Changelog

All notable changes to Open Audio Analyzer are recorded here. The format is defined in CLAUDE.md; the short version is that 📐 Measurement always comes first, because a change to a reported number can invalidate a decision somebody already made about a master.

This project follows Semantic Versioning.

Unreleased

0.10.0 — 2026-08-22

✨ Added

• A skin editor. Settings → Appearance → Edit skin opens all thirteen colour roles with a picker on each, and the canvas behind the panel repaints as you drag — as does a tablet mirroring the session. Every role prints its contrast ratio against the surface it is read on, and the ones below the floor that role is held to are marked with the reason; a low-contrast palette still saves, because refusing to save what is already on screen is worse than a warning. Skins remain plain JSON files and hand-editing them is unchanged.
• Skins can be deleted from the editor, which previously meant finding the file.
• Reset, beside Edit on Settings → Meters → Delivery target, deletes every target you have saved and puts the built-in six back. It asks first, in a dialog naming what goes, and afterwards says how many files it removed. Targets you wrote by hand are gone from the disk, so this is also how a correction to a built-in is undone without going looking for its JSON file.
• The VU Meter marks where the needle reached. A mark rides the scale at the loudest point of the last second and a half and then falls back to the needle, because a 300 ms movement is slow by design and the loudest moment of a phrase is over before the needle has finished describing it. It is the same quantity the needle draws, held rather than measured a second time.
• The VU Meter prints its reading and its reference along the bottom of the face: the deflection in VU on the left, and which dBFS level reads as 0 VU on the right. The second one is the genuinely confusing thing about a VU meter and the face now says it out loud. A module too small for either drops it and keeps the dial.
• The Spectrogram and the Oscilloscope can be drawn in full colour. Colour: Full RGB in either module’s menu replaces the skin’s one-hue ramp, and what it paints is the quantity that module’s axes leave uncoloured. On the Spectrogram that is the level: the spectrogram rainbow, indigo through green and yellow to red and white, which separates far more steps of level than one hue can. On the Oscilloscope it is the band balance, because a waveform has no frequency axis at all: red, green and blue are its bass, its mids and its highs, split at 200 Hz and 2 kHz, so a kick is red, a hat is blue and a full-spectrum hit is white, the way DJ software has coloured audio for twenty years. That balance is taken in decibels against the loudest of the three bands rather than in power — bass carries an order of magnitude more power than air does, and a power mix draws every piece of music red — and it is recorded with the column, so the picture holds its colours instead of being repainted by whatever is playing now. A column whose source published no spectrum keeps the accent rather than being handed a colour that would claim a balance nobody measured. Off by default in both, because a rainbow reads as more precise than the measurement behind it, it collapses for the eight percent of men who cannot separate red from green, and on the Spectrogram it brings its own dark ground into a light skin. Nothing measured changes at either setting: no reading, no report and no byte on the wire moves.

⚡ Changed

• A menu marks the value it holds by recessing that row, not by brightening it. Every dropdown in the application — the delivery target, the signal source, and a module’s dozen settings — used to print its current value as the lightest row in the menu, which put the emphasis on the one choice pressing cannot change and left the options you can still take looking greyed out. The chosen row now sits in a darker well and the alternatives sit on the menu’s own surface, which is how selection reads everywhere else in the interface. Nothing about what the menus contain or choose changed.
• The signal source in the status bar is a bordered chip, the same shape the delivery target beside it has always had. It is the menu people open most and it was drawn as a dot and a bare word in a row of bordered controls, which reads as a caption rather than as something you can click. The dot stays, inside the chip: bright while something is being metered, dim on Silence.
• The OAA wordmark is gone from the status bar. The window carries the application’s name; the bar carries what changes while you work, and three capitals that never change were the one item in it saying nothing about the signal. Every width at which the bar drops an item moved with it — the controls at the narrow end now leave 25 to 40 px earlier than they did, because a bordered chip cannot give back the width a bare word could. All of them still have keyboard shortcuts except the three remote controls, which is unchanged.
• The sample rate no longer prints flush against the DAW’s playhead. The gap between the source group and the readings is a fixed one now rather than whatever the window had left over, so it is there at every width instead of only on a wide one.
• The VU Meter has a face. The scale is a band along the rim rather than a hairline arc, red above 0 VU over a tinted ring, and the needle is a taper under a machined hub — so how far up the scale the programme is can be read from the rim before the needle is read at all. Nothing about the movement, the ballistics or the reading changed.
• The Loudness Distribution fits its loudness axis to the programme instead of always drawing −60 to 0 LUFS. A mastered programme’s gated distribution occupies eight to fifteen of those sixty decibels, so the picture the LRA reading is taken from was squeezed into a fifth of the module with the rest left blank; the axis now holds every occupied bin, the gated range and the delivery target, rounds outwards to whole ticks, and stays where it is until the distribution grows past it. No reading changes — the same bins are drawn against a shorter stretch of the same scale. Scale: Full range in the module’s menu restores the published axis, which is the one to pick when two distributions are being compared side by side.
• Open Audio Analyzer has a new logo, and every icon it ships follows it. A white waveform on a teal ramp, replacing the four teal bars on graphite. The app icon, the installers’ icons, the Android launcher icon, the layered icon macOS and iOS render, the favicon, the documentation site’s header, the README and the social card all change together. The tile is also a rounder shape than it was, because the one macOS 26 and iOS 26 draw is rounder than the one every version before them drew. Nothing about what the application measures or displays is affected.
• The icon reads less well at 16 px than the one it replaces, and that is a real cost rather than an oversight. Four bars at four heights survived being two pixels wide; a waveform with nine excursions across twelve pixels does not, and no stroke weight rescues it. The line is held at one device pixel so that it stays white instead of smearing to grey, and the sizes the icon is now designed around are 32 px and up. Where this shows is the Windows Explorer list view and a Linux panel at its smallest setting.
• Precision Instrument and Daylight can no longer be replaced. A skin file naming one of their two ids used to shadow the built-in; it is now ignored, and the skin editor offers Save as new rather than an in-place save. The two shipped skins are what proves the thirteen colour roles are semantic rather than aliases for particular colours, and a reference point a file on disk can quietly redefine is not one. If you had such a file it is inert now — rename its id and it comes back as a skin of its own. Delivery targets are unaffected and still shadow their built-ins, because a target is a claim about somebody else’s published specification and has to be correctable without a release.

🐛 Fixed

• The VU Meter no longer draws a needle for a reading nobody measured. The per-channel arrays are NaN while a remote display’s link to the host is quiet, and the dial rested at the bottom of its scale with exactly the confidence it reads a quiet passage with; the face now shows no needle and the reading is an em dash. Only a tablet display could reach this.
• The VU Meter’s scale numbers no longer overprint each other on a small module: −5 and −3 sit close together on a face that crowds towards the bottom in voltage, and at the narrowest widths they collided into one smear. The face now sheds the numbers it has no room for, keeping 0 and both ends.
• A panel that scrolls draws one scrollbar rather than two. Settings, and every other panel tall enough to scroll, carried the skin’s own thumb on its right edge and a second, wider grey one immediately inside it, which faded in whenever the panel was scrolled. The second was supplied by Flutter, which decorates a scrolling region on macOS, Windows and Linux without being asked; the panel now declines it. Tablets never showed it.
• The Super Meter’s target tick sits in its slot again. The slot is a surround laid down under the tick so that its contrast is identical on all three rings, whatever each ring is currently drawing; it was being stroked at exactly the tick’s own width, so the tick covered it completely and was read against the arc after all. It stands 1.5 px out either side now, which is what the constant describing it always said. The tick itself is unchanged.

🚧 Internal

• The skin editor’s suite waits for the effect of a save rather than for the file. Its helper stopped as soon as the skin’s file existed, which is one event-loop turn before the editor is told the write finished — so the four assertions about what saving then selects raced the disk, and won on a Mac and lost on the Linux runner every time.
• flutter analyze no longer walks website/tools/. The two Flutter web targets that render the real modules for the site now share the mock MeterSource as a package rather than each holding a copy, and a package:oaa_mock import needs package resolution where the relative one it replaced needed none — so on any machine that had not run pub get inside each of them, which is every CI runner, the repository’s analyze gate failed on eight unresolved references in code that no gate builds. They are excluded in analysis_options.yaml and analysed by their own.
• The remote display’s skin frame is rate-limited to one every 150 ms, with the last value always delivered. Dragging a colour in the new editor produces a palette per pointer move, and those frames are queued rather than dropped — unthrottled they grew that queue for as long as the pointer was down.
• The artwork and the program that renders it swapped places. assets/brand/oaa-logo.svg is now the drawing, hand-edited, and packaging/icon/make_icons.dart reads it; previously the mark was four rectangles described as numbers in that program and every vector in the repository was a hand-copied transcription of them, kept in step by a rule saying the mark changed there first and was brought across afterwards. That rule does not scale to a path with three hundred control points. The program now also writes packaging/icon/oaa.svg, the rest of assets/brand/ and the website’s icons, which removes the last four files anybody kept by hand.
• make_icons.dart grew a path rasteriser — an SVG path parser, a cubic flattener and a scanline fill — because the mark is a stroked cubic path and the old renderer only knew rectangles. Still no dependencies. The stroke is drawn as the union of a rectangle per segment and a disc per vertex rather than by offsetting the outline, which is the same shape and is not a research problem.
• The tile’s corner is measured rather than remembered. Three macOS 26 system icons were rendered through NSWorkspace, thresholded to a silhouette and fitted: a corner of a third of the side on a superellipse of exponent 2.7, where the shape Apple drew from iOS 7 to iOS 18 was 22.37% on an exponent near 4. The first cut of this release used the remembered numbers and looked visibly square next to everything else in the Dock.
• The documentation site serves its favicon as a file instead of inlining it into every page as a data URI. The mark was a few hundred bytes when that choice was made and is three kilobytes now, which across twenty pages was a hundred kilobytes of the same icon.

0.9.0 — 2026-08-22

📐 Measurement

• 7.1 material reads 0.35 LU quieter, and correctly. The +1.5 dB surround weight was applied to every channel past the LFE, so a 7.1 stream had it on the rear pair as well as the side pair. BS.1770 gives it to the surround pair only — Report ITU-R BS.2217’s channel table states 7.1 as 1.00 / 1.00 / 1.00 / N/A / 1.41 / 1.41 / 1.00 / 1.00 — and the ITU’s own two 7.1 compliance files read 0.35 LU high against a ±0.1 tolerance until this was fixed; they now read −23.000 and −24.000 exactly. Only 7.1 and wider is affected: mono, stereo, quad, 5.0 and 5.1 never reached the arm that was wrong, and every reading from them is unchanged. Anything measured from a 7.1 master — its LUFS-I, LUFS-M, LUFS-S, LRA and the delivery verdict drawn from them — was 0.35 LU too loud and is worth re-measuring.
• Max M and Max S are found on a 10 ms grid, where momentary and short-term loudness previously advanced only every 100 ms. The official EBU test vectors caught this: Tech 3341 tests 13 and 14 slide a 400 ms tone — exactly one momentary window long — through twenty files in 20 ms steps and require Max M within ±0.1 LU of −23.0 every time. On a 100 ms grid a tone offset by anything else never lies inside one window whole, so sixteen of the twenty read low, by up to 0.45 LU, and test 14 by 0.70. Every Max M and Max S therefore rises rather than falls: on the EBU’s two authentic programme segments Max M rises by 0.16 and 0.11 LU and Max S by 0.02 and 0.01, and a transient that fell between the old grid points can rise by up to 0.7 LU. LUFS-I, LRA and its percentiles, TP Max and Peak Max are unchanged — identical to three decimals across all 70 files of the test set — because both gating windows are still filed every 100 ms, which is the 75% overlap BS.1770 asks for. Re-read a delivery decision that turned on Max M or Max S against a ceiling; one that turned on the integrated numbers stands.
• The spectrum analyser draws its bands tilted, at 4.5 dB per octave by default. The curve is rotated about 1 kHz, so 20 Hz is drawn 25.4 dB lower than it measures and 20 kHz 19.4 dB higher — 44.8 dB between the ends of the range, where previously the drawn level was the measured level at every frequency. It is a view: the offset is added per band at the moment of drawing, and Spectrum and Spectrum peak are published and sent over the wire exactly as the engine measured them, so nothing needs re-measuring. What changes is that a level read off the analyser’s own dB scale is only true at 1 kHz now, which is why the module prints the tilt it is drawing at and why Tilt: 0 dB/oct — where the scale is true everywhere — prints nothing.
• The analyser’s peak-hold line holds the drawn curve rather than the raw bands. It is the highest the curve has been, held for a second and a half and then let down at 12 dB a second, which is the schedule the engine’s own per-band hold already followed. Two consequences: the line moves with the curve instead of jumping to a peak the curve is still easing towards, and on Response: Slow it reads lower than it used to — by as much as the pole smoothed the transient away — because the curve it is holding never went there. Response: Fast holds every published frame and is the setting to find a click with. Spectrum peak itself is unchanged.
• The histogram draws both of its bands averaged over a second, where it previously drew every 100 ms column exactly as measured. Smoothing in its menu chooses Off — the old picture, byte for byte — Light at 0.5 s, Normal at 1.0 s or Broad at 2.0 s, and Normal is the default. The window is centred rather than trailing, so nothing moves along the time axis; what it costs is height on a short event. A transient the momentary band reached for a single column now reads lower by roughly the distance between that column and its neighbours — a 6 LU spike over a steady body reads about 1 LU above it on Normal — and the newest column at the right edge lags the live meters by up to half the window, settling as it ages. Nothing measured changes: the ring holds the columns as they arrived, the smoothing is applied on the way out every frame, and reports, the wire protocol and every other loudness module are untouched. Choose Off to find the loudest 100 ms in a programme.
• The loudness distribution’s fill no longer double-draws itself. Each of the 120 published bins was stroked half a pixel wider than its own spacing to keep butt caps from leaving seams, so every overlap was composited twice through the translucent gradient — a brighter line at all 119 bin boundaries, and on a default-sized module nearly a fifth of the fill drawn at double alpha. The bins are unchanged and none of them moves; what changes is that the drawn area is now a shape rather than a row of overlapping lines, and a column covering more than one bin takes the loudest of them rather than blending them, the way the engine already maps transform bins into spectrum bands.

