cfdd24d shipped the "?;" tolerance with no changelog entry, against the
project's own rule. Placed before the CW entry: the link had to survive being
busy before anything keyed through it could matter.
CW over CAT never worked on a Kenwood. The KY implementation was written against
Elecraft's, which accepts a string of any length, so "KY OH5CX;" went out and a
TS-590SG answered "?;". OpsLog read that as "this radio refuses CW over CAT" and
told the operator to fit a serial keyer — advice that was wrong, and expensive.
The TS-590 manual is explicit: P2 has a fixed length of 24, blanks are filled
with spaces, and those spaces are not keyed. So the fix costs nothing on air; it
is simply the shape the command has. Elecraft stays variable-length, where
padding would key the trailing spaces as word gaps.
The semicolon is also gone from the allowed CW characters. It TERMINATES a CAT
frame, the manual forbids it in P2, and one in a macro would have closed the
command early and left the rest of the message to be read as commands.
Found from a log and a manual page, not from a rig: nobody here owns a Kenwood.
What is proven is the frame shape; that a TS-590SG then keys it still needs the
operator to confirm.
A Kenwood answers "?;" while it is busy, and the TS-590SG does it for a moment
after RX; — which is exactly when the poll resumes. One rejected IF; then tore
down the whole CAT link.
The code already knew this. Its own comment says a "?;" to IF; is "a transient
busy, NOT unsupported" and that latching it off "would read as lost the rig".
That reasoning was applied to the unsupported-command map and stopped there: the
call still returned an error, the poll loop still turned it into connected=false,
and the link still dropped.
What it cost was not cosmetic. WSJT-X keys through shared CAT, so it lost the
rig mid-sequence; Hamlib then sent "F 9223372036854775808.000000" - an
uninitialised 2^63 - which OpsLog refused and logged. The operator sees a
frightening frequency error whose actual cause is three lines earlier.
A "?;" is now sentinel-wrapped so the poll can tell "the rig declined" from "the
serial link is gone", and three consecutive ones are ridden out on the last
known state before the link is called dead. Under a second of tolerance: enough
for the rig to finish, far too short to hide an unplugged cable. A serial fault
is untouched and still drops immediately.
Tested against the emulated rig already in this package, with a hook that
answers "?;" on demand: the link survives inside the grace and still reports the
last good frequency, the count resets on recovery so a later busy spell gets the
full allowance, and a rig refusing forever is still reported as broken.
N0CALL is what WSJT-X transmits under when its owner never set a callsign. It
has a letter, a digit and an ordinary shape, so nothing rejected it: it was
spotted, coloured, counted as a new WPX prefix, and now that CQ grids are read
it would have put a grid into the worked index under a callsign nobody holds.
The obvious fix - adding it to looksLikeCall's reject list - was wrong, and the
test caught it before it shipped. That function answers "could this token be a
callsign at all", and the CQ grammar uses it to decide whether the word after CQ
is a modifier (DX, NA, a zone) or the call itself. Teaching it that N0CALL is
not a callsign made "CQ N0CALL JN36" skip a slot and return JN36. Shape and
policy are different questions and now live in different functions.
Chasing that turned up the real defect behind it: ANY unrecognised word after CQ
made the parser skip a slot, and a four-character grid passes every shape test a
callsign does. "CQ FOO JN36" returned JN36 as the sender - logged, spotted and
coloured as a station. A grid in the callsign slot is now refused outright.
An operator reporting "nothing happens for two or three seconds after I click"
was impossible to answer from opslog.log, because the first line applog writes
already sits inside startup(). Everything spent loading the binary and creating
the WebView2 environment happened before the log begins.
In one of the reported logs, startup() to window-visible is 251 ms end to end -
so whatever the wait is, it is not OpsLog initialising. It just could not be
proven, only asserted.
processStart is stamped on the first instruction of main. Two lines now bracket
the launch: how long was spent before startup() ran at all, and click-to-window,
which is the only number that matches what the operator experiences.
Finishes the half of this that was already computing on the backend and reaching
nobody. A Grid column in the Geo group, NEW GRID as a badge in Status, a filled
cell in the marker's own colour, and a filter chip.
new_grid joins lib/spotMarkers rather than getting colours of its own, so the
badge, the cell fill and the chip cannot drift apart - and the per-marker colour
setting will drive it with the rest from one table. Magenta: the last hue in the
categorical set not already spoken for, and one that does not read as a status,
because a new square is never urgent the way a new entity is.
