chore: simplify image configuration

Drop the custom firewall, PS1, oomd and nftables overrides, trim sysctl
settings, and add the Secure Boot certificate.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Job79andClaude Opus 5.5 committed 2026-11-16 19:40:51 +01:00
1 parent 9245b26500
commit bbd0364792
14 files changed
+52 -82

No files matched your search

+7 -2
View File
@@ -19,11 +19,16 @@ jobs:
AUTH: "Authorization: token ${{ github.token }}"
steps:
- name: Install tools
run: dnf install -yq mkosi
run: dnf install -yq mkosi sbsigntools
- name: Check out
uses: builtin:checkout
# Signs systemd-boot and the UKI for Secure Boot (mkosi.conf, [Validation]).
- name: Build
run: mkosi build
env:
KEY: ${{ secrets.SECUREBOOT_KEY }}
run: |
(umask 077 && printf '%s\n' "$KEY" > mkosi.key)
mkosi build
- name: Compress
working-directory: mkosi.output
run: |
+2
View File
@@ -1,2 +1,4 @@
# Build outputs.
mkosi.output/
# The Secure Boot private key (the SECUREBOOT_KEY secret); mkosi.crt is its certificate.
mkosi.key
+8 -5
View File
@@ -4,13 +4,13 @@ A VM that runs [Claude Code](https://claude.com/claude-code) in the background,
- **Resets at reboot:** the OS is read-only, so Claude can't change it, and Claude's home, `/tmp` and all containers are wiped at every boot. Only the setup answers and Claude's login survive (`/var/lib/agent`).
- **Claude as non-root, tools in containers:** Claude runs as the user `claude`, without root, and runs its toolchains in containers with rootless podman (also as `docker`). `/etc/claude-code/CLAUDE.md` tells it about the machine; `/etc/claude-code/managed-settings.json` lets it run without asking permission. The host has only git, ssh and basic command-line tools.
- **Self-updating:** like [git-runner-mkosi](../git-runner-mkosi): images are rebuilt weekly (also picking up the latest Claude Code) and on every push. Hosts download updates with systemd-sysupdate, check their signature (`SIGNING_KEY` secret), write them into a second root partition (A/B) and reboot into them daily at about 04:10. That reboot also resets Claude's work.
- **Self-updating:** like [git-runner-mkosi](../git-runner-mkosi): images are rebuilt weekly (also picking up the latest Claude Code) and on every push. Hosts download updates with systemd-sysupdate, check their signature (`SIGNING_KEY` secret), write them into a second root partition (A/B) and reboot into them daily at about 04:10. That reboot also resets Claude's work. A new version gets three tries to boot; if none gets through, systemd-boot goes back to the previous one.
- **Read-only system:** systemd-boot starts a UKI that mounts the root filesystem, /etc and /usr included, from a read-only erofs partition checked by dm-verity. All state (Claude's home, its login, the root password, logs) lives on `/var`, its own partition.
- **Hardened:** a firewall that drops all incoming traffic but ssh, KSPP sysctls and kernel arguments, and systemd-oomd kills Claude's work, not the system, when it runs the VM out of memory.
- **Secure Boot:** systemd-boot and the UKIs are signed with our own key (`mkosi.crt`; the private key is the `SECUREBOOT_KEY` secret). systemd-boot enrolls it at first boot if the VM's firmware has no keys yet; the firmware then boots nothing else.
## Install
Download [the disk image](https://git.plabble.org/api/packages/Job79/generic/agent/latest/agent.qcow2) (built by the `os` workflow), resize it to the disk size you want (/var grows to fill it) and boot it with UEFI. The screen and the serial console ask for a root password, a hostname and an ssh public key for the user `claude`.
Download [the disk image](https://git.plabble.org/api/packages/Job79/generic/agent/latest/agent.qcow2) (built by the `os` workflow), resize it to the disk size you want (/var grows to fill it) and boot it with UEFI in Secure Boot setup mode (no keys enrolled; in Proxmox, an EFI disk with "Pre-Enroll keys" unchecked), so it enrolls our key. The screen and the serial console ask for a root password, a hostname and an ssh public key for the user `claude`.
Then log Claude in once:
@@ -29,11 +29,14 @@ As soon as Claude is logged in, `claude-remote.service` starts `claude remote-co
mkosi build
```
`mkosi.output/` then holds the disk image (`agent_<version>.raw`), and the UKI and root partitions that the `os` workflow publishes as an update. To boot the disk image (needs `qemu-kvm-core edk2-ovmf`; ssh on port 2222):
The build signs with `mkosi.key`: the `SECUREBOOT_KEY` secret, or for a local test a key of your own (`mkosi genkey` writes `mkosi.key` and `mkosi.crt`; don't commit them).