✨ Added

• The oscilloscope locks to the DAW. Sync: Tempo makes the width a musical division — 4 bars down to 1/32, straight, triplet or dotted — and puts every sample in the column its position in the bar falls in, so each pass overwrites the last in place and a kick is drawn in the same spot every bar. The graticule divides into beats rather than tenths while it is locked, and the corner says which division is on screen. It uses the host’s own playhead, tempo and time signature, which the plugin already forwards: MIDI clock carries neither a bar position nor the accuracy, and is not involved. A source with no playhead — a sound card, or a DAW that reports none — draws the free window and labels it as the free window, because a display that said 1 bar over something else would be worse than one that is not synced.
• The spectrum analyser has a tilt, Tilt in its menu: 0, 1.5, 3, 4.5 or 6 dB per octave, rotated about 1 kHz. Programme material falls with frequency at something like 3 to 4.5 dB an octave, so an untilted analyser draws every mix ever made as the same ramp — bottom octaves against the ceiling, top octaves crushed into the floor — and spends most of its height on the one part of the picture that carries no information. Tilted, a mix is roughly horizontal and what is left is the deviation. The analyser prints the tilt it is drawing at, because a dB scale quietly rotated across its width is worse than no scale at all.
• The oscilloscope draws a stereo signal either as two lanes or with both channels around one centre line, from Stereo in its menu. Overlaid, the two are told apart by weight rather than by colour, and the trace gets the whole height of the module — which is the arrangement that shows what the channels are doing differently.
• The oscilloscope can be triggered by a transient. Trigger: Transient makes the display wait, armed, until the signal rises through a level you set, draw forward across the whole width once from that sample, and hold what it caught until the next crossing. It works at every time base — the roll above 200 ms is replaced by the sweep rather than left in place — so the attack of one kick can be looked at instead of a picture that lands somewhere different every pass. Trigger: Off, which is the default and what the module did before, is unchanged: a rising zero crossing below 200 ms, a rolling display above it, and always something on screen. A threshold nothing reaches leaves the last capture where it is, which is the mode working rather than failing. Trigger stays in the module’s menu under Sync: Tempo, greyed rather than dropped: a bar-locked window is placed by the bar line and has no use for it, and a row that vanishes is a row somebody hunts for. Grid is now greyed under Sync: Free for the same reason instead of being dropped — there is no triplet of a millisecond.
• The oscilloscope’s height and trigger threshold are sliders on the module, HEIGHT and THRESHOLD in a strip along the bottom of the plot, and the threshold is drawn across the lane at the height it is set to. Both are numbers over a wide range that are chosen by watching the waveform while they move, and a menu that closes over the waveform on every step cannot be used for that. The height goes from 1x to 32x, continuously rather than in named steps — a reverb tail thirty decibels under full scale is a flat line at 1x, and the setting that fits the material is rarely a round multiple. The strip is dropped on a module too short to spare the room, like the graticule and the lane letters, and is not drawn on a remote display, which has no layout to write to. Nothing measured changes: a sample is still marked as clipped only if it reached full scale, and a trace that runs off its lane is a trace that is zoomed.
• AUTO beside the threshold takes it from the loudest transient. It follows the largest positive excursion of the mid signal over the last two to four seconds — the quantity the trigger itself compares against — and sets the threshold six decibels under it, which is half the amplitude and so still inside the attack: the sweep starts before the transient rather than on top of it, which is what a threshold set exactly at the peak would do. The slider shows where the level has got to and stops answering while the box is checked; unchecking it keeps the number Auto found rather than snapping back to whatever was dragged before. Silence moves nothing — a passage nobody played is not a measurement, and a threshold dropped to the floor would arm the trigger on the noise underneath it.
• The histogram has a Smoothing setting — Off, Light, Normal or Broad. See the Measurement note above for what each one does to the picture and why the window is centred rather than trailing.
• The loudness distribution prints LRA. The module drew the picture behind the number for two phases without ever showing the number, so reading one off it meant putting a Number Box beside it. It is now on the bracket across the top, which is the distance the reading is.

⚡ Changed

• macOS 14.2 is now the stated requirement for the application, the VST3 and the Audio Unit, where the application claimed 10.15 and the plug-ins 11.0. Nothing is lost that worked: neither could load below 14.2 at all — see the Fixed entry. The installer refuses the volume under 14.2 rather than gating its plug-in rows, because there is no longer a version that can run one component and not another.
• Nothing crosses the loudness distribution’s plot except the target. The 10th and 95th percentiles were two full-height lines at the reading weight, and on steady material they and the dashed target line piled up within a few pixels of each other over the bars. The percentiles are now a bracket across the top strip carrying the LRA reading, with short end marks at the top and at the axis and the gated range shaded between them — so the distribution is the only thing in the plot, and the target, which is the one line the user chose, is the only line through it. The 10% and 95% labels are gone: the bracket’s ends are those percentiles and the number on it is the distance between them.
• The loudness distribution’s annotations are painted over the bars rather than under them. The percentile marks and their labels were drawn before the fill, so the translucent gradient was composited on top of the part of the module meant to be read first.
• The settings panel says what the source list actually offers. The note under the device picker still read “input devices only” and told macOS users to install BlackHole, sitting directly beneath a picker that has offered System Output since 0.8.0 — so the one sentence read at the moment of choosing said the feature was not there. It now names what each platform does, and says that macOS asks permission and reads silence if it is declined.
• Upgrading from 0.5.0 or earlier is documented as revoking every macOS permission, not just Local Network. The bundle identifier moved in 0.6.0 and macOS keys permissions to the identifier, so Microphone, Camera and System Audio Recording were revoked as well. System Audio Recording is the one worth checking first, because Apple’s refusal of it is silent — see the install page.
• The page switcher on a remote display’s link bar sits at the right-hand end of the bar, beside Disconnect, where it used to sit between the host’s name and its playhead readout. It no longer moves when the host gains or loses a DAW, or reports a tempo where another reports only a clock.
• The iPad display needs iPadOS 15 or later, where it previously needed 13. Apple stops accepting uploads built against anything below 15 in Spring 2027 and warns on every one until then, so the floor moves now rather than in the release that would otherwise have been refused. No iPad loses the display by this. iPadOS 13, 14 and 15 support exactly the same devices — every model back to the iPad Air 2 and the mini 4 — so what this excludes is an iPad that has not been updated, not an iPad that cannot be.
• Disconnecting a remote display lands back on the view it was opened from, rather than leaving the host picker on the screen. Leaving a display used to put it in its no-host state, and that state is the panel that asks which machine to attach to — so the one control marked as the way out asked a question instead of answering one. A display with nothing behind it, which is a tablet that opened straight into one, still lands on the picker, because there is nowhere else for it to go.
• The phase scope’s correlation bar and the super meter’s three rings have rounded ends. On the bar the fill is clipped to the same corners, so ±1 fills the rounded tip rather than sitting square over it. On the gauge it is the scale that is rounded, at both extremes of the sweep: a reading’s own moving end stays square, because a round tip is half a ring of ink standing past the number it points at, which on that scale is most of a decibel nobody measured. The M, S and I names step a little further round the open end to keep their clearance from ink the rings did not have before.
• Everything drawn against the delivery target is red above it. The LUFS meter’s momentary and short-term bars and the super meter’s three arcs are now cut at the target and drawn in over past it, where they were one colour from the floor to the reading. The histogram and the loudness distribution already split at the target and already used over, so their picture is unchanged. One rule in four modules, and red is the single mark for past the number you set. The cut is a clip at the target rather than a verdict on the whole shape, so what it shows is how much of the reading is over — a momentary peak 3 LU above a −14 target is three quarters grey with a red cap, not a red bar.
• The histogram’s outline turns red past the delivery target, where it stayed the accent colour across the whole programme. The fill under it already split there, so the line the eye lands on ran one colour straight through the boundary and disagreed with the area it bounds. Cut at the target like everything else, on the same pixel row as the fill.
• The super meter’s innermost ring no longer turns entirely red when the mix is over its target. The ring takes its verdict’s colour for its whole length — which is what makes an in-spec reading a green ring rather than a green tip — and once over-target became a red verdict that meant the arc ran red from the bottom of the scale to the reading, with its own target tick stranded in the middle of it and the same colour on both sides. At −11 against a −14 target, three quarters of the arc claimed to be over when a quarter of it was. It is now drawn like the other two rings: neutral up to the target, red past it, so the red segment is the size of the miss. In spec it is still green end to end, and the centre readout is red either way.
• The super meter’s three target ticks read as one mark, and are drawn at 3 px rather than 2. The same grey on three rings is not the same appearance: the backdrop at the target is the track on a ring that has not reached it, that ring’s fill on one that has, and red on the far side of the line — measured at L* 24, 39 and 51 in one ordinary frame, so a single grey stood +36, +21 and +9 above what surrounded it and read as three different greys. No choice of shade fixes that, because the backdrop moves with the audio, so each tick is now laid in a slot of the meter’s own track colour — the mark reads as cut back to the empty scale, and a ring that has not reached the target shows no notch at all, because there is nothing there to cut. Local contrast is +36 on all three rings at every reading and under every skin. The extra weight is because the mark is radial: antialiasing spreads it across two pixel columns where an axis-aligned rule lands on whole ones, and it now has red on one side of it as well.

🐛 Fixed

• A mix over its loudness target is red in the Number Box and the Alert Meter, where it used to be drawn as a reading with no opinion. LUFS-I past the target and its tolerance returned a neutral state, so the delivery report called the same mix a failure while the meter beside it printed the number in plain text — and the bars and arcs that cut at the target now paint that region red, which made the disagreement plain. Under the target is still neutral: quiet is not over. Amber is unchanged and still means approaching — true peak inside the last decibel before the ceiling is the case it exists for. The rule now has a test; it had none, which is how this stood for eight phases.

• The macOS application and its plug-ins load again on every macOS they claim to support, and the floor is now 14.2. The engine holds a strong reference to CATapDescription — the Core Audio class behind System Output — and a strong reference to a class that does not exist is a library dyld cannot resolve. Below macOS 14.2 the entire engine library therefore failed to load, so the application died at launch and a DAW found the plug-in absent, on every version between the old floor and 14.2. Nothing degraded to “no system capture”; it was all or nothing. The application previously declared 10.15 and the plug-ins 11.0, both of which were promises the binaries could not keep, and the library itself was being compiled at 13.0 by a build flag that ignored either setting. All three are 14.2 now, engine/src/oaa_tap.h fails the build if any of them is lowered again, and the five unguarded-availability warnings the old floor produced are gone.

• The super meter’s target ticks read as marks rather than as rendering artefacts. All three are drawn at the emphasis weight now. They are radial, so unlike every other target mark in the application they cannot be drawn with antialiasing off and land on whole pixels — at the mark weight the same 1.5 px was spread across two pixel columns at about half the alpha each, which is a deliberate annotation with half the contrast of the ones it was matched to.

• Backspace, Delete and the arrow keys work inside the tab rename field. Renaming a tab, the keys that edit text did nothing at all: Backspace and Delete are bound to “delete the selected module” and the arrows to “move the selected module”, and the guard that was supposed to stand these aside while a text field has focus was checked one layer too late. CallbackShortcuts reports a key handled the moment its activator matches, whatever the callback then decides, so the keystroke was consumed and never reached Flutter’s own text editing bindings. Typing was unaffected throughout, which is why this was easy to miss and infuriating to hit: a name could be entered but not corrected. The guarded chords are now absent from the map while a field has focus rather than present and declining.

• An alert meter’s latch can be cleared again after the source has gone quiet. The latch holds the worst reading since the last reset and is cleared by the elapsed clock running backwards, which is what a reset looks like. A link that goes quiet reports its elapsed time as “not a number”, that value was stored, and every later comparison against it was false — so the lamp stayed lit for the rest of the session, across every source selected afterwards. It now ignores a non-number rather than remembering one.

• A remote display no longer receives one frame of the previous source’s audio. The waveform collected for the tablet was not dropped when the desktop switched sources, so up to one publish interval of the old programme was sent spliced onto the front of the new one’s first frame — a join the display had no way to know about.

• The oscilloscope’s trace no longer comes apart at short time bases. From about 20 ms to 100 ms across the width, one or two samples land in each column of the display, so each was drawn as a dot a pixel tall with nothing joining it to its neighbours — a waveform that arrived on screen as a dashed line. Columns now reach to the one before them, which draws the connecting line a per-sample trace would have drawn anyway, and the picture is snapped to whole pixels so its density no longer flickers sample by sample. Below 10 ms, where the display is one point per sample, the trace is a single joined polyline instead of a segment per sample; independent segments were composited independently and the line read as chewed.

• The Super Meter’s arcs no longer step ten times a second. Momentary, short-term and integrated loudness advance on the engine’s 100 ms gating grid, and the meters repaint at about 47 Hz — so four frames in five drew the arcs exactly where the frame before had left them and the fifth jumped, which on a gauge that size reads as an instrument stuttering. The arcs now travel between readings over 50 ms, which is half the gap between them and an eighth of the shortest window any of the three measures. No reading changes: every number the module prints, and the pass or fail colour it takes, is the measurement drawn the frame it arrives.

• The playhead and the elapsed clock in the status bar are packed left, like the pair on a remote display’s link bar, and the gap between them is the same seam as every other one in the row. Both used to sit in boxes reserved for the widest thing they could ever print — a drop-frame timecode, eleven glyphs — so a host counting bars left 56 px of nothing beside its own counter, and the elapsed clock another 14. The boxes are the width of what is in them now.