The band map leaves it to the cluster, as it already leaves the prefix. The pill
is 22 px tall and its accent strip stops being readable past three segments.
A row carrying a new grid is no longer "dull", or the dimming would grey out the
one thing worth looking at.
The column is off by default, like the other Geo columns: it is only ever filled
for stations this receiver decoded over the UDP link, so for an operator who
does not run digital it would be a permanently empty column.
"No colour on worked" blanked the spot status as well as the worked-call flag,
on the theory that a spot bringing no novelty should stop painting entirely.
But the status is exactly what the cluster list reads to DIM a row. Blanking it
made isDull() fall through its "unresolved, never dim" guard, so every quiet grey
row came back at full brightness. The option meant to calm the list was the one
making it shout. It also forced a `muted` flag to exist purely so tooltips could
say that an empty status did not, this time, mean "entity not resolved".
It now removes the blue already-worked mark and nothing else, which is all it
ever needed to do: a worked entity already renders with no colour of its own and
is already dimmed. The flag, its two tooltip strings and the special-cased band
map styling all go with it.
The amplifier PUSHES status frames ("S0|state=…") on the same socket it answers
commands on, and it pushes them constantly once in OPERATE - power, SWR and
temperature all move while transmitting.
command() read exactly one line and took whatever arrived first as its answer.
One pushed frame therefore put the stream permanently one reply behind: every
later command read the PREVIOUS command's answer, and eventually one waited out
the 3 s deadline, failed, and dropped the connection. Hence a fresh TCP
connection and a fresh authentication every few seconds in the log.
The stall in transmit is the same defect seen from the other side: command()
holds the connection mutex across that whole dead wait, so SetOperate and every
other control queued behind up to three seconds of nothing.
It only appeared remotely and in OPERATE because that is when there is anything
to push - on a LAN with an idle amplifier the race hardly ever opens. Which is
why it did not do this yesterday.
readReplyLocked now reads until the R<id>| that belongs to the command, feeding
every frame it passes to parse() on the way - a pushed status is fresher than
the one we were about to ask for, so nothing is wasted.
authLocked used the same one-line read and worked around this by re-sending the
whole handshake, which is visible in the log as "auth reply=S0|state=IDLE
(try 1)" followed by a second attempt. It goes through the same reader now.
IOTA is an official ADIF field, and almost everything for it was already here:
the qso table has carried my_iota since the first migration, the insert and scan
handle it, and both ADIF import and export write it. The one missing link was
the station profile, so the reference had to be typed on every contact of an
island activation, or added afterwards in a bulk edit.
Migration 0025 adds the column; it joins my_sota_ref and my_pota_ref in the
profile, in the Station Information panel, and in the same stamp-if-empty block
that fills the other My* fields on a logged QSO.
Uppercased and trimmed on save. ADIF spells it EU-005, and an operator typing
"eu-005" would otherwise put a reference no award matcher recognises on every
QSO of an activation - the kind of mistake you only find months later.
AddQSO emitted qso:logged but never invalidated the cluster status snapshot.
The frontend did its part - it re-queried every visible spot two seconds later -
and ClusterSpotStatuses answered out of a snapshot built before the contact, so
it returned exactly the same "new band" as before. For ever, until an import, an
edit or a profile switch happened to invalidate it for another reason.
Reported twice: an E51 and then a ZD7 that stayed yellow on the band map with
the QSO plainly in the log. The first report was answered by fixing the
visibility gate, which was a real bug of its own and hid this one.
Confirmed against the operator's MySQL logbook rather than guessed: ZD7BG has
dxcc=250 on all eleven QSOs and 0 of 29579 rows lack a DXCC number, which ruled
out the missing-entity-number theory and pointed here.
Deliberately not invalidateAwardStats(): that also drops the award matrices,
which are expensive on a large log, and a contest run would pay for it once per
QSO. This index is a few DISTINCT scans and it is what the spot colours read.
Also fixes the new worked-grid query, which asked for a column named
"gridsquare". The column is "grid" - gridsquare_ext is a different one - so NEW
GRID could never have worked on either backend. Verified against the real
database: 12516 distinct grid|mode pairs.
RBN spotters report as "VU2OY-#" and cluster nodes as "DL1ABC-2". That string
went straight into the DXCC prefix matcher, which saw an unknown callsign and
gave up, so every spot came back with an empty continent.
