So the largest deciding factor for a model is always going to be what it's trained on.

Even if it knows how to string escape, it's not going to do it successfully.

You see this very clear when you tell it to try and execute commands over ssh.

We need protocols that pipe data directly to the shells it's using rather than trust everything through bash tunnels.

ultimately, something like python or yaml will win because they are defacto token efficient and well known to the LLMs.