Right now GH-AW sandboxes the agent a lot, but I think in the general case we need the ability to:
- Run containers (possibly privileged even)
- Run (nested) VMs
What I'm thinking here is we change our "working" agents to always provision two GHA runners, one relatively smaller one that runs the agent, and a "sidecar" where the agent can interactively log into it over ssh (think tmate style) and it has the default full sudo access to the machine. This "development sidecar" VM can be a larger runner too.
This is an instance of the general recurring pattern I've seen of "sandboxed agent but with remote development workspaces".
Right now GH-AW sandboxes the agent a lot, but I think in the general case we need the ability to:
What I'm thinking here is we change our "working" agents to always provision two GHA runners, one relatively smaller one that runs the agent, and a "sidecar" where the agent can interactively log into it over ssh (think tmate style) and it has the default full
sudoaccess to the machine. This "development sidecar" VM can be a larger runner too.This is an instance of the general recurring pattern I've seen of "sandboxed agent but with remote development workspaces".