It ran into a panel nobody was showing: the button never set showLog, so
the tab sat on 'Pick a service' while a 25000-QSO upload came and went.
Now it opens the console like every other action here, and the console
has something to say: the callsign it is scoped to and how many of the
logbook's QSOs that leaves (a two-callsign database sends one of them,
which is the whole explanation for a count that looks short), the
exported record count and file size, and HamQTH's own reply.
Plus the two facts that make the site's own number readable: HamQTH
accepts the file and imports it in the BACKGROUND, and it reports what
it made of each record by e-mail — never in the reply we get. An
operator comparing counts a minute later is not looking at a failure.
HAMLOG.online was wired into all three when it arrived; HamQTH shipped
with only its uploader and its editor row, so the log could not be asked
'which contacts have never gone there' — the question that precedes
every backlog upload — and a log uploaded to HamQTH by hand could not be
marked, leaving OpsLog offering to send every one of those contacts
again. Sent side only, both whitelists, both panels, two hidden-by-
default columns.
The importer has always read an absent BAND_RX/FREQ_RX as 'not split, so
RX = TX'. The logging paths never did, so what a contact carried
depended on which door it came in through — and it shows outside
OpsLog: the record forwarded over UDP is written from the QSO as logged,
so a receiver that reads BAND_RX (Log4OM does) found nothing there.
Filled at AddQSO, the one funnel every path goes through, so the
database, the export and the forwarded copy agree. A genuine split
contact keeps what it was given — table-tested both ways.
v0.27.6 was tagged at 88f35e2, and the disconnected-pill fix and the
HamQTH whole-log upload came after it: their entries were promising a
released build features it does not carry. They open 0.27.7, and the
.6 pill entry goes back to describing what actually shipped.
The per-QSO API is the only correct way to send a SELECTION, and at the
pace it must be driven a 14k backlog costs the better part of an hour.
HamQTH's other endpoint takes a whole log as one file — and REPLACES
what is on the site with it: its documentation says plainly that partial
uploads do not exist. So it is offered as its own deliberate act behind
a confirmation, never as the batch path behind 'send these', where it
would delete every QSO the operator had not selected.
Scoped to the callsign this profile uploads as, so a database holding
two operators' contacts cannot push one into the other's log; tar.gz
above 12 MB because the ceiling is 20 and a six-figure log passes it as
text; and every QSO not already stamped is marked sent afterwards, in
bulk, so the backlog list agrees with reality.
The pills listed live SESSIONS, so disconnecting one made it vanish —
along with the only way back. They now list every enabled server, and a
server with no session shows as disconnected: the state you click to
undo. The address, known only while a session exists, drops out of the
tooltip rather than reading ':0'.
The colour already said connected; the word beside it said it again, in
the widest part of a row that has to hold five servers. It goes, and the
pills earn the job nothing else in the app had: reaching ONE cluster
without opening Settings. Click connects or drops that session — a
connecting one included, so a server retrying into a dead host can be
stopped. State, retry count, address and last error move into the
tooltip, where they cost no width.
Tried it, looked at it, took it out. A headline names no band, so the
only honest verdict is entity-level — and a pane where every row reads
NEW DXCC or worked says nothing an operator can act on. Two words
repeated forty times is noise wearing the clothes of information.
The announcements pane keeps its badge, where the announced bands and
modes make the verdict specific enough to be worth a colour. The news
pane keeps what it is good at: what is happening, and a Watch button.
Changelog entry withdrawn with it — it never shipped.
ctrl+wheel is the WINDOW zoom everywhere else in OpsLog, and one gesture
that resizes the whole app over one panel and one band map over another
is a gesture nobody can trust. The + and - buttons keep the zoom, and
the footer hint stops advertising what is gone.
The DX-World headlines now carry a chase badge like the announcements —
but ENTITY-LEVEL only. A headline names no band, and asking the slot
question without one answers 'new band' for every entity ever worked:
the one mistake that costs a QSO. So the news says NEW DXCC or worked,
and the sharper verdicts stay where the bands and modes are known.
A legend is a reference, not a running display: once its colours are
learnt it spends four lines of a short screen saying nothing new. A
chevron in the footer folds it, the same chevron brings it back, and the
choice survives a restart.
