If you get your action weight from a function call, something like:

action_heal.get_weight(battle_context) => float;

You can use the information in battle_context (remaining HP, for instance) to dynamically change how likely each action is to be chosen, which seemed to be something you wanted to look into.

You can even return a weight of 0 if the battle_context would suggest the action would be useless.

For something much more complicated, you might want to look into a concept called a Behavior Tree. Even if you don't use one, you might learn a lot from how they're constructed. Video game AI, robotics, and all kinds of other stuff use them pretty often.

If you want to see an example of a what a really simple behavior tree is capable of (so simple it might be more accurately called a Decision Tree), you should look into Final Fantasy 12's Gambit system. This game actually lets players program their own party member AI to do certain things when certain conditions are met, and it ends up being a really useful if barebones model that I've based a few AI systems on. If you wanted to combine that with the weighted random-choice trick to add a little more decision variety, I'm sure there's a cool way to do that.