feat: Add MVP Demo for LLM UI Action Planning

[abrichr]

Description:

This PR introduces the initial Minimum Viable Product (MVP) demo for planning UI actions using an LLM.

Key Features:

• Synthetic UI: Generates a sample UI (synthetic_ui.py) to mock visual parser output for rapid prototyping.
• LLM Planning: Takes the synthetic UI elements and a user goal, prompts an LLM (Anthropic via completions.py), and gets a structured action plan (core.py using Pydantic). The plan includes reasoning, action type, target element ID, and text-to-type.
• Visualization: Highlights the LLM-chosen target element on the synthetic UI image.
• Runnable Demo: Includes a top-level demo.py script to execute the flow.

Purpose:

Provides a working baseline demonstrating the core concept of LLM-driven UI action planning based on visual context.

To Run:

1. Ensure ANTHROPIC_API_KEY is set in your environment or .env file.
2. Run python demo.py.
3. Check the demo_output/ directory for generated images.
   
   

[abrichr]

This gives us a solid baseline with the multi-step demo, tests, and CI.

The next logical step is definitely moving towards interacting with real UIs. Thinking about what's next, the main paths seem to be:

1. Integrate Real OmniParser:

   • Replace synthetic_ui.py with actual screen capture (utils.take_screenshot).
   • Get the omniparser/client.py working to call a running OmniParser instance (needs URL config).
   • Map the OmniParser JSON response back to our List[UIElement].

2. Implement Real Action Execution:

   • Use the MouseController/KeyboardController to actually perform the planned click/type.
   • Handle coordinate conversion (normalized bounds -> absolute screen pixels).
   • Map the LLMActionPlan (action, text_to_type) to controller calls.
   • Replace the simulate_action call in the loop.

3. Refine LLM Interaction / Error Handling:

   • Improve prompt robustness for real UIs.
   • Handle LLM errors (bad JSON, invalid element IDs).
   • Handle action execution errors.

Suggest we tackle #1 (Real OmniParser integration) first. We need the system to perceive the real UI state before we can reliably plan and execute actions on it.