Hi @ea5wa — thanks for putting this together! CI is red on all three platforms (build, check-macos, check-windows). I dug through the diff and the most likely root cause is the new sources are unconditionally added to the cross-platform source lists, but they are Windows-only code.

The big one — Windows-only sources added to cross-platform targets

In CMakeLists.txt you added these to CORE_SOURCES and GUI_SOURCES:

• src/core/WfmDemodulator.cpp
• src/core/WaveOutWriter.cpp
• src/core/WaveInReader.cpp
• src/gui/WfmDeviceDialog.cpp

All of them #include <windows.h> and <mmsystem.h> and call waveInOpen / waveOutOpen / etc. Those headers don't exist on macOS or Linux, so check-macos and the Linux build job will fail at the first #include.

Fix: wrap those entries in if(WIN32) ... endif() like the project already does elsewhere — see the existing pattern at CMakeLists.txt:1190-1199 where winmm/CoreAudio/ALSA are selected per platform. Something like:

if(WIN32)
    list(APPEND CORE_SOURCES
        src/core/WfmDemodulator.cpp
        src/core/WaveOutWriter.cpp
        src/core/WaveInReader.cpp
    )
    list(APPEND GUI_SOURCES
        src/gui/WfmDeviceDialog.cpp
    )
endif()

Good news — winmm is already linked on Windows at CMakeLists.txt:1192, so the waveIn* / waveOut* symbols should resolve once the build reaches link time.

A few smaller cleanups while you're in there

1. Hardcoded debug log path — WaveOutWriter.cpp, WaveInReader.cpp, and WfmDemodulator.cpp all write to C:/Users/reigc/wfm_debug.txt. That's your local machine; on any other Windows box QFile::open will silently fail (or, worse, succeed). Worth removing the wfmLog helper or routing through qDebug() / qCDebug() before merge.

2. Stray file in .gitignore — Resumen para empezar de nuevo.txt looks like a personal scratch file. Safe to drop that .gitignore entry from the PR.

3. check-windows is also failing — I couldn't pull the raw compiler output from here, but with the platform guards in place the Windows job will run further and any remaining error will be much easier to read. If it still fails after that, ping me with the log and I'll take another look.

Really appreciate the contribution — WFM demod via DAX IQ is a great addition. Once the sources are Win32-gated this should turn green on all three platforms.

_{🤖 aethersdr-agent · cost: $5.2795 · model: claude-opus-4-7}

Hi @ea5wa — thank you for the work, the on-air verification against 9k6 satellite telemetry through hs-soundmodem is exactly the kind of end-to-end test that earns review time.

I want to redirect on architecture before you invest more cycles. @aethersdr-agent already nailed the immediate CI diagnosis (Win32-only sources added to cross-platform CMake lists, three hardcoded `C:/Users/reigc/wfm_debug.txt` paths, stray `.gitignore` entry), but adding `if(WIN32)` gates isn't the right fix here.

The project rule

Every feature in AetherSDR should be cross-platform unless it specifically addresses a platform-only problem. AetherSDR exists to give Linux/macOS operators feature-parity with the Windows-only SmartSDR client — see AGENTS.md "Project Goal". A Windows-only feature inside AetherSDR re-creates the gap the project is trying to close.

Examples already in tree that ARE genuinely platform-specific because they address a platform-only problem: `MacNRFilter` (macOS-only neural NR using Apple's framework), the Windows WASAPI 24k-artifact policy (#2120), the macOS Bluetooth-HFP guard (#2615). They're each solving a platform-shaped bug, not just convenient to ship that way.

Why WFM doesn't qualify

Looking at the diff, the Win32 dependency is only in the audio output stage — `WaveOutWriter` wraps `waveOut` to deliver Int16 PCM to a Virtual Audio Cable. The actual demodulation chain in `WfmDemodulator` is pure math (atan2 discriminator + 75 µs de-emphasis + volume scaling) and would run identically on every platform.

The use case ("route demodulated audio into another application") also exists on every platform:

• Windows — Virtual Audio Cable, VB-Audio Cable
• macOS — BlackHole, Loopback, Soundflower
• Linux — PipeWire null-sinks, JACK, PulseAudio loopback modules, ALSA `snd-aloop`

All three loopback drivers register as ordinary audio output devices, so they'd show up in `QMediaDevices::audioOutputs()` with no special handling.