The filter then matched nothing at all - and worse, silently: unresolved spots
are deliberately never dropped, because the status arrives a moment after the
row and filtering meanwhile makes the list flicker. So selecting AS left the
Europeans and the Americans exactly where they were, with nothing to say why.
A real callsign never contains a hyphen, so cutting at the first one is safe.
The panel had grown two shapes for the same kind of choice. Some filters were
checkboxes, some were chips, and New counties only was a checkbox duplicating a
chip - which I added, and which was the worst of it: a control that exists twice
is not more discoverable, it is one control the operator has to recognise twice.
The rule now:
SWITCH a behaviour that is on or off, and narrows nothing by a property of
the station: hide worked, group duplicates, the two display options,
LoTW users only.
CHIPS pick any number from a set; none picked means all. Status, mode,
spotter continent. Selected is solid, unselected is the same chip
faded, so the palette keeps teaching the colour code while switched
off.
Nothing appears in both shapes. The duplicate county checkbox is gone.
Spotter continent became a chip row: seven two-letter codes fit on two lines,
they read as a set the way Status and Mode do, and several can be picked at
once - which "EU or NA" needs and a dropdown cannot express.
Both Lock buttons now sit in the heading of the section they lock, instead of
floating between sections, and every multi-select section clears the same way
through the same heading slot. One section helper and one chip helper, so the
next filter cannot drift.
The panel was also entirely hardcoded English - Search call, Hide worked, Bands,
Status - against the project's own bilingual rule. All of it goes through t()
now, both locales.
The L badge is a single letter, not a word: the call column is 120 px and holds
a callsign. It keeps the muted blue of a confirmation and never a status colour,
because whether a station uploads to LoTW says nothing about whether the spot is
worth chasing - those are different questions and must not share a palette.
The spotter's continent is not the DX's. It asks whether anyone near you is
hearing the band at all, which is why it earns its own control rather than
reusing the existing Continent column. The spotter callsign now travels with the
status query so the backend can resolve it against the one DXCC prefix table,
instead of a second continent rule appearing in the frontend.
Both AND with the status chips rather than joining their OR: "a new band, and
from Europe" is the question being asked. An unresolved spot is never dropped by
them - the status arrives a moment after the row, and filtering meanwhile made
the list flicker.
New counties only is the SAME state as the NEW COUNTY chip, reached a second
way, not a second filter. County chasing is a mode you switch into, and hunting
for one chip among eight is not how you switch into it.
Backend groundwork; the columns and filters that consume it come next.
GRIDS. A CQ is the one WSJT-X message that carries a locator, and wsjtSender was
throwing that token away. It is now returned, validated as a real field+square,
and remembered per callsign. This is the ONLY grid source available: a DX-cluster
line carries the spotter's grid at best and never the DX's, and a per-callsign
QRZ lookup under an RBN firehose is not a trade worth making. So grids are known
for the stations this receiver decoded - which is exactly the FT8/FT4 watering
hole an operator is looking at while grid chasing.
RR73 is why the grid is validated rather than pattern-matched. R is inside A-R
and 73 inside 00-99, so a sign-off satisfies the Maidenhead shape exactly and
would have planted a grid that does not exist into the index, silently.
NEW GRID keys on "GRID|MODE" with the mode put through the same normMode as
everything else, so the "group digital modes" option decides whether a grid
worked on FT8 is still new on FT4 - one rule, no branch. Grids are truncated to
four characters: a log holds a mix of JN36 and JN36QU, and without that the same
square is new forever, once per subsquare.
On cost, which was the condition: one more DISTINCT scan when the status
snapshot is rebuilt, then map lookups per spot. The same shape as the county and
POTA sets it sits beside, and the snapshot exists precisely so a spot batch never
touches the logbook.
Spotter continent and the LoTW flag come from tables already in memory - the
DXCC prefix table and ARRL's user list - so they cost a lookup each. The spotter
continent answers a different question from the DX's: whether anyone near you is
hearing this at all.
parseFloat accepts a numeric PREFIX. "2026-08-10" therefore became the number
2026, every date in the same year compared equal, and since the sort is stable
nothing moved: the Date column looked as though it were simply not sortable.
The same trap caught every callsign starting with a digit - 8B81SU and 8P9AB
both read as 8 - and the times, where "09:56" became 9. The numeric test is now
anchored to the whole cell, so anything that is not entirely a number is
compared as text, which is exactly right for an ISO date.
