v0.3.1

Fixed

• Pointer actions answer once the page has handled the event. With the
  firefox-39 engine, a press, a move or a release returns after the page's
  listeners have run. Before, the next call could reach the page first: under
  load a short drag-and-drop lost its drop, and a script read right after a
  double click could count one click short.

Changed

• A click that opens alert() is handled by the engine. The engine ends
  its own wait when a dialog opens, so this package no longer carries a hook
  for it.

v0.3.0

Changed

• The Juggler client comes from invisible-core (38.34.0,
  invisible_core.juggler), the one copy every invisible_ wrapper shares.
  This package carried its own, taken from invisible-playwright on
  2026-09-24, and it now has the fixes made there since:
  • typing into a field waits the moment a person takes after focusing it,
    and longer while the field is still changing, so a page that writes into
    the field when it gets the focus no longer overwrites what is typed;
  • a picked file reaches the input after the hesitation a person takes to
    choose it;
  • an action that needs no point on the screen (setting the files of a
    hidden <input type=file> behind a styled button, focusing or
    dispatching an event on a hidden element) no longer waits for the element
    to become visible until it times out;
  • every page of a session numbers its own clicks, fields and typed strings,
    so two tabs never draw the same movement.
• A preference value that is not a bool, an int, a float or a string is
  refused with a TypeError naming it, instead of being written as its
  text.
• The license is MIT only. The two files taken from Playwright's driver
  (Apache-2.0) moved to invisible-core with the client.

v0.2.2

Changed

• The window a page reads is the one Windows Firefox draws. Through
  invisible-core 38.33.0 and the firefox-38 engine. A maximized window
  answers screenX -8 and an outerWidth 16 wider than screen.availWidth
  at 100%, as Windows Firefox does, where it answered 0 and the work area; its
  content starts where retail's does at every display scale; a popup answers
  its own size, frame and position, and one opened without a position is
  placed where Windows Firefox places it.

Removed

• The pins screen.chrome_w, screen.chrome_h, screen.window_x and
  screen.window_y. The window follows screen.dpr, which takes 1, 1.25,
  1.5 or 2.

v0.2.1

Changed

• The screen a page reads is in CSS pixels. Through invisible-core
  37.33.0 and the firefox-37 engine. A 1920x1080 panel at 125% reports
  screen.width 1536 and screen.height 864, as Firefox on that monitor
  does. Before, it reported the panel's device pixels. The default
  viewport is derived the same way.

Fixed

• On Linux the virtual display keeps X access control on. The Xvfb
  display opened for a headed session without a screen used to run with
  access control off (-ac), so any local process could connect to it and
  read or drive the browser window. It now gets a private session cookie,
  handed only to the browser, and removed when the session stops.

v0.2.0

Changed

• The language of locale="auto" is the real Firefox of the egress
  country, by share. Through invisible-core 36.33.0 (the required version
  is exact, as before). A country used to map to one country tag (de-DE)
  that no Firefox build has, so a German egress sent de-DE,de;q=0.9,...
  where a German Firefox sends de,en-US;q=0.9,en;q=0.8, and every session
  of a country spoke the same language. Now each country carries the Firefox
  builds people there run and their share, from Mozilla's Firefox Public
  Data Report (Belgium: French 40%, Dutch 31%, US English 22%), and the egress
  IP picks one: the same address always gets the same language, and the
  addresses of a country spread over its builds as measured.
• An explicit locale is navigator.language. locale="en-AU" reports
  en-AU, en-US, en (it reported en-US, en), and "fr-FR" reports
  fr-FR, fr, en-US, en (it reported fr first): the requested tag goes
  first and Firefox's own table gives the tail.

v0.1.3

invisible-selenium 0.1.3 is on PyPI.

Artifacts and checksums: https://pypi.org/project/invisible-selenium/0.1.3/

v0.1.2

Changed

• Values a user picks are committed by Firefox itself. Select, a click
  on an <option>, send_keys on a file input and clear() on a date or time
  field hand the choice to the engine's native input commands
  (Page.selectOptions, Page.setUserInput, Page.setFileInputFiles), which
  take the same paths as the dropdown, the date box and the file picker. The
  page gets input and change with the shape a user's pick gives, inside
  shadow roots too, and nothing when the choice did not change. The command
  this used to call, Page.dispatchTrustedInputEvents, is gone from the
  engine, so this needs the engine that ships the new commands and lands with
  the core pin that seals it.

Fixed

• A click lands where it was asked on a page with a saved zoom, and on
  Linux behind a SOCKS proxy real sites see WebRTC working: both are fixes in
  the firefox-35 engine.

Requires

• invisible-core 35.32.0, which seals the firefox-35 engine. This version
  does not run on firefox-34, and earlier versions do not run on firefox-35.

v0.1.1

Fixed

• clear() leaves the field, and change comes from Firefox, once. It
  deleted the text and then asked the engine for a change while the field
  still had focus: cancelable and composed, followed by Firefox's own at the
  next blur, and sent even to a contenteditable, which never fires one. Inside
  a shadow root that request failed and clear() raised after emptying the
  field. It now runs Selenium's unfocusing step after the Delete, so change
  comes from the blur, as it does for a user.
• On Linux, a headless session starts on its own virtual display. The
  display counted as ready as soon as its lockfile existed, which Xvfb writes
  before it opens its sockets, so sessions started together could land on
  each other's display and the browser failed with cannot open display.
  The display is also no longer reachable over TCP.

Requires

• invisible-core 34.32.0, which carries the display fix. Same engine
  (firefox-34).

v0.1.0

First version: a replica of invisible_playwright 0.25.7 with Selenium's contract.

• webdriver.Firefox launches the same patched Firefox with the same profile,
  proxy, geography and hidden-surface handling as InvisiblePlaywright, and
  drives it over Juggler: no geckodriver, no Marionette, navigator.webdriver
  stays false.
• Selenium's public API: WebDriver, WebElement, ShadowRoot, SwitchTo,
  Alert, ActionChains, ScrollOrigin, By, Keys, WebDriverWait,
  expected_conditions, Select, Timeouts and the exception hierarchy.
• Every click, hover and drag travels along a path drawn from the session seed;
  keys carry the session's typing rhythm; input events are trusted.
• execute_script runs in the page's world and reads its result back through
  the engine's Debugger, without running a serializer the page can observe.

On PyPI: https://pypi.org/project/invisible-selenium/0.1.0/