• The rule between items in a menu is no longer drawn in pure white. Material’s outlineVariant — the colour a Material 3 divider actually reads — was never named in the theme, and a ColorScheme role left unset falls back to a foreground colour: #FFFFFF under a dark skin and #000000 under a light one. So the rules in the signal-source menu were brighter than any colour a skin defines and brighter than the text beside them. They are the hairline every other division in the application uses, at the graticule weight rather than the border weight so they still read against the raised surface a menu is drawn on.

• A module’s settings no longer look like entries that cannot be chosen. The Metric, Response and Time base rows in a module’s menu were drawn in muted text to mark them as settings rather than actions, directly above a Duplicate at full brightness — which is what a disabled item looks like everywhere else in the interface, including in the menu one row up. They are the same colour as every other row now, and a rule separates them from Duplicate and Delete instead.

• A module setting reads Time base: 50 ms where it read Time base — 50 ms. An em dash between a setting and its value is the punctuation this interface uses for a value that was never measured, and a menu row is the one place that reading is available.

• The oscilloscope on a remote display draws a continuous waveform. It could not before, at either of the two slower link rates, and the reason was structural rather than a slip: a snapshot carried one analysis block — 1,024 frames, 21.3 ms at 48 kHz — while a link at 30 Hz stands for 1,600 frames of audio and one at 15 Hz for 3,200. The module worked out how much audio had elapsed, correctly concluded its buffer was no longer contiguous, and cleared it. On every single frame. Nothing logged anything and no test failed; it simply would not draw, and only the 60 fps rate escaped it. The host now accumulates what it measured between two sends and says how much that is, so the trace is continuous at 15, 30 and 60. Above 48 kHz on the slowest rate more audio can elapse than one frame may carry, and the shortfall is drawn as the gap it is rather than filled in.

• A remote display honours the system’s reduce-motion preference. It ignored it entirely and redrew at 60 fps whatever the tablet had been asked for — the desktop reads that preference in its status bar, and a display has no status bar. It now caps at 30 fps when the platform asks, like every other screen. Choosing the display’s own refresh rate from the tablet is still not possible; the setting exists but nothing on that screen writes it.

• The phase scope costs less than half what it did to draw, which is felt on a tablet before anywhere else. It was 9.6 ms of rasterising per frame on an Apple Silicon Mac — over half a 60 fps budget for one module — and is now 3.3 ms, from switching off antialiasing that a 1.4 px dot in a cloud of tens of thousands gains nothing from, and from drawing the dimmest half of the trail more sparsely than the half you actually read. The trail is the same length, fades over the same time, and is still computed exactly rather than compounded through an 8-bit surface. The figure it draws is unchanged; the faintest edge of the cloud is very slightly sparser.

• A remote display uses less than half the network it did. The measurement frame is 7,648 bytes where it was 15,056, so a tablet at the default 30 Hz link rate now costs about 230 kB/s instead of 452 kB/s, and 406 kB/s instead of 904 kB/s at 60 Hz. The five arrays that exist only to be drawn — the spectrum, its peak hold, its pan, the scope and the histogram — travel as fixed point instead of 32-bit floats, at a resolution between two and four orders of magnitude finer than the pixels they land on. No number you read changes: every scalar, and the per-channel peak, RMS, VU and clip figures, are carried exactly as before. This is wire protocol version 4, so a tablet and a desktop must be on the same release to link — a mismatch is refused at connect with a sentence naming both. A plugin already installed in your DAW keeps working with a newer app, as it did before.

  The waveform is the one thing that got bigger: it is now carried at the rate it was measured rather than one block a frame (see the oscilloscope entry under Fixed), which costs about 190 kB/s at 48 kHz whatever the link rate. Net of that, 30 Hz is still down by a third and 60 Hz by more than half; 15 Hz is the one rate that costs slightly more than it did, and it is the rate that was most broken.

🚧 Internal

• packages/oaa_engine/test/vectors_test.dart runs the two official vector sets through the decoder and the engine: the EBU Loudness Test Set — every case in Table 1 of Tech 3341 and of Tech 3342 — and the compliance material of Report ITU-R BS.2217. 112 cases, all passing, and each group skips unless OAA_VECTORS or OAA_VECTORS_ITU names an unzipped copy. Not a gate, because neither set may be redistributed here and fetching 811 MB would put the network in front of the one suite that must never be flaky; the generated cases in conformance_test.dart remain the gate, and now assert the 7.1 weights too, since no official file can be committed to hold them. Both measurement entries above are what the first run found.
• The six ITU files wider than 7.1 — 10, 12 and 24 channels — are asserted to be refused rather than measured. The engine carries eight channels and has no weights for those layouts, so a number would be worse than an error.
• Decoding a snapshot was measured and cleared as a cause of a slow tablet: 4 µs a frame against a 33,000 µs budget at the default link rate. It was briefly rewritten as nine block copies, eleven times faster and worth nothing; protocol version 4 then made that impossible anyway, because fixed-point arrays have to be converted element by element rather than moved. Recorded because the wire was the obvious suspect and was the wrong one — the cost was in a single module’s rasterising, two benchmarks away.
• tool/bench_wire.dart and tool/bench_modules.dart measure the two halves of what a remote display costs — decoding a frame, and drawing it — with recording and rasterising reported separately, because they are different costs on different threads and the interesting one is not always the one being measured.
• tool/bench_gpu.dart measures the same modules on a real GPU, off the engine’s own frame timings in a profile build. It exists because the software rasteriser flutter test uses overstated the phase scope by two and a half times and reordered the table under it — the module that needed the work was still the one at the top, but no figure from a software backend was worth quoting.
• Every benchmark now reports how many pixels it inked, and fails or says so loudly when that is none. The harness had drifted from the wire layout it was writing by hand and stopped drawing altogether, while still reporting timings that were low, stable and entirely plausible; it was caught by somebody looking at the window rather than by anything in the suite. The material is now built through the real encoder and decoder — tool/bench_material.dart — so there is no second implementation of the protocol to drift.
• website/ holds open-audio-analyzer.com: a static Astro site, deployed by hand to Cloudflare Workers, in this repository rather than beside it because its content is derived from the code next to it. The numbers on its front page are measured at build time by website/scripts/measure.mjs rather than typed, and the fourteen thumbnails in its module catalogue are photographs of the real widgets, taken from package:oaa by website/tools/module-renderer — an approximation of a measurement display is the one picture this project should not publish. No job in ci.yml builds it or deploys it, and nothing in a release contains it.

0.8.0 — 2026-08-22

✨ Added

• An Oscilloscope module: the waveform itself, in a lane per channel. Its one control is a time base, from 5 ms to 5 s, and it sets both how much time the width holds and how the window is found — below 200 ms the display is triggered on a rising zero crossing, so a periodic signal stands still and can be read, and above it the display rolls right to left the way a DAW draws a waveform. Anything that reached full scale is drawn in the over colour, so a clipped passage is a red band rather than a flat top somebody has to notice. It measures nothing the previous thirteen modules did not: the samples were already published, and it works on a tablet unchanged.
• The oscilloscope leaves a gap where audio was measured and never published — a file pushed through faster than real time, or a device that overran. The alternative is drawing the samples that happen to be at hand across the whole stretch, which is a picture of a programme that never played, at the right amplitude and in the right place.

⚡ Changed

• The default layout’s Spectrum tab carries the oscilloscope across its full width. The analyser and the three displays below it are a row shorter each to make room. A layout you have already arranged is untouched — the canvas as you left it is what reopens.

🚧 Internal

• The meter clock now acquires the engine’s snapshot on every tick and throttles only the repaint. The frame-rate setting still does what it says — at 30 fps the meters rasterise thirty times a second — but a measurement is no longer lost between two of them. The engine’s snapshot has one slot, so anything a reader does not take before the next publish is gone; that cost the old arrangement one measurement in three at 30 fps against a 47 Hz publish rate. The oscilloscope is the first module for which that is a hole rather than a coarser picture, and it reads the new unthrottled channel.

0.7.0 — 2026-08-22

✨ Added

• Pairing a display by camera. The sending machine shows a QR code under the code button beside PUBLISH in its status bar, and a tablet reads it under Scan a QR code on the screen it opens with. It carries the address and the port and nothing else, so a display that scans one can still only watch. This is the route that works where discovery does not — a venue or a guest network that blocks multicast previously left somebody reading four numbers off a laptop across the room and typing them into a tablet.
• The pairing code names which network interface it is for, on a machine with more than one. A laptop on Wi-Fi with a dock has two addresses and only one of them may be reachable from the tablet, which is not knowable from the sending side.
• Scanning is offered on Android, iPadOS and macOS. Windows and Linux have no camera implementation, so the row is absent there rather than present and refusing; the typed address is offered on every platform, as before.
• The macOS, Windows and Linux downloads now install the plug-in for you, with a checkbox that starts ticked. macOS offers the VST3 and the Audio Unit as separate rows, Windows and Linux the VST3; unticking leaves the application on its own. Previously the plug-in was a bare archive and the instructions for placing it by hand ran to a screenful, most of which existed because a bundle in the wrong folder and a bundle that failed to load look identical from inside a DAW.
• A plug-in installed by the macOS package needs no xattr step. Files placed by an installer are not quarantined, so the Gatekeeper refusal that has no override — the one a DAW reports as “could not verify … is free of malware” — does not arise.
• The Windows installer registers an uninstaller under Settings → Apps, which removes the plug-in as well as the application.
• A third Linux download, a tarball, beside the AppImage and the flatpak. It is the only one of the three that carries the VST3: unpack it and run ./install.sh, which needs no root and writes under your home directory unless you ask it for --system.
• The Linux tarball’s install.sh is its uninstaller too, and takes --vst3, --no-vst3, --system and --uninstall for use without a prompt.

⚡ Changed

• The remote display is reached from the status bar in one press, from either end. REMOTE and the panel behind it are gone. In their place the bar carries a PUBLISH switch, a button beside it showing the pairing code, and ATTACH, which opens Show another machine directly. Publishing no longer takes three presses and a question about which end of the link this machine is — a question anybody pressing the button had already answered.
• The remote display’s name, port, update rate and pairing code are now in Settings under Publish, with the rest of what the application remembers, instead of in a modal of their own. The section also states whether it is publishing and how many displays are attached, which the switch has no room to print.
• The name and port commit when the edit ends — on Enter or on leaving the field — rather than from an Apply button. Nothing else in the settings panel has one.
• The pairing code’s button is disabled, rather than absent, while publishing is off. A code nothing is listening at is a display that scans, connects and times out, which reads as a broken feature rather than as a switch that is not on.
• The pairing-code mark is redrawn on Material’s qr_code geometry: three finders and a scatter of data cells. The previous mark was five filled cells in a 3x3 grid, which reads as a dice face.
• A module’s title bar and its corner resize grip now accept a finger from a larger area than they draw: 40 px down from the top edge instead of the 24 px bar, and a 32 px corner square instead of a 16 px one. Both were well under the 44 pt and 48 dp the platforms ask for, which made a module fiddly to move and harder still to resize on a tablet. Nothing on screen changed size, and neither did anything for a mouse, a trackpad or a stylus-free pointer — the larger targets are admitted to touch and stylus only.
• A tap on a module’s resize grip selects the module. It previously did nothing: the grip is opaque, so it took the tap from the module underneath it and had nothing of its own to do with it.
• A typed address that is not an address now says so, instead of the Connect button appearing to do nothing. A mistyped port — 192.168.1.20:70000 — was previously read as a host name with a colon in it and dialled, which failed several seconds later as a name lookup, and an unbracketed IPv6 literal was split at its last colon into a host and a port that could never connect.
• The macOS download is an installer package rather than a disk image, and the Windows download is an installer executable rather than an msix. Neither original format can install a plug-in: a disk image has no install step at all, and an msix’s virtualised filesystem cannot write the shared VST3 directory.
• The Windows installer is not yet signed, so Windows shows “Windows protected your PC” and the user clicks More info → Run anyway. The msix it replaces was refused outright when unsigned, so this is a state a user can get past rather than one they cannot.