Sorting had no way back either. Each row now carries the index it was published
at, and a third click on a header restores that order. Headers also show an
arrow: with no indicator, a column that silently refused to sort was
indistinguishable from one that had sorted into the same order.
Blank cells sink in both directions rather than leading the ascending sort. An
empty field is missing data, not the smallest value.
Tested where it can be: the page ships its own script, so a regression is silent
- the table still renders, it just sorts wrongly.
Colour moves from the text to the CELL. A filled Band cell means new band, a
filled Pfx cell means new prefix, a filled County cell means new county. This is
not the pills coming back: a pill is a box inside the cell with its own height
and padding, so it pushed the text off the row baseline. A background has no
geometry - the text does not move a pixel - and it reads from across the room,
which a tinted glyph does not. Worked-call stays text-only: it is not a novelty,
and filling it would wash most of the rows.
The colours still come from the semantic tokens and lib/spotMarkers, so a fact
keeps one colour across the grid, the band map and the filter chips - and the
per-marker colour setting will drive all of it from one table.
US County column. The backend already resolved the county from the offline ULS
store to decide NewCounty and then threw it away; it now returns it, which costs
nothing.
The Locator column is renamed Spotter locator. It always held the SPOTTER's grid
- cluster.go says so - so a column labelled Locator next to a DX callsign was
reading as the DX's grid and was mostly empty besides. The DX grid is not in the
feed at any price worth paying under an RBN firehose.
Both display options move out of Preferences into the cluster filter panel,
beside Hide worked. They are changed while working a run, not set up once, and
a preferences dialog reopened every ten minutes is a filter in the wrong place.
The slot-highlight option reused the new-slot status, which already meant
something else and something narrower: the ENTITY was worked on this band and
on this mode, but never on the two together. What the option surfaces is a fact
about the CALLSIGN - never worked on this band and mode - and the entity may be
long since confirmed there.
Overloading it cost exactly what it always costs. Forty consecutive rows read
NEW SLOT, the genuine new-slot rows drowned in them, and the Status column
stopped carrying information at all.
new-call is its own status now, with its own badge, its own tooltip spelling out
that it is the callsign and not the entity, and its own filter chip. It keeps
the caution colour for the moment: it sits at the same "worth a look, not a new
one" level as new-slot, and the colour rework will separate them.
The status filter had to move onto the displayed status too. It read the raw
backend entry, where new-call by construction never appears, so the chip would
have selected nothing at all.
The "colour the stations not worked on this band and mode" option did nothing.
The backend computed WorkedSlot correctly and the generated bindings declared
it, but the status map in App.tsx is assembled field by field from the reply,
and worked_slot was not on that list. The extra key was simply dropped, so the
frontend saw undefined, the === false test never matched, and every spot stayed
muted. Silent by construction: nothing warns that a copied-out struct is missing
a field. All three assembly sites now carry it, and the state type names it so
the compiler catches the next one.
The band map also memoised the transformed status map on spotStatusRaw alone,
so toggling an option changed nothing until a poll happened to hand over a fresh
object. The options are read at render time and are part of the dependencies.
Muting emptied the status, which dropped the spot onto statusStyle's default
branch. That branch paints bg-primary/60 - the theme's burnt orange at 60% over
a dark card, i.e. brown - and it was written for the rare unresolved entity,
where a hint of the accent colour is right because the operator should look.
With the mute option on, most of the map is muted. So the whole band map turned
brown and the spots that were meant to stop competing for attention were the
only ones tinted with the app's own "look here" colour.
Muted now lands on QUIET_STYLE: card background, grey border, grey accent bar
and leader. That is the same style the 'worked' status already used, so it is
now one constant rather than two copies - a muted spot and a worked spot are
the same statement and must not drift apart.
Three defects in the options added in the previous commit, all found on air.
Order. Mute returned early, so with both options on the mute swallowed exactly
the spots the slot option existed to surface: the mute test is entity-level, and
an unworked callsign inside a worked entity passes it. Slot promotion now runs
first, and protects itself for free since bringsNothingNew is false on new-slot.
The two were documented as composing - "mute what is done, light up what is
not" - and they did not.
Empty status. Muting emptied the status, but an empty status already meant
"entity not resolved" in the band map, so every muted spot claimed its entity
was unknown with the country printed two words earlier in the same tooltip. A
muted flag now records WHY it went empty; blanking still drives the colour, the
badges and the ranking.
