The hub no longer probes a node's Docker on tcp :2375, holds compose files
for a node, or runs compose against another machine's Docker. hubStacks,
nodeDocker, taskOnHub and the hub-side /api/nodes/{node}/docker/... routes
are gone; the generic node proxy hands those requests to the node's own
Ham Lab. Stack.remote / held_by and compose.Spec.Remote are removed.
New stacks default to <data dir>/stacks on every machine. Folders with two
or more deployed stacks are still offered, after it, never as default.
The ~/stacks fallback is gone. A missing home is skipped without a sudo find.
A schedule saved earlier with a tcp target is refused with a message
instead of being run against this machine's Docker.
PORTABILITY.md 3.1, 3.2, 3.3 and 9.2 done; 3.5 checked, no change needed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
spice-html5 opens a websocket per channel (main, display, inputs, cursor, playback), and
each one was making Ham Lab open a new console operation, which Incus refuses as already
connected. Now the first channel opens the operation and holds its control websocket; every
further channel is one more data websocket on the same operation, which Incus allows; the
operation is let go when the last channel closes. A stale operation from a previous Ham Lab
process is taken over with force.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- Library: an ISO shelf in virt/isos: upload (the Files uploader), fetch by URL with an
optional sha256, import a file from the machine by copying it in, rename, delete (refused
while attached), attach to a VM as a CD-ROM live or stopped with boot-first, eject, and
New VM from ISO: an empty VM with the ISO as first boot device, Secure Boot on or off.
- The VGA console: Ham Lab asks Incus for the VM's SPICE console over its unix socket and
bridges the display websocket to the browser, where spice-html5 (vendored from upstream
b372aab, LGPL, loaded on demand) draws it; Ctrl+Alt+Del, fit-to-window.
- Exports: every instance as a tarball in virt/exports/<name>/, kept to a count, as a cron
line in /etc/cron.d/hamlab-virt calling hamlab virt-export, settings and the last run
as files beside the data; Export now per instance or for all.
- Disks tab per instance: the root disk, attached ISOs with Eject, bind mounts in and out
(shift for containers, virtiofs for VMs).
- instance.* sensors (running, down, state, cpu, mem) and the VM-or-container-down alert
default, offered once to installs seeded before it existed.
- cloud-init: Use my SSH keys writes a cloud-config from the account's public keys.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Found on the first live test on grid: incus config set refused 'description' and the whole
limits save with it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
PORTABILITY.md sweeps the whole repo for assumptions tied to this lab's hosts,
network, hardware, distro, init system, orchestrator history and privilege
posture. Nothing is changed yet; this is the list to work from, graded A/B/C
with the general form of each.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The tab reports one of four states for the selected machine (ready, repair, installable,
unavailable) from what it finds — /dev/kvm, the package manager's incus candidate, the
pool drivers the kernel has — and acts on that machine only. Install is one followed task:
packages, a pool whose backing file sits in Ham Lab's own virt/ folder (zfs pool on a file,
btrfs on a loop mount, or a plain dir), a bridge on a subnet picked clear of the machine's
routes and Docker's pools (never 10.x on a /8 LAN), the DOCKER-USER rules that keep
Docker's FORWARD drop off the bridge, the incus-admin group for Ham Lab's account, and a
restart. Instances list with live CPU and memory, create from the images: catalogue or a
local image (container or VM, limits, cloud-init), start/stop/restart/force-stop/rename/
delete, console and shell as terminal tabs, snapshots with Incus's own schedule and expiry,
limits editing. Images: local store, catalogue search and download (exported into the
library too), import a tarball or a VM disk from a path, publish an instance.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
git describe had started picking the nearest release tag (v0.1.0-137-gc129f65,
fetched from Gitea), so a build came out as v0.1.0-137-gc129f65-36-g7d0b87b,
which the version compare reads as build 137: older than the 169 it replaces.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Mac's Asahi GPU exposes no busy counter, clock, sensor or memory reading, so
its box was a graph that could never draw beside a column of dashes. A GPU with
nothing to read now gets one line (driver, "nothing measurable" or the stall,
and whether a process has it open), its row shrinks to fit, and the Gpu% column
and gpu sort are dropped when no GPU on the machine can light them. QoL #3 from
bug bash 6 (BUGS.md).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A session unused for seven days is dropped, the Sessions list marks a day's idleness and
can sign all of those out at once. The hub keeps a KVM's own cookies from any browser that
has them and lends them to one that hasn't, on the wire and back to the page. The stack
editor and the new-stack dialog use CodeEditor: line numbers, YAML and .env colour under a
real textarea, Tab/Shift+Tab/Enter indentation, and a key bar on touch screens.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Found by a headless pass over every page on every machine at desktop and phone widths, then
every interaction on the hub. BUGS.md has the list, what was checked and found fine, and the
quality-of-life notes to discuss.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The router has no lspci, so its GeminiLake chip came up as 'Intel GPU
3185' beside the identical chip's 'Intel UHD Graphics 600' on hamtv.
