The validation tail from the 2026-09-03 run. Each is small; together they are
the difference between a field that checks its input and one that looks like it
does.
Subnets accepted any string containing a slash, so 'notanip/xx' was stored -
and _devices_in_subnet then swallowed the ValueError from ip_network and
returned an empty device list, so the subnet looked real and permanently empty.
It is parsed with the same library that has to read it later, a prefix length is
required (ipaddress reads a bare 10.1.1.0 as a /32, which is legal and never
what somebody typing a subnet meant), and the stored form is normalised so
10.1.1.5/24 and 10.1.1.0/24 cannot become two rows nobody can tell apart.
VLAN numbers had no range check: 99999 and -5 were stored. Worse, a non-numeric
value reached a duplicate probe that compares against an integer column, which
MySQL coerces to 0 - so 'xyz' matched an existing VLAN 0 and the 409 named a row
that had nothing to do with it. 1-4094 per 802.1Q, checked before the probe.
Deleting a VLAN that subnets still reference now answers 409 like every sibling
in-use guard, rather than 400, which told the caller their request was malformed
when the request was fine and the state was the problem.
An appversion sha256 was stored raw, so an over-long value was truncated to the
column width and came back as 64 plausible hex characters - indistinguishable
from a real digest, and wrong. A digest is a claim the delivery path checks
against, so it is 64 hex characters or it is refused. The releasedate goes
through the strict parser for the same reason it did everywhere else.
PUT /api/settings/<key> ignored the declared valuetype. "A6-not-a-number" went
into an integer key verbatim and every consumer then read 0, because
get_typed_value swallows the ValueError - the setting looked set and the
behaviour was the default.
/api/maplevels told a caller who omitted the level name that 'buildingname and
buildingid are required'; buildingname is a field that handler never reads. It
now names the fields it actually wants.
Restoring a guide revision that is already current writes nothing, because
save() dedups on the content hash - and the response still said "Restored
version N as a new draft version", describing a revision that does not exist
while the unchanged revision list disagreed.
GET /api/backups/revisions/<id>/diff?against=<id> accepted any revision id, so
pointing it at another asset's revision rendered that asset's configuration as
the "before" side of this asset's history: a diff that reads as a change on this
machine and describes a different one. Same asset and same kind, or refused.
The default branch was always asset-scoped; only the explicit one was not.
Two frontend fixes: the per-page selector rendered blank on any list defaulting
to a size not in [10,20,50,100] (network devices uses 25), so the operator could
not see the current page size; and an emptied PC alias was dropped from the
payload instead of sent, so it could not be cleared - the form came back showing
the alias that had just been deleted.
Twenty-two tests. The backups guard verified by planting.
Scope: product
Three more findings from the 2026-09-03 run.
Werkzeug answered 405, 415 and a malformed JSON body itself, in HTML. A client
written against the documented {status,data,meta} contract got a page of markup
with no error code, which is indistinguishable from the server having fallen
over. The 405 had a second edge: on a route the caller cannot otherwise see, a
method-not-allowed CONFIRMS the path exists, so a gated kiosk route meant to
read as absent announced itself by refusing the wrong verb. All three now
answer in the envelope, and only under /api/ - everything else is the SPA and
keeps the framework's behaviour, deep links included.
The Toner Report rendered "0 Printers Checked / 0 Low / 0 Critical" and "No
printers match the selected filter" while the API it had just called said
supply monitoring was unreachable. Zeros read as a clean bill of health for a
fleet nobody has looked at. The page now carries the reason the API gave and
suppresses the counts, because a number nobody measured is not a result. Same
defect shape as the supplies card fixed on 2026-08-29; `unavailable` was
already in the payload and the page ignored it.
Backup revisions badged "current" per (asset, kind). NTLARS settings describe a
MACHINE but are read off the controlling PC, so a machine with two reporting
PCs has two independent chains of the same kind - and only the higher revision
id was badged, so the second PC's newest revision showed as history and lost
its last-verified stamp. The machine looked like one source re-reporting rather
than two. sourcehostname is the chain; a revision without one forms its own,
which keeps the old behaviour for everything that has only ever had one source.
Nine tests. The envelope ones verified by planting: remove the three handlers
and three fail.
Scope: product
The NTLARS/DNC registry carries an FTP password in cleartext and it was stored
verbatim. That put it in the backuprevisions table, on the revision panel which
renders values, and in the download route which hands back a ready-to-import
.reg - so anyone holding backups.view could read it off a web page.
Redaction happens at PARSE, so the secret never reaches the database at all.
Redacting on display would have left it at rest and one query away.