The architectural change

Replace `WaveOutWriter` with `QAudioSink`. It's part of Qt 6 Multimedia (already a dependency), works on Win32/CoreAudio/PipeWire/PulseAudio/ALSA, and the existing AetherSDR audio stack already routes through it — see `AudioEngine::startRxStream()` for the canonical pattern (opens a `QAudioSink` on a selected `QAudioDevice` at a negotiated format, writes float PCM to its `QIODevice*` from a timer).

You'll also want to replace `WfmDeviceDialog` with a standard Qt device picker — enumerate `QMediaDevices::audioOutputs()`, store the selected ID in `AppSettings` under a `"WFM"` nested-JSON root (Constitution Principle V — flat `"WfmAudioDevice"` keys are non-compliant; see `docs/audio-pipeline.md` or `CwDecodeSettings.h` for the pattern).

Once that's done:

• The four new `.cpp` files can go back into the unconditional `CORE_SOURCES` / `GUI_SOURCES` lists (no platform guard needed because nothing is Win32-only anymore).
• The WFM toggle button can stay visible on every platform.
• macOS and Linux operators get the same feature you've already validated on Windows.
• The hardcoded debug path goes away because you can switch to `qCDebug(lcAudio)` (project's logging category in `src/core/LogManager.h`).

Bonus: there's a foundation coming that fits this exactly

We just landed #3397 (and #3398 is in flight) — the consolidated audio sink/rate negotiation factory (`AudioFormatNegotiator` + `AudioDeviceNegotiator`, tracked in #3306). Once #3398 lands, WFM's output stage can go through that helper for free: it'll handle the 48 kHz preferred rate, the Int16-only-device fallback for VAC clones, and the per-OS preferred-format ladder without you having to encode any of it in WFM-specific code.

You don't have to wait for #3398 to refactor — `QAudioSink` direct works today. But if the wait is easier, that's also fine.

What I'd like next

• Refactor the output stage to `QAudioSink`.
• Drop the `.gitignore` line and the hardcoded debug path.
• Move the device-pick + "remember choice" settings into nested JSON (Principle V).
• Verify on at least one of {macOS, Linux} in addition to your existing Windows pass. If you don't have a non-Windows test box, I can run the Linux side once the refactor is up.

I'm happy to point at more of the existing patterns or sketch the `QAudioSink` wrapper if it'd help. The demod chain itself is great work — I just want it usable for everyone.

73, KK7GWY

Agree with @ten9876's direction here — the QAudioSink refactor is the right call over my earlier Win32-gating note (disregard that bit). Making the output stage cross-platform is the whole point: macOS and Linux operators get the feature too, and it slots straight into the existing audio stack. The demod chain itself is genuinely good work — once the output moves to QAudioSink + nested-JSON settings per Principle V, this'll be in solid shape. 73

Thanks for the detailed architectural guidance

Refactor complete in ef5b0e9:

WaveOutWriter — rewritten using QAudioSink (Qt6 Multimedia). Device is now selected by QAudioDevice::id() (the persistent opaque byte string from QMediaDevices::audioOutputs()) rather than a human-readable name fragment, so it survives device renames and reordering.

WaveInReader — removed entirely. The VITA-49 DaxIqModel::iqSamplesReady path is cross-platform and already delivers IQ samples on all platforms. The Win32 DAX waveIn path was redundant.

WfmDeviceDialog — rewritten using QMediaDevices::audioOutputs() and QAudioDevice::id(). Works on Windows, macOS, and Linux. Virtual loopback devices (Hi-Fi Cable, BlackHole, PipeWire null-sink) show up automatically as ordinary output devices.

WfmSettings — new header-only settings class following Principle V: all WFM settings live under a single "WFM" nested-JSON key in AppSettings. Includes migrateLegacy() to transparently migrate the old flat "WfmAudioDevice" key for existing testers.

WfmDemodulator — simplified to the single cross-platform VITA-49 path; all debug output now goes through qCDebug(lcAudio).

CMakeLists — WFM sources are back in the unconditional CORE_SOURCES / GUI_SOURCES lists. No if(WIN32) guard needed.

I don't have a macOS or Linux test box available right now — happy to re-verify on Windows once CI confirms the other platforms are green. If you're able to run the Linux side I'd appreciate it.

Compiled without errors. Tomorrow I'll test it on windows 11

73 de EA5WA Juan Carlos

I'm having problems to have a good 48kHz output with QAudioSink. At the moment, no valid output for UZ7HO soundmodem......

