[dddddaviddddd]

I believe the author's idea is to do dev work from a Github account that only has access to the fork, but not to the main repo. Then, as a contributor, you'd open PRs from your fork to the main repo. I think this would only work if your Github account doesn't have write access to the main repo, though. I know you can use 'deployment keys' to give read-access to a single repo using an SSH key, but not sure if you can otherwise restrict access to a single repo with write access. Essentially, though, you'd want to find a way to give the remote host the most limited possible privileges to your Github account.

[throwaway173738]

You could also just set the development machine up as a remote on the repo on your local host and then pull, diff, and merge locally. Then the llm agent doesn’t have access to any github account at all.

[kstenerud]

I use an overlay copy of my workdir, then the sandboxed LLM doesn't get any of my secrets, can do its own commits, and I pull the ones back that I want.

[thwarted]

Oh, a separate GitHub account that has its own forks of the repos the agent is working on. Yeah, that's probably the most secure, isolated, and safest. The merge to the canonical repo then needs to go through a human, or at least separately controlled, process via a GitHub pull request.

[dezgeg]

Maybe this is doable with scoped API keys instead of SSH keys?

[dolmen]

On a GitHub project, agents must just be considered untrusted external contributors.