The channel list had grown in the order channels were added, so finding
one meant remembering when it arrived. Paper QSL and LoTW lead — they
are the two that carry an ARRL award — and the rest sort by name. One
ordering function feeds both the picker and the status table, so the two
cannot drift apart; the three extras-backed channels (OpsLog card,
HAMLOG.online, HamQTH) now sit in that same list instead of being
appended by hand after it.
The channel picker and the status table both list it now, sent only —
HamQTH publishes no confirmation feed, so its received column shows a
dash rather than an N that would read as 'not confirmed yet'. Backed by
the same APP_OPSLOG_HAMQTH_SENT extras the uploader stamps.
The news pane had clickable callsign chips while an announcement had a
Watch button — two gestures for one act. The chips become plain labels
(they still show which calls are already watched) and the news card gets
the announcement's own Watch button and Open link.
A DXpeditions tab holding the two feeds the DX world announces itself
on: NG3K's ADXO (structured — dates, entity, calls, bands, modes, QSL
route) and DX-World's headlines. What a logger can say that a news
reader cannot is whether the operation is worth chasing, so every
announcement is put through the SAME verdict the cluster paints on a
spot — one badge, strongest wins, dimmed when the need is only a missing
QSL — with an 'only what I need' filter. One click watches every
callsign of an operation; the news headlines are mined for callsigns so
they can be watched the same way.
Parsers ported from DXHunter's and pinned with table tests against real
feed text: the 'as PJ2/W2APF' mining, the DXCC-prefix-first
normalisation without which no spot ever matches, both ADXO date forms,
and the '160-6m' span whose unit is written once (DXHunter read that as
6m alone and lost the low end).
Watchlist membership now announces itself app-wide, on the same card as
a new version: the bindings emit watchlist:changed, so a call added from
the cluster or this tab is confirmed where the operator is actually
looking — the panel's own inline message never was.
HamQTH joins the Confirmations page with a sent side only — the site
publishes no confirmation feed, so there is nothing to receive — and it
defaults to R, the same 'still to upload' the other online services get.
That default could not have worked as written. HamQTH and HAMLOG.online
keep their sent state in an ADIF extra, and eligibility blocked on the
key being PRESENT — so an operator setting the natural R default would
have stamped every new QSO 'already gone' and silently disabled the very
auto-upload the default arms. Eligibility now reads the VALUE: the ADIF
pending statuses mean pending, anything else means sent. Guard-tested
both ways.
Chase POTA already withdrew its badge, filter chip and column when
unchecked; US counties and new prefixes now have the same switch, and
unchecking Chase new grids finally withdraws the NEW GRID badge too (the
backend stopped learning squares but the display kept shouting the ones
it knew — the setting is mirrored to localStorage so the display layer
can gate synchronously). All withdrawal happens at the display layer:
the facts keep being computed, so re-ticking a box needs no rescan.
TQSL validates MY_CNTY against the ADIF secondary-subdivision list,
which is the US county enumeration — 'XX,County' with a two-letter
state. A Canadian profile produced 'ONTARIO,Kawartha' (the export joined
MY_STATE onto the county wholesale) and TQSL refused the whole record.
Two layers: adifCounty only prefixes a two-letter state, so exports stop
manufacturing the invalid shape; and UploadLoTW scrubs any MY_CNTY that
is not the US shape before signing — MY_STATE and MY_GRIDSQUARE already
locate the station for LoTW, and the US form survives for the county
hunters. Table-tested.
qso_realtime.php with the account's callbook credentials (which the
external-services config falls back to when its own are blank), one ADIF
record per QSO, prg=OpsLog. HTTP status IS the answer: 200 saved, 400
rejected (a duplicate counts as delivered, like HRDLog's insert 0), 403
credentials. Sent-state lives in APP_OPSLOG_HAMQTH_SENT extras like
HAMLOG.online — ADIF names no HamQTH field. Auto-upload on log, on-close
batch, right-click Send to, QSL Manager backlog, and a Test button that
authenticates against the callbook login, which cannot touch the log.
Fixes a real mis-route on the way: manual 'Send to HAMLOG.online' had no
branch in runManualUpload and fell through to QRZ.com — the selection was
uploaded to the wrong service with the QRZ key. Both extras-stamped
services now have their own branch.
A database switch relaunches OpsLog, and the child regularly won the
race against the old instance's teardown — so following the app's own
instructions produced 'OpsLog is already running'. The relaunch now
passes --relaunch, which gets the same 20-second mutex patience the
post-update restart has always had. Opens the 0.27.6 block.
