce it, create a page with lots of links to PDF files and tasks Skyvern to click all of them.
It'll only visit the first PDF.

[n-sviridenko]

Btw. we also need to download PDFs (and access them later #1347) instead of just viewing them.

[n-sviridenko]

Was also trying to see how downloading invoices is currently implemented https://docs.skyvern.com/getting-started/skyvern-in-action#log-into-a-portal-and-download-invoices

[n-sviridenko]

I assume implementation is:

1. subscribing to opened pages (before await locator.click)
2. if a new one just got opened
3. try fetching it (HEAD request, but w/ cookies of the current session, e.g. via calling fetch)
4. if content type from headers is PDF, then come back to prev page and download programmatically using the same URL and cookies from current session

Place to change:

skyvern/skyvern/webeye/actions/handler.py

Line 390 in 92d9495

async def handle_click_to_download_file_action(

[suchintan]

We actually have this implemented in Skyvern cloud similar to what you did here: https://github.com/Skyvern-AI/skyvern/pull/1349/files but I think we forgot to bring it to the open source repo

@LawyZheng can you please review https://github.com/Skyvern-AI/skyvern/pull/1349/files and make any augmentations? We should open source our PDF handling logic

[LawyZheng]

i just left some comments a few minutes ago.

[suchintan]

@LawyZheng let's get our closed source changes for PDFs ported over to the open source version

[LawyZheng]

Fixed in #1363