* 'main' of https://github.com/browser-use/browser-use: (269 commits)
fix for follow_up_tasks example
fix for current page detection grabbing a chrome extension service worker
add support for close_tab action
disable cross-origin iframes support for now
fix: update hatch build includes for correct package structure
fix the simple example
typofix
fix error: name 'browser_class' is not defined
Updated readme
streamlined to match other examples
Added env variables in example
Update custom-functions.mdx
fix spaces vs tabs
Update supported-models.mdx
fix chrome connecting to extension context by accident
fix broken BrowserConfig chrome_instance_path kwarg
simplify test task
fix removing contenteditable check by accident
tweak salesforce test cross-origin iframe task
ignore invisible and irrelevant cross-origin-iframes
...
- Enhanced model name matching for DeepSeek models using 'in' operator
- Improved think tag removal to handle stray closing tags
- Updated message conversion and plan processing for DeepSeek models
Fixed an issue where the event loop was closing before all HTTPX connections were properly terminated, causing resource warnings and potential memory leaks. The problem occurred because browser-use was shutting down threads before Playwright had fully closed its HTTP connections. Now, before shutting down the agent, all HTTPX processes are explicitly closed, ensuring a clean shutdown and preventing unclosed transport warnings.
- Modify SystemPrompt class to support optional system message override and extension
- Update Agent and AgentSettings to include new system prompt customization parameters
- Update documentation and example to demonstrate new system prompt customization methods
- Remove deprecated system_prompt_class parameter
- Provide two ways to customize system prompt: extend or override completely
- Add example for setting success to false when task is not fully completed
- Provide more explicit guidance on reporting task completion status
- Improve readability of last step instructions
- Update last step instructions to emphasize single action requirement
- Clarify guidance for using the "done" action
- Improve wording to highlight task completion and success status reporting
- Refactor last step warning message to be more concise and direct
- Clarify instructions for using the "done" function
- Improve guidance on task completion and success status reporting
- Update last step warning message in service.py to provide clearer, more explicit instructions for the "done" action
- Modify test_models.py to comment out DeepSeek model and adjust model testing configuration
- Enhance guidance on task completion, success status, and action requirements
- Add type hints and type ignore comments in service.py for image message handling
- Update system prompt to provide clearer guidance on done action usage
- Refine action description for done action in controller service
- Enhance documentation for task completion scenarios
- Adds error entry to history when max steps limit is reached
- Ensures consistent error tracking in Agent History List
- Improves error visibility for debugging purposes
Fixes#439
- Update service to conditionally limit actions based on max_actions_per_step
- Modify views to use json_schema_extra for action validation
- Improve action list handling to prevent unnecessary slicing
- Add support for 'raw' and 'json_mode' tool calling methods
- Introduce new methods for setting message context and tool calling method
- Modify input message conversion for specific model names
- Update tool calling method detection logic
- Extend ToolCallingMethod type to include new methods
- Extract common task completion logging into a new `log_completion()` method
- Simplify duplicate logging code in `run()` and `run_with_max_steps()` methods
- Maintain existing logging behavior for task completion status
- Add support for creating a done-only action model
- Modify logging to distinguish between successful and unfinished tasks
- Update agent to switch to done action model when max steps are reached
- Improve system prompt to clarify done action usage in max steps scenario
- Extend registry action model creation to support filtering actions
- Update telemetry event to include more comprehensive agent performance metrics
- Modify `AgentEndTelemetryEvent` to capture total input tokens and duration
- Refactor `is_successful()` method to return `None` if agent is not done
- Adjust `ActionResult` default success value to `None`
- Capture token count before generating model output
- Store tokens in a local variable to ensure accurate input token tracking
- Modify metadata logging to use the pre-computed token count
- Update `_add_message_with_tokens` and `add_message` methods to support `None` as default position
- Modify last step warning message in agent service
- Add logging for last step completion
- Improve type hints for message position parameters
- Introduce warning message for the last step in the agent
- Add `success` parameter to `ActionResult` and `DoneAction`
- Modify `AgentStepInfo` to check for last step
- Update `AgentHistoryList` methods to handle optional values and success status
Carlos A. Planchón