feat(sandbox): add pi sandbox image

[vegarsti]

Adds a new sandbox Docker image with Pi installed. We don't pin the version in the npm install command, since there are new versions very often. The network policy document is a trimmed version of the base policy and with additional GitHub release domains to allow Pi downloading binaries at startup (fd and ripgrep).

As an aside, I'm working on making it easier to authenticate headlessly in pi, such as authenticating a Codex subscription via a device code (which is what the Codex CLI does). This will make it smoother to set up providers in an OpenShell-sandboxed Pi. Feel free to follow along here: earendil-works/pi#3424

[vegarsti]

FYI @drew

[drew]

Sorry hit approve too early. Looks good though, just a couple doc requests. Can you update the main README to include pi in the list of sandboxes?

[vegarsti]

Thanks a lot @drew! Done!

[vegarsti]

Will you add it to https://docs.nvidia.com/openshell/about/supported-agents? 🙏🏻

[vegarsti]

Will you add it to https://docs.nvidia.com/openshell/about/supported-agents? 🙏🏻

Oh nevermind, it's open so here's a PR: NVIDIA/OpenShell#1563

[MichaelYochpaz]

Hey folks, I noticed Pi's sandbox uses ghcr.io/nvidia/openshell-community/sandboxes/base as a base image, and this base image includes Claude Code, Codex, OpenCode and Copilot inside it, meaning that when you use the Pi sandbox, you get all 5 harnesses, with a much bigger image size.

Is there a reason behind this? Shouldn't the base image just be a very minimal, thin image without any harnesses, that harness-specific images (like this one) extend and use as a base?