Wording. The muted tooltip said "already worked" of a station never worked -
LZ8NG on a 421st Bulgarian. What is worked is the ENTITY on this band and mode,
so that is what it says now.
new-mode had no case in the band map's statusLabel and fell through to "entity
not resolved" too. Pre-existing, same one-line switch.
Two options that change what the eye is pulled towards, both applying to the
cluster list AND the band map.
Mute worked before: a spot that brings nothing new loses its colour and its
badges. It stays in the list - the operator asked for less noise, not less
information. "Brings nothing new" reuses the dimming rule the cluster list
already had rather than inventing a second notion of done, so it keeps obeying
the same-slot option and the digital-mode grouping for free. A new-band or
new-slot status is NOT muted: having worked that callsign once on another band
says nothing about the band in front of you.
Highlight unworked in this slot: colours any callsign not yet worked on this
band and this mode, whatever the entity says. For an operator filling slots a
common entity on a fresh band+mode is the whole point, and the entity-level
status flatly calls it worked. It reuses the existing new-slot status, so no new
colour, no new legend, no new badge - both panels already knew how to draw it.
WorkedSlot is computed independently of the same-slot preference: it is what
this option reads, and it must not change meaning because a different option was
toggled. The slot index is now built when either option needs it, and the status
cache is keyed on both so a toggle invalidates it.
The rules live in one module used by both panels. Marker colours already taught
us what happens when the two derive the same thing separately.
Observation, not prediction, and it needs no new data source: the cluster event
worker already enriches every spot with the great-circle distance and bearing
from the operator's grid, which is exactly what a single-hop Es detection rests
on.
The signature is four or more DISTINCT stations at 500-2400 km inside a 90
degree bearing sector within twelve minutes. Each constraint earns its place:
distinct callsigns because one station spotted by six skimmers is six spots and
one station; the lower bound because a 6 m contact under 500 km is ordinary
tropo and says nothing about the ionosphere; the upper bound because past one
hop the bearing test stops meaning anything; and the sector because a real Es
cloud illuminates a direction, which is what separates an opening from a merely
busy evening.
Fed AFTER the Historical guard in the worker. A SH/DX reply replays a hundred
past spots in a second - precisely the shape of a burst - and would announce an
opening that ended hours ago.
Season LABELS, it never gates. Both hemispheres get a summer peak and a lesser
winter one, and an opening outside those is announced with "unusual for the
season" attached: the rare one is the one an operator must not hear about last.
One announcement per band per opening (45 minute quiet period). An opening runs
for hours and produces hundreds of spots; one alert is information, forty is
noise.
Opened as soon as 0.24.2 is tagged, so the next change has somewhere to go and cannot end up described under a version that already shipped - which is exactly what happened to 0.24.1. Invisible until the build reports 0.24.3: GetChangelog drops every entry newer than appVersion.
v0.24.1 was tagged and shipped mid-session; the six commits that followed kept appending to its block, so the released version appeared to describe changes it does not contain. 0.24.1 is restored to exactly the five entries that shipped - verified against git show v0.24.1:changelog.json - and the five later ones open 0.24.2.
The band row was hardwired to 160-6 m, so an IC-9700 - the one radio in the range with no HF at all - showed ten dead buttons and none of 2 m, 70 cm or 23 cm. bandsFor(model) follows the model, exactly as attOptions already did for the attenuator steps. bandOfHz's labels had to move with it: it returned '70' where the button says '70cm', so the current-band highlight could never have matched, and it knew nothing of 23 cm.
Settings -> Web publishing. Renders the last N QSOs with the columns the
operator picks, writes the file locally, and optionally uploads it by FTP or
explicit FTPS.
Local write FIRST, upload second, always. A network failure then leaves a good
file on disk that can be published another way, instead of a truncated one on
the server. The local write itself goes to a temp file and renames over the
target, so a reader — or a syncing client — never sees a half-written page.
The HTML page is fully self-contained: inline CSS, inline sort script, no font,
no CDN, no external request at all. It has to work on hosting that blocks
third-party requests, and a page about someone's hobby should not report its
readers to anyone.
Columns are a curated set, not "every ADIF field". This is published to the
public: RST and QSL status belong on it, the operator's home address does not.
