Week 0x2A

The best way to prove out the new tree-walking task-focused loop will be to get the server sandboxes working. So, we spent a bunch of iterations getting delegation up and running. Sample tasks were working, it felt good to see background work get done.

Then I scaled up to a meaningful job and it became clear that Builder Bear had misunderstood and was provisioning a sandbox per task. Spinning up a container and connecting an armature just for one small action. :facepalm:

Mostly resolved that and then put a relay in place to allow some networking. But the Hugging Face incident where an OpenAI agent escaped their sandbox made me think that even tighter controls would be more future-ready. So now there is an option to deny the sandbox networking entirely and pre-load dependancies (starting with Rust’s Cargo toolchain). It’s overkill, but it’s my project and I learn by overengineering.