🐛 Fixed

• Publishing to a tablet no longer stops following the desk when the window is narrowed. The status bar drops whole items rather than squeezing them, and everything that fed the display host ran inside the item it dropped — so past that width the socket went on streaming measurements while layout, skin and delivery-target changes stopped arriving, and a changed name, port or rate was never adopted. Nothing said so, and the tablet looked healthy throughout.
• The status bar no longer overflows its own row while publishing at narrow window widths. The remote button’s label grew from REMOTE to REMOTE · ON when it was switched on — 27 px the width gate below it had never been measured against, because the sweep that proves the bar fits never published. The switch that replaced it is the same width in both states.
• A warning that no display will find this machine no longer flashes on screen for a second every time publishing is switched on. The first announcement of a session is often refused before the multicast join settles and succeeds on the next one a second later; the notice now waits for the responder’s opening burst to finish before it says anything, and clears immediately.
• Publishing that could not start — the usual cause is a second copy of Open Audio Analyzer already running — now says so on screen. The switch flicked on and back off under the pointer with the reason recorded nowhere anybody was looking.
• A desktop that cannot announce itself on the network now says so, on the Remote row and in the panel behind it, instead of reading Publishing while no tablet ever lists it. The advertisement had no error reporting at all: its bind error, its send error and its socket’s error stream were all discarded, and the last of those is the one a refused local-network permission arrives on. Nothing about the machine looked wrong either, because the permission blocks the announcement and not the port — a display given the address by hand connected and worked perfectly throughout.
• macOS users upgrading from 0.5.0 or earlier have to allow local network access again. 0.6.0 changed the bundle identifier from dev.openaudioanalyzer.oaa to com.openaudioanalyzer.oaa, and macOS keys the permission to the identifier — so the entry granted to the old application no longer applies to this one and tablets stop finding the desk. Allow Open Audio Analyzer under System Settings → Privacy & Security → Local Network. Presets, skins and paired hosts are unaffected; they are keyed by name, not by identifier.
• A panel now moves clear of a tablet’s software keyboard. Every panel stayed centred on the whole screen while the keyboard covered the bottom third of it, so the address field in Show another machine — the last row above the footer — was behind the keyboard together with the caret and everything typed into it. Panels make room for the keyboard now, and the field being typed into is scrolled in front of whoever is typing.
• The DAW’s playhead no longer floats loose in the middle of the status bar. Its box reserves the width of a timecode, so a host that counts bars and beats instead — 1|1.0 — filled a third of it and left the rest as a hole on the right, while on the left the sample rate and channel count printed flush against the reading at every window between about 1160 px and 1310 px. The playhead now sits one gap from the elapsed clock it is read beside, and a group’s width away from the format readout at every window size.
• A tablet’s page tabs now sit beside the name of the machine they belong to. The playhead was in front of them, and it reserves room for a timecode, a tempo and a time signature whether or not the host at the other end keeps all three — so on a host counting bars the tabs stood 88 px clear of the reading with nothing between the two, and on one reporting a clock and no tempo 190 px, reading as a control floating in the middle of the bar. The tabs come first now, and nothing that appears, vanishes or reserves unused room can move them.
• The Linux VST3 now loads on the same distributions as the Linux application. It was built against a newer glibc than everything shipped beside it, so on an older distribution it was simply absent from the DAW’s plug-in browser with nothing logged — indistinguishable from having installed it in the wrong place. Affects every release that shipped a Linux plug-in.

🔥 Removed

• The Remote panel, and the Send these meters panel behind it. Both ends of the link are controls in the status bar now, and what those panels configured is in Settings → Publish.
• The macOS .dmg and the Windows .msix. Their replacements are above, and each installs a superset of what the one it replaces did.

🚧 Internal

• Turning the remote display off publishes its state before it closes its sockets rather than after. Both closes suspend and dispose does not wait for them, so a service that had been publishing wrote to a disposed ValueNotifier from inside a microtask — visible only in the console, and only ever reached by a service that had actually been switched on.
• A release no longer carries artifact.tar, an unlabelled tarball of the documentation site that has been attached to every release since 0.3.0. The publish step downloads every artefact the run produced, and the Pages action gives its payload that fixed name.
• The signing keychain accepts a .p12 carrying both Developer ID certificates, grants productbuild and productsign use of the key, and verifies what it imported against the identities the job asked for instead of printing a summary that could not show an installer certificate at all.
• The macOS package turns off bundle relocation. Left on, the Installer asks Spotlight where the application already lives and writes the update there — so a stale copy in ~/Downloads would receive it while /Applications kept the old version, and the install would report success.
• The five packaging jobs in ci.yml are named for the platform and the artefact they produce — macos-pkg, windows-installer, linux-tarball, linux-appimage and linux-flatpak — where a format, a platform and a container had been three naming schemes for one family of jobs.
• The plugin and CLI artefacts are named by platform rather than by runner image. Three installer jobs download the plugin bundles by name, and the name was oaa-plugin-ubuntu-22.04 — the image is pinned for glibc reasons and moving it would have failed those downloads rather than the job that moved it.
• OAA_IOS_TEAM_ID now reaches the iPad build. The Runner target’s Release configuration set its own DEVELOPMENT_TEAM, which outranks the xcconfig make_ipa.sh writes, so a profile created under any other team was installed and valid and Xcode looked for it under the project’s team instead — the archive failed and no build could reach TestFlight. make_ipa.sh also compares the profile’s team against the build’s before archiving, as its fourth pre-build check.
• The test that holds the two ends of the link to different advice no longer asserts it on the one path that writes no advice at all. Outside macOS a refused multicast send is reported as its errno, by both ends, because the errno is the only fact there is — so the assertion passed on the machine it was written on and failed the Linux job.

0.6.0 — 2026-08-22

✨ Added

• macOS releases can be notarised, which is what lets a downloaded plugin or dmg be opened with no further steps. A Developer ID signature now also carries the hardened runtime and a secure timestamp, both of which Apple requires before it will notarise anything and neither of which was being requested.
• Tagging a release now builds the iPadOS display and uploads it to TestFlight, so a tablet no longer has to be built from source to be updated. The upload runs after the release is published rather than before it, which means a TestFlight build always corresponds to a release that exists. The IPA itself is deliberately not attached to the release: an App Store build provisions no devices, so a downloaded copy could not be installed by anyone.
• The iPadOS build declares that it uses no encryption. Without that declaration every TestFlight build waits in “Missing Compliance” until somebody answers the export question by hand, which is the one step an automatic upload cannot take.

⚡ Changed

• The application identifier is now com.openaudioanalyzer.oaa, and the plugin’s is com.openaudioanalyzer.oaa.plugin. Both were under dev., which reversed a domain nobody held; these reverse open-audio-analyzer.com. The hyphens do not survive the trip, and cannot: an Android applicationId and a Kotlin package are Java identifiers and reject -, while Apple’s CFBundleIdentifier accepts only letters, digits, - and . and so rejects the _ that Android would want in its place. Stripping them is the one form that is legal on all six platforms. The Linux application id, the AppStream and flatpak ids, the msix identity and the hicolor icon filenames all move with it.

🐛 Fixed

• The macOS plugin bundles are built for Apple silicon and Intel, and load on macOS 11 and later. Every release up to 0.5.0 shipped a bundle built for whichever machine the release ran on — Apple silicon only, and requiring that machine’s own macOS version or newer — which a DAW reports in exactly the way it reports a plugin that was never installed.
• The instructions for a downloaded macOS plugin now describe what macOS 15 and later actually do when the quarantine flag is still on it: a modal saying the plugin cannot be verified free of malware, or that it will damage your computer, with nothing in System Settings to override it — the “Open Anyway” button there is only offered for a blocked launch, and loading a plugin into a DAW is a library load. The fix has not changed and is still xattr -dr com.apple.quarantine; what was wrong was the description of the symptom, which called the failure silent and sent people looking for a different problem.
• The README no longer claims the macOS build is distributed “signed with a Developer ID and notarised”. It has never been either.

🚧 Internal

• CI imports a signing certificate before anything signs. OAA_SIGNING_IDENTITY was being handed to codesign on runners whose keychain was empty, and OAA_NOTARY_PROFILE named a credential that only exists on a developer’s own machine, so neither secret could have worked on a runner — and neither failed a job, because they had never been set and the scripts took their no-credential branches instead.
• packaging/ios/make_ipa.sh and packaging/ios/testflight.sh, and the ipa and testflight jobs in ci.yml. Manual signing is injected through ios/Flutter/Release.xcconfig so the Xcode project stays on automatic signing for local development; the script reads the signing authority back off the archive, because flutter build ipa exits 0 on an export that silently fell back to automatic signing.

0.5.0 — 2026-08-22

✨ Added

• Double-clicking the status bar on macOS zooms the window again, as double-clicking a title bar does everywhere else on the Mac — the status bar is this window’s title bar, and it had answered nothing since 0.3.0. It obeys the system’s “double-click a window’s title bar to” setting, so a Mac set to minimise minimises and one set to do nothing does nothing, and it waits the double-click interval that Mac was configured with. No control in the row answers any later for it: what Flutter recognises there is still a single click, and the pair is counted in the runner.

⚡ Changed

• The iOS and Android home screen label is OAA, the wordmark the status bar already draws. It was Open Audio Analyzer, which no tablet ever actually showed: iOS fits an icon label to the width of one icon by tightening the tracking before it truncates, and tracking is spent on the spaces between words first — a 19-character name arrived on the home screen as OpenAudioAnal…, run together and cut short anyway. Apple’s guidance is twelve characters. The product’s name is unchanged everywhere it fits: the macOS, Windows and Linux applications, both permission prompts, and every configuration directory — nothing a tablet has saved moves.

🐛 Fixed

• The canvas no longer gets slower the longer a spectrogram is on screen. The spectrogram redrew its whole history as run-length marks on every published frame, and its cost was budgeted against smooth columns — real material jitters a brightness step between most adjacent rows, so a large display was re-recording and re-rasterising 150,000–230,000 marks about 47 times a second, a cost that ramped for the half-minute the history took to fill after a mount and then stayed. Every meter repaints from the same clock, so it dragged the whole canvas — most visibly the analyser — and the remote display paid the same price on weaker hardware, more of it at the 60 fps link rate. The history is now kept as pixels and drawn as one image, a quarter of a millisecond per published frame whatever the signal does; the numbers are in tool/bench_spectrogram.dart.
• On a remote display, the tab control now sits beside the name of the machine being watched instead of stranded in the middle of the link bar. The bar kept room for the host’s playhead whether or not the host had one, so on every host with no DAW behind it — a desktop metering a device or a file, which is most of them — 248 px of nothing stood between the name and the tabs. The slot appears with the playhead and does not move again while the link lasts, so a reading that goes stale does not shift the tabs out from under your finger.

🚧 Internal

• tool/bench_spectrogram.dart is formatted the way the test job’s dart format --set-exit-if-changed step demands. It was not, so that job failed on every push after the benchmark landed while the analyzer, the suites and the benchmark itself all stayed green. That step is now listed with the other gates in README.md and CLAUDE.md, which is why it was never run by hand.

0.4.1 — 2026-08-22

🐛 Fixed

• The macOS plugin bundles now carry a valid code signature. All three — the VST3, the Audio Unit and the Standalone — shipped with an invalid one in every release up to 0.4.0: the VST3’s resource seal was computed before JUCE wrote moduleinfo.json into the bundle, and the other two were never bundle-signed at all. On Apple Silicon that is enough for a DAW to leave the plugin out of its browser without logging anything, and for auval to refuse the Audio Unit. Each bundle is now signed after everything that writes into it and verified on the spot, so the build fails rather than producing a bundle it cannot verify. auval -v aufx OaaM OaaA passes. If the plugin never appeared in your DAW, this is why — and on macOS also strip com.apple.quarantine from a copy you downloaded, which no build can do for you.

✨ Added

• The documentation site’s Install page now has an In a DAW section: which folder the VST3 and the Audio Unit have to be copied into on each platform, the xattr line that a downloaded copy needs on macOS, and the Ableton Live preference that has to be on before Live looks in the folder at all. The page had linked to that section since the site was first published, and the section did not exist.

🚧 Internal

• OAA_CODESIGN_IDENTITY signs the macOS plugin bundles with a Developer ID instead of ad-hoc. Releases stay ad-hoc.
• Signing runs from a target that depends on the format target rather than from POST_BUILD, which is not last: a MACOSX_PACKAGE_LOCATION resource is copied by a sibling rule that make may run after the link, and on a from-scratch build it does. Only visible with the bundle deleted first.

0.4.0 — 2026-08-21

✨ Added

• An Android tablet now finds hosts on the network by itself, like every other platform. It held a multicast socket that never received an answer, because Android’s Wi-Fi driver discards multicast for an app that is not holding a WifiManager.MulticastLock — silently, with the search still saying it was looking. The lock is held while a search is running and given back when it stops, so it costs nothing once a display is attached.
• An Android tablet now remembers its layout, its skin and the host it was last attached to. Nothing in an Android process names a directory it may write to, so the configuration directory is asked for over a platform channel and lives in the app’s own files directory; every launch before this one started from the defaults.
• Metering your Mac’s own output no longer needs a loopback driver. On macOS 14.2 and later, System Output is the first entry in the source menu, named after the output device it is metering. It is a Core Audio process tap, so there is nothing to install, nothing to reroute and no password prompt — the audio still reaches your speakers while it is being measured. macOS may ask for permission to record system audio the first time. Below 14.2 the entry is absent and BlackHole or Loopback is still the answer; Windows and Linux are unaffected, since WASAPI loopback and a PipeWire monitor source already appear in the list.

⚡ Changed

• The wire protocol is at version 3, and a receiver now accepts any version it knows rather than only its own. A plugin built against 0.3.0 keeps working with a newer app, which under the old equality check it would not have: a plugin lives in the DAW’s plugin folder and stays there across app upgrades, so the mismatch was the ordinary case, and what it looked like from the plugin was a port that accepts and hangs up forever — indistinguishable from the defect where the port was never bound at all. The rule is one-way on purpose: a peer newer than this build is still refused, because a later version may have moved a table and misreading a measurement table is how a meter draws a confident wrong number. A tablet still running 0.3.0 or earlier will refuse a newer desktop and say so, rather than drawing anything; update both ends. Version 3 adds a frame type and moves no byte of any existing table — held by a test that decodes the frozen version-2 golden and diffs it against the version-3 one, where exactly four bytes differ and all four are version fields.
• A host search that is running but cannot receive says so instead of “Looking for hosts on this network…”, which is the face a search that is about to succeed wears.

🐛 Fixed

• The plugin’s status panel no longer claims a playhead from a host that is not giving one. The line was drawn from whether a transport had ever been published, and the plugin publishes one for every audio block — the empty one that means “this host said nothing” included — so it read as a playhead the moment audio started flowing, on precisely the host it exists to warn about. It now follows what the host is reporting now, and says “no playhead from host” again if a host stops.

