* chore(workarounds): re-evaluate stale hardware/kernel fixes
Prune workarounds whose root cause is gone, fix one vendor typo, and
document the ones we keep:
- drop `transparent_hugepage=madvise` on shoryu/shushou/youko: it was tied to
the old fixed 1G hugetlb pool, and the VMs now use ordinary memory
- shoryu: `hardware.cpu.intel.updateMicrocode` -> `amd` (it is an AMD APU)
- ai: drop `i915.enable_guc=3` (GuC/HuC submission is the default on modern
i915/Xe)
- kernel-hardening: name the Dirty Frag CVEs (CVE-2026-43284 / CVE-2026-43500)
and the re-evaluation condition
- document re-eval conditions for the Insta360 resume fix, the darwin
auto-optimise-store workaround, and the youko USB-SATA bridge mitigations
* chore(workarounds): keep the Dirty Frag module blacklist as defence in depth
* chore(ai): correct the i915.enable_guc rationale (verified against the kernel)
* chore(darwin): drop the auto-optimise-store workaround (nix#7273 fixed)
* fix(ai): relax NVMe power-management workarounds and enable smartd
Logs show no NVMe errors, so allow APST again (capped at 100 ms), restore
the default io_timeout, and drop pcie_aspm=off. Add smartd for drive
health visibility.
* feat(storage): add drive/SSD SMART monitoring on physical hosts
node_exporter's nvme collector only exposes node_nvme_info and namespace
capacity, not SMART health, so SSD wear/lifespan and drive temperature were
not observable anywhere.
- enable services.prometheus.exporters.smartctl on physical hosts, gated on
modules.hardwareTools (off on QEMU/MicroVM guests), bound to the host's LAN
address on port 9633
- firewall: drop 9633 from anything but the monitoring host
- youko: scrape smartctl-<host> for the physical hosts
- alert rules (storage.yml): SMART health, NVMe wear, available spare, media
errors, critical warning, and drive temperature (warn 70 / crit 80 C)
- Grafana dashboard "Drive SMART (smartctl)"
* feat(idols-ai): persist /var/tmp and small state dirs, wipe /var/tmp at boot
Move the strata model staging area (~3G) off the root tmpfs onto the
persistent btrfs volume to free RAM, preserve btrbk/cups/logrotate
state, and wipe /var/tmp at boot so it only accumulates between
reboots.
* chore(idols-ai): add nosuid,nodev to root tmpfs and tmp btrfs mounts
Hardens /, /tmp and /snapshots; safe because /dev, /run, /nix and
/run/setuid-programs are separate mounts, so nothing there needs suid
bits or device nodes.
* chore(idols-ai): add noatime to the @tmp btrfs mount
Temp files do not need access times; matches the @nix subvolume style
and saves btrfs metadata writes.
* chore(idols-ai): harden the /var/tmp bind mount with nosuid,nodev
World-writable staging data should not honor suid bits or device
nodes, same as /tmp and /snapshots.
* fix(idols-ai): preserve logrotate state as a file, fix persisted dir ownership
- /var/lib/logrotate.status is a regular file, not a directory: move it from
preservation directories to files. A missing state entry makes logrotate
treat every log as freshly rotated, so it must survive reboots.
- keep /var/lib/btrbk owned by btrbk:btrbk 0750 so preservation does not
override the btrbk module's rule with root:root.
- set /var/tmp mode 1777 explicitly instead of relying on systemd's tmp.conf
ordering.
- rewrite comments to describe the paths generically.
* chore(idols-ai): cap journal, wipe coredumps at boot, persist upower/noctalia
- services.journald.settings.Journal: SystemMaxUse=10G, RuntimeMaxUse=256M,
matching the cap already set on idols-kana/idols-ruby.
- wipe /var/lib/systemd/coredump at boot: it lives in the persisted
/var/lib/systemd and crash dumps can contain secrets copied from memory.
- persist /var/lib/upower (device battery history) and
~/.local/state/noctalia (shell settings, notification/clipboard history).