A trimmed pci.ids — the graphics devices of Intel, AMD and NVIDIA, and the
virtual GPUs, 115 KB — rides in the binary and is consulted first; lspci
is the fallback for a card newer than the table. scripts/pci-ids-gpu.sh
regenerates it from the hub's copy.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
On grid the sampler sat 42 minutes in runtime_pollWait on the Arc A380's
hwmon fan1_input: kernfs accepts epoll, so os.ReadFile made the fd
non-blocking, i915 answered EAGAIN once, and the poller waited for a
readiness kernfs only signals on sysfs_notify. Every graph on the machine
froze while the API carried on.
readStr now reads with plain blocking syscalls. Each GPU's sysfs reads run
under a three-second deadline; a device that does not answer is marked
stalled, logged, and left alone until restart, with the box saying so.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The gpu box sat in an auto-height grid row. The canvas inside it was sized
from its box and also sized the box, so every resize fed the next and the
Processes tab grew for as long as it was open. The canvas is now absolute
within its box, and the gpu row has a floor instead of auto.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The desktop grid placement, the stats panel, the engine label width and
the client rows' columns were never in the stylesheet: on a phone the
caption sat on the stats head, 'enhance' ran into its meter, and a client
row came out as one word. Restored, with the graph box as tall as the
cpu's so its caption always has a line of its own.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A Cache section between the pools and the schedules. One card for the ARC:
size against its ceiling with the target and floor, sparklines of size and
hit rate from the sampler's history, what it holds (MRU/MFU/dirty, data vs
metadata, compression, the ghost lists), the prefetcher and dbuf cache, the
memory-pressure counters, and the since-boot breakdown by demand/prefetch
and data/metadata. Then a card per pool for L2ARC, SLOG, special vdev,
dedup table and spares — the devices and their fill where they exist, and
where they don't, plainly none and what that means (the ZIL writing to the
pool's own vdevs, with how much it has written).
Live numbers come from the sampler, which reads arcstats every tick and
publishes zfs.arc.* sensors; the fuller picture is read on request, with
zpool list -Hpv and zfs get run without sudo and cached for thirty seconds.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Every DRM device with a real GPU driver gets a box under the cpu box, laid out
like it: the busy curve, a meter for the device and one per engine, temperature,
clock, power and fan where the card has them, and the processes holding it open
with their share. The process table gains a Gpu% column and sort.
Device busy is Intel's RC6 idle residency inverted (what intel_gpu_top shows,
and readable without root) or amdgpu's gpu_busy_percent. Per-process shares are
DRM fdinfo deltas, as nvtop reads them; other users' fdinfo needs root, so a
daemon running as a user keeps one root child (hamlab drm-clients) over a pipe
rather than paying a sudo a second. Everything is also a gpu.* sensor.
What cannot be measured is left out, not shown as zero: an Arc A380 on i915 6.12
has no temperature sensor, an integrated GPU wears the package temperature
marked *, and Asahi has neither a busy counter nor fdinfo yet.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
On a tunnel this machine configures — today that means OpenWrt's UCI, which is
where the server here lives — the VPN tab can add a device, rename it and take
it away. A wg-quick host is reported as unmanaged and shows no buttons, because
half-supporting a second configuration format is worse than an honest refusal.
Adding generates the keypair and a pre-shared key on the machine, picks the
lowest free address in the tunnel's range, and hands back the complete client
profile with a QR code for the WireGuard apps. The private key in it is
generated for that one call and stored nowhere, so the profile is shown until
dismissed and then gone — the same bargain as any password this console makes.
Nothing restarts the network, which matters when the interface belongs to the
router the house is behind. `uci commit` writes the file and applies nothing;
`wg set` changes one peer on the live interface and touches nothing else.
The order of the two is chosen so they cannot disagree quietly. Adding writes
the configuration first and rolls it back if the live tunnel refuses, because a
device that works only after the next reboot is the worst of the three outcomes.
Removing does the live tunnel first, since that is the moment access actually
ends, and says plainly if the configuration half then fails.
Removing asks for the device's name to be typed back, at both ends. It is the
one action here that cannot be undone and it sits one tap from a list of devices
that all look alike.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
On a machine with wireguard-tools installed but no tunnel up — core, which is
the machine the console opens on — the tunnel list was a nil slice, which is
null in JSON, and the page took its length. That throws, React unmounts the
whole tree, and what is left is a blank window with nothing to go on.