The Confirmations page's Club Log row gains its received side, applied to
every new QSO the way QRZ.com's already is — so the match download's Y
stands out against a ledger of explicit Ns.
The catalog's order is the grid's column order, which appends new columns
at the end — so the two ClubLog rows sat at opposite ends of the Uploads
box. Sorted for display only; the grid keeps its own order.
startAmps tore down and rebuilt every amplifier client on any settings
save, and closing a COM port makes the Windows driver drop DTR/RTS. On a
KPA500 those lines ARE the power switch — so 'save' meant 'amplifier off
twenty seconds later', exactly F1TPL's report, and the fix that held the
lines while OPEN couldn't survive a close. An amplifier whose config is
unchanged now keeps its running client across saves; the others still
rebuild.
Also: isolate on the FT Map panel — Leaflet's z-1000 panes were painting
over the menu bar and the Preferences dialog, as the grid map's own
comment warned.
The 0031 clublog ALTER hit split settings databases (their qso table
moved to the logbook years of QSOs ago) and the whole open failed —
silently, because the error went to a println the GUI subsystem
discards. The app then ran with no settings store: every panel showed
defaults, 'db not initialized' in Preferences, and operators read it as
their database being lost. Nothing was ever touched: the failed
migration rolled back on every attempt.
The SQLite migration path now tolerates what the MySQL path always has —
plus the one case it never meets: ALTER/CREATE INDEX/DROP on a table
this database legitimately does not hold. And a failed open is written
to the rotating log, where the next such morning can actually be
diagnosed.
A stash pop during a test-on-parent check left conflict markers inside
changelog.json and the next commit carried them. The 0.27.4 release had
already shipped with its seven entries, so the two Club Log entries that
landed after it open the 0.27.5 block, per the released-block rule.
Two new QSO columns, clublog_qso_download_status/date — a Club Log MATCH,
the service's own confirmation (both stations uploaded the QSO, paired
within 15 minutes). Full promoted-column lockstep: migration 0031, repo
insert/scan, ADIF dictionary + import + export (app-defined
CLUBLOG_QSO_DOWNLOAD_*), table columns, filter builder, bulk edit and the
QSO editor's Club Log row.
The QSL Manager's Club Log entry now actually downloads: getmatches.php
with the existing account settings and the embedded application key,
incremental via the match-completion date filter, matched call+band+mode
±15 min with a mode-blind fallback because Club Log reports 'false' for
modes it cannot infer. Matches always exist on both sides, so unmatched
ones are listed rather than skeleton-added.
And Super Check Partial can merge Club Log's weekly SCP list (~180k calls
worked on the air in the last 3 years) with MASTER.SCP — an opt-in
checkbox under the SCP setting.
The hard maxBounds added for the single-world look fought the zoom: with
the world smaller than the window the clamp dragged every zoom step into
a corner. The noWrap tiles already guarantee one world on their own, so
the bounds go — on both maps.
An FTx menu between View and Tools gathers FT Decodes, FT Map and the
Grid squares map — three views of the same UDP feed that were buried in
Tools. And the watchlist spot rows swap the elastic country column for
fixed widths, so band, mode and frequency read as columns instead of
starting wherever each country name happened to end.
The FT map wrapped the planet into a row of copies and showed Leaflet's
hard-coded grey beyond it. noWrap tiles inside hard bounds now, like the
grid-square map already did — and the container background override moves to
style.css so EVERY map's surround follows the theme's card surface instead
of #ddd (the grid map's per-div fix was losing to leaflet.css specificity).
A new FT Map tab (Tools menu, beside FT decodes): every station decoded in
the last half hour drawn as a great-circle arc from the operator's own
square, coloured by band the way PSK Reporter taught everyone to read it,
fading with age, one arc per call (its freshest sighting), a tooltip with
call/grid/SNR, the MainMap basemap picker and a legend of only the bands on
screen.
Performance was the open question and it is answered structurally: the panel
mounts only while its tab is ACTIVE (zero cost otherwise), the map renders
canvas (one element, not one node per arc), arcs are capped at 300, and
redraws follow the decode list — every period, not every frame.
spotEntityStatus says 'worked' when every grain is in the log; the
unconfirmed check compared its all-ledger verdict against the empty string,
so no band/mode/slot need ever dimmed however many unconfirmed QSOs stood
behind it — only the entity-level NEW, tested by map presence, behaved. The
check now runs for exactly the four needy statuses and asks for 'worked'.
