Building a custom agent harness to test Omarchy in QEMU
- YT :: https://www.youtube.com/watch?v=GOu_05J_fdc
- Original title :: 🚨🚨 Omarchy Automation Investigation (Building Custom Harness) 🚨🚨
A two-and-a-half hour live build stream. Prime works on "oligarchy", a homegrown harness that boots Omarchy ISOs in QEMU VMs and lets coding agents drive the install to the desktop, with Linear as the test-definition and result store. Most of the session is fighting his own generated code — argument parsing, Drizzle migrations, unhelpful error output — plus a long defensive tangent about the Omarchy/DHH backlash and a Linear-versus-Jira riff.
What the harness is
A server wrapping QEMU (he calls the client "Kimu") plus a client that mirrors QEMU commands. Give the server an ISO URL and it downloads and caches the image on the machine; concurrent requests for the same ISO enter a downloading phase so two agents cannot clobber each other — demonstrated live with two agents sharing one download. A Cloudflare tunnel exposes the local machines to the internet so the client can also run as Cursor cloud agents (Tailscale would work, he just already uses a lot of Cloudflare). PlanetScale plus Drizzle for the database, Sentry for anything that 500s or crashes unnoticed, since the point is a harness robust enough to trust. The driving prompt is roughly "launch a QEMU server, launch the client, use the latest Omarchy ISO, and drive it until you reach the desktop."
Deliberately no unit tests: the test is that the whole program runs end to end. Also on screen: "screenplay", the terminal UI project that has been eating his video-production time — webcam in the terminal at 30fps, DNS visualizations, blinking status lights he is annoyed do not animate.
Interfaces: Vim or mobile, nothing else
Strong stated preference that everything be drivable from Neovim/CLI or from a phone, with no desktop web UI in between — "yo, run a build; what was the status of the build?" on the go, and fast keyboard-driven listing in the editor otherwise. He keeps a service-less second phone purely for recording and prompting agents.
Linear as the controller
Each test definition should create its own pending test result, with each run linked to exactly one definition — he catches the agent creating a single ticket holding all six definitions. Linear is the coordinator because something has to be, and he wants to try its recurring agent workflows and looping to kick runs off against his server. Disclosed as a paid partner, but the Jira comparison is unpaid and unkind: years of JQL, burndown charts, ceremony around reporters and components; he has not opened Jira in years.
Fighting the generated code
The recurring theme. The CLI parses one flag with a space and another with an equals sign, so he asks for a single generic schema-driven parser — or better, an off-the-shelf one like yargs — and to delete the ad-hoc parsing. Pending Drizzle migration files pile up and have to be collapsed to one that matches the remote database. An insert into test_runs fails with no stack trace at all, which he calls undebuggable and fixes by demanding more error detail. Effect keeps coming up (TJ pushing it in chat) and he admits he does not understand its generators and yields well enough to justify the code, so he has Grok research it and at least insist on schemas.
Model notes in passing: Grok is his workhorse and cheap; Fable's output he finds unpleasant to read and too expensive to reach for casually; he has not tried DeepSeek or the new GLM models. On MCP: he is close to not standing it, but concedes it depends — a Sentry MCP that pulls the latest issue and explains it is genuinely useful, so neither "all MCPs good" nor "all bad" holds. Harnesses like this do gobble tokens, and one agent burned ~100k arguing about a UUID.
The Omarchy tangent
A chunk of the stream is him defending the project against chat. Claims about it being "a big sloppy security hole" he intends to actually check by reading through the Omarchy code with Grok's security harness. On the DHH hostility: he has made videos mocking DHH's ideas himself and does not agree with everything, but $12M raised to improve the Arch ecosystem, billionaires funding open source, and paying maintainers so Hyprland can get better are good outcomes, and there have never been more people newly excited about Linux. His own stake: he wants to be able to walk into companies and not have to use a Mac.
He ends short of the goal — local runs still awkward even though the client is meant to be remote from the server, and the Linear loop research untouched — and floats a follow-up stream.