v0.27.0

Changes

• Standardize time parameters to seconds: All time-related parameters now use seconds consistently

  • typing_delay: 80ms → 0.05s
  • wait_for_load/wait_for_network_idle timeout: 10000ms → 10.0s
  • keyboard_type delay: 80ms → 0.05s

• Add raise_on_timeout parameter: Agent.wait_for_load() and Agent.wait_for_network_idle() now accept raise_on_timeout (default: True)

  • When False, silently returns after timeout instead of raising TimeoutError
  • Useful for lenient waits where you want to continue regardless

Example

# All time params now in seconds
agent = await wt.create_agent(
    llm=llm,
    wait_after_action=1.0,   # seconds
    typing_delay=0.05,       # seconds (was 80ms)
)

# Lenient wait - continue even if timeout
await agent.wait_for_network_idle(timeout=5.0, raise_on_timeout=False)

v0.26.0

Changes

• Live form value parsing: DOM context now captures live input values (inputValue, textValue), checkbox/radio states (inputChecked), and select option states (optionSelected) from CDP snapshots
• Configurable typing delay: Added typing_delay parameter to Agent (default: 80ms)
• Centralized constants: Added constants.py with DEFAULT_WAIT_AFTER_ACTION and DEFAULT_TYPING_DELAY

Bug Fixes

• Fixed issue where user-typed values (e.g., search queries) weren't visible in DOM context
• Fixed 17 deselected tests by adding missing @pytest.mark.unit markers

v0.25.1

Bug Fixes

• Fix TypeTool clear parameter not working: The clear parameter in the type tool was using element.fill("") which would lose focus before keyboard_type. Now passes clear to keyboard_type which uses JavaScript to clear the active element while maintaining focus.

Other Changes

• Add demo video example and GIF to README
• Add videos/ to gitignore

v0.24.0

Changes

• TypeTool and TypeAtTool: New tools that click-then-type in one action for more human-like interaction
• SelectTool: New tool for dropdown selection
• Simplified message types: Single Message class with Role enum instead of separate UserMessage/AssistantMessage classes
• Tools receive wait_after_action directly: Cleaner tool initialization without browser.wait() calls
• ToolParams base class: Rejects extra parameters for stricter validation
• Fixed clear functionality: Improved action timeout handling

Breaking Changes

• Removed test recording/replay infrastructure (webtask.testing module)
• Removed e2e test workflow from CI

v0.23.2

Changes

• Fast-fail timeout for element actions: Element actions (click, fill, type, upload) now use a 100ms timeout instead of Playwright's 30s default. This prevents long waits when clicking non-existent or hidden elements while still performing actionability checks.

Details

• Added DEFAULT_ACTION_TIMEOUT = 100ms in playwright_element.py
• Applied timeout to click, fill, type, upload_file methods
• Removed context-level set_default_timeout from playwright_browser.py
• Deleted unused constants.py

v0.23.1

Changes

• Add ToolParams base class with extra="forbid" to reject invalid LLM tool parameters
• All tool Params classes now inherit from ToolParams
• Export Tool and ToolParams from webtask.llm
• Add test for extra field rejection
• Move TODO from README to docs/todo.md

v0.23.0

Changes

Human-like interaction

• Remove FillTool and TypeTool (element-level typing)
• Remove TypeTextAtTool from pixel tools
• Add KeyboardTypeTool (page-level keyboard typing)
• User flow: click element to focus, then type on keyboard

Improved scrolling

• ScrollDocumentTool now scrolls 50% of viewport (maintains context, won't cut elements in half)
• Uses JavaScript window.scrollBy() instead of PageUp/PageDown

Tool organization

• DOM mode: click, upload + keyboard tools
• Pixel mode: click_at, hover_at, scroll_at, scroll_document, drag_and_drop + keyboard tools
• Keyboard tools (both modes): type, key_combination

v0.22.1

Changes

• Keep li, ul, ol elements as semantic HTML tags
• Dropdown menu items now get element IDs (e.g., li-0, li-1)
• Add tests for semantic knowledge functions

v0.22.0

Changes

• Simplify agent modes from three (text/visual/full) to two (dom/pixel)
• dom mode (default): Uses element IDs for interaction, includes DOM + screenshots
• pixel mode: Uses screen coordinates, includes screenshots only
• Move ScrollDocumentTool to pixel mode only
• Both modes now get screenshots for better LLM context

v0.21.4

What's New

• Added [END OF PAGE] marker: DOM context now ends with a clear marker to help the LLM understand where page content ends, preventing it from echoing back DOM elements in responses.