Working in the VM

Once the artwork boots in an era-correct machine, the research workbench (rvmc) is where you verify it, fix it, and freeze known-good states. This is the part that is actively being refined on the agentic side — the manual controls below are solid; the agent that drives them is improving.

Golden images

The catalog ships era-correct base images, cloned per project so your work runs on the machine of its time. Working today, each with a working guest agent (an in-VM helper): Windows 2000, XP / Server 2003, 7, 8.1, 10 (22H2), plus Mac OS 9.2.2 (PowerPC) with a MacPython agent over emulated networking. In progress: Windows 11 (and the Windows 10/11 agent transport), Windows 98 SE (the agent runs; period networking is being finalised), and Mac OS X Tiger (10.4) — so don’t rely on 98 SE or 11 shipping a fully working guest agent yet.

The guest agent is what lets the controller act inside the guest: run commands, read/write files, send keys, take screenshots, read the event log. If a golden can’t carry a guest agent for its OS, that’s called out — the agent is a hard requirement of a usable golden.

The screen

Open the VM’s screen (noVNC) to see and drive the work directly — mouse, keyboard, resolution. For old guests where the cursor or input behaves oddly, per-OS fixes are applied (e.g. a dot-cursor for XP/QXL, absolute-mouse handling for classic Mac).

Network capture

The most important research affordance: watch what the work actually reaches at runtime. Start a capture, run the work, and inspect the requests it makes — which confirms (or corrects) the autopsy’s static network findings with live behavior. A dead host seen here is a concrete restoration target.

For works that expect a server that no longer exists, a wayback-cache proxy can serve archived responses to the guest, so a work that phones home can be revived against an archive instead of the live (dead) internet.

Save-points

Freeze a known-good running state as a save-point — the work loaded, positioned, and behaving correctly. Save-points are how you capture “this is how it should look and run” so it can be returned to exactly, and they are the raw material a future exhibition instance would be built from.

Data-link CD

The original artwork plus its helper runtimes reach the guest on a generated data-link CD — no manual copying into the VM. Inside the guest the same files are also served over a local HTTP address.

The agent

The workbench exposes its capabilities to an LLM agent (via MCP) so an analyst can drive a VM the way a researcher would — start it, insert the CD, run the entry point, watch the network, take screenshots — and now read the autopsy report to decide which machine and what to watch for. This agentic layer is the current area of refinement; the underlying controls it calls are stable.

You supply the agent’s LLM — Bring your own AI.

Toward exhibition

When a work runs verifiably and you’ve captured a good save-point, it’s ready in principle to be locked into a long-running exhibition instance (evmctl). That automatic hand-off isn’t built yet; today you have a reproducible, documented, running research VM — which is the substance the exhibition stage will freeze.