nvidia-smi on a busy card can take past the 3 s deadline, which latched the card
as stalled until restart. A late read now carries the last readings and is used
when it lands; no second read starts while one is out; only one out for 15 s
(past both nvidia-smi timeouts) marks the card stalled, and its return clears it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The kernel numbers the whole machine's threads in one run and deals them out a
socket at a time, twice: nexus's first CPU listed C0–C9 and C20–C29, its second
C10–C19 and C30–C39. Each CPU now lists its own twenty as C0–C19; the kernel's
number is in the row's tooltip.
Co-Authored-By: Claude <noreply@anthropic.com>
Its sensors were gpu.<name>.usage_2 and so on; the page looks under gpu.<name>_2. Each
card now gets its own stem, as the page already assumed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The sensor feed dropped which socket a coretemp reading came from, so every core matched
the first CPU's chip. coretemp sensors now carry a package label, and history folding
keeps the two sockets' same-named cores apart.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Per-package RAPL watts in the sample, the process table and as cpu.package.N.power
sensors. The socket rows collapse with an arrow; the fold is remembered per browser.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hardware counts CPUs, cores and threads per CPU. Top groups threads under a row
per socket, reads each core's temperature from its own socket's coretemp chip,
and the CPU temperature is the hotter package.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A verify pass reads the pool's copy of every file the mover moved back against the SHA-256 the
mover took of the landing copy (its ledger, named in its status file): match, mismatch, edited
since, gone, waiting, unreadable, unhashed. 'New moves' checks what moved since the last pass,
'everything' the whole ledger. Passes run as 'hamlab landing-verify' in their own session — the
same command a schedule's line in /etc/cron.d/hamlab-landing runs — so a console restart does
not cut one short. Stale files: anything waiting on the landing branch over an hour, listed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A txg records its bytes only when it syncs and the mover reports every 2 s, so a per-sample delta
read 0 most of the time. Rates are now taken over the gap between syncs (or mover updates) and held
until the next one.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Live from arcstats (dirty), dmu_tx (throttled writes), each pool's txgs (bytes synced to disk)
and objset kstats (bytes from programs). A landing tier — a mergerfs pool whose first branch is
fast storage and whose last is a ZFS dataset — is detected from mergerfs's branch xattr; a mover
status file in /run/landing-mover/ adds its queue, the file in flight and totals. Sensors
zfs.write.*, zfs.<pool>.write_bps|logical_bps and zfs.<pool>.landing.* carry the history.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every theme's full-window overlays are off while a video is open, the picture
is never filtered or blended, the player's lettering is plain. --blur, as in To
Do and Notes: the header (the top row of the app), dialogs, the full-screen
player's note and an expanded KVM start below iOS 26's scroll-edge blur in a
Home Screen app; zero on desktop and in a Safari tab.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The setting's device makes every video preview: on the file's own machine
when it is there (the hub tells it by header), else on the device's machine,
which reads the file through the hub by a one-file token, and the preview is
kept in the file's machine's own cache (PUT /api/files/thumb, refused if the
file changed). Images and covers stay on the file's machine's CPU. If the
device's machine can't, the file's makes it as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Same fix as Ham Share's: a trailing change event read the bar's old position
back as a drag in progress.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tonemap_vaapi refuses HLG, and the hardware path never honoured a rotation
(ffmpeg's autorotate is for frames in memory). HDR10 is still tone-mapped,
on the card and, failing that, in software before the upload.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every video plays YouTube's way: hls.js reads the connection and picks the
quality, or one is picked by hand, or Original plays the file as it is. The
qualities (a ladder by size, H.264) are made on the fly in 3 s pieces from
wherever the viewer seeks, one ffmpeg per play, paused when far enough ahead.
360 videos open in a look-around view (drag, pinch or scroll to zoom from
8 to 300 degrees, phone motion).
Settings -> Transcode device picks the machine and device that makes them,
for every machine's files: a machine transcoding another's file reads it
through the hub by a token for that one file. Replaces the codec/size picker
and the single-quality re-encoder.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Graphs stepped once a history point (1 s at a 5 min window) while the numbers changed every
tick; they now slide a fraction of a column each tick. The process table (and the GPU boxes,
MHz and watts it carries) was polled on its own timer; it is now read per sample and committed
with it. GPU clients are walked every sample, and the 1–2 s floors on the gallery, Network,
Docker and ZFS ARC live views are gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The proprietary driver's DRM card says DRIVER=nvidia, which the driver list did not
know, so the card was never enumerated; and it exposes no sysfs busy counter or
fdinfo engine totals. Read it with nvidia-smi instead (--query-gpu per sample,
pmon -c 1 -s um for clients), by PCI address. Parsers unit-tested, including an
older pmon column layout and [N/A] fields.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A re-encode's playlist is now the whole film on a fixed 4 s grid. A segment
that isn't on disk and isn't coming soon moves the encode there; runs stop
short of segments already made, and gaps are filled after. Every run is cut
the same way (-ss, -output_ts_offset start+10, frag_discont), so segments
from different runs join. Repackaging keeps its growing playlist.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The registry's answer was cached and served as-is, so the stack view and the
machine-wide list kept calling an image behind after an update had pulled it.