Dimming the filled chip to a transparent background left its light
foreground text on nothing — an invisible badge in the row. Unconfirmed now
renders in the category's own hue, dashed border and text, like the extras
beside it.
The frontend caches statuses and only asks about unknown keys, so flipping
the chase mode left half the screen judged under the old rules until a
restart — which read as the mode not working. The backend now broadcasts an
invalidation and the panels re-ask as rows repaint. The entity badges also
say their name with a question mark, dashed, when the need is only a missing
QSL — the 50% opacity was invisible on a filled chip.
The unconf label was hard-coded to GRID? for every category — a prefix
worked-but-unconfirmed showed as a second grid badge beside the real one,
which read as impossible. Each badge now says its own name plus a question
mark. And the global Chase block moves above the Chase POTA/SOTA/grid
checkboxes, where its operator expected the master switch.
my_dxcc/my_cq_zone/my_itu_zone go through BulkSetIntField, not the string
column map — the guard flagged them as unwritable ever since they were
added, which is the drift it exists to catch, one package over.
The gap between worked and confirmed becomes visible everywhere. A Chase
setting on the DX Cluster page picks the mode; in new-plus-unconfirmed the
verdicts for DXCC/band/mode/slot, prefix, county and state are judged against
CONFIRMED-only ledgers built from the operator's chosen sources (LoTW, card,
eQSL, QRZ.com — HRDLog has no confirmation field in ADIF to offer), and a
need that exists only because a QSL never arrived is flagged Unconf: the
cluster and decode badges render it dimmed, the convention the grid hunt
introduced. The grid's own Chase selector folds into the global one; its
square-matching scope stays where it was, being grid-specific. One snapshot
key change rebuilds the cache when any of it flips.
The WAS gap made visible: a worked-states set joins the cluster status cache
(distinct states over the three US entities), the status entries carry a
NewState flag wherever the state resolves (log county first, ULS licence
second), and the decode panel gains the badge in its status column plus the
filter chip alongside NEW/BAND/MODE/GRID. Colour from the shared marker
palette, so the cluster can adopt it later without a second hue.
Between the locator and the country, per its operator: the two-letter badge
anchors the eye and the spelled-out name answers it — 'SD' alone is a quiz
for a European. The status-cell badge from the first pass moves here.
Plain text, not a badge — it is a fact, not an action — placed first so a
WAS chaser can scan the column edge. Shown whenever the status cache knows
the state (log county data or the licence address).
The voice keyer grows to F1-F12 — the UI was already data-driven, so the
count constant, the F-key range and the labels are the whole change.
And the preferences stop stuttering on a busy station: every spot batch,
decode batch and CAT snapshot re-rendered the entire App tree behind the
dialog — cluster grid, decodes panel, thousands of nodes — and the pointer
janked over the very panel the operator was trying to use. While Settings is
open the flushes park their batches in the pending refs (bounded) and CAT
repaints at most every two seconds; closing the dialog drains everything.
The panel now mirrors the decoder's own window line for line — the operator
compares the two side by side, and the SNR sort scrambled that
correspondence. The report is still right there in its column.
Three faults, one report. The slow poll asks ^TP — the KPA1500's ATU, which
a KPA500 (no ATU) never answers — and ask() dropped the whole connection on
any read timeout: a two-second stall and a reconnect every slow cycle, which
is why buttons lagged and the status read wrong. Worse, every serial reopen
toggled DTR/RTS — and those lines are the KPA500's POWER SWITCH (that is how
the Elecraft utility turns it on), so the amplifier obediently switched off
twenty seconds after its operator pressed nothing but Standby.
Silence is no longer a dead link (write errors still are), ^TP is never asked
again after one silence, the control lines are asserted once and held, and
the baud field becomes a list of the rates these amplifiers actually speak.
A connect that lasts 50 ms leaves nothing to say which answer the client
disliked — WSJT-X refusing to attach to the shared CAT was undiagnosable
without the full wire trace. The handshake is now always logged; steady-state
polling stays behind the trace switch.
Configure (message 15), mode field only, every other field sent as
no-change: an FT4 spot clicked while WSJT-X sits in FT8 lands the operator
ready to decode instead of staring at gibberish. Sent to every instance
heard this session — one already in the right mode treats it as a no-op —
and only for modes the decoder actually speaks (FT8/FT4/JT65/JT9/MSK144/
Q65/FST4); CW and SSB are none of its business. Toggle in the Connections
panel, on by default.