🚧 Internal

• Groundwork for the Elapsed and Timecode LUFS modes, which are not yet offered by any module — this is the protocol and the engine underneath them, and the modules and their menu are the change after this one. docs/WIRE.md gains 0x0020 SET_LUFS_MODE, the first frame that travels from consumer to producer, permitted on the ingest port only because that one binds loopback where whatever connects is already running as this user; the display port binds every interface and stays read-only until somebody designs authentication for it. Three implementations moved together, as that file requires.
• The engine can reset itself when the signal returns after a silence (oaa_engine_set_silence_reset, ABI 5), which is what the System mode is made of. It lives in engine/ rather than above it because silence is a property of audio and not of a host — one implementation serves a plugin and a sound card, where two would eventually disagree about when a track began. The gate runs before the block it judges, so a track whose loudest sample is in its first block keeps that peak rather than having it cleared by its own reset. Off by default, and off for file analysis, which must measure a file whole. The two transport-driven modes need no engine API at all: they are the producer declining to push, so engine/ still does not learn what a DAW is.
• The plugin’s answer to a host that supplies no transport is now held by a test. plugin/test/transport_capture_test.cpp hosts the AudioProcessor directly, which is the only way to reach either of the two branches behind it: no VST3 or Audio Unit host can express “no playhead” or “no position”, so neither a DAW nor the fake DAW can ask for them. It runs in the gated plugin job’s ctest, and both it and the transport box’s own test were verified by breaking the code under them and watching them fail.
• Five documents caught up with protocol version 3’s two goldens. docs/WIRE.md still gave the SNAPSHOT payload size “at protocol version 2” in the one document whose header declares version 3; packages/oaa_wire/AGENTS.md named wire_v2.bin as the golden its codec test decodes, when the test reads both and wire_v3.bin is the one tracking the current serialiser; and two comments in ci.yml credited the C++ producing side with writing wire_v2.bin, which it no longer does. The one that could have cost something: the regeneration command in plugin/test/wire_fixture.cpp still wrote to wire_v2.bin — the frozen file whose whole purpose is to hold bytes produced before version 3 promised to move no table. Following it would have destroyed the evidence and left every test green.
• Two claims about that branch are corrected here rather than in the 0.3.0 notes, which are released. It is not reached by the Standalone build — JUCE 8’s AudioProcessorPlayer installs a counting playhead whenever the processor it is given has none — and it is unreachable through an Audio Unit as well, not only through VST3. The plugin’s own handling was correct throughout; only the note about what exercised it was wrong. README.md no longer lists it under Known gaps, because it now has a test instead of a paragraph.

0.3.0 — 2026-08-21

📐 Measurement

• The clip indicator now catches clipping. clip is the longest run of consecutive full-scale samples since the last reset, latched until Reset; it was the run still in progress at the block boundary, which is zero for every clip that ended inside the block — almost all of them. A run of 40 full-scale samples in the middle of a 1024-frame block published 0. The Digital Meter’s clip lamp is drawn from this, so it was dark for real clipping, which reads as proof that nothing clipped. No other reading changes; nothing that was reported as clipped is now reported as clean.
• Crest is now taken over the block being measured rather than from the values the meters draw. Both operands were the displayed peak and RMS, which carry a 1.5 s hold and a 300 ms averager, so the figure described the ballistics: a single block of DC at 0.9 read 11.6 dB where the answer is 0, and 0.43 s after a transient it read 17.8 dB and was still climbing. A steady sine reads 3.0103 dB either way, which is why the test suite never caught it. Re-measure anything whose crest you recorded from a moving signal — readings on transient material fall, typically by several dB, and were previously too high. Multichannel now reports the peakiest channel rather than the loudest peak minus the loudest RMS, which could describe no channel at all.

✨ Added

• A DAW’s playhead now reaches a tablet. The desktop decoded the plugin’s transport frame and kept it: bars, beats, tempo, time signature and timecode stopped at the desktop, so a remote display showed a plugin’s meters beside no position at all. It is relayed to every attached display now — on change rather than with every measurement, so a parked session costs nothing, and replayed when a display attaches so that one joining a parked session is not left blank until somebody presses play. A jump in the playhead survives the hop: the flag is an edge delivered once, and a relay publishing thirty times a second against a DAW’s ninety-odd blocks accumulates it rather than sampling it.
• The status bar and a tablet’s link bar both show the host’s transport: the position in the most precise unit the host gave — timecode, else bar and beat, else its own clock — with the tempo and time signature where there is room for them, and brighter while the transport is rolling than while it is parked. The app had been decoding all of it and showing none of it. A value the host did not supply is not drawn at all, rather than printed as a plausible zero.
• The app accepts plugin connections. PluginLink was written, tested and never constructed, so port 47822 was never bound: a VST3 or AU inserted in a DAW retried against nothing forever while the README said the desktop app meters what the DAW is playing. Inserting a plugin now puts it on the canvas — the act of inserting it is the act of choosing it — and removing it hands the canvas back to the local source. A port that cannot be bound is reported in the window rather than being silent, because the usual cause is a second copy of Open Audio Analyzer already running.
• oaa --target reads your own delivery targets, not only the six built-in ones. The app has always merged calibrations/*.json over the built-ins by id; the CLI knew nothing about them, so a corrected atsc-a85.json changed the verdict in the window and left the exit code — the one a release pipeline believes — judging against ours. oaa --list-targets shows them, and --config-dir points at a directory other than the default.
• The VST3 and the Audio Unit are published with each release, as one archive per platform holding the plugin bundles. They are not bundled inside the desktop installers yet.
• The plugin is compiled on Linux, macOS and Windows as part of a release, and on demand. Nothing built it before: the only plugin check in CI compared two text files and never invoked CMake, so a JUCE dependency fetched by tag, a C++ wire producer that has to agree with the Dart one byte for byte, and a version that moves every release were all held together by whoever last built it by hand. It is not built on every push, because three parallel JUCE builds cost more than a push asks for.

⚡ Changed

• The application is called Open Audio Analyzer everywhere it names itself. The window title, the macOS menu bar, the Windows version resource and the Android launcher label all said oaa, which is the repository’s short name and not the product’s. The executable moves with them: Open Audio Analyzer.app on macOS, OpenAudioAnalyzer.exe on Windows and open-audio-analyzer on Linux, where a space is not available — the Flutter-generated CMake makes the executable name a CMake target name, and CMake rejects one containing a space. The oaa CLI keeps its name; it is a command, and a command with spaces in it is not one. Bundle identifiers are unchanged, so no configuration moves and no host loses track of the plugin.
• The plugin’s bundle identifier is dev.openaudioanalyzer.oaa.plugin. It was io.github.jonasgrunau.bel: the rename to Open Audio Analyzer moved every other identifier and missed this one, because nothing built the plugin to notice. A host caches a plugin by that string, so it moves before the first published build rather than after.

🐛 Fixed

• A relocate is reported once rather than twice. The plugin marks the single audio block on which the playhead jumps, and that flag was carried both in the accumulator that exists to deliver it exactly once and in the transport payload, which is sampled — so a machine loaded enough for two frames to leave inside one audio block delivered one relocate as two, and a three-lap loop reported four. docs/WIRE.md lets a consumer count relocations by counting flagged frames, so the count was wrong; no reading changes, because nothing yet acts on the flag. Held by a new deterministic test rather than by a loaded machine.
• The “audio was lost” notice counts the frames the metered source discarded. It read the local engine’s counter while the flag that raises it comes from whatever is on the canvas, so a plugin that overran produced “Audio was lost — 0 frames were discarded”: a warning that contradicts itself, about a real loss of audio, carrying the number somebody would put in a bug report.
• The remote display no longer reads a destroyed engine. Changing the audio source or device while publishing to a tablet left the publish timer holding the engine it was built with, acquiring through a freed handle thirty times a second and sending 15 kB of returned heap to the tablet as a measurement. The service now follows the engine, and lives beside it rather than inside the status-bar button — which the status bar drops below 620 px of window width, so narrowing the window silently tore down an active session.
• A disposed engine reports itself unavailable instead of reading freed memory. Every scalar reading returns NaN, false or zero and refresh returns false, so a holder that keeps one a frame too long draws em dashes rather than plausible numbers. The array views cannot be guarded and are documented as invalid the moment dispose returns.
• Cancelling a file analysis no longer leaves the previous one running. The native cancel flag was freed while the worker isolate was still reading it, so the next analysis reallocated those four bytes as zero and the old worker read “not cancelled” and went on decoding its whole file — competing for the frame budget the isolate exists to protect. The flag is released once the isolate has actually exited.
• Resizing a module that a stored layout had left smaller than its own minimum, against the right or bottom edge, no longer throws. Rects are now pinned to the canvas and to the module’s minimum as they are read, whatever wrote them.
• The remote display’s advertised name can be cleared, not only replaced. Emptying the field restored the previous name, so “use this machine’s name” was unreachable once a name had been set.
• A file analysis that cannot start no longer leaks its open decoder.
• A preset named CON, AUX, NUL, PRN, COM1–COM9 or LPT1–LPT9 can be saved on Windows, where those are reserved with any extension.
• Stopping a pushed engine clears its running flag. It never started a thread, so oaa_engine_stop returned early and the snapshot reported a stopped engine as running for the rest of its life.
• A capture device’s id is no longer truncated to an odd number of hex digits with its terminator landing by luck. Trailing zero bytes are dropped before encoding, which keeps every id a real backend produces well inside the field.
• The plugin drops a wrong-length frame rather than sending it. The check was an assert, and the plugin is built Release everywhere including CI, so NDEBUG removed it from the only build that exercises the C++ producer.
• The plugin no longer tells the app the playhead relocated while the transport is parked. A stopped DAW still runs its graph and reports the position it sits at, unchanged, every block; the plugin compared that against “one block further on” and raised its discontinuity flag on every published frame for as long as the transport was stopped — measured at 140 frames out of 140. It is now only evaluated while the transport is rolling. Nothing in the application consumes the flag yet, so no reading changes today; the Elapsed and Timecode LUFS modes are what would have been affected.
• The plugin’s name arrives at the app as text rather than as mojibake. The HELLO frame’s producer name was built with juce::String’s const char* constructor, which reads its argument one byte to one codepoint, so the em dash in “Open Audio Analyzer plugin — “ was mangled on the way in and re-encoded on the way out: the app’s title bar read Open Audio Analyzer plugin â<80><94>. docs/WIRE.md specifies that field as UTF-8, which makes it a protocol defect rather than a typographical one. JUCE asserts on this and the assert is compiled out of the Release build, which is why nothing caught it.
• A relocate now reaches the app instead of usually being missed. The flag marks the single audio block on which the playhead jumped, and the streaming thread samples the transport once per published frame — every second block at a 512-frame buffer, one in sixteen at 64 — so it was set on one publish and read from another: three loop laps in a row delivered it zero times out of 186 frames. Edge flags now accumulate outside the seqlock and are delivered once each, which makes the count independent of buffer size. Verified at 64, 128 and 512 frames.

🚧 Internal

• The plugin’s framework-free C++ tests run on every push. plugin/’s JUCE fetch is conditional now, so -DOAA_BUILD_PLUGIN=OFF configures the directory in under a second and builds liboaa, the wire fixture and the transport box test in three — where the full job takes about ten minutes and, being gated to releases and manual runs, had been the only place any of it ran. It closes a real hole as well as a slow one: the wire golden is only worth having from both ends, and only the decoding half had been checked between releases.
• The fake DAW no longer invents relocates on a slow machine. An offline run handed the transport a read-ahead thread, and when that thread falls behind, JUCE’s buffering source returns silence while the reported playhead stops advancing — a host claiming to play with a position that sits still, which is a discontinuity by any definition, and the plugin flagged it. The plugin was right; the instrument was wrong. Offline runs read the file synchronously now, which is what makes their timeline independent of the scheduler; the device path keeps its buffer, where a stall is a click rather than a wrong number.
• resolveConfigRoot, ConfigDir, ConfigFile and slugify moved from the app into oaa_core, which is what lets the CLI read the same delivery targets the app writes. They are pure functions, so the package keeps its “no I/O” rule.
• A fake DAW, in plugin/host/. It plays an audio file through the VST3 or the Audio Unit and hands it a transport — tempo, time signature, timecode frame rate, loop points, the record flag, and the playhead itself, which can be switched off. None of the plugin’s playhead handling had ever been run before: the only host that could reach it was a real DAW driven by a person. It is built by the same CMake run as the plugin and shipped nowhere.
• The plugin is driven end to end by a test. plugin/host/ also runs headless — no window, no sound card — and packages/oaa_wire/test/plugin_e2e_test.dart spawns it, listens on the port the app listens on, and decodes what arrives. It is the only coverage of prepareToPlay, the FIFO, the playhead, the engine, the streaming thread and the socket at once; the byte-for-byte golden beside it is produced by a fixture that links no JUCE. It generates its own audio, so nothing in CI downloads anything, and it skips rather than fails without a built plugin.
• A DAW’s meters are held against what a tablet shows, in one test. test/plugin_to_display_e2e_test.dart runs the same fake DAW through the application’s own plugin ingest and display host, attaches a display client — which is what a tablet runs — and compares twenty-nine readings field by field, plus the playhead: the tempo, the meter, the timecode and the bar the host was told to be at. What a display receives is a re-encode of a snapshot the app decoded off the plugin’s socket, so a field dropped in the middle left both halves’ suites green and the tablet showing a dash. It skips without a built plugin, and CI runs it on the Linux leg of the plugin job.
• One thing the fake DAW found is written down rather than patched, under Known gaps in README.md: the plugin’s “host supplies no transport” branch cannot be reached through VST3 at all, because the format has no way to say it. That is a property of VST3 rather than a defect, and the branch is reached by the Standalone build. The other finding recorded that way on the first run — that the discontinuity flag usually did not survive the trip to the app — was fixed inside this release instead, and is under Fixed above; it moved a value docs/WIRE.md describes, which is why it was a decision before it was a fix.
• tool/fetch_test_audio.dart downloads the Creative Commons music the application is looked at with — a tone produces a spectrogram that is one bright line and a stereo cloud that is a dot, both correct and neither informative. Two CC BY 3.0 post-rock tracks from Wikimedia Commons, chosen by measuring four candidates with oaa rather than by reading titles: the default is a loud master with a real 10.3 LU range, a true peak above its sample peak, and a stereo field that moves. It resumes a partial download, verifies the length and signature, writes the attribution the licence asks for beside the audio, and is run by hand. No test depends on it.
• One workflow instead of three. ci.yml runs the tests, the documentation site, the installers and the release as jobs gated by event, so a push produces one run rather than two and a tag no longer produces a third that names neither. The reason it was split — that packaging must not slow the signal everybody waits on — is kept by not running those jobs on a push, which is a condition rather than a file. Two things the split had made impossible: a release can now depend on the test jobs, so a tag cannot publish from a red commit, and workflow_dispatch builds every installer without publishing anything.