What is deliberately NOT done is dropping the value silently. This codec already
refuses to discard an unparseable value, precisely because a backup that looks
complete and restores a machine with a setting missing is the worst outcome. A
redacted value keeps its name, is listed in the projection, and renders as a
COMMENT naming what must be set by hand. Emitting a placeholder as the password
would produce a machine that looks restored and cannot connect, and the
technician would go debugging the network.
scripts/redact_backup_secrets.py scrubs revisions recorded before this, dry run
by default. It leaves contenthash alone: the hash identifies the configuration a
bay reported, so rewriting it would make every bay's next post look like a
change and insert a duplicate revision across the fleet.
Separately, sourcehostname is now normalized at ingest. samesource() maps '' to
IS NULL for the query while the insert stored the raw '', so rows written by a
client posting an empty hostname could never match their own next post: every
post inserted, the table grew without bound, and the history filled with
identical entries nobody could read. Whitespace did the same thing more quietly.
The plugin shipped five settings and no page for any of them. The core settings
catalog is a hardcoded list a plugin cannot extend, so backups_shareroot and the
collection interval were reachable only through the API or a flask shell on the
server - which is not a configurable setting in any useful sense, and it is how
today went: the setting simply was not there to set.
The page is contributed the ADR-010 way rather than by editing core - a card
from get_settings_cards plus a route in the plugin's own routes.js, both of
which disappear when the plugin is disabled. Filed under Machines, next to
Machine Types, because that is where the assets these back up live.
It also states the two things the settings do not say for themselves: that
health is judged on when a backup was last CONFIRMED rather than when it last
changed, and that retention prunes database revisions only - files on the share
stay, so enabling it on a share-backed kind leaves orphaned snapshot
directories.
Snapshot paths are now shown as full UNC text in both places: a Snapshot row on
the Marker Contents card, and the existing path on the history page, which
gains a copy button. Deliberately NOT a link - browsers refuse to navigate from
an https page to file://, so a clickable path looks actionable and silently
does nothing, which is worse than no link. Copy, then paste into Explorer.
The path comes off the revision, never composed from the configured share root:
a snapshot written before the layout changed still lives where it was put.
A 'share' kind keeps its bytes on the file share and gives ShopDB only a hash
and a UNC path. That is right for an opaque vendor file, but it means the app
cannot answer the question a tech actually arrives with - what does this part
marker hold, and which pattern disappeared last month. Reading the share from
the app server is not an option: the download routes refuse to mount it on
purpose, because a permissions slip there becomes an unexplained empty panel.
So the collector now accepts an optional manifest alongside the hash: the
index, not the content. Names, sizes, per-file hashes, and the marking system
each pattern was authored for. Advisory throughout - a client that sends none
behaves exactly as before, and dedup still keys on contenthash.
The marker asset gains a Marker Contents card built from it: counts by type,
the controller family, and a diff against the previous snapshot in the same
source chain. Removed files are named rather than counted, because "1 removed"
only tells someone to go and look. Files edited in place are named too - the
file set can be identical while the content hash moved, and reporting no
change there would contradict the reason the revision exists.
The manifest is deliberately NOT covered by contenthash, which attests the
bytes on the share. A client that starts describing itself therefore posts an
unchanged hash, so a no-op backfills a manifest the row does not yet have -
otherwise the card would appear only on markers that had recently changed, and
stay hidden on every healthy one. The same no-op refreshes sharepath, so a
snapshot that moved on the share stops being pointed at where it used to be.
Manifests are validated at ingest rather than at render: a malformed one would
otherwise become the latest revision and fail every view of the asset page
until the content next changed, outliving the client bug that caused it.
sharedir() for the marker kind is overridden to the layout the site actually
uses, so what ShopDB displays and validates matches what the collector writes.
get_setting on BasePlugin namespaces what it reads to plugin.backups.<key>,
while get_settings_defaults declares - and the settings page writes - the bare
key. So the retention read never found the operator's value and always fell back
to 0, and 0 means keep everything. Retention was configurable in the UI and did
nothing. It is the only place in the codebase using the namespaced helper.
The share root also stops shipping one site's file server as its default. That
put a site's internal topology in a bundled plugin and in the public mirror, and
pointed a second site at a server it cannot reach. Blank now, per ADR-015, and a
share kind with no configured root returns nothing rather than composing a path
from somebody else's hostname - a path built on an empty root is not a lesser
answer, it is a wrong one.
lastseenat already recorded it and the API already returned it; nothing
displayed it, so from the UI a healthy machine still looked abandoned - one
revision from last spring and no sign anything had looked at it since.
The history page gains a Last verified column beside Captured, and the asset
panel a Verified field. Only the CURRENT revision carries one: an older
revision was superseded, so saying it was verified today would be false - what
was verified is the configuration the PC holds now.
A current revision with no check yet says "not yet checked" rather than showing
a blank or borrowing the captured date. That state is real and temporary: the
column is new, so every chain reports it until its PC next posts.
