By the LLM, either in a separate pass: "are any of these plugins relevant..." or giving the LLM the option "if you're unable to respond confidently then pick a plugin instead"
Embeddings and a vector database
UX needs to be seamless and quick, without slowing down the response
Implementation:
Ideally just a JSON file
easy to be customised (eg add a plugin repo on $HOME)
Plugin ideas:
git, GitHub cli
Todo.txt
Jira cli or other cli tools
Frameworks (eg Symfony)
editor, which encourages use of sed, etc
Clipea, which gets it to manage itself
general, which makes it a general assistant, so you can ask anything (maybe this gets its own shell alias)
docker, which tells it context like ps or presence of docker compose
net, which gives it ip/route/DBs info. Not sure if useful
web, which tells it which web scraping tools to use
history. Take a look at recent she'll history
different llm models
explanation mode that explains commands and their flags
Plugins would need to prove their worth. They'd only be useful if they performed better than the base setup
Recipies are only useful if they
save typing a long prompt
can't be replicated by something simple on the shell, like "ls | clipea"
solve common problems
it should be easy to test whether a recipe is needed / effective by comparing performance to when it's disabled
Consider security, especially if they're automatic and pull in more environment info.
Plugins would allow:
Plugins could be activated at run time. TBC:
--plugin git,@gitor similarUX needs to be seamless and quick, without slowing down the response
Implementation:
Plugin ideas:
Plugins would need to prove their worth. They'd only be useful if they performed better than the base setup
Recipies are only useful if they
Consider security, especially if they're automatic and pull in more environment info.