Two fixes, because either alone leaves the failure possible. The parse returns
an empty list rather than nil, with a test that pins the marshalled shape; and
the page treats a missing list as an empty one, so a node still running an older
build cannot blank the console either.
And the reason a small mistake had such a large blast radius: nothing caught it.
Each page now renders inside a boundary, so a fault costs that page and says
what it was, while the rail, the switcher and every other tab carry on. The
terminals are deliberately left outside it — they hold live WebSockets, and
remounting them to recover from somebody else's error would drop the sessions.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two faults in the VPN tab, both found by running it against the real machines.
omitempty does nothing for a zero time.Time, so a peer that has never handshaken
was serialised as the year one and the page would have shown mbp14 as last seen
on 1 January 0001 instead of "never connected". The field is a pointer now, and
a test pins the JSON.
And peer names on a wg-quick host were never going to appear: /etc/wireguard is
root-only because it holds private keys, and the file was being opened as the
daemon's own user. It now falls back to reading it as root, the same way the
kernel state beside it is already read. Only comments and public keys are taken
from it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Beside Network, per machine. Every WireGuard interface on the machine in the
switcher, with its address, listen port and public key, and under it the peers:
name, allowed addresses, last handshake, where they connect from, and what they
have moved.
Per machine is the point rather than a limitation. The tunnel devices dial into
and a tunnel a machine dials out through are different boxes here, and putting
them on one page would invite the question of which box you were looking at.
The kernel is the truth and the configuration only supplies names: wg show all
dump has the handshakes and the counters and knows nothing but public keys, so
the # comment above a peer in a wg-quick file, or a UCI section's description on
OpenWrt, is read alongside purely to label them. Whether a tunnel is dialled or
dialling is worked out from the shape of its peers, and says neither when it
cannot tell.
Online means a handshake in the last three minutes, and the tab says so in as
many words: WireGuard is silent when idle, so there is no connection to be in.
Read-only, and no secrets leave: the private key is the second column of the
dump and is read past, pre-shared keys are noted as present and never read.
The Network tab keeps listing tunnels as the interfaces they are, and hands off
to here for anything about them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Where Mitchell had already dragged it. An account with its own rail order is
unaffected either way; this is the slot a fresh one gets.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Cloudflare rate-limits by volume, and the block it handed out while this was
being built lasted over twenty minutes — too fragile to be the only thing behind
a button. So the download phase tries its sources in order: the anycast endpoint
first, then a plain 100 MB file on Linode's Seattle mirror, which has no opinion
about how much has been pulled today.
Only a refusal moves to the next one; anything else would fail the same way
twice. Every result names the source it used, and the page says "from
linode-seattle" rather than "via SEA" when it was the fallback — that mirror is
in one fixed place, so somewhere else it reads low, and a figure that quietly
came from a different continent would be a lie.
Upload stays on Cloudflare, which has never been the one to refuse.
Also: the too-short-window guard was refusing to divide 512 MiB by ten
milliseconds on a link fast enough to hit the runaway budget before the first
tick. It now guards against dividing by zero and nothing more.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Per machine like every other tab: the interfaces as the kernel has them
(addresses, MAC, MTU, negotiated link speed, carrier, totals, live rates from
the sampler's own sensors), the default route and resolvers, what the internet
sees of this connection, and a speed test with the last fifty results kept.
The test runs in the daemon, not the browser, so the figure is the line of the
machine in the switcher — the router's, if you switch to gate — and it carries
on when the phone locks. It measures against Cloudflare's public speed
endpoints, so there is nothing to install on any host, which is what makes it
work the same on Debian and on OpenWrt where flash is scarce.
What the measurement does and does not count:
- Several streams, HTTP/2 deliberately off. One stream over one connection
measures a congestion window, and HTTP/2 would put every stream back onto
one connection and measure the same thing.
- The first second of each phase is thrown away as the ramp.
- A phase ends on 64 MiB as well as on four seconds, so a fast line says what
it is without moving a gigabyte. A minimum window keeps a very fast line
from filling the cap before there is anything to divide by.
- Latency is the round trip less the far end's own Server-Timing. A zero-byte
response is mostly Cloudflare's worker thinking: left in, it reported 31 ms
for a line the edge itself measures at 3.5. Jitter is the mean change
between consecutive probes, so one slow probe cannot run away with it.
- The far end rate-limits by volume. So: a 20 s cooldown counted from the last
attempt rather than the last success, streams that retry twice, the rate
computed over the window in which bytes were actually still arriving, and a
plain sentence when it was refused. Coasting to the end of the clock with
nothing arriving reported a third of the truth.