`mkosi.output/` then holds the disk image (`agent_<version>.raw`), and the UKI and root partitions that the `os` workflow publishes as an update. To boot the disk image with Secure Boot (needs `qemu-kvm-core edk2-ovmf`; `OVMF_VARS.fd` has no keys, so it enrolls ours; ssh on port 2222):
```sh
qemu-img convert -O qcow2 mkosi.output/agent_<version>.raw test.qcow2 && qemu-img resize test.qcow2 40G
cp /usr/share/edk2/ovmf/OVMF_VARS.fd vars.fd
qemu-kvm -m 8G -smp 4 -cpu host -nic user,model=virtio-net-pci,hostfwd=tcp::2222-:22 -drive file=test.qcow2,if=virtio -device virtio-balloon-pci,free-page-reporting=on \
-drive if=pflash,format=raw,readonly=on,file=/usr/share/edk2/ovmf/OVMF_CODE.fd -drive if=pflash,format=raw,file=vars.fd
-machine q35,smm=on -global driver=cfi.pflash01,property=secure,value=on \
-drive if=pflash,format=raw,readonly=on,file=/usr/share/edk2/ovmf/OVMF_CODE.secboot.fd -drive if=pflash,format=raw,file=vars.fd
```
+10 -7
View File
@@ -25,7 +25,6 @@ Bootloader=systemd-boot
# - systemd-pam login sessions (pam_systemd): Claude's user manager and its
# D-Bus, which rootless podman needs
# - openssh-server ssh login for the user (keys only)
# - nftables-services the firewall
# - zram-generator-defaults compressed swap in RAM
# - podman, podman-docker containers, for all of Claude's toolchains (also as docker)
# - git-core, openssh-clients Claude Code uses git; it pushes over ssh
@@ -44,7 +43,6 @@ Packages=
shadow-utils
systemd-pam
openssh-server
nftables-services
zram-generator-defaults
podman
podman-docker
@@ -67,7 +65,7 @@ UnifiedKernelImageFormat=%i_%v
# (mkosi.extra/etc/machine-id is empty, so no boot counts as the first), no systemd-firstboot
# questions (agent-setup asks its own), freeing memory unused for an hour (damon_reclaim, in µs; the
# watermarks make it run whatever memory is free), so the host can take it back (virtio-balloon
# free page reporting), and KSPP hardening (the last four).
# free page reporting).
KernelCommandLine=
console=tty0
console=ttyS0,115200n8
@@ -79,10 +77,6 @@ KernelCommandLine=
damon_reclaim.wmarks_high=1000
damon_reclaim.wmarks_mid=999
damon_reclaim.wmarks_low=0
slab_nomerge
init_on_alloc=1
randomize_kstack_offset=on
page_alloc.shuffle=1
# Kernel modules in the initrd: only what finds and mounts the disks of a VM.
KernelInitrdModules=
virtio_blk
@@ -122,3 +116,12 @@ RemoveFiles=
/usr/lib/systemd/system/systemd-userdbd.*
/usr/lib/systemd/system/systemd-nsresourced.*
/usr/lib/systemd/system/sshd-keygen*
[Validation]
# Signs systemd-boot and the UKI for Secure Boot, and puts our keys on the ESP, where systemd-boot
# enrolls them at first boot if the firmware is in setup mode (no keys enrolled; in Proxmox, an EFI
# disk without pre-enrolled keys). mkosi.crt is the certificate; mkosi.key, its private key, is the
# SECUREBOOT_KEY secret of this repository, which the os workflow writes to mkosi.key.
SecureBoot=yes
SecureBootKey=mkosi.key
SecureBootCertificate=mkosi.crt
+18
View File
@@ -0,0 +1,18 @@
-----BEGIN CERTIFICATE-----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-----END CERTIFICATE-----
-1
View File
@@ -12,4 +12,3 @@ You run on a small Fedora VM made for you, as the user `claude`, without root.
- **Everything resets at reboot**: your home (/var/home/claude), /tmp and all containers and
images. Only your Claude login survives. Keep results somewhere else (push them to git, for
example). The VM may reboot at about 04:10 to update.
- Incoming traffic is blocked except ssh; to serve something, the user needs a tunnel.