0.2.0 — 2026-08-19

📐 Measurement

• The VU meter reads differently, and lower on most material. It was a one-pole smoother over mean square — an RMS meter with a 300 ms time constant. It is now what a VU movement actually is: average-responding and RMS-calibrated, through a second-order mechanism. A steady sine still reads its own RMS exactly, so a calibration tone is unchanged; anything peakier now reads lower, by 9 dB on a signal with a 10% duty cycle and typically 1 to 4 dB on dense modern masters. If you have been matching levels by VU against Open Audio Analyzer’s previous readings, re-check them. The needle also overshoots by about 1.2% on a transient, which is inside the tolerance the standard allows and is most of why a VU feels like a VU.

• The spectrum is now measured. 512 log-spaced bands from 20 Hz to 20 kHz, from a 4096-point Hann window per channel at a 1024-sample hop, zero-padded to a 16384-point transform. A band wide enough to contain bins takes the loudest of them rather than their average, so that a narrow resonance survives the mapping; a band too narrow to contain one — everything below about 216 Hz at 48 kHz — reads the transform between its two nearest bins. Levels are window-compensated: a full-scale sine on a bin centre reads 0.0 dBFS, verified on every push, and one falling between two bin centres reads within 0.3 dB of its own level rather than up to 1.4 dB low. Frequency resolution is that of the 4096-point window and the padding does not change it: two tones closer than 11.7 Hz still merge. What changes is that the bottom three octaves are drawn as the curve the transform measured instead of as a staircase of up to twenty-five identical bands.

• Per-band stereo position and the raw stereo sample stream are now published, which is what the stereo cloud and the phase scope draw.

• The short-term loudness distribution is now published, together with the 10th and 95th percentiles LRA is the difference of and the relative gate they were taken above — the same population the LRA number is computed from, so a distribution drawn from it cannot disagree with the number beside it.

• LRA itself is unchanged. It is now computed through the published percentiles rather than beside them, which is a refactor, not a new number.

• Loudness is now measured. LUFS-M, LUFS-S, LUFS-I and LRA previously read as a dash and now report real values, verified against the EBU Tech 3341 cases within the standard’s ±0.1 LU on every push, on all three desktop platforms. K-weighting follows ITU-R BS.1770-4 with coefficients designed at the stream’s own sample rate, so 44.1, 88.2, 96 and 192 kHz are correct rather than approximately correct.

• True peak is now measured, with the 4× polyphase oversampler from BS.1770-4 Annex 2. It reads higher than sample peak — by up to about 3 dB on dense, limited material — because that is what the signal actually reaches between samples. A master that previously passed a −1 dBTP ceiling on sample peak alone may now fail it. Re-measure before delivery.

• DR-S, DR-I, PLR and PSR are now defined, being differences of the above. See docs/METRICS.md for the formulas; none of them is Decibel’s proprietary TrueDyn and none pretends to be.

• LFE is excluded from loudness and surround channels are weighted +1.5 dB, per BS.1770-4. Channel layout is inferred from the channel count, and the four-channel case is read as quad (L R Ls Rs) rather than L R C LFE — see METRICS.md, and expect this to be replaced by real layout metadata when a device or file source can supply it.

• A device’s own sample rate and channel count are adopted, not converted to. Open Audio Analyzer measures the stream the hardware produced; resampling in front of the measurement would move inter-sample peaks and shift the K-weighted energy, and the resulting numbers would still look plausible.

• An interface wider than 7.1 is refused rather than measured eight channels at a time and reported as programme loudness.

• An unknown device id fails instead of falling back to the default. A preset naming an interface that is not plugged in would otherwise silently meter the laptop microphone.

• Files can now be measured, and they read exactly as the live meters do. Analysing a file decodes it and pushes the blocks through the same oaa_analyse a capture device drives — there is no second DSP path — so the numbers are identical rather than merely close. A test asserts that equality on the same samples analysed both ways, to the bit. A file is measured at its own sample rate and channel count; nothing resamples or remixes, because a converter in front of a measurement changes the measurement.

• A file report states maxima, not final values. Momentary and short-term loudness, correlation and per-channel peak describe an instant, so reading them once at the end would report the fade-out and call it the programme. They are watched across the whole file instead. Integrated loudness, LRA and the “max since reset” peaks are integrating quantities and are read at the end, which is correct for them.

• Sample values are not clamped on the way out of the decoder. A float WAV may legitimately hold values beyond ±1.0, and those are precisely the overshoots true-peak metering exists to find.

• The Spectrum Analyzer draws an average of the bands rather than the last transform, and it now says which. The engine publishes about 47 transforms a second and the module drew every one untouched, which flickers hard enough that the shape of a balance is difficult to read. A new Response setting in the module’s menu chooses Fast — no averaging, exactly what it did before — Normal at 120 ms, or Slow at 500 ms; Normal is the default, so an existing analyser reads calmer than it did and a band’s drawn level now lags a change by about that much. A short peak reads lower on Normal than it did on the frame it happened, by as much as the difference between the peak and what surrounds it. Nothing measured changed: the peak-hold line above the curve is never averaged at any setting, reports and the wire protocol carry the bands as measured, and the spectrogram and stereo cloud draw them as published.

✨ Added

• A remote display. Turn on publishing in the desktop’s status bar and a tablet on the same network shows the same meters — the same modules, the same layout, the same skin and the same delivery target, rendered by the same painters from measurements sent over the network rather than reimplemented. Hosts are found by name over mDNS (_oaa._tcp), and an address can always be typed instead, because multicast is the first thing a guest network blocks.

• A remote display says when it has stopped being current. Two seconds without a measurement and every reading becomes a dash and the link is marked stale, rather than leaving the last picture on screen. A frozen meter is indistinguishable from a quiet passage, so a display left running after its host slept would otherwise show a confident, detailed reading of a signal that had stopped existing.

• The display is watch-only, and off until switched on. It cannot reset, retarget or reconfigure the machine it is watching, and the port is closed until a human opens it. There is no password on the connection; anyone who can reach the port can read the measurements and the layout, which is why it is offered rather than assumed.

• A headless VST3 and Audio Unit plugin. Insert Open Audio Analyzer on any track, bus or master and the desktop app meters what the DAW is playing, through the same engine and the same painters as a live input — so a plugin reading and a device reading of the same audio cannot disagree. The plugin draws no meters itself; it measures and streams, and a small status panel says whether it is connected. Built for macOS, Windows and Linux, plus a standalone target for testing the link without opening a host.

• The DAW’s transport reaches the app: play and record state, playhead in seconds, samples and quarter-notes, tempo, time signature, loop points and SMPTE timecode with its frame rate. Every field carries a “the host supplied this” flag, because hosts differ enormously in what they report and a missing tempo arriving as zero is indistinguishable from a real one — an absent value renders as an em dash rather than a plausible number. The Elapsed and Timecode LUFS modes are not built yet; this is the measurement they will be counted from, and no module offers them today.

• The plugin reports when the playhead jumps — a relocate, a loop, a scrub. Anything integrating across that boundary is averaging two passes of the same music into one number and nothing about the result looks wrong, so it is stated rather than inferred. Acting on it — restarting an integration when the transport moves — needs an app-to-plugin control frame, and so wire protocol 2.

• The app listens for plugins on port 47822, loopback only. Several inserts may be connected at once; the most recently connected is the one shown, on the grounds that inserting a plugin is the act of selecting it.

• All twelve modules. The eleven that said NOT BUILT YET now measure something: a LUFS Meter (momentary and short-term as bars, integrated as a rule they pass through, the target as a band); a Digital Meter (per channel to 7.1, RMS as the column and peak as a floating tick, so the gap between them is the crest factor); a Super Meter (the three integrations on concentric arcs); a VU Meter (0 VU at the calibration’s reference level, not at digital full scale); an Alert Meter (one metric with its worst reading latched until reset); a Validator (three delivery checks and a verdict); a Histogram (short-term loudness over time, with the momentary window banded above it and everything past the delivery target in the over colour); a Loudness Distribution (how often the programme sat at each loudness, with the two percentiles LRA is the distance between, drawn from the same blocks the number is computed from); a Spectrum Analyzer; a Spectrogram; a Phase Scope; and a Stereo Cloud (per-band stereo position, which answers which part of a mix folds badly rather than only that it does).

• Open Audio Analyzer opens on a working meter bridge, with the frequency displays on a second tab, instead of six readings on an empty grid.

• The canvas is arrangeable. A 24-column by 16-row grid: drag a module by its title bar to move it, drag the corner grip to resize, alt-drag to duplicate, right-click or double-click empty space to add one, right-click a module for its options. Modules may not overlap and an illegal drop is refused rather than nudged elsewhere, with the target cells shown live while the pointer is down.

• Tabs, with rename, duplicate and delete, switchable with the number keys.

• Undo and redo, over every layout edit, from the keyboard or the tab strip.

• A Number Box shows any of the sixteen measurements, chosen per module from its menu.

• Inter and JetBrains Mono are bundled instead of requested from the system, so digit width and tracking are identical on macOS, Windows and Linux. Both are SIL OFL 1.1 and their licences ship alongside them.

• Capture from real audio devices. Inputs are enumerated and selectable from the source menu in the status bar. On Windows, WASAPI loopback meters system output with no setup; on macOS and Linux a virtual loopback device (BlackHole, a PipeWire monitor) appears in the same list — see the README.

• OaaSource.push and oaa_engine_push(): audio supplied synchronously by the caller, with no thread and no clock. It makes the engine a pure function of the samples it was given, which is what the conformance suite needs and what file analysis will be built on.

• Lost audio is reported rather than hidden. If analysis falls behind, the capture callback drops frames and the count is published; the app shows a warning saying the integrated reading can no longer be trusted. Integrated loudness averages every block since the reset, so dropped audio does not make it stale — it makes it an average of a different programme than the one that played.

• Offline file analysis. Drop a file on the analysis panel, or pick one, and it is measured end to end: WAV, AIFF, RF64, Wave64, FLAC and MP3. The run happens on a worker isolate so the live meters keep their frame budget, it reports progress, and it can be cancelled — cleanly, releasing the engine and the open file rather than leaking them.

• A report panel, showing the source, every programme-wide measurement, a short-term loudness graph with the integrated level marked, and the delivery verdict against the selected target.

• Reports export as text, JSON and CSV. Text is the human summary; JSON carries every measurement, the target and the verdict under stable field names; CSV is the loudness timeline, one row per point, for a spreadsheet or a plot. An unmeasured value is an em dash, a null and an empty cell respectively — never a zero, which is a legitimate reading for correlation and several dB quantities and so cannot double as “no data”.

• A oaa command-line analyser, so a loudness check can be a step in a release pipeline instead of something somebody remembers to do. It runs the same engine and the same decoder as the app. oaa --target streaming-14 master.wav exits 2 when the file misses its delivery spec, 1 on a file it cannot read and 0 when all is well, which is what lets a build fail on a master that is 2 LU too loud.

• Open Audio Analyzer remembers what you set up. The frame rate, the delivery target, the skin, the signal source and the arrangement on the canvas all survive quitting. The window reopens on the layout it was closed on, listening to the device it was listening to.

• Presets. Save the arrangement under a name, open it again later, delete it. One JSON file per preset in a documented directory, so a preset can be sent to somebody, dropped in from a forum post or kept in version control — and one corrupt file costs one preset rather than the library. A preset optionally carries the delivery target and skin it was saved with; leaving either out means “follow whatever is selected”, which is what makes a layout reusable across jobs.

• A delivery-target editor. Any spec Open Audio Analyzer does not ship — a label’s house standard, a game platform’s submission requirement — is now twenty seconds of typing rather than a feature request. User targets appear beside the built-ins everywhere, and a user file carrying a built-in’s id replaces that built-in, including in presets that already name it. Deleting the file brings the original back.

• Skins. The palette is thirteen named roles in a JSON file. A skin may set as few as one of them and inherit the rest, so changing the accent colour is a three-line file. Ships with Precision Instrument and Daylight, a light palette for a room with a window in it. “Duplicate for editing” writes the active palette out in full as a starting point, and “Reload from disk” picks up an edit without a restart.

• A settings panel, reachable from the status bar: signal and capture device, refresh rate and delivery target, skins, and whether the last layout is restored at launch. It also names the directory everything is kept in.

• The remote display’s name, port and frame rate are remembered. Whether it is publishing is not, deliberately: configuration is worth remembering, and the decision to open a port with no password on it is worth asking for every session — a laptop carried somewhere else must not start advertising itself because somebody enabled it once at home.

• The macOS build is no longer sandboxed. It was, which put your settings, presets, skins and delivery targets inside ~/Library/Containers/dev.openaudioanalyzer.oaa/Data/… instead of ~/Library/Application Support/Open Audio Analyzer, and stopped OAA_CONFIG_DIR from pointing anywhere outside that container. Configuration you cannot find is configuration you cannot edit, mail to somebody or keep in version control, which is most of the point of it being files. Open Audio Analyzer gives up Mac App Store eligibility, which was never planned; notarising the dmg does not need the sandbox. If you ran an earlier build, your existing configuration is in the container path above — move it across, or it will look as though Open Audio Analyzer forgot everything.

• The settings panel prints the configuration directory as selectable text rather than a path you would have to retype.