Running strata with qwen3.8-flash-next keeps ~53G of model weights
resident, leaving little RAM for the desktop side and pushing ~13G of
anonymous pages into the 20G swapfile. With swap that full, a burst
risks OOM-killing processes. Disk space is not a constraint (1.9T
btrfs, ~1.4T free), so grow the swapfile to the classic 1x-RAM size.
Keep zram off and zswap at 10%: zram's compressed pool would steal page
cache that the mmap'd weights rely on.
Manage ~/.config/go-musicfox/config.toml declaratively: the upstream default plus the mpv engine, ~/Music download dir, 2 GiB cache and cover art.
The app rewrites the file with a temp file plus rename, which would replace a plain symlink, so use force = true like fcitx5's profile: the repo copy is restored on every rebuild and runtime edits are ephemeral. Runtime data and state stay on the persistent volume.
* docs(skills): point nix-config-secrets at the private repo under ~/codes
Record where the private nix-secrets checkout lives and that agents should
do the mechanical secret work themselves, handing back only the
interactive/sudo agenix operations and pushes.
* docs(agents): add related repositories to nix-config
* chore(agents): adopt ~/src for source checkouts
Move third-party/source checkouts from ~/codes/src to ~/src so the home
layout reads as three siblings: ~/codes (personal), ~/work (work),
~/src (checkouts). Preserve and back up the new path, and update the
global rule and the two nixpkgs skill references.
* docs(agents): tighten the code layout rule
* docs(agents): fold the work-confidentiality note into the layout line
* docs(agents): base the confidentiality rule on repo visibility
* docs(agents): drop the stale KubeVirt mention
* fix(youko): add ROCm OpenCL so Jellyfin can tone-map HDR
Jellyfin's HDR/DoVi transcodes use ffmpeg's `tonemap_opencl`, but only
Mesa was installed via `hardware.graphics.enable`, so no OpenCL ICD was
registered. ffmpeg then failed to create an OpenCL device ("Failed to
get number of OpenCL platforms: -1001") and aborted with exit code 237,
leaving HDR titles unplayable.
Add `rocmPackages.clr.icd` to `hardware.graphics.extraPackages`; it
ships the ICD for the Barceló iGPU (reported as `gfx90c`) under
/run/opengl-driver/etc/OpenCL/vendors, where `ocl-icd` (built with
`--enable-custom-vendordir`) finds it.
The Barceló iGPU only manages ~0.8x realtime for a 4K HDR -> 1080p
tone-mapped transcode, so playback can still buffer; this only makes
such titles playable at all.
* fix(youko): allow /dev/kfd so Jellyfin OpenCL can reach the GPU
The jellyfin module's `DeviceAllow` whitelist only permits the VA-API
render node (`/dev/dri/renderD128`). ROCm's OpenCL runtime opens
`/dev/kfd`, so even with the ICD installed OpenCL device creation failed
("No matching devices found") and HDR transcodes still aborted with exit
code 237.
Add `/dev/kfd rw` to the whitelist; list definitions merge with the
module's entry. Verified under the unit's sandbox (PrivateUsers,
PrivateTmp) that OpenCL init then succeeds as the jellyfin user.
davfs2 caches the whole opened file regardless of cache_size, so opening media on /mnt/fileshare downloads it fully and blocks the mount. The share is no longer used.
* fix(youko): stop homepage from probing the authenticated transmission RPC
The siteMonitor hit /transmission/rpc unauthenticated, so every probe returned 401 and counted against transmission's anti-brute-force threshold.
* feat(youko): tune transmission and firewall its network namespace
Raise the upload limit and queues, relax required encryption to preferred, and add a default-drop nftables input chain inside the netns so the peer port stays open while everything else is closed. Allowlist the nftables 'iif' keyword for typos.
* feat(youko): export transmission metrics and add a Grafana dashboard
nixpkgs has no transmission exporter, so add a small stdlib Python exporter that queries the authenticated netns RPC on loopback, scrape it from VictoriaMetrics, and ship a Transmission dashboard.
* fix(youko): cap transmission upload at 2500 kB/s
30 MB/s ignored the 50 Mbps uplink; keep seeding around 20 Mbps.
* Revert "feat(youko): export transmission metrics and add a Grafana dashboard"
This reverts commit 9df7a26526.