73 de EA5WA Juan Carlos

Now is working....
implementation with a fully cross-platform Qt6 Multimedia stack.

Signal path

FlexRadio hardware
  → SmartSDR DAX IQ (exposes IQ as a Windows audio capture device)
  → QAudioSource  (reads "DAX IQ RX 1" as if it were a microphone)
  → FM phase-difference discriminator
  → WfmRingBuffer (thread-safe circular buffer, 1 s capacity)
  → QAudioSink    (timer-driven push mode, 10 ms ticks)
  → HiFi Cable / VAC → hs-soundmodem

Key design decisions

• QAudioSource instead of VITA-49: SmartSDR DAX takes exclusive ownership of the IQ stream when SmartSDR is running. Issuing a second stream create type=dax_iq command put the stream into a Busy state and no data arrived. Reading the DAX capture audio device directly (same approach SmartSDR's own DAX client uses) avoids the conflict entirely.
• Timer-driven push mode: WASAPI pull mode (where the sink calls readData()) proved unreliable on Windows. A 10 ms QTimer calls bytesFree() and pushes exactly that many bytes per tick, letting the driver control the effective data rate without blocking or flooding the sink.
• Float32 PCM output: Int16 is not universally supported by virtual audio cables (HiFi Cable, VB-Audio). Float32 is accepted by all Qt backends (WASAPI, CoreAudio, PipeWire/PulseAudio) and by the VAC drivers tested.
• Linux/macOS fallback: If no DAX audio capture device is found, the demodulator falls back to the VITA-49 DaxIqModel::iqSamplesReady signal path, keeping the feature usable without SmartSDR DAX.
• Spectral inversion correction: SmartSDR DAX delivers IQ with upper-sideband mixing, which inverts the baseband spectrum. The discriminator output is negated (-atan2(...)) to compensate. This needs verification with a live satellite pass — if the polarity turns out to be already correct, the negation should be reverted.

Files changed

Known pending item

Waterfall polarity seems inverted, I have to confirm it with a real satellite pass tomorrow. A UI toggle (flip polarity without recompiling) may be added as a follow-up if needed.

73 de EA5WA Juan Carlos

From my side is OK.

Could anyone check it?

73 de EA5WA Juan Carlos

Summary of changes

src/core/WfmDsp.h + src/core/WfmDsp.cpp (new, 288 lines)
The heart of the change: the equivalent of SkyRoof's Slicer as a pure C++ class — no Qt audio, no I/O, cold-testable. Chain: phase-continuous NCO → twin r8brain resamplers (I/Q phase-matched) down to exactly 48000 Hz, flat to 0.95·Nyquist (the same useful-bandwidth ratio SkyRoof uses) → atan2 discriminator → 95-tap linear-phase FIR, fc = 20 kHz. Deliberately no de-emphasis, squelch or EQ in the data path. Why: everything HS-SoundModem needs was scattered or missing; isolating it lets the DSP be proven correct with unit tests instead of being tuned by eye against the waterfall.

src/core/WfmDemodulator.h + .cpp (rewritten, −328/+~250)
Now only the device plumbing, delegating DSP to WfmDsp. Two functional changes:

1. Capture the DAX IQ endpoint at its native rate. It previously forced 48 kHz, so Windows silently resampled whatever DAX actually delivered (24/48/96/192 k) — a 24 k stream upsampled by the OS mixer has no energy above ±12 kHz: that was the real origin of the "10–13.5 kHz notch" that earlier commits tried to patch with EQ compensators (now removed).
2. Idle-input gate (−80 dBFS with hysteresis): the atan2 discriminator is amplitude-invariant and turned the dither of a stopped stream into full-scale noise; it now writes silence while keeping the DSP state and output cadence alive.

src/core/WaveOutWriter.cpp (12 lines)
pushData now zero-fills whenever the ring buffer falls short — the exact policy of SkyRoof's RingBuffer.ReadBytes. Why: when the IQ stream stopped, no writes arrived and HiFi Cable looped its internal buffer — the periodic pattern seen on the soundmodem waterfall. A virtual audio cable must never starve; it now receives continuous silence (verified: waterfall goes black on IQ stop).