The great-circle arc to a polar entity (Franz Josef Land) peaks near 88°N;
fitting the raw arc points framed a band of tile-less white above 85°, where
Web Mercator ends. The camera's bounds are clamped to ±85° — the line still
draws to wherever it goes, only the framing stays where there is a map.
WSJT-X restates the same DX Call in every Status packet, and a spot clicked
in OpsLog put that call into the last-UDP memory (deliberately, for another
case) — so the guard let the decoder's next restatement straight through and
the entry snapped back to old news. The decoder's stream now only counts when
its value CHANGES; a clear over there resets the edge so re-selecting the
same station still lands. Explicit picks (N1MM, remote, spot relays) are
untouched.
Message 13 paints callsigns in WSJT-X/JTDX's own Band Activity window with
verdicts from the same cluster status cache that colours the spot grid:
watchlist pink, new DXCC green, new band for its entity orange. Deduplicated
per instance+call+verdict, one datagram each; a verdict that lapses (the
operator worked them) clears that call, and turning the option off clears
everything via the protocol's CLEARALL!. Off by default, switch in the
Connections panel.
Message 7 asks a program heard for the first time this session to replay the
decodes already on its screen, so the FT decodes panel starts full instead of
empty until the next period. Replayed lines arrive marked not-new and are
shown but never auto-answered — the auto-caller now checks, on top of its
30-second freshness gate.
The operator preferred the original's aesthetics. Adopted: the signature pink
callsign (a fixed identity colour, like the cluster palette), counters as
coloured pills, a quieter card surface with a hover, a stronger needed bar on
spot rows, and the lightning bolt back on the DXpedition badge. Everything
else stays on theme tokens.
The operator measured the whole table: 10 W showed 83, 12 W showed 100, 13 W
showed 8 — the bargraph is relative to a meter RANGE that flips at 12 W
(0-12 QRP, 0-120 above). A bar percentage was never watts; with the PC
setting choosing the range, it converts, and the console prints the watts
beside the bar.
One placeholder per id in a single IN (…) hits SQLite's bound-variable cap:
bulk field set died at 168 000 QSOs with 'too many SQL variables' (10 000
passed). One chunker now serves bulk set (text/int/extra/frequency), the
delete, the post-upload markers and the export-selection iterator — the last
collecting and sorting once at the end so its chronological contract holds
across chunks. 500 ids a statement keeps every backend far from any limit.
Tried on the real IC-7760: the magic packet wakes nothing; Icom's own
mechanism is the network-standby session plus the CI-V power-on, which
OpsLog already does. The MAC store and the WOL sender go, and the changelog
entry with them. The CI-V pump now says once per session that the client
pings are armed — invisible pings read as absent pings.
wfview sends a ping (0x07, 0x15 bytes, its own seq and a monotonic
timestamp) every 500 ms on every stream, and so does RS-BA1. OpsLog only
ever REPLIED to the rig's pings — and the loaner IC-7760 stopped serving
CI-V data about a minute into nearly every session while the transport
stayed alive: the rig had concluded nobody was listening. Scope, TX audio
and idle numbering were each eliminated in turn before wfview's source
settled it. Pings now go out on control, CI-V and audio alike.
Two lookups race when a second callsign is typed before the first answered,
and the slow one landed last: the grid showed the previous call's QSOs under
the new call's name — LU5AVM's contact filed under LW8ETV, screenshot in
hand. A request token now discards any response that is not the newest.
A 7760 switched fully off takes its LAN server down, and no session means no
console and no ON button — the one thing remote operation cannot tolerate.
The rig's MAC is learned from every login and persisted; when a dial finds
nobody home, the magic packet goes out to it before the next attempt. The
rig wakes into standby, the session opens, and the ON button does the rest.
The generic Freq write lands on the TX VFO when split is engaged: an FT-2000
moved B and left A — the receiver — behind, so a spot click QSYed the
transmitter and nothing audible changed; without split the same write moved A.
With split on (and the operator on MAIN), FreqA and FreqB are both written,
so the radio arrives on the spot whole, split left as the operator had it.
