This moves setting the navigator.webdriver flag from after WebContent
creates the WebDriver connection, to its own IPC to be triggered from
WebDriver. This is closer to the spec, but mostly serves as an easy
test to validate the connection.
First, this moves the WebDriver socket to the /tmp/websocket/ directory,
as WebDriver now creates multiple sockets per session. Those sockets are
now created with Core::LocalServer rather than manually setting up the
listening sockets (this was an existing FIXME which resolved some issues
I was hitting with creating a second listening socket).
WebDriver passes both socket paths to Browser via command line. Browser
continues to connect itself via one socket path, then forwards the other
socket path to the WebContent process created by the OOPWV. WebContent
then connects to WebDriver over this path.
WebContent will temporarily set the navigator.webdriver flag to true
after connecting to WebDriver. This will soon be moved to its own IPC to
be sent by WebDriver.
This is essentially an ErrorOr<JsonValue, Web::WebDriver::Error> class.
Unfortunately, that ErrorOr would not be default-constructible, which is
required for the generated IPC classes. So this is a thin wrapper around
a Variant<JsonValue, Web::WebDriver::Error> to emulate ErrorOr.
Instead of verbosely repeating the NOTE and to_int() checks, move this
step to a helper. When we have real connected elements, we can then just
fix this helper.
This doesn't follow the spec to a tee. Our OutOfProcessWebView already
has a bitmap that can be used as the window screenshot. Therefore, we
can bypass the steps that assume we need to access the window's frame
buffer in-flight.
We also don't create an HTMLCanvasElement. We would need a Document in
the WebDriver process to do so. Instead, we can still run the encoding
steps exactly as-is using the screenshot bitmap.
Some endpoints, like /session/{id}/screenshot, will require sending
large data to the client. We won't be able to write all of the data in
one shot, so loop over the data until we've sent it all (or fail).
Otherwise, we end up propagating those dependencies into targets that
link against that library, which creates unnecessary link-time
dependencies.
Also included are changes to readd now missing dependencies to tools
that actually need them.
- It's not a "window object" (in the JS sense), it's a simple struct
storing a handle and is_open property
- We usually omit 'get' for getters
- The new name makes it more clear that this is looked up using the
m_current_window_handle as a key