src/gui/MainWindow.cpp (39 lines, in activateWFM/deactivateWFM)
SkyRoof's tracking policy (FrequencyWidget): centre the pan on the slice once at WFM activation; from then on Doppler rides the demodulator's NCO (setFreqOffsetHz, phase-continuous retune, no clicks). The pan is recentred only if the slice would leave the usable IQ window (±14.8 kHz), which is rare at 70 cm. Why: without the correction, a Doppler offset beyond ~8 kHz produced a DC term that hard-clipped the demodulated audio entirely — half the pass lost.

src/core/RigctlProtocol.cpp + src/core/TciProtocol.cpp (35 lines)
In-span retunes use autopan=0; only band changes recentre. Why: Doppler steps from SatPC32 every few seconds kept yanking the pan. This dovetails with the above: fixed pan + NCO.

tests/wfm_dsp_test.cpp (new) + CMakeLists.txt (+17)
13 checks pinning the contract: the NCO removes the DC term exactly (1.25 → 0.000) with the FSK tone intact (measured 0.375 = theoretical), 96 k→48 k preserves amplitude/phase, chunked processing is bit-identical to single-shot (no clicks at block edges), and the offset-window policy values. CMake adds WfmDsp.cpp to the main target and registers wfm_dsp_test with CTest, following the tx_mic_channel_normalizer_test pattern.

Verified on air: flat soundmodem waterfall (no EQ compensators), clean silence on IQ stop, G3RUH 9600 bd decoding through a full Doppler swing.

73 de EA5WA Juan Carlos

Working perfectly:

Update on the red check-windows: the CI fix is already on main — Jeremy landed #3509 before I got mine up (closed #3525 as duplicate). So no waiting needed: just merge/rebase onto current main and check-windows should go green, along with picking up the VfoWidget conflict resolution from my review above.

@ea5wa — I pushed a small maintainer fix-up to your branch (c811b72) addressing the latest agent review. Triage of the five items:

Fixed in the push:

• build(deps): Bump actions/checkout from 4 to 6 #3 (dead "WFM" branch) — removed. "WFM" is never an entry in m_modeCombo, so the branch was unreachable; selecting a real mode already tears down the overlay via the unconditional wfmActivated(false) emit, which is now the only path. No behaviour change.
• APD status indicators need refinement #4 (toggle desync) — fixed. The RxApplet and VfoWidget WFM buttons were independent, so activating from one left the other stale and a mode-change teardown left a button green while the demod was off. MainWindow now pushes the authoritative demod state onto both surfaces (reflectWfmButtons()) on every activate/deactivate — including the dialog-cancel and start-failure paths, so the button that triggered an aborted activation un-sticks. The setWfmActive() setters self-gate on their slice and block the button signal, so there's no feedback loop.
• Implement XVTR band sub-menu #5 (while (true)) — collapsed to a plain block; there was no retry path so it always returned on the first pass.

Left as-is (with rationale):

• PHONE applet: DEXP commands rejected on firmware v1.4.0.0 #1 (CAT autopan=0 reaches beyond WFM) — kept in this PR. It's coupled to WFM's in-span Doppler retune, the logic is correct (null pan → falls back to the old recenter path, preserving rigctld set_freq shifts panadapter center instead of tuning the slice #536), and it's already explicitly documented under "Behaviour change (CAT)" in CHANGELOG.md — which is the bot's stated minimum-acceptable resolution. Flagging it here for maintainer visibility: this is a conscious cross-cutting change to set_freq/VFO behaviour for all rigctl/TCI users, not a rider.
• build(deps): Bump github/codeql-action from 3 to 4 #2 (HIDAPI guard) — kept. The #ifdef HAVE_HIDAPI around m_hidEncoder->isTMate2() is correct and needed for non-HIDAPI builds (m_hidEncoder is HIDAPI-gated); the branch is already rebased onto current main. Dropping it would only tidy the diff at the cost of a non-HIDAPI build break, so it stays.

One thing to confirm (not changed): with the dead branch gone, a mode change on the WFM slice (e.g. FM↔NFM) still tears down the demod via wfmActivated(false). That's defensible (a real demod mode supersedes the software overlay) and matches VfoWidget's existing leave-FM behaviour — just confirm it's what you intend.

Built clean locally and the full suite (34 tests, incl. wfm_dsp_test) passes. The earlier findings (Win32-gated sources, hardcoded debug path, stray .gitignore) were already resolved by your QAudioSink rewrite. Nice work on this — the headless WfmDsp + real unit tests are the strong part. 73