Two triggers, both debounced through one path: a QSO is logged, or the optional
timer fires. Fifteen seconds of coalescing means a run of contacts produces one
upload rather than one per QSO, and nobody reading a web page can tell the
difference.
The config is one JSON blob under a single settings key, and that key is marked
sensitive: the FTP password lives inside it, so the whole blob is encrypted at
rest with the others. A locked vault reads back empty, which correctly reads as
"not configured" — publishing must not run with a password it cannot decrypt.
The post-QSO status refresh was gated on the DX-cluster list being on screen; the band map draws the same statuses and was left out. Working a station in the ordinary layout - Recent QSOs left, docked band map right - left its pill NEW until the cluster tab happened to be opened. Reported on an E51 that stayed orange with the QSO already in the log.
OSM blocked OpsLog and every user's map filled with "Access blocked / 403"
tiles at once. Their policy is explicit: an application must identify itself
with a proper, unique User-Agent on every tile request. A program that draws
its maps inside a web view cannot do that — the browser sets that header and
Leaflet fetches tiles as plain <img> loads — so this is not something to tune,
it is a service we are not entitled to use.
Two layers still pointed there: the Street basemap and the locator map. Street
moves to Esri World Street Map, the locator map to Carto, both key-free and
both under terms that cover a redistributed application. OpenStreetMap keeps
its credit: Carto's tiles are built from OSM data.
The locator map also gains the tile restraint the world map already had —
updateWhenIdle, no fetching mid-zoom, one ring of buffer. Fetching as fast as
the pointer moves is another item on OSM's block list, and the replacement
providers are under no more obligation to tolerate it than OSM was.
The basemap preference key is untouched, so an operator who had picked "Street"
still gets a street map, from the new source.
A new county and a new park both drew --success, so the two were the same green
in the cluster list and, since the band map copied it, in the band map too.
County takes violet; POTA keeps green, the association being worth something.
Violet is a chart hue rather than a semantic token because every token was
already spoken for: red, orange and yellow are the entity statuses, blue is
"callsign already worked", green is now POTA. It is defined in both the light
and dark chart groups, so all eleven themes have it.
Changed in BOTH panels at once — separating them in one place would have left
the two views contradicting each other about the same fact, which is worse than
sharing a colour. To make that impossible to get wrong again, the marker
definitions move to lib/spotMarkers: key, colour and label in one table that the
cluster list and the band map both read. That table is what a per-marker colour
setting will drive, which is why it is a table and not three constants.
The band map deliberately shows three of the four markers. A new PREFIX stays a
cluster-list badge: the pill is 22 px tall and a fourth segment on its strip
turns it into a colour code nobody reads at a glance.
The band map coloured only the entity status and dropped the other three
markers on the floor — not for want of data: they were already in the status
entry it receives, the local type simply never declared them, so a new park on
an entity you have worked was indistinguishable from any other worked spot.
Colours are the DX-cluster list's, taken from it rather than chosen: --info for
a worked callsign, --success for a new county and a new park. The same fact must
not be blue in one panel and green in the next.
They STACK, they do not replace. These markers are orthogonal to the entity
status — worked entity plus new park is an ordinary combination — and the
cluster list already spells them out side by side rather than letting one win.
So the pill keeps the status colour and the markers take its left strip, split
into one segment each. That strip previously repeated the pill's own colour and
carried no information at all, which is what made it the right place.
Note that a new county and a new park share --success in the cluster list, so
they share it here too. Splitting them means changing both views together, which
is the moment to do it — when the per-marker colour setting arrives.
The input was bound straight to the number and committed on every keystroke, so
it could not be EMPTIED: clearing it yields "", Number("") is 0, 0 fails the
"> 0" guard, the state never moved, and value={qsoLimit} snapped the old figure
straight back. Going from 200000 down to 100 was a fight against the field — and
read as the setting refusing to persist, which it was not: writeUiPref stores it
and syncPortablePrefs makes the database authoritative at boot. It also wrote
the preference once per keystroke (1, then 10, then 100).
Raw text in local state, committed on blur or Enter, per the controlled-input
note in CLAUDE.md.
Opening the Awards panel pulls every QSO into a cached slice and kept it for the rest of the session: the cache was only ever invalidated by a logbook change, never by disuse. Each cached QSO is a 1896-byte struct plus its strings AND a decoded map of its ADIF extras - one map allocation per QSO. Harmless at 30k rows, several hundred megabytes at 132k, which is where it was reported. A janitor drops it after 15 minutes without a reader and calls FreeOSMemory, because Go hands pages back lazily and the whole point is that the operator sees the memory return. 15 minutes is deliberately generous: an awards session recomputes every few seconds and re-pulling a large remote logbook costs seconds. The heap size is now logged when the snapshot is built and when it is released - a memory report was unanswerable without a number.
