Improve the safety of the auto-approver

[darkogj]

Roo states this:

Going with this logic, they probably did NOT pay a lot of attention of the safety of the commands they execute. So you probably want to put a lot more safety measures (modifying the system prompt, etc.) if you plan on turning auto-approve on by default.

[janpaul123]

Ideas:

• Reading and writing of files seems to now be possible arbitrarily for the whole system. We can scope this to the current project, but add a way for the model to create a new project, when the user wants a new app from scratch.
• Executing terminal commands is harder to scope.
  • Our current default is to require explicit approval, which might be a bit too conservative.
  • Could use Dev Containers (well-supported in VS Code) but this requires Docker to be installed, and doesn't work for many use cases. Remote Dev Containers would be a nice feature, but again, not suitable for all use cases.
  • We could use some light checks, such as asking the model (separately) which directories would be affected by running a command, and requiring approval if this is outside the current project dir.

[gnrlbzik]

Scoping to the current project seems legit. VSCode already does this.

[iRonin]

I can see auto-approve docs but can't see the feature in my editor - is it in?

[adamhill]

There was also a PR over in the Rooniverse from @KJ7LNW for formal logic based permissions that I can't find right now, that was kinda cool

[iRonin]

Now that the Deny list is in, what I would like to see is the *-pattern or regex support in Approve/Deny lists.