Hosts used to trust the router as the only boundary (servers ran with the
firewall off; service ports were opened globally), so any LAN device could
reach every host service. Move filtering to the hosts as defence in depth
behind the router, which keeps its firewall on and uses a port mapping for
the few ports we actually need.
- base: nftables backend; deny inbound by default; trust the home LAN by
source and the tailnet by interface (restrict the tailnet with ACLs);
suppress per-service `openFirewall` global ports; allow node_exporter only
from the monitoring host. The local podman bridge, the tailnet and the pod
CIDRs are accepted only after the node_exporter drop.
- servers: stop disabling the firewall; drop redundant global port opens
(caddy 80/443, NFS 2049, Jellyfin, localsend/test ports); delete the
redundant desktop firewall module; aarch64/riscv64 import the firewall
explicitly (akane was missing it).
- k3s: Cilium pods (10.0.0.0/8, fd00::/8) may reach host services; drop the
flannel-era cni0/flannel.1 trust.
- youko: postgres listens on localhost + the LAN with pg_hba limited to the
LAN; the NFS export stays LAN-scoped; remove the unauthenticated v2ray
proxy.
- tests: `security-firewall` asserts every host runs the shared firewall, has
the LAN CIDR rule and no global ports.
* feat(youko): add Jellyfin media server
Jellyfin reads the existing Transmission download tree on the public HDD
(via the shared `fileshare` group) and transcodes with the AMD Barceló
iGPU through VA-API. Encoding settings are declared in Nix
(`forceEncodingConfig`), so the web UI cannot silently drift.
Replace the per-service public ACME certificates with a single
`*.writefor.fun` wildcard (plus the apex) shared by every public vhost,
so adding a service needs no new ACME entry; a wildcard also keeps
individual hostnames out of public CT logs.
Pin the jellyfin uid/gid to avoid the dynamic-id drift fixed for the
other services in #320.
* fix(youko): enable graphics for Jellyfin VA-API
VA-API needs Mesa's radeonsi VA driver under /run/opengl-driver, but
youko is headless so hardware.graphics was off. Without it Jellyfin's
ffmpeg fails with 'va_openDriver() returns -1'. Enable it so the iGPU
can be used for hardware transcoding.
* fix(youko): use VA-API hardware encoding for Jellyfin
Enabling hardware decoding alone left encoding on the CPU (libx264), which
pins ~3 cores per 720p stream. Set transcoding.enableHardwareEncoding so
the iGPU (Vega VCN) does the H.264 encode too.
* feat(youko): harden Jellyfin proxy trust, groups, and HEVC encode
Add jellyfin to the render/video groups for robust VA-API access (was
relying on the world-writable render node), enable HEVC hardware encoding
(Vega's VCN supports it; AV1 does not), and patch network.xml on start to
trust Caddy as KnownProxies and advertise the public URL.
* fix(youko): run Jellyfin network.xml patch every start; allow LAN discovery
The module's preStart exits early once encoding.xml matches, so the
mkAfter network.xml patch only ran when encoding settings changed. Run it
with mkBefore instead. Also allow Jellyfin app auto-discovery (UDP 7359 /
TCP 8096) from the home network only.
Service users/groups get a dynamically allocated uid/gid, which shifts when the user set changes: adding `immich` moved gitea 993:985 -> 996:993 and rustfs 987:977 -> 988:983. Files on the persistent volumes keep the old numeric owner, so gitea and rustfs lost access to their data and failed to start after a restart. Pin every dynamically allocated service id on youko to its current value. The ids are host-specific, so this is a host file.
Add Immich via the native nixpkgs services.immich module on youko, reusing the host PostgreSQL 16 (pgvector + vectorchord) and a shared Valkey instance for Redis-compatible caching and BullMQ queues (Dragonfly rejected BullMQ's dynamic Lua keys). Media lives on the encrypted HDD at /data/apps/immich; Caddy publishes immich.writefor.fun with a publicly-trusted Let's Encrypt cert (security.acme + Cloudflare DNS-01) so the mobile app connects. OTEL metrics are scraped by VictoriaMetrics, and /var/cache is now preserved on the tmpfs-root hosts so the ML model cache survives reboots.