Every line of cluster traffic went straight to state: a spread copy of a 2000-element array, a slice copy, and a React render, per line. An RBN feed sends hundreds a second, so that was tens of MB/s of garbage in the renderer and hundreds of re-renders - and it ran whether the console was open or not, so the cost was paid for a panel nobody was looking at. Reported on an old PC where the UI had stopped responding. Lines are staged in a ref and flushed on a 200 ms timer, the same shape the spot handler already uses. The staging buffer is bounded too, so a burst longer than the console can show is not carried in full just to be sliced away on commit.
Several operators hit "stage current exe: rename …\OpsLog.exe …\OpsLog.exe.old:
Accès refusé" and could not update again. Two separate causes, both ours to
handle.
The staging name was fixed. os.Rename replaces its target, so a single leftover
".old" that could not be deleted — a scanner holding it open is the usual
reason, and the pre-existing os.Remove was best-effort and ignored — made every
later update fail with that error, permanently, recoverable only by deleting the
file by hand. Staging now uses a unique ".old-<nanos>", which no leftover can
block, and the startup cleanup sweeps the pattern instead of one name.
Renaming a running image is legal on Windows, but some endpoint protection
(Bitdefender's ransomware remediation among them) blocks it outright, and no
retry gets past that. So the swap is deferred: the new build is parked beside
the old one and a detached helper moves it into place after this process exits,
when the file is no longer a running image. It keeps trying for ten seconds,
since a scanner tends to let go a beat after the process dies.
A short retry stays in front of both, for the ordinary case of a scanner holding
the file it has just watched being written.
If even the deferred move fails, OpsLog restarts on the CURRENT version rather
than leaving the operator with nothing — someone mid-QSO losing their logger is
worse than an update that waits — and only a successful swap passes
--post-update, so the download survives for the next attempt instead of being
swept by the cleanup.
Both band maps were pinned to a hardcoded width — 300px docked beside the
tables, 260px per card in the Band map tab. On a busy band the map could not be
given more room, and on a quiet one the log could not take it back.
The docked map becomes a resizable grid column with the grip in the gap between
the panes, so the handle costs no space; the tab cards share one width with the
grip on their right edge. Side-by-side columns of different widths read as a
mistake rather than a choice, which is why the tab has one width and not one
per card. Double-click either grip to return to the default.
The drag measures from the pointer's START position rather than the container,
so the same helper serves both edges — the docked map sits on the left or the
right depending on the operator's setting, and the grip is on its inner edge
either way. Pointer capture, like the main splitter: without it the map or the
grid under the cursor swallows the moves.
Both widths are persisted through writeUiPref and registered as portable, so
they travel with the data folder like the main splitter and the rest of the
layout.
The seven existing themes are warm beige, cool grey, sage grey, slate, warm
dark, graphite and black — every one neutral, six of the seven accented orange.
Picking a theme changed the shade of grey and little else.
These colour the SURFACES, not just the accent, and each takes a different
primary so the picker tells them apart at a glance: violet on deep indigo, cyan
on deep teal, magenta on aubergine, indigo on a crisp cool white.
The semantic colours stay recognisable as themselves. A logger is read for
hours and "red means a problem" cannot become a decorative choice, so the hue
budget went on the surfaces and the primary. Where a theme's primary would have
collided with a meaning, the MEANING kept its identity and the ornament moved:
Ocean's info is blue rather than cyan, Plum's danger is red rather than rose,
and the matrix entity ramp shifts to cyan under Indigo and teal under Nordic so
a "confirmed" cell never reads as a button.
Everything else follows for free: the grids resolve var(--…) at runtime, so
they re-skin with no re-render. The three shared registrations that do NOT
follow automatically are done — the chart palette (light vs dark ramp) and the
date-picker icon inversion, which is keyed on an explicit list of dark themes.
Prefix exists so a field holding a bare value counts for an award whose codes
carry a letter: a French operator writes "74" in STATE, DDFM's codes are "D74".