• The capture device is reopened by name when its id no longer matches. Device ids are not stable across reboots on any of the three platforms, so an id-only lookup silently drops you back to the test tone after a restart with no explanation.

• Keyboard shortcuts, and a sheet that lists them. Press ? or F1, or the ? in the status bar. Arrow keys nudge the selected module a cell and Shift+arrows resize it; Ctrl/Cmd with N, D, Z, T, R, O, P and , add a module, duplicate, undo, open a tab, restart the measurement, analyse a file, open presets and open settings. Open Audio Analyzer draws its own chrome and so has no menu bar to read a chord off, which is why the sheet exists.

• --config-dir names where Open Audio Analyzer keeps settings, presets, delivery targets and skins, and beats the OAA_CONFIG_DIR environment variable. On macOS it is the only one of the two that works on an installed .app: passing an environment variable means launching the binary inside the bundle, which changes how the system attributes the microphone request, so the variable and device capture could not be used in the same run.

• --open-panel=<name> opens one panel once the window is up, for settings, presets, calibration, report or shortcuts. Debug builds only; a release build says so rather than ignoring it.

• Installers for all four desktop targets — dmg, msix, AppImage and flatpak — plus the oaa analyser as a standalone binary, published on every tag.

• A documentation site, built from the Markdown in this repository and published from main. The keyboard page is generated from the same table the application binds, and a test fails if it has drifted.

• An application icon, at every size the four installers ask for.

• The icon now covers iOS and Android too. Both platforms were still shipping Flutter’s default logo — the blue chevron on white — so the app installed on a phone or tablet under somebody else’s mark. iOS gets the full layered icon described below; Android gets an adaptive icon, so the launcher masks it to whatever shape it uses and Android 13’s themed home screen has a monochrome layer to tint instead of shrinking the icon inside a grey circle. There is a Play Store icon in packaging/android/ for the console to be given by hand.

• The macOS and iOS icon is layered, so the system lights it. Both platforms now get an AppIcon.icon document — the graphite ground and the bars as separate layers — instead of a folder of pre-composited PNGs. macOS 26 and iOS 26 render it with their own specular highlight and shadow, and derive the dark and tinted appearances from it; on a themed or dark home screen the icon now follows instead of staying light. Older systems are unaffected: the same document still produces a classic .icns back to macOS 10.15 and flat icons back to iOS 13.

⚡ Changed

• The icon’s bars no longer climb in order. They were 0.34, 0.55, 0.74, 1.00 — which is the cellular signal glyph, drawn that way in the status bar of every phone the icon was about to appear on, and the shape is what the eye reads rather than the colour. They are now 0.62, 1.00, 0.44, 0.80: up, peak, valley, up, which is a meter. The bars also have slightly rounded corners, the tile’s ground is a diagonal gradient from hairline to background instead of flat graphite, and the hairline border is gone — it was a one-pixel detail that iOS masked off, Android cropped, and 16 px could not draw. Every platform’s artwork is regenerated from the same change.

• The application is now called Open Audio Analyzer. It was Bel. Nothing it measures changed — every reading is identical to the previous build — but the name it installs under, the directory it keeps your configuration in, the command-line binary and the protocol the tablet and the plugin speak all moved with it. The four below are the ones that can cost you something.

• Your settings, presets, delivery targets and skins are not carried across. Configuration now lives in ~/Library/Application Support/Open Audio Analyzer on macOS, in $XDG_CONFIG_HOME/oaa or ~/.config/oaa on Linux, and in the matching path on Windows. The old directory is left untouched and is never read; copy its contents over before the first launch to keep what you had.

• This installs beside the previous version rather than over it. The application identifier is now dev.openaudioanalyzer.oaa, so every installer and package manager treats this as a new application. Remove the old one by hand if you do not want both.

• A tablet or a plugin from an earlier release will not connect. The wire protocol is at version 2: the frame magic spells the new name, and a host advertises _oaa._tcp rather than _bel._tcp. A mismatched peer is refused at the handshake rather than drawing wrong numbers, which is the failure that matters — but both ends have to be updated together.

• The command-line analyser is oaa, not bel. Any script or CI step that calls it needs the new name; its arguments and its exit codes are unchanged.

• The status bar wordmark reads OAA, and the gap between it and the source picker is one step tighter. OAA sets about 3.4 px wider than the mark it replaced, which on its own was enough to run that row past its edge at 1000 px with the longest delivery-target name.

• Every number is set in Google Sans Code instead of JetBrains Mono. The advance is 0.6 em in both faces, so nothing moves and no readout changes width; the digits are a little smaller on the body and rounder in the bowls. One character the old face carried is missing from the new one — ∞, which a reading only shows if it is not finite, and nothing the engine produces is — so it now falls back to Inter, which Open Audio Analyzer already bundles, rather than to whatever the host offers. Both faces remain SIL OFL 1.1 and their licences still ship with every package.

• Undo and redo in the tab strip carry a mirrored arrow beside the word. The two words differ by one letter in the middle, and a mirrored pair says which way it goes before either has been read — but the arrow alone left the row’s two most-used controls unnamed, so it now punctuates UNDO and REDO the way the plus punctuates + MODULE. The arrows are drawn, like every other mark in Open Audio Analyzer, so they are the same on every platform and cost no dependency. The keyboard shortcuts are unchanged.

• The delivery target in the status bar is built like the buttons beside it, one step quieter. Same height and the same capitals, because it opens a menu on a click exactly as the four buttons to its right do and a control that can be pressed should look like one. Its border stays the fainter of the two hairlines, which is what still tells the thing that reports a setting from the four that do something. The target’s own name is unchanged everywhere else: the menu, the settings panel and every report print it as it was typed.

• A remote display’s modules have no menu button. The title bars drew one and it did nothing, because there is nothing on that screen a viewer is allowed to change. It is gone rather than disabled.

• The stereo cloud’s centre line is drawn over the frequency axis. The three horizontal guides crossed in front of it, which broke the one line the module exists to mark into what looked like a dashed one.

• A rule separates the tab strip’s two kinds of action. UNDO and REDO step back and forward through what has been done; + MODULE does something new. They were four controls of the same size, colour and weight in one run.

• A remote display’s link bar has room to breathe, and its tabs are beside the host’s name. The bar is 48 px rather than 40, so the tab picker and Disconnect are not pressed against the rule under them, and the tabs now follow the name of the machine being watched instead of sitting in the far corner — which on a tablet is the most awkward place on the screen for the control the viewer touches most. Disconnect stays on the right, alone, where it is not hit by accident.

• Both pluses in the tab strip are larger. At the size the words use they read as specks rather than as controls. The one beside the tabs is the larger of the two, because it is the only symbol in the strip carrying an action with no word to help it; the one in + MODULE is set between the two, visible without competing with the word it belongs to.

• The keyboard sheet is one screen again. Seventeen shortcuts in a single narrow column were taller than the panel, so the sheet scrolled and cut the line explaining that Ctrl and Cmd are interchangeable in half. It is now two columns on a wider panel — Canvas and Measurement on the left, Tabs and Configuration on the right — with the rows further apart and the whole list on screen at once, down to the smallest window Open Audio Analyzer supports. Below that it stacks back into one column rather than clipping. Measurement now comes before Tabs on the documentation site’s keyboard page as well, which is the same ordering.

• The remote panels are marked rather than only worded. Sending and receiving were two rows of the same shape whose only difference was the sentence in them; each now carries a mark — a machine broadcasting, a screen on a stand — and a chevron on the rows that open a panel rather than choose in place. Every host a search finds wears the broadcast mark too, and it brightens under the pointer instead of sitting in the same grey as its address. A note that is a warning — the link has no password, this device cannot search the network, publishing failed — now has a warning mark in the margin beside it, so it is a different kind of line rather than a differently coloured one. The marks are drawn rather than typeset, so none of them can arrive as a tofu box on a platform whose fonts differ. Sending and receiving also sit further apart than the rows of a list do, since they are two directions to choose between rather than entries to pick from.

• A machine’s name, its port and the Apply that commits them are one line. They were three stacked rows, so a two-field form read as three separate settings and the button that finishes it sat a row below either field. They now share a line, which the panel has room for.

• A segmented control sets its choices in capitals, like every button beside it. Source, refresh rate, the remote update rate and a display’s tab picker were the only controls in the interface labelled in sentence case, which read as a line of prose in a box rather than as something to press. What a menu or a field holds — a device, a delivery target, a name somebody typed — is unchanged, because that is a value rather than the control’s own word.

• The Histogram’s loudness line is always on screen. It used to be drawn only over the columns the programme had reached, so an empty module — before the first audio, and for as long as you looked at it after a reset — had nothing in it at all, and a part-filled one had a line over part of its width. The line now rests on the floor of the scale and runs the full width of the plot, rising where the programme starts. Resting is not a reading: the floor is the bottom of the scale, nothing is filled beneath it, and it is drawn where measured silence would put it.

• Nothing is a double click any more. Renaming a tab, adding a module by clicking empty canvas, and zooming the window from the status bar were all double clicks; the first two are now a long press — which also gives a tablet a tab’s rename, duplicate and delete for the first time — and the window is zoomed with the green window button, as it always could be. See 🐛 Fixed for why a double click was worth removing.

• Moving a module dims the rest of the canvas, and the placement grid now has a border. The grid is ruled inside a rounded border that sits one gutter outside the modules, with the same corner radius they have, and for as long as the pointer is down every module except the one in hand is washed toward the canvas colour. A drop target among a dozen meters that are all still moving was something you had to hunt for; the cells, the module being carried and where it will land are now the only things at full contrast. Nothing is measured differently — the meters underneath keep updating throughout.

• The phase scope’s trail no longer smears. It was a picture faded and redrawn on every frame, so a moving dot was resampled once per frame and spread outwards; it is now the last forty frames of samples, each drawn at the brightness its age has earned. A dot fades where it was rather than blurring, and the decay is exact instead of compounding through the rounding of an 8-bit surface — the tail is fractionally longer for the same reason.

• The stereo cloud is accumulated as numbers rather than as a picture, in cells of two logical pixels, with each band’s dot spread across the four cells it falls between. The shape, the fade and the brightness are the same; a close look finds dots on a two-pixel grid where they were previously at arbitrary positions.

• A module’s readings now grow with the module. The bars, the arcs and the VU face have always been sized off the tile they are in; four modules sized their numbers off a constant instead, so the LUFS meter’s LUFS-I and LRA, the Super Meter’s centre readout, the Alert Meter’s value and the Validator’s measured column stayed the same size whether the module had a corner of a laptop screen or a quarter of a 27” one. The labels beside them — a scale’s ticks, a column heading, PASS and FAIL — deliberately do not scale; they are the same size in every module on the canvas.

• The smallest supported window is now 960x768, up from 720x480 (macOS; the other platforms set no minimum). The canvas is a fixed 24x16 cells at every window size, so at 480 px tall a two-row module had 12 px of body left after its title bar and margin — less than a digit. 768 is the height at which the smallest module in the default preset still has room for its number.

• A module too small to draw in now says so instead of showing an empty panel. The size a module needs is in pixels, not grid cells, so it was never something the cell minimums could enforce; each painter checked it privately and drew nothing when it failed. The thresholds are now declared on ModuleKind and the frame substitutes the “too small” placeholder, which is what it was always for.

• The status bar drops items in a stated order as the window narrows, one at a time and each at the width below which the rest stop fitting: first the sample-rate readout, then the OAA wordmark, then ANALYSE FILE, then REMOTE, then the ? button. All four buttons keep their keyboard shortcuts, which are listed in the ? sheet and in docs/site/keyboard.md. What never drops is the source, the elapsed clock, the delivery target, SETTINGS and RESET. There was one gate before this, it was 20 px too generous, and nothing had measured it.

• The settings panel has been reworked, and every other panel with it. Sections are now ruled off from each other with a hairline instead of by whitespace alone, so Signal, Meters, Appearance and Session read as four groups rather than one column of grey; a row’s explanation runs the full width of the panel underneath its control instead of wrapping in whatever space the control left over; and the loopback advice under Capture device is two lines rather than four. The same changes reach the preset browser, the delivery-target editor, the analysis report, the remote-display panel and the keyboard sheet, because all six are built from the same primitives.

• REMOTE now asks which end of the link this machine is. Pressing it opens a chooser — send these meters, or show another machine’s — and each answer opens its own panel. Publishing used to be the whole of what the button offered: turning this screen into a display was a footer button on the publishing dialog marked “Use as display”, which is the row a panel reserves for the ways out of it. The chooser also says whether this machine is already publishing and how many displays are attached, so that is answerable without opening anything further.

• The remote display’s connect screen is a panel like every other panel. It was the one screen in Open Audio Analyzer that had never been near the design system — an unstyled list, a stock text field with a rounded outline and Material text buttons, on the hardware the feature exists for. Discovered hosts are panel rows now, the typed address is an Open Audio Analyzer field with Connect in the footer, and it is the same panel the desktop opens rather than a second implementation of it. The strip across the top of a live display gets the status bar’s fill and hairline, with the tab picker and Disconnect as controls rather than as text. Disconnecting returns to the picker instead of to a dead screen.

• Every boxed control in a panel is the same height. Buttons, menus, segmented controls and text fields derived their heights independently and came out at 30, 32, 31.4 and 28.9 px, so no two standing side by side in a row ever quite lined up. They are all 32 px now.

• A menu looks like a menu. The capture-device and delivery-target pickers were bordered labels in caption grey — fainter than the buttons beside them, which reads as disabled — with nothing to say they opened anything. They now carry a caret, show their value in the same weight as the rest of the panel, highlight under the pointer, mark the current choice in the open menu, and can be reached and opened from the keyboard, which neither of them could before. The menu also no longer arrives with a Material drop shadow.

• The selected skin is visible as selected. Selection in a panel list was carried by a single step of grey on a 1 px border; it is now a raised fill as well.

