Bringing back System 7 in three days (with help from agents)

I’ve wanted to reimplement Mac System 7 for a long time. I love the era — the Chicago font, the menu bar that does its own thing, the desktop patterns, the sound. But the sheer surface area of System 7 is intimidating. People have spent years in Ghidra on smaller systems.

This summer I tried something I’d been putting off: I let an LLM do the grunt work alongside Ghidra, with me in the loop for review and verification. Three days later, I had a bootable prototype with a working Finder running under QEMU.

The framework is in the Zenodo paper, but the short version:

• Provenance over heroics. Every claim about what a function does is traceable back to the binary location and the analysis pass that produced it. No tool gets to “remember” something; every fact is an artifact in a tree.
• Verification by screenshot. Instead of trying to score “accuracy” of reconstructed code abstractly, the pipeline boots the result in QEMU and diffs against captured frames of the real System 7.1 — Chicago font rendering with proper kerning, menu-bar behaviors, desktop patterns, icon display.
• Humans gate every commit. The LLM proposes; I review. The structure of the workflow makes review fast (because the artifacts are well-shaped), not absent.

The result is on GitHub at Kelsidavis/System7. The README puts current state at “~94% core functionality complete” and calls it a proof of concept — there’s still a long way to go before it’s something you’d hand to a non-developer. But the workflow itself is the point.

Hackaday wrote it up a few days after I posted the paper, which was both flattering and terrifying.

I think this kind of structured, verifiable, AI-assisted reverse engineering is going to be how a lot of dead software gets brought back over the next few years. And not just for nostalgia — there are real embedded systems out there with no source, no documentation, and no living authors. We need methods for that.

If you want the long form, the paper is here.