-2
View File
@@ -1,2 +0,0 @@
# The prompt: hostname and directory on a purple bar (the arrow needs a Powerline/Nerd font).
export PS1="\[\e[30;45m\] \h | \w \[\e[0;35m\]\[\e[m\] "
-18
View File
@@ -1,18 +0,0 @@
# Firewall: drop all incoming traffic except replies, loopback, ICMP, DHCPv6 and ssh, and forward
# nothing (rootless podman's networks live in Claude's own network namespace). Loaded by
# nftables.service.
destroy table inet firewall
table inet firewall {
chain input {
type filter hook input priority filter; policy drop;
ct state established,related accept
ct state invalid drop
iif lo accept
meta l4proto { icmp, ipv6-icmp } accept
ip6 saddr fe80::/10 udp dport 546 accept
tcp dport 22 accept
}
chain forward {
type filter hook forward priority filter; policy drop;
}
}
@@ -1,31 +1,3 @@
# Kernel hardening, mostly the KSPP recommendations (Talos sets these too), and one limit raised
# for containers. Left out on purpose: user.max_user_namespaces=0 breaks rootless podman, and
# kernel.modules_disabled=1 breaks loading the network modules podman needs.
# Hide kernel addresses (they help exploits bypass KASLR), even from root.
kernel.kptr_restrict = 2
# Only root may read the kernel log; it leaks addresses.
kernel.dmesg_restrict = 1
# No perf events for non-root users.
kernel.perf_event_paranoid = 3
# Harden the BPF JIT against JIT spraying.
net.core.bpf_jit_harden = 2
# No BPF for non-root users, and no switching it back on at runtime.
kernel.unprivileged_bpf_disabled = 1
# No kexec: no booting another kernel past Secure Boot.
kernel.kexec_load_disabled = 1
# No core dumps of setuid programs; they can hold secrets.
fs.suid_dumpable = 0
# No loading TTY line disciplines on demand; old ones had exploitable bugs.
dev.tty.ldisc_autoload = 0
# userfaultfd only for root; it makes kernel race exploits reliable.
vm.unprivileged_userfaultfd = 0
# No opening FIFOs and files owned by others in sticky directories like /tmp.
fs.protected_fifos = 2
fs.protected_regular = 2
# ptrace only on your own descendants.
kernel.yama.ptrace_scope = 1
# Containers (file watchers, systemd in containers) run out of Fedora's 128 inotify instances
# per user; 8192 is common on Kubernetes nodes.
fs.inotify.max_user_instances = 8192
@@ -2,8 +2,6 @@
# login on the screen (tty1, and tty2-6 through its autovt@ alias); this file sorts first, so it wins.
enable systemd-networkd.service
enable systemd-resolved.service
enable systemd-oomd.service
enable nftables.service
enable sshd.service
enable getty@.service
enable agent-setup@.service tty1 ttyS0
@@ -8,8 +8,6 @@ After=network-online.target user@1000.service agent-boot.service
[Service]
User=claude
WorkingDirectory=~
# In user.slice, so systemd-oomd kills it (or Claude's containers), not the system (user.slice.d).
Slice=user.slice
Environment=CLAUDE_CONFIG_DIR=/var/lib/agent/claude XDG_RUNTIME_DIR=/run/user/1000
ExecStart=/usr/bin/claude remote-control --name %H
Restart=always
@@ -1,6 +0,0 @@
# Load our firewall (/usr/lib/agent/firewall.nft) instead of /etc/sysconfig/nftables.conf.
[Service]
ExecStart=
ExecStart=/usr/sbin/nft -f /usr/lib/agent/firewall.nft
ExecReload=
ExecReload=/usr/sbin/nft -f /usr/lib/agent/firewall.nft
@@ -1,7 +0,0 @@
# systemd-oomd for Claude and its containers: when they run the VM out of memory, kill the cgroup
# responsible before the VM thrashes on zram swap or the kernel's OOM killer picks sshd.
[Slice]
# Memory pressure above 60% for 30s: kill the cgroup reclaiming the most memory.
ManagedOOMMemoryPressure=kill
# Swap over 90% full: kill the cgroup using the most swap.
ManagedOOMSwap=kill
@@ -1,4 +1,8 @@
# Installs the next UKI, which systemd-boot boots by default. Verify=yes: see 10-root.transfer.
# It gets three tries (boot counting, the +@l-@d in its name; no Mode=0444, systemd-boot can't
# rename a read-only file): systemd-bless-boot marks it good once a boot gets through
# (boot-complete.target); if none does, systemd-boot goes back to the previous UKI, and with it
# the previous root partition (each UKI holds its root hash).
[Transfer]
ProtectVersion=%A
Verify=yes
@@ -12,6 +16,7 @@ MatchPattern=agent_@v.efi
Type=regular-file
Path=/EFI/Linux
PathRelativeTo=boot
MatchPattern=agent_@v.efi
Mode=0444
MatchPattern=agent_@v+@l-@d.efi agent_@v+@l.efi agent_@v.efi
TriesLeft=3
TriesDone=0
InstancesMax=2