The stale-backup card could not be built as designed, and the reason is more
important than the card. Dedup means an unchanged configuration writes no
revision, so collectedat moves only on a CHANGE. A machine stable for six
months has a six-month-old newest revision and is perfectly healthy. Keying a
staleness card on revision age would have flagged most of the fleet - exactly
the noise that makes a board worth ignoring.
Underneath that: ShopDB could not distinguish those cases at all. On a no-op
the server returned "unchanged" and wrote nothing, so "we checked yesterday and
it matched" was discarded. That fact is the one thing a backup system must be
able to prove, and the only record of it was a line in a log file on the PC.
lastseenat records the check rather than the change. Touched on every matching
post including the no-op; set on creation, since a new revision has by
definition just been seen; backfilled from collectedat or createdat so existing
rows start from the last moment the config can be PROVEN current, rather than
from now - claiming a check that never happened would be worse than silence.
The card keys on it, one row per CHAIN rather than per asset: a machine with
two part markers can have one still reporting while the other stopped, and a
per-asset view would report the machine as fine. It stays deliberately silent
about assets never backed up, because whether one SHOULD be is a question only
the manifest can answer, and guessing would list a hundred healthy machines.
The rule lives in services/staleness.py rather than the route, so it is
testable without an auth layer in the way - the same split retention.py uses.
Threshold is backups_staledays, default 3, and 0 disables the card.
Backup-NtlarsSettings reads backups_intervalhours from /api/settings/public,
because it runs before it holds any credential, and its Get-IntervalHours falls
back to 24 on any failure. The plugin never declared the key public, so the
endpoint did not return it, the fallback fired on every PC, and the setting
looked configurable in the UI while changing nothing. The fleet log shows the
symptom plainly: "Throttled: last attempt under 24h ago", every cycle,
regardless of what the setting said.
Declared public. A collection cadence is not a secret. backups_shareroot stays
private - it is internal topology - and the test asserts both directions so a
later edit cannot quietly widen it.
Same defect as the 3D parts kiosk label prefix already in this changelog: a
logged-out reader against an allowlist its key was not on. Worth noticing that
the pattern has now bitten twice.
Dedup compared a posted config against the latest revision for the ASSET,
which is only correct when a machine number means one PC. Several PCs share one
here: the part markers on 0613, 0615 and WJPRT are separate devices, differing
by COM port, filed under one machine number. Each marker's post therefore
differed from whichever marker had posted last, nothing ever deduped, and the
table grew by one row per PC per collection cycle.
A chain is now (asset, kind, source hostname). An unchanged config is a no-op
again, and each PC keeps its own history against the machine. NULL sources -
rows written before the column was populated, and hand-loaded ones - form their
own chain via IS NULL; `column == None` never matches in SQL, so without that
those rows would have re-posted forever.
Two consumers assumed the old key and are fixed with it. Retention pruned per
asset, so a busy marker's revisions could evict a quiet marker's only backup;
it now prunes each chain separately, protecting the newest and oldest of each.
The revision diff compared against the previous revision on the machine, which
across two markers reported one device's COM port as a change on the other; it
now compares within the source's own chain.
scripts/collapse_duplicate_backup_revisions.py cleans up what the old rule
wrote. It removes only a revision whose hash repeats the one before it in the
same chain - rows the fixed code would never have written - and keeps every
genuine change, every chain's newest and oldest, and every source. Dry run by
default. Its --report mode explains what grew each chain, which separates a
legitimately shared machine number from two PCs wrongly carrying the same one,
and from a value inside the config that changes on its own.
A line of the form "Name=" with nothing after the sign is not strictly legal,
but it occurs in real exports - the part marker's WJPRT.reg has KRelay1 like
this. The parser raised on it, which failed the whole file, which meant that
machine could never be backed up at all. Read it as an empty string so the
value name is still preserved.
Backup timestamps read wrong because of two faults stacked, which is why it
looked like a single offset.
The API serialised naive ISO ("2026-08-07T12:00:00"), with nothing saying the
value was UTC. JavaScript's new Date() parses that as BROWSER-LOCAL, so every
timestamp shifted by the viewer's offset before any timezone formatting ran.
Every datetime this plugin stores is naive UTC, so the wire format now carries
a trailing Z.
The history view then formatted with toLocaleString(), i.e. the viewer's zone,
ignoring the site_timezone setting entirely. It now loads that setting and
formats through the shared formatInZone helper, matching NotificationsList.
The panel list label is built server-side with strftime, so a client cannot
correct it afterwards. It now converts to the site zone using the same Setting
lookup the notifications plugin uses - without that it showed UTC, four hours
out at West Jefferson.