Def's own doc says exactly that. But searchOne applied the prefix AFTER looking
the token up in the reference list — after the step that had just failed — so
the bare form matched nothing and the prefix decorated an empty result. The only
way through was a regex, in a mode where the operator had chosen "code" and
explicitly not "pattern".
The token lookup now tries the prefixed form when the bare one is not a known
reference. The list stays the authority: an unknown number still matches
nothing, so the prefix completes references rather than inventing them.
Second bug in the same pass: the blanket prefix also hit codes that came
straight OUT of the list, so a field already holding "D74" produced "DD74" as
soon as a prefix was configured — for both the token lookup and the
description matcher. Branches that yield whole codes are now excluded from it;
the ones that yield a raw capture (regex, whole-field split) still get it.
A forwarder (W&P, seen in the field in front of MSHV) prepends the origin as
plain text before re-broadcasting:
"127.0.0.1:2237|" + <the original, untouched WSJT-X packet>
That puts the magic 15 bytes in, so every datagram failed on "bad magic
0x3132372e" — those four bytes being ASCII "127." — and an operator running
MSHV behind the relay saw no decodes, no callsigns and no auto-logged QSOs.
No new service type: what follows the header IS a WSJT-X packet, so the parser
and everything downstream apply unchanged, and a separate type would duplicate
decode, status and logged-ADIF handling to strip 15 bytes. ParseWSJT skips the
header instead, which also covers any other relay that wraps traffic this way.
The match is deliberately narrow — the magic must fall within the first 64
bytes AND every byte before it must be printable ASCII. A corrupt or truncated
packet that merely contains those four bytes somewhere is not resurrected into
a QSO; it fails exactly as it did before.
Test data is the real captured datagram, header included.
"WSJT parse error: bad magic 0x3132372e" named neither the sender nor the
payload, so there was nothing to act on — even though those four bytes are
ASCII "127.", i.e. some program broadcasting an address on a port expecting
WSJT-X binary.
The line now carries the remote address, the size, a printable preview and a
hex dump of the first 96 bytes. Text senders are readable at a glance; a
genuinely binary payload still shows its bytes.
And it stops after five. The reported case wrote that line about 150 times a
second: a permanently misconfigured port would fill the 10 MB rotating log with
one repeated sentence and bury every other piece of evidence — the log's whole
purpose. The fifth line names the two things worth checking, the sender and the
service type.
N1MM's parse error goes through the same path; it had no packet detail either.
The award editor offered five confirmation sources and Def's own doc comment
named five, but confirmed() had cases for three. "qrzcom" and "custom" fell
through the switch, so ticking either marked nothing as confirmed — the exact
failure the GrantCodes comment in this struct warns about: a checkbox that does
nothing is worse than no checkbox, because it is trusted.
QRZ.com reads qrzcom_qso_download_status, not the upload one: uploading a QSO
is us telling QRZ about it, which confirms nothing.
Custom names a field. Rather than a checkbox per external source, the Def gains
ConfirmField + ConfirmValue: any QSO field or ADIF extras key, and optionally
the comma-separated values that count. That one shape covers the three cases
asked for — the OpsLog card marker (APP_OPSLOG_QSL_RCVD), an arbitrary ADIF
tag, and a tag stamped by an imported club list — because all three end up as a
field on the QSO.
An empty ConfirmValue means any non-empty content confirms: the OpsLog marker
stores the date the card arrived, not a Y/N flag. A custom source naming NO
field confirms nothing, deliberately — the opposite default would silently mark
a whole logbook confirmed.
Windows parks a minimised window at -32000,-32000 with a stub size and reports
it as not-maximised, so closing OpsLog from the taskbar while minimised stored
exactly that. Seen in a log: "window: saving -32000,-32000 237x39
maximised=false".
The restore side already rejects both the impossible corner and the
below-minimum size, so nothing opened off-screen — but it fell back to the
default placement, and the operator silently lost the size, the position and the
maximised state they had set.
saveWindowState now keeps what was already stored when the window is minimised.
Two tests for that: the Wails flag, and the -32000 corner, because a window that
is mid-close can sit at that corner with the flag already cleared. A poisoned
window.json repairs itself on the next close of an on-screen window.
The stamp answers "PSE QSL", so it reads as its counterpart or not at all: a
bare "TNX" next to a QSL message says thanks for something unnamed.
Changed at the source of the {qso.pse_tnx} token, so every card picks it up
with no template edit. The live indicator in the QSO editor and the hint beside
it follow.
rouggy