The goal is speech recognition on the NPU for vinput, and sherpa-onnx cannot
deliver it: every model it loads keeps a dynamic time axis, which the NPU
compiler rejects outright. Qwen3-ASR and SenseVoice both core-dumped in
vpux-compiler ("Got non broadcastable dimensions pair"), and reshaping the ONNX
inputs does not help because the dynamic dimensions sit inside the graph
(WA-037). Whisper is the only family OpenVINO publishes fully static IRs for,
and its fixed 30 s window is exactly what the compiler accepts.
The host now runs an OpenAI-compatible /v1/audio/transcriptions service backed
by whisper-large-v3-turbo int8 on the NPU: 1.5 s for 11 s of audio against
2.1 s on CPU, with the compiled graph cached in /var/cache so only a cold cache
pays the ~60 s NPU compile. vinput on the desktop sends raw PCM to it through a
small provider script.
Only the NPU driver and compiler still come from the unmerged branch in the
`nixpkgs-npu` input (NixOS/nixpkgs#570230), whose nixpkgs also happens to be
the one where openvino-genai 2026.4.0.0 builds at all.
Assisted-by: DeepSeek Harness 0.2.1-rc.2 (deepseek-v4.1-flash)
CPU warns at 90 C, GPU at 95 C and other components at 75 C; per-core probes
are excluded so a full-core compile no longer raises 26 alerts.
Assisted-by: DeepSeek Harness (deepseek-flash)
The Explore link named a datasource that does not exist, and rules without an
alertgroup label rendered an empty value.
Assisted-by: DeepSeek Harness (deepseek-flash)
changes() produced one series per pg_settings_* metric with identical labels,
so VictoriaMetrics rejected the rule with "duplicate output timeseries" whenever
a setting changed.
Assisted-by: DeepSeek Harness (deepseek-flash)
CPUTIN and AUXTIN* read 101 C or saturate at -128/127 C with no sensor wired,
so the alert fired while the CPU package stayed near 53 C.
Assisted-by: DeepSeek Harness (deepseek-flash)
Aquamarine builds for ai and shoukei across the three platforms, with both
daemons logging in as root. They authenticate with their passphrase-free SSH
host keys, and ssh_known_hosts now carries each builder's address.
Assisted-by: DeepSeek Harness 0.2.0-rc.2 (deepseek-flash)
On a tmpfs root the flake/fetch cache is empty at every boot, so
`nix run <flake-ref>` re-downloads its input and fails until DNS is up: a reboot
test showed llama-swap's first start failing on that. Other caches write into
RAM meanwhile.
Assisted-by: DeepSeek Harness (deepseek-flash)
disko creates the swapfile only when it formats the pool, so the live 96G files
keep their size until they are recreated with btrfs filesystem mkswapfile; that
frees 64G per host.
Assisted-by: DeepSeek Harness 0.2.0-rc.2 (deepseek-flash)
* fix(ssh): retry and multiplex LAN connections
* fix(ai): shrink the neighbour-probe fallback to ~1s
Assisted-by: DeepSeek Harness 0.2.0-rc.2 (deepseek-flash)
The tuned package ships tuned-ppd.service with WantedBy=graphical.target, so
activation starts it without /etc/tuned/ppd.conf and switch-to-configuration exits 4.
Assisted-by: DeepSeek Harness (deepseek-flash)
systemd ignores `path` inside [Service], so the unit ran with the default PATH
instead of the coreutils one it asks for.
Assisted-by: DeepSeek Harness (deepseek-flash)
Enabling the module runs `openrgb --server` at boot and installs its udev rules,
so the GUI and the CLI reach the MSI controller without root.
Assisted-by: DeepSeek Harness (deepseek-flash)
The module loads i2c-dev and grants /dev/i2c-* to the i2c group and locally
logged-in users, so ddcutil no longer needs root.
Assisted-by: DeepSeek Harness (deepseek-flash)
The gateway LAN IPv6 transparent path stalls on TCP flows while IPv4 forwarding remains healthy. Withholding AAAA answers keeps LAN clients on the working fake-IPv4 TProxy path until the IPv6 path is fixed
Assisted-by: DeepSeek Harness (gpt-6.1-sol)
A deploy from an aarch64 machine built the x86_64 system locally and died on the first platform-specific derivation; colmena 0.5.0 can build on the target instead.
Assisted-by: DeepSeek Harness (deepseek-flash)
The second NVMe and its LUKS volume leave with the Z890 board, so the data-disk disko module is deleted; `ai` must activate this change before the disk is pulled, or its initrd blocks on the missing volume.
Assisted-by: DeepSeek Harness (deepseek-flash)
The Z890 board and its root SSD move here from `ai`, and the retired VM's name and 192.168.5.101 are reused.
Assisted-by: DeepSeek Harness (deepseek-flash)
--redact produced a "shareable" template but replaced only the secret and the proxy-subscription urls: rule-provider tokens, private domains, the LAN controller address and the routing rules all stayed. Nothing in the repo consumed it.
Assisted-by: DeepSeek Harness (Qwen3.8-Flash-Next-IQ3_S-ista-text)
Adds the review-comments-and-docs skill (1282-word body) and two evaluation suites: a 21-case
measured suite for the skill, and a harness that runs the two concurrency scenarios previously left
unrun in agents/evals/global-rules.md.
The worker MicroVMs ran `home-manager-ryan.service`, which failed on two of them
(`/nix/var/nix/temproots` permission denied, and no `/home/ryan`), so
`HostSystemdServiceCrashed` fired for days. Mirror the masters and extend their
regression test to all six nodes.
Assisted-by: DeepSeek Harness 0.2.0-rc.2 (deepseek-flash)
* flake.lock: Update
nixpkgs only. The new revision builds VictoriaMetrics with cgo
(CGO_ENABLED=1); its pure-Go zstd decoder rejected a stored values block, so
every small-part merge panicked on youko.
* fix(colmena): address the VM hosts by name, not by their IP
`~/.ssh/config` pins the passphrase-protected builder key with IdentitiesOnly
for 192.168.*, so IP-addressed ssh cannot authenticate without a terminal. The
name resolves through the aliases in /etc/ssh/ssh_config.
* fix(monitoring): repair two alert rules VictoriaMetrics rejects
`on(uid)` matches on a label kube-state-metrics no longer emits, and
`!= ON(__name__, instance)` treats every pg_settings series as a duplicate.
They now join on (namespace, pod) and compare with `changes(...[5m])`.
* fix(monitoring): keep disk latency alerts off mechanical disks
Both rules averaged per-op latency with no device or sample guard, so the 4 TB
USB HDD volumes paged constantly (>100 ms/write in 26% of the last 30 days).
They now require a non-rotational physical device and >=5 ops/s.
Assisted-by: DeepSeek Harness 0.2.0-rc.2 (deepseek-flash)
ryan4yin | 二花
Ryan Yin