The last result is published as net.internet.* sensors, each stamped with when
the test ran, so the line can be a widget, a graph, an alert or a Prometheus
series like anything else. The public address is refreshed on its own quarter
hour, which is worth having here: it is a DHCP lease and it moves.
Tests are offline — the window, the byte cap, the refusal and the cancellation
run against a local server; the route parser's byte order is pinned by a case
that would have caught the bug it did. The live measurement is behind
HAMLAB_NET_TEST=1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A guest password is typed by hand off a message on a phone. Five tries is not
many when a hyphen lands in the wrong place, and being told to wait is worse
than the brute-force risk here: past twenty the doubling lockout still applies,
the global backstop is unchanged, and the password is ~78 bits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A guest password is read off a message and typed by hand, and from the login
form a mistyped character and a wrong password look identical. Revealed, the
field is a text field, so it also turns off capitalisation and autocorrect,
which a phone keyboard would otherwise offer for it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mitchell, on the read-only version built earlier today: scratch it. He hands
these to people he trusts and revokes them otherwise, so a console with half
its tabs missing solved a problem he doesn't have.
So the allow-list, the middleware check and every "not for a guest" branch in
the UI are gone. What is left is the part that was the point: a throwaway
account that expires by itself, is checked on every request, swept once a
minute, and can be ended in one tap. A guest now gets the whole console on
every machine, terminal included.
Said once, since it is his call and he has made it: that means a guest has a
root-capable shell on every host, and can make an account or a token of its own
that outlives the guest login.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Users -> Guest logins makes a throwaway user name and password, good for half
an hour by default, to hand to somebody who wants to see the console. It is a
normal account with `guest: true` and an expiry, so login, sessions and the
saved look need no special case, and it is left out of the account list, out of
"the last account can't be removed", and out of the one account an empty user
name means.
What a guest may do is one allow-list, checked in the auth middleware, rather
than a flag ~190 handlers have to remember. The list is matched by asking the
router which handler a request would reach, so it cannot drift from the routes
it names, and it is checked against the scoped path — /api/nodes/grid/procs is
/api/procs on grid — because the hub proxies to a node with a token that is
full control of that node. The node cannot tell a guest from its owner, so the
only place it can be stopped is here, before the request leaves.
A guest gets the dashboard, containers, stacks, disks, hardware, ZFS and usage,
on this machine and on any managed one, and nothing else: no terminal (attaching
is a GET, and it is a root shell), no files, no container logs or environments,
no stack files, no alert channels, no cron, no tokens, no settings, no power.
Expiry is enforced on every request and swept once a minute, so a console left
open on a visitor's phone stops within the minute rather than at their next
click; revoking signs them out at once. The console hides what a guest cannot
use and says how long is left where the power button would be.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The binary was already static and musl-clean; the only thing standing between
Ham Lab and an OpenWrt box was that setup.go knew systemd and OpenRC and
nothing else, so an install there ended at "nothing starts it at boot."
Adds an embedded procd init script, detected on /sbin/procd + /etc/rc.common
ahead of the OpenRC branch. Both backends write /etc/init.d/hamlab, so
uninstall tells them apart by USE_PROCD=1 rather than by path.
Verified on the new router: procd supervises the instance, S95/K10 symlinks
land, and the hub drives its Docker and stacks like any other node. The
Updates tab needed nothing — apk-tools 3 answers the same query Alpine's does.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Recreate is `up -d --force-recreate`. It exists because Deploy and Update
both leave a container alone when its definition has not changed, which is
precisely the state a container is in when Docker will not start it — pinned
to a network that has since been recreated, answering "network … not found"
to everything. siyuan landed there today and there was no way out from the
console.
The Updates tab is a page, not a list with a detail pane, so on a wide window
it was being squeezed into the right-hand half with an empty list column
taking the left — which read as a phone layout on a desktop. It now sits
under the tab strip at full width. The list-and-detail tabs are unchanged.
The rule needed :not(.has-sel) to outrank the wide-window rule that grows the
list to fill the row.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Updates tab could tell you eleven images were behind and offer nothing to
do about it; updating meant knowing which stack each image belonged to and
going to find it. Now each row has Update and the header has Update all.
Updating is per stack, because that is the unit compose deals in, so a row
acts on whichever stacks use that image — postgres:15-alpine is wikijs's,
postgres:17 is vikunja's — and Update all walks the affected stacks through a
queue one at a time, with each task's output on screen, rather than
recreating half the machine at once. A stack that finishes says "updated just
now" rather than re-checking, since every check is a registry request and
Docker Hub counts them.
