Ms website explorer plugin

[msvana]

This branch contains a bit of a different approach - I separately try to perform Planning, Execution and Reflection. There is no single chat sequence. Instead there is always a standalone prompt for each action. These prompts share a "memory" of what was done in the past as well as current context.

Here are some observations:

• No matter what I do, the agent prefers using search over browsing through categories. I tried to tell it to choose the most deterministic, least risky path.
• There is still a lot of repetition. The agent tends to repeat the same step several times. The issue might be that the HTML provided by Playwright does not provide information about input field values.
• I also tried experimenting with creating 3 different strategies (a version of what is called Tree of Thoughts) and selecting the best one. Didn't really help.
• Still has issues with abstract reasoning. I am thinking a lot about making the agent think about how to translate requests such as "I want to find a PC for machine learning" to "I want to find a PC with at least 32GB of RAM, and an recent NVIDIA GPU")