Tests cover the wire format and that 16:30Z renders as 12:30 in
America/New_York.
DNC Info becomes a single tabbed card - General, eFocas, Serial, NTSHR and
MARK - instead of a flat wall of every value. On 3204 that is 11 rows visible
rather than 22, and on 0600 eleven rather than 27, which also stops the card
unbalancing the detail page's two-column layout.
Everything DNC now lives on that one card, so the Part Marker panel is gone:
its settings are the MARK tab. The partmarker KIND is untouched and still
stores, dedupes and serves revisions - they are listed on the backup history
page - it simply contributes no card of its own, which on 145 of 147 machines
would have been an empty box.
Downloads are named for the machine: 3204.reg, and 3204-wow6432node.reg for
the dialect that imports outside NTLARS. The view had been rebuilding the name
from sourcefilename and producing 3204.reg-wow6432node.reg, so the revision now
carries assetnumber and both ends agree. That needed a viewonly relationship to
Asset - no backref, so the core asset side gains no dependency on this plugin.
The history page was hardcoded to light colours (#e0e0e0, #f4f9ff, #666) and
rendered as a white table on a dark page. It now uses the palette variables
throughout, per frontend/CLAUDE.md. The current-revision tint is a color-mix
against --primary so it reads in both themes rather than a baked light blue
that disappears on dark, and the diff columns are headed as well as red/green,
since colour alone does not survive a colourblind reader.
Every machine was getting a Part Marker Configuration panel, and only two of
the 147 known machines are part markers. A kind applies to an asset TYPE, but
whether a given asset ever carries that kind of backup is a property of the
individual machine, so type alone cannot decide what to show.
The generic renderer already handled this: a list panel is visible when it has
rows OR declares empty text. Declaring emptytext on both kinds defeated it and
forced them to render everywhere. emptytext now defaults to None on the base
class, neither bundled kind sets one, and the panel builder OMITS the key
rather than emitting null - a present-but-null 'empty' would still have kept
the panel on screen.
The DNC Info card carried empty text that could never be displayed, since
keyvalue visibility is decided purely on field count. Removed rather than left
to mislead.
A lathe now shows DNC Info and NTLARS history; 0600 and 0614 additionally show
Part Marker once something collects for them; a machine with no NTLARS data
shows no backup panels at all. The emptytext mechanism stays available for a
kind that genuinely wants to say "expected here, nothing yet".
Adds a kind-pluggable backups plugin. Configuration captured from a PC is
filed against the MACHINE it controls, with a revision history and download
back to the native format.
NTLARS/DNC is the first kind. Settings live in the controlling PC's registry
but describe the machine, so revisions attach to the machine's asset and carry
no foreign key to the PC: history survives a PC being replaced or deleted, and
sourcehostname records the handover.
Storage splits by kind. Parseable kinds store a dialect-neutral JSON
projection in ShopDB and re-render on download; opaque vendor formats (part
marker and similar) keep their bytes on the SFLD share with ShopDB holding
metadata and the UNC pointer.
Two .reg dialects exist in the wild: NTLARS's own Save... export omits the
WOW6432Node path segment, scripted exports include it. Parsing strips whichever
root matched, so a stored revision commits to neither and download offers both
(NTLARS Load... by default, WOW6432Node for direct reg import). Getting this
backwards is silent, so the dedup hash deliberately excludes sourcedialect and
both dialects of one config dedup to a single revision.
Dedup is load-bearing: the collector runs every GE-Enforce cycle across the
fleet, so a revision is inserted only when the content hash differs from that
asset's latest for that kind.
A freshly imaged PC opens NTLARS with a blank General tab. Recording that would
make an empty config the newest revision exactly when someone needs the last
good one, so a blank MachineNo is rejected rather than accepted as a change.
Two of the 320 known-good backups on the share already have that shape.
DNC Info card summarises the latest revision on the machine page: General
(Cnc, NcIF, HostType), eFocas, Serial, NTSHR when populated (only 18 of 147
machines), and MARK when the machine is a marker. MARK is gated on Cnc=MARKER
or the ShopDB machine type, not on the MARK key having content: MARK carries
serial defaults on 145 of 147 machines and DncPatterns reads YES on 103
including ordinary lathes, so neither identifies a marker.
The info card is owned by the kind (BackupKind.infopanel/buildinfo) and served
by a generic endpoint, so the expected successor to DNC ships its own card by
adding a class rather than changing the plugin or the panel wiring.
Also: schedule and retention settings with a prune that never drops the newest
or the oldest revision, and scripts/import_ntlars_backups.py to seed history
from the existing per-machine .reg files (144 of 147 resolve to assets).
Codec verified against all 320 real backups: round-trips clean through both
dialects. Bay-side generation verified on Windows against reg.exe export.