The cache now keeps the digest the registry named, and an image marked behind
is re-checked locally on every read: once this machine holds that digest it
is current, at no registry cost. The page re-reads the machine-wide answer
after a stack task, a queued update, or a schedule run.
Also: stack keys can no longer collide with the machine key, and the dead
"/"+stack forget in the schedule runner is gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ham Lab is meant to be lightweight; VMs, passthrough and their consoles
are too much. Removes the Virtualization tab, the incus package, the
/api/virt routes, the instance.* sensors and their alert default, the
virt-export command, the VGA console and the vendored spice-html5, and
VIRTUALIZATION.md. A saved instance-down rule is dropped on load.
Everything removed is kept, as it was, on the virtualization branch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The sweep's battery hit the route three times; the second and third runs only collided
with the first's restart timer.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
From the race detector, staticcheck, a garbage-in battery over every safe route and a
headless pass through every tab on two machines. Fixed:
- docker inspect/health of a container that isn't there answered 500; now 404, and a
daemon that is down 503 (dockerErr, with a test).
- /api/virt/instances/{name} on a machine without Incus answered 502 "incus not found";
it now says Incus isn't ready, like its siblings.
- Deleting a ZFS snapshot with a bad name answered 500; a bad name is 400, a snapshot
that doesn't exist 404.
- The task list read a task's Running flag without the task's lock while the waiter
goroutine cleared it: a data race, now read under the lock.
- The Virtualization export config and run record were written in place; a crash
mid-write would have left a half file. Atomic now, like the config and state files.
- An unused function in kvm.go, one in incus/subnet.go, an unused field in files/move.go
and a capitalised error string, from staticcheck.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Found by walking every handler and the UI with odd inputs, then a battery of requests
against the live hub. Fixed:
- A save over the 8 MiB cap was cut short in place: the file was truncated, then the
first 8 MiB written, and the caller told 204. Now the body is read whole first and a
body over the cap is a 413 with the file untouched.
- A link to a file showed the link's own time; now the file's, so the row reads right
and the preview key follows the file.
- Reading a link to nothing answered 200 with an empty body (the open failed after the
headers in root mode); now a 404 that names the target.
- A play request with an unknown codec or size was answered by inspecting the file
first, so an audio-only file said "no video stream" instead of what was wrong; the
choice is checked before any work, and is a 400.
- A stream that ended in failure was handed back to the next identical request for ten
minutes; a failed one is now dropped and a fresh one made.
- A player that left while its stream was still finding a first segment made the loop
try the next recipe; it stops. A stream closed mid-start is not started again.
- Ending a root-mode ffmpeg could wait forever: the daemon cannot signal root's
processes, so the wait is bounded and the helper's own grace does the rest.
- A page going away (a phone closing the tab) releases its stream by a beacon, rather
than ten minutes later by the sweeper.
- The path box resolves like a shell — relative to the folder shown, . and .. applied,
trailing slashes dropped — in the tab and in the Move picker, with a test. A typed path
that is a file opens it; a remembered folder that is gone falls back to / once.
- A dropped folder arrived as a nameless empty file; it is refused with a word. The
upload toast counts what was actually uploaded.
- The viewer names a device, pipe or socket, and a dangling link, instead of calling
them binary files.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The viewer tries the file as it is first. When the browser gives up — Safari and an MKV,
HEVC without a hardware decoder — or on a tap, a picker: codec (H.264, HEVC, AV1, or the
original codec repackaged) and size (original, 1080p, 720p, 480p), each codec labelled
with the engine this machine would use. POST /api/files/play runs ffmpeg into an HLS
event playlist of fragmented-MP4 segments under <data dir>/transcode/<id>/; Safari plays
HLS itself and everything else goes through hls.js, loaded only then. The recipes are
tried best first: a card on the bus before the integrated GPU, decoding on the same card
before decoding on the CPU and uploading, the software encoders last; a recipe that
produces no first segment is dropped for the next. The seekable range grows as ffmpeg
gets ahead; status says how far. Streams are reused for the same request, closed when
the player goes, swept after ten idle minutes, and the folder is emptied at start. Root
mode runs ffmpeg as root through a helper that ends when the daemon lets go of it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The package was named for its first job; it now also decides what a file is for the
Files tab (image, video, audio) and reads its streams, which the viewer will need. Inspect
is the richer probe — codec, profile, pixel format, size, frame rate, HDR transfer, the
audio beside it, container and duration — run as root through sudo when asked.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ServeContent seeks to the end for the size before it seeks to the range, which was
throwing away the stream Check had just opened.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The viewer hands a video or audio file to the browser untouched — container, codec and
all — and the browser plays what it can; when it can't, the viewer says so and offers the
download. What the server adds is the part a player needs: /api/files/read now goes
through http.ServeContent for a real file of known size, so Range requests (seeking) and
HEAD work, and a download can resume. The root backend's helper takes an offset for that,
and a Seeker opens the backend once, at the range asked for, instead of streaming from
the top. Audio files are a preview kind too: their cover art is the thumbnail.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The kernel answers EXDEV for two paths on different filesystems before it looks for the
source, so the cross-filesystem path was taken for a missing file.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
