I would suggest not to use bash as we do not know whether the user is actually using bash or another shell (such as zsh, fish, ...). I would suggest to either use the more generic shell (maybe even request the command to be POSIX compliant).

Alternatively one may retrieve information about the shell in use, e.g. from the SHELL env variable.

@heiderich @xenia-lang I'd like to get this PR merged—the idea seems useful to me.
What do you think about getting the current shell name as well as the OS to cover all cases?

e.g. using

get_shell()
  if [ -n "$SHELL" ]; then
    current_shell="$(basename "$SHELL")"
  else
    current_shell=""

and the adapting the prompt to:

  role="You translate the given input into a shell command for $os. The user is running $current_shell. You may not use natural language, but only the command as an answer."

Sounds good

@heiderich @xenia-lang I'd like to get this PR merged—the idea seems useful to me.
What do you think about getting the current shell name as well as the OS to cover all cases?

e.g. using

get_shell()
  if [ -n "$SHELL" ]; then
    current_shell="$(basename "$SHELL")"
  else
    current_shell=""

and the adapting the prompt to:

  role="You translate the given input into a shell command for $os. The user is running $current_shell. You may not use natural language, but only the command as an answer."

Does the assignment to os in os=get_os have any effect at all?
If I read the definition of get_os correctly, then it doesn't actually return any value, but directly assign the OS name to the os variable.

To return something, shouldn't it be:

get_os() {
  unameOut=$(uname -s)
  case "$unameOut" in
    Darwin)
      echo "MacOS"
      ;;
    Linux)
      echo "Linux"
      ;;
    *)
      echo "Linux"
      ;;
  esac
}

And then the function call would be os=$(get_os)