Use JSON Schema for all requests

[nalbion]

After "removing" function calling, the new implementation requiring responses in JSON schema seems to be more reliable against several models including

• openai/gpt-3.5-turbo-16k
• mistralai/mistral-7b-instruct
• and apparently most local models

It seems too much to ask for an LLM to respond with EVERYTHING_CLEAR without getting extra spaces & \ escaping the underscore. They all seem quite happy responding in JSON though.

The new implementation should also allow for an arbitrary number of tools (see also #140) to be made available such as:

• add_task - adds a task to the backlog (CRUD)
• list_directory - lists files in a directory
• read_file - reads a file (start/end line optional)
• http_request - sends a GET, POST etc

[nalbion]

See also #266