The inference tables knew FT8 on some bands and nothing on others: a 10.136
spot with a bare comment read as generic DATA — which is what every
skimmerless spot of a DXpedition's 30 m FT8 slot did. The standard FT8/FT4
windows are added on 30 m, 60 m, 80 m (FT4), 17 m (FT4) and 12 m, in BOTH
tables (backend band plan and frontend spot.ts), ahead of the wide segments
they sit inside — order is the mechanism there.
Every idle went out with seq 0, seven a second, interleaved with
properly-numbered data — where RS-BA1 and wfview draw idle sequence numbers
from the same counter as everything else. A rig that follows the sequence
tolerates the zeros for a minute or so and then stops serving CI-V data on
the stream: the shape of every dropout the loaner IC-7760 has shown, with the
scope and the TX path both since eliminated as causes. Idles now take the
next tracked seq and sit in the retransmit buffer like any other packet.
The 0x26 01 copy was fire-and-forget, and a real 7760's sub VFO kept
yesterday's mode with nothing to show why — refused, or sent with a stale
mode byte. The mode is read first when unknown, and both outcomes are
logged.
The console's thin home-made bars were the one holdout — every other panel
(Flex, Elecraft, the amps, the tuner) draws the same MeterBar. And the mode
badge said SSB, which on any band leaves the operator guessing which
sideband the rig is on: the panel now shows USB or LSB (the log keeps ADIF's
SSB).
It keeps 'not repeating' on a network rig and every guess at which guard was
tripping proved wrong. Each round now logs play / playing-flag / end / gap,
and every exit names its reason.
The blind kill at connect still left a real 7760 streaming — the selector
byte differs by model, and the result was discarded so a quiet failure looked
like a cure. Both forms go out now (a NAK costs one frame) and the outcomes
are logged. The power table also gains its third measured anchor: a real
200 W reads full deflection, so raw 213 is 200 W — not the printed face's
250.
Over the network the rig's mode read fails now and then and the CAT push
blanks the mode for a poll or two; the auto-CQ loop treated that blank as
'not a phone mode' and stopped after the first call. Only a KNOWN non-phone
mode stops the loop now.
The CI-V meters are point samples: between two SSB syllables a poll lands on
0 W and a perfect SWR, so a 150 W transmission read 0-10 W most of the time
and the SWR sat at 1.0. Power now decays like a needle and SWR holds its peak
then snaps to the live value — the same meterPeak the Yaesu and Kenwood
consoles already use. The readout column also widens so '180 W' stays on one
line.
TUNE (1C 01 02) started a cycle; nothing sent the 0 that puts the tuner back
through. An ATU chip beside SPLIT now toggles it, and the tuner state is read
on the slow front-panel beat so the radio's own TUNER button stays in sync.
Added on a report that they were missing; they were below the fold, not
absent. The useful parts stay: the MOX button routes the PC microphone on a
network station, split aligns the TX VFO's mode, and SetIcomSplitOffset
remains available for a future use.
The first attempt sent it only after CI-V answered — but the waveform flood
is what stops CI-V answering, so a rig wedged streaming never received its
cure. It now goes out at the end of Connect, before anything waits on a
reply: an unanswered set costs one frame, and the rig acts on what it decodes
whether or not we hear the acknowledgement.
The 0x27 output flag lives in the RADIO and survives our sessions: a scope
enabled by an older build (or another program) kept flooding every new
session with 700-byte frames — and on the IC-7760 that flood is what kills
the CI-V link. One 'waveform output off' as soon as the rig answers, once per
session; the front-panel display is deliberately untouched.
The voice keyer and the live microphone each numbered their packets from 1.
A message played after a talk session re-used sequence numbers the rig had
already seen and was discarded wholesale: the PTT keyed for the full length
of the message and none of it was modulated. PlayTX now feeds the same framer
and counters as the microphone.
Every Icom streams its waveform differently: the IC-7851 controls a scope it
never streams, and a real IC-7760 stops answering CI-V altogether a few
frames in — taking CAT and the audio down with it, which is this week's whole
dropout saga. Chasing a per-model frame layout for a decoration is not worth
a console that drops the link; the radio's own scope is better anyway. The
backend scope plumbing stays (harmless while nothing enables it); the UI and
its enable path go.
Keying alone is right on USB, where the operator talks into the radio's own
mic; over the LAN it transmitted silence. The console button now goes through
one binding that picks the road: network To-radio → the same key-and-stream
path as the Talk button, USB → key and nothing more.
