# Agent notes How to build and test this image as a coding agent. Follow [STYLE.md](STYLE.md) for every change. ## Container `mkosi vm` needs systemd on the host. A dev container has none, so run mkosi in a podman container that: - boots systemd (`--systemd=always`, with `/sbin/init` as its command) - is privileged and has `/dev/kvm` - has `mkosi qemu-kvm-core edk2-ovmf systemd-repart openssh-clients` and `$(mkosi dependencies)` installed - mounts the repository, such as at `/repo` Then run mkosi with `podman exec mkosi ...`. podman resolves host paths on the host, not in the dev container, so `~` and the scratchpad don't exist there. Mount paths inside the repository instead, such as `./mkosi.output/...`, which is ignored by git. ## Test VM The `[Runtime]` section of `mkosi.conf` sets up the test VM. It boots a throwaway copy (`Ephemeral=yes`), so every boot is a first boot. The setup takes its answers from credentials: root's password is `test`, and the `claude` user gets the keys in `~/.ssh/*.pub` of whoever runs mkosi (`mkosi.credentials/agent.ssh-key`). - If the container has no `~/.ssh`, make a throwaway key in `mkosi.output/test-ssh/` and mount it at `/root/.ssh`. - `mkosi vm` needs a terminal. Without one, run it with `podman exec -d -t` and send its output to a file. - ssh runs inside the container, which forwards port 2222 to the VM: `podman exec ssh -p 2222 claude@localhost`. Pass `-F /dev/null` if ssh rejects the mounted config because of its owner. - For root, run `su -` over ssh. ## Image version `mkosi.version` prints the build time, so every mkosi run has a new version. `mkosi vm` then looks for an image that was never built and fails with "has not been built yet". Either rebuild with `mkosi -f vm`, or boot an existing build by pinning its version: ```sh mkosi --image-version 202611171816 vm # the version in mkosi.output/agent_.raw ``` Pinning saves a full rebuild when only testing. ## Boot time Measure it over ssh with `systemd-analyze`, `systemd-analyze blame` and `systemd-analyze critical-chain`. In `blame`, the `.device` units only show when udev announced the device, not time spent, so filter them out. The test VM has one CPU and always does a first boot (setup, growfs), so a real VM boots faster.