The s3-console Caddy vhost redirected every request to / to /rustfs/console/. The RustFS console frontend signs its own S3/STS requests to / (POST / for STS, GET /?x-id=ListBuckets), so the 302 broke the console session and login failed server-side with SignatureDoesNotMatch ("missing signed header: content-type"). Restrict the redirect to unsigned browser GET /; signed requests now reach RustFS unchanged.
Deploy RustFS in place of the retired MinIO and move the S3 consumers over.
- flake.nix / flake.lock / outputs/default.nix: add a nixpkgs-small
(nixos-unstable-small) input for the rustfs 1.0.0 package (with the web
console), instantiated as pkgs-small in genSpecialArgs alongside the other
nixpkgs instances; drop the rustfs flake input. The rest of the system stays
on the default nixpkgs.
- hosts/12kingdoms-youko/homelab-services/rustfs.nix: use nixpkgs' own
services.rustfs module, with package = pkgs-small.rustfs and
environmentFile = the agenix rustfs.env secret.
- caddy.nix: s3.writefor.fun (S3 API) and s3-console.writefor.fun (console,
redirects / to /rustfs/console/).
- secrets/nixos.nix: declare a single rustfs.env (dotenv) secret holding
RUSTFS_ACCESS_KEY / RUSTFS_SECRET_KEY, replacing the separate key files.
- infra/minio -> infra/rustfs: buckets are managed with the hashicorp/aws
provider against the RustFS S3 endpoint; IAM/ILM are applied with rc.
Crush hardcodes banned bash commands, blocking many operations. Drop the package, its AGENTS.md/CRUSH.md install target, env vars, and preservation entries.
Keep the six node configuration intact while removing the failing user activation from the three control-plane MicroVMs. Update the regression test and current K3s configuration notes.
kubelet refuses to self-assign the reserved node-role.kubernetes.io/* labels, so the worker --node-label made the agent crash-loop. Drop it and rely on the masters' control-plane taint for placement; document that the taint only lands via registerWithTaints on first registration, so it was applied by hand here.
Add worker-1 (4 vCPU / 16 GiB) on shoryu, and worker-2 (4 vCPU / 16 GiB) plus worker-3 (2 vCPU / 8 GiB) on shushou, as microVMs. They come with a genK3sAgentModule and the per-host/output/home-manager wiring, and are labelled node-role.kubernetes.io/worker=true. The masters keep their control-plane:NoSchedule taint, so workloads land on the workers.
Reserve 192.168.5.192/27 as a dedicated kube-vip block, outside the router's DHCP pool (.2-.99): one /29 per cluster, k3s-test-1 gets .192/29 with the API VIP at .193.
Advertise only youko itself (.183/32) and that block over Tailscale, instead of the whole /24. The /24 route shadowed hosts that are already reachable directly, e.g. idols-ai at 192.168.5.100.
Deploy prerequisites (not in this repo): approve the new Tailscale routes; point kube-vip at the new VIP; move the *.writefor.fun wildcard from .101 to .183 and update scripts/darwin_set_proxy.py, since .101 is no longer advertised.
Give the idols-ruby and idols-kana VMs a headless computer-use
environment: Xvfb plus i3, driven by cua-driver (and computer-use-linux
as a fallback). Includes the system module (AT-SPI, fonts, Mesa, Clash
Verge proxy, proxy-region timezone), the home i3 module (Xvfb/i3/x11vnc
services, browser flags), host wiring and an eval test.
A failed backup was silent. Post to alertmanager (which routes above-info
to telegram), and have caddy answer 404 instead of its empty 200 for
*.writefor.fun names that are not services.
A desktop holds both the plaintext and the repository password, so blocking
deletion on the server protects little while preventing the automatic
retention the client needs. Immutability belongs in the offsite copy.
The design (layers, password groups, retention, restore steps) was only in
code comments. Also make the timer configurable: a whole hour of random
delay defeated any staggering, and the desktop's push, youko's own run and
btrbk all otherwise contend for the same disks.