The MOX and SPLIT controls lived at the bottom of the Transmit card, below
the fold on most layouts — a console whose PTT needs scrolling is not a
console. A strip under the mode row now carries the PTT and Split as
OFF/+1k/+5k/+10k, DXpedition style, on a new SetIcomSplitOffset that takes
the chosen TX offset (0 keeps the old CW/SSB convention). The mode-alignment
from the previous fix rides along on every engage.
Its enable condition predated the network audio: with From-radio empty — the
normal network setup — the button was greyed out, and an operator who had
stopped listening had no way to start again from this panel. The stale 'later'
hints on both buttons said the network paths did not exist; they do.
A radio in standby is silent on CI-V by design; only the transport chats.
The quiet-recovery added for the mid-session deafness treated that silence
as the fault it usually is and tore the session down every 15 s — taking the
console and its ON button with it, so a radio that was off when OpsLog
started could never be turned on. The recovery now arms only once the
session has heard at least one CI-V payload.
The sink — decode + recorder feed — was built only on the success path of
starting the monitor, so speakers-off (the remembered preference) or an
output device that failed to open silently disabled the WHOLE stream: the
session connected audio=false, and a ticked RX-audio checkbox did nothing.
The sink now exists whenever the stream is on; only the monitor obeys the
speaker choice.
'Monitor already running' surfaced as the speakers checkbox unticking itself
the instant it was ticked, whenever the UI's idea of the state and the real
monitor had drifted apart. Stop-then-start makes the button mean what it
says.
The TX-audio entry points type-assert the transport; a nil port failed the
assertion and produced the sound-card message, sending an operator whose
radio was simply off — or on a new DHCP address — hunting through audio
devices.
The unselected VFO was read only when split was on, so a dual-receiver rig's
second dial sat blank on the console. It is now read on the same slow beat
regardless. And engaging split copies the main's mode (and data flag) onto
the TX VFO with 0x26 01 — a split whose TX VFO still speaks yesterday's mode
transmits FM into a CW pileup; rigs that predate 0x26 NAK it harmlessly.
The talk button learns the network road the voice keyer already took: when
To-radio is the radio itself, the microphone is captured and re-framed into
the rig's 320-sample packets, the capture callback serving as the clock — the
mic delivers in real time, so no ring and no pacer. Counters and the frame
remainder live on the audio stream, so a talk session spans calls. PTT keys
before and releases after, through the same code as the USB path.
With this, a network Icom is a complete remote station over three UDP ports:
RX audio to the headset, live voice and recorded messages back, CW through
the rig's keyer, CAT for everything else.
AudioStartMonitor refused before reaching the network branch when From-radio
was empty — which is the NORMAL network setup, so the new speakers checkbox
un-ticked itself the instant it was ticked. The capture-device requirement now
applies only to the USB path that actually captures.
Same remembered preference as the console's speaker button, surfaced where
the stream itself is enabled — tick RX audio and decide in the same breath
whether it reaches the speakers. Applied immediately, no save needed.
Muting the network RX audio took a trip through Settings → Audio → Stop
listening — two panels away from the console the operator is actually using.
A speaker button now sits beside ON/OFF (network only, like them): one click
to listen or mute, the recorder and the voice keyer unaffected, the choice
remembered across reconnects and launches.
The RX stream and the speakers were welded together: every reconnect and
every launch restarted the monitor, so an operator sitting in the room with
the radio could not have the stream (which the recorder and the voice keyer
need) without also hearing it twice. Stop listening now persists, the Listen
button turns it back on, and reconnects respect the choice.
The audio session opens its TX side alongside RX (txenable in the conninfo,
16 kHz both ways), the stream gains a paced sender — 320 samples every 20 ms,
the frame mirroring what the rig itself sends on this socket, IDs swapped —
and the network Icom takes the same NetworkPlayer slot a TCI radio does: pick
'Radio (network audio)' as the To-radio device and the voice keyer needs no
cable and no virtual sound card. PTT brackets the message through the same
code as every other audio path.
The capture shows the spiral: a session goes deaf on CI-V (28 unanswered
commands, transport alive), the fast-fail rebuilds it, the new session answers
for one second — one state read got through — and is then strangled when the
rig purges the half-open old one. Twenty seconds a lap, no audio throughout,
and only the third or fourth session survives. When the previous session died
less than 90 s after connecting, the redial now pauses 20 s so the purge
happens BEFORE the new session, not on top of it. A long-lived session still
reconnects immediately.
rouggy