Also: cron expressions read as English. "daily at 3:15am", with 15 3 * * *
kept beside it. Jobs.tsx had a smaller version of this for its own list; it
now uses the shared one, which handles lists, ranges, steps, names,
weekdays/weekends, midnight and noon, and says nothing at all when it is not
certain — the caller falls back to the expression.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The board, the chassis and the firmware on it, which the System card was
reducing to two lines. Maker, model, version, serial and asset tag; chassis
and where the board sits in it; the firmware's vendor, version, revision,
date, ROM size and whether the machine boots UEFI — that last one is a line
in SMBIOS's characteristics list rather than a field, so the parser now keeps
those lists too.
Expansion slots come with it, with the device found at each occupied slot's
PCI address rather than the address alone. They are only shown when the
firmware tracks occupancy at all: core's NucBox lists five PCIe x1 slots it
does not physically have, all "Available", copied out of a reference design.
grid's board reports real addresses and real occupancy, which is the case
worth printing.
A machine with no SMBIOS gets a card that says so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Mitchell's placement. The rail order is the account's anyway — this is only
where a page starts for someone who has never dragged it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Per machine, like every other page. Four tiles — processor, memory, storage,
graphics — over a card for each.
Memory is the part that had nowhere to live before: every slot, with the
module's size, type, form factor, rated and running speed, maker, part
number, serial, rank and voltage, and the empty slots marked empty. A module
whose maker is only a JEDEC id gets a name when it is one we know and keeps
the number when it is not, because a wrong name is worse than a code.
Everything is read from what Debian already has: dmidecode for the parts only
the firmware knows, lscpu, lspci, sysfs. None of it changes under a running
kernel, so it is read once and kept for ten minutes; Re-read is for the one
thing that does change, a drive plugged in.
Every reader is allowed to come back empty. A machine with no dmidecode, no
root to read it with, or no PCI bus at all still gets a page, with a note
saying what was missing and why. Checked against the Mac mini, which is all
three cases at once: no SMBIOS memory table, a GPU only the device tree knows
about, and cores that are not all alike — lscpu counts one cluster as "cores
per socket" there, so without SMT the thread count is the honest answer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two gaps the Docker tab had against Dockhand and against its own Services
tab next door.
Logs: one `docker logs -f` per viewer over a WebSocket, the shape the
Services tab's journal tail already uses, instead of asking again every two
seconds. stdout and stderr both, with stderr marked so the page can colour
what went wrong. Docker replays the backlog down two pipes in no particular
order, so the opening 300 ms is held and sorted by the stamp Docker wrote;
after that the log is live and in order. A reconnect resumes from the last
line rather than from the top — `--since` is inclusive and two lines can
share a nanosecond, so the client drops what it has already shown by line,
not by "later than". A container that has stopped ends the stream at once,
so reconnects back off to 15 s.
Networks: remove and prune, which images and volumes have had all along.
The overview now says which containers are attached to each network, so the
✕ appears only on one nothing uses, and never on bridge, host or none.
`network rm` prints the real reason and then an aggregate "exit status 1",
so the message shown is the first line, not the last.
Prune is not exercised against `core`: several networks here belong to
stacks that are merely down, and pruning would give them new subnets on
their next deploy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A view shows one machine: the one in the switcher. Seeing another machine's
state means switching to it, not a hub-side page that reaches across — and
this was one. Recorded as a decision so it does not get rebuilt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Inspect's values are long paths, network names and env lists. In a two-column
grid on one line each they were half truncated and read as columns of
unrelated things. One per row, label above value, wrapping instead of an
ellipsis. Mounts are one block rather than a repeated "Mount" label.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every stack on core and grid is labelled with a path inside Dockhand's
filesystem — /app/data/stacks/… — because Dockhand ran compose from inside its
own container. Ham Lab resolves those by walking the bind mounts of running
containers, which works right up until the container doing the mounting is
removed. Retiring Dockhand would, in one step, make seventeen stacks on core
and five on grid unfindable.
So the mounts are written down as they are seen, and a path no live container
explains is looked for in what was seen before. A remembered mount resolves
with no container to run in, which is the right answer: compose runs on the
host, against the host's own copy of the file.
The same problem bites a scheduled update, which captures how to invoke
compose when it is made so the cron job never has to rediscover the world at
3am. repoint() checks the named container is still running and falls back the
same way if it is not — and pointedly leaves the spec alone if nothing
explains the path, so it fails loudly rather than deploying a guess.
Out in cron the record is read and never added to: the daemon owns state.json
and would clobber anything written from there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