• The unfilled part of a meter is now visible. meter_track sat at 1.10:1 against the panel behind it on the dark skin and 1.22:1 on the light one — close enough to the surface that a bar showed its own fill and nothing else, and the Super Meter’s three arcs could only be located by whichever one happened to be lit. Both shipped skins now hold it at roughly 1.6:1, and still about 2.5:1 below meter_fill so the reading stays the figure and the track stays the ground. A user skin that sets meter_track keeps its own value; a skin that inherits it gets the new one.

• Modules have room to breathe, and the same amount of it on every side. The inset between a module’s border and its meter went from 8 px to 12 px, and the title bar’s inset moved with it so the title still starts where the meter starts. Meters draw to the edges of what the frame hands them — a painter that adds a second inset of its own is a module that sits differently from the other eleven.

• A dB scale no longer reserves more room than it uses. The gutter beside a graticule was a flat 30 px whatever the labels said, so a meter with short labels sat visibly off-centre in its module — thirteen pixels of empty reserve on the scale side against nothing on the other. It is now measured from the labels themselves. Affects the LUFS meter, the digital meter, the spectrum analyser and the histogram.

• The VU dial is sized to the module rather than to a fixed sweep. The face opened 70° whatever shape the tile was, which in a wide tile drew the whole instrument across the middle half of the width and left the rest bare. The sweep now opens as far as the box allows, between 70° and 110°, and the dial is centred on what is actually drawn.

• The frame rate is no longer in the status bar. It was a second way to reach one setting, sitting in a row otherwise reserved for what changes while you work and what a reading has to be read against. It is chosen once for a machine; it lives in the settings panel and only there. The delivery target stays in the bar, because every PASS and FAIL on the canvas is a verdict against it.

• A panel that scrolls now says so. Settings is taller than the 760 px a panel is allowed and ended on a row the viewport happened to cut in half, with nothing to suggest there was more below it. A scrollbar appears when the content overflows and stays hidden when it does not.

• The REMOTE button now looks like the rest of the status bar. It was the one stock Material button in the row — borderless where its four neighbours are bordered, Material-sized rather than bar-sized, ink-rippled, and not reachable by keyboard. It now carries the same border, padding, type and focus ring as ANALYSE FILE, SETTINGS and RESET, and states what it is doing on hover. Publishing is shown by the label brightening rather than by the signal hue, which in this row means “in spec” and nothing else. The panel behind it is built from the same pieces as the other panels.

• The macOS window has no title bar of its own. The status bar now runs to the top edge of the window, and the close, minimise and zoom buttons sit inside it on the same row as the source and the elapsed clock. The strip of system grey above a bar of panel grey, and the window title printed in a font Open Audio Analyzer does not choose, are both gone. Dragging the status bar moves the window and double-clicking it zooms, as dragging and double-clicking the title bar did. Windows and Linux are unchanged.

• A light skin no longer runs under dark window buttons on macOS. The window follows the skin’s light flag, so Daylight gets light chrome and the two stop disagreeing about which way up the room is.

• The signal hue now means one thing on the canvas. Teal previously marked both “in spec” and “selected”, so a selected module was outlined in the same colour as a reading that had passed its target, a few pixels from the reading itself. Selection is now a brighter, heavier border; the active tab, the highlighted menu row, the resize grip, the listening indicator and the drop preview all moved to neutral values. Teal now appears on the canvas only when something is within its delivery target.

• The em dash that means “not measured” is drawn in a legible colour. It shared a value with scale ticks and disabled controls, at 2.81:1 against the panel against readings at 15:1 — the one mark in the interface that says a quantity was not measured was the hardest one to see.

• Selection, hover and focus borders are visible. hairline_strong was 1.47:1 against the panel — a role whose whole purpose is to be seen, set to a value that could not be, which is why callers reached past it for the accent. It is now roughly 3:1 in both shipped skins. A user skin that sets hairline_strong keeps its own value and is unaffected.

• RESET now states its scope on hover: it restarts the measurement, and leaves the layout, target and skin alone.

• The meters cap themselves at 30 fps when the system asks for reduced motion, and the settings panel says so rather than showing a rate nothing is running at. Open Audio Analyzer has no decorative animation to switch off — what moves is the measurement — so a lower redraw rate is the only honest reading of that preference. No reading is withheld.

• The main view is an arrangeable canvas instead of a fixed wall showing every metric at once. It opens on six readings — LUFS-M, LUFS-S, LUFS-I, LRA, TP Max and Peak Max — and the rest of the canvas is yours.

• A module’s default measurement now follows its kind rather than always being integrated loudness, so a freshly placed Alert Meter watches true peak — which is what an alert is for. A saved layout that omitted the key reads back differently for alert meters.

• The signal source is a persisted setting rather than something the status bar holds. Two controls can change it now — the bar and the settings panel — and both write to the same place, which is also what lets the next launch reopen it.

• The interface no longer has a single compile-time palette. Everything that draws takes its colours from the active skin.

• The delivery-target menu lists the user’s own targets alongside the built-in ones.

• The DAW plugin ships as VST3 and Audio Unit rather than the CLAP the plan named. CLAP’s SDK is the nicest of the three, but Ableton Live does not host CLAP, and a metering plugin that cannot be inserted in Live is one most people cannot use.

• Keyboard shortcuts work wherever focus is. They were installed inside the canvas and stopped working the moment focus left it — clicking the source picker was enough to silently disable undo. They now wrap the whole workspace.

• The tab strip’s + sits beside the last tab and scrolls with the tabs, rather than being carried to the far end of the strip.

🐛 Fixed

• A remote display recovers from a dropped link, instead of showing one frame of invented measurement and then never coming back. A connection that died partway through a measurement left the head of that frame in the display’s frame reader, which nothing cleared — so the retry reassembled the dead stream’s bytes onto the live one. That splice is exactly the right length to decode, so the meters drew a detailed, confident reading nobody took, lost sync on whatever followed, and dropped the link again, carrying the leftovers into every attempt after that. A tablet that lost a single frame to a Wi-Fi hiccup stopped recovering until the app was restarted — Disconnect and reattach did not clear it either. Nothing about how Open Audio Analyzer measures has changed and no past reading needs re-checking: the desktop and the CLI were never on this path.

• The status bar’s controls are one height. The delivery target read 25.4 px tall against 22 px for the four buttons beside it, because each shape derived its height from its own text style plus its own padding rather than being given one. The borders in the row now start and end on the same pixel.

• A warning mark sits in the middle of the note it marks. It was pinned to the top of the block, so beside the sending panel’s two lines of orange it looked as though it belonged to the first line rather than to the warning.

• A long capture session no longer dies with a bus error. The sub-block ring the loudness measurements are built from carried a write cursor beside it, and the store into the ring trusted that cursor. One was found holding a float bit pattern after 36 minutes of capture, which turned an ordinary 100 ms boundary into an 8-byte write 229 GB past the engine and killed the process. The row is now derived from the sub-block count where it is used, so the write lands inside the ring whatever the counter holds. Every reading is unchanged, bit for bit, and the EBU conformance vectors still pass. What put a float there has not been identified; this stops the engine turning it into a wild store.

• A capture device is released when the engine fails to start. If an allocation failed after the device had been started — sizing the analysis buffer, or the spectrum’s transforms — the engine was freed while the real-time callback was still writing into the ring inside it, handing that callback a dangling pointer into a block the allocator was about to reuse. Startup now tears the device down on every failing path.

• An iPad can now find hosts on the network. It never could: iOS and iPadOS refuse an app the multicast socket Open Audio Analyzer browses with unless it carries an entitlement Apple grants per developer on request, so every send was rejected and nothing was ever delivered — on the hardware the remote display exists for. The tablet now searches through the system’s own Bonjour responder, which needs no entitlement, and finds the same hosts by the same name. Typing an address still works and still always will.

• A host on a network that hands out a domain is visible again. Most routers give DHCP clients one, which makes the machine’s name studio-mac.fritz.box rather than studio-mac.local — and Open Audio Analyzer advertised that whole string where DNS-SD allows a single label, so the announcement came out as six labels instead of four. Open Audio Analyzer’s own browser read it anyway, so an Open Audio Analyzer desktop found an Open Audio Analyzer desktop and nothing looked wrong; every browser built on the system responder — an iPad, dns-sd, anything Apple — dropped the record, and the host answered every query on the wire while appearing to no one. The instance is now one label whatever the machine is called, and the name you read in the list is unchanged.

• Open Audio Analyzer no longer advertises an address record for the machine’s own name. The name a Mac answers to belongs to the system responder, which defends it: an address record it did not publish, for a name it owns, is a conflict, and the loser of a conflict renames itself. On a machine with a second network interface the two sets differ, so the loser would have been the user’s computer. Open Audio Analyzer publishes its own name for the service to point at.

• A search that cannot run says why. A device that could not search showed Looking for hosts on this network… indefinitely, which is exactly what a network with no hosts on it looks like. It now names the reason where there is one to name — a refused Local Network permission on macOS points at System Settings — and says plainly that it cannot search where there is not.

• A laptop in a dock is listed at an address that answers. A machine with a second network interface — an empty dock, a Thunderbolt bridge — announces every address it has, and the browser kept whichever arrived last. Half the time that was the interface’s self-assigned 169.254 address, which nothing on the network can reach, so the host appeared in the list and the connection timed out. A routable address is now preferred, and a host that moves replaces its old address rather than adding to it.

• A panel follows a change of skin while it is open. Choosing one in Settings → Appearance repainted the canvas, the window chrome and the panel’s own text fields and menus, but left the panel’s surface, hairlines, labels and the dimming over the canvas in the previous skin until the panel was closed and reopened — so the one place a skin is chosen was the one place it could not be seen, and the panel was drawn in two skins at once meanwhile.

• A two-finger trackpad gesture no longer drags things. Right-clicking a module’s title bar on a trackpad flashed the placement grid on screen, because a two-finger tap is how macOS sends that click and every drag in the application accepted a trackpad gesture as one: a two-finger scroll over a title bar moved the module, over the corner grip resized it, and over the status bar it started dragging the window. A drag now begins from a button press and from nothing else. Clicking and dragging on a trackpad is unaffected — that is a mouse as far as the system is concerned.

• The top of a meter’s scale is no longer cut in half. The 0 on the LUFS meter was drawn as its own bottom half, and the same line of code clipped the spectrum analyser’s top label and the first and last labels on the histogram’s frequency axis. An end label now sits fully inside the meter; its gridline has not moved.

• The M and S under the LUFS meter’s bars are drawn at all. They were centre-aligned in an unconstrained line box, which put each letter half a megapixel to the right of the meter — so the two bars had nothing naming them and the space for the names was still reserved beneath them.

• The LUFS meter’s two readings are no longer printed with their last digit missing. They were sized off the module’s height alone, so a tall narrow meter asked for digits wider than the column under the bar they belong to and the reading stopped drawing where it ran out of room: -17.6 was shown as -17., which reads as a different number rather than as a clipped one. Both now take the largest size that fits the column as well as the height, and are hidden — as they already were on a short module — when that size would be too small to read.

• The super meter’s ring names sit beside the arcs they name. M, S and I led their arcs by a fixed angle, which is a fixed fraction of each radius, so the outer name stood nearly twice as far from its arc as the inner one did from its and the three read as a diagonal drifting off the gauge. They are now the same distance from every arc.

• Buttons no longer take a third of a second to respond. Every control in the status bar on macOS, every tab, and clicking empty canvas to clear the selection fired 300 ms after the click that pressed them. Each sat under a gesture that also recognised a double click, and Flutter’s double-tap recogniser holds the gesture arena from the first tap until it times out — so the button’s own tap could not be resolved until the wait for a second click had expired. The double clicks are gone (see ⚡ Changed) and the controls answer on release.

• Open Audio Analyzer crashed after a few minutes with a spectrogram, phase scope or stereo cloud on the canvas, and grew without bound until it did. One report showed the application holding 266 GB before macOS stopped it. Those three modules accumulated their history into an image taken with toImageSync, which retains the display list that drew it for as long as the image lives — so every frame’s image pinned the frame before it, back to the first one, and disposing the handle released none of it. The application then died on the raster thread when that chain was finally dropped and its destructors recursed once per retained frame, 3,286 deep in the report that found this. All three now keep their history as data and redraw it, which costs memory proportional to the module’s size rather than to how long Open Audio Analyzer has been open.

• A spectrogram opened before any audio arrived began with a column of full scale, and the stereo cloud with a bright line down its centre. Both were reading the zeroed arrays of a source that has not published yet, which as dB is 0 dBFS on every band. A source that has measured nothing now draws nothing.

• The LUFS meter’s two readouts sat under the wrong things. They were laid out against the middle of the module, but the bars begin after the scale’s gutter — so LUFS-I printed under the scale numbers and LRA under the middle of the momentary bar. Both now line up with the bars they sit below. No reading changed; the two numbers are the same numbers, in the right column.

• Six Number Boxes rendered as empty panels on a small window. On anything under about 700 px tall a two-row Number Box had less body than a digit is tall, and the painter’s own size check meant it drew nothing rather than saying so. Every module now shows the “too small” placeholder instead, and the supported minimum window is large enough that the default preset never reaches it.

• The status bar ran 121 px past the edge of the window at the smallest size the window could be dragged to. In a debug build that is a striped overflow warning; in a release build the controls past the edge are simply not there.

• The signal source overflowed its own place in the bar whenever the name of the capture device came close to filling the room the bar had left for it — a striped warning across the source label in a debug build, and a name clipped without a mark in a release one. The name now shortens with an ellipsis to whatever the row can spare, as the delivery target beside it already did.

• The Stereo Cloud looked broken on a mono source. A built-in laptop microphone has one channel, per-band stereo position needs two, and the engine reports mono as dead centre — so every band plotted at the middle of the display and the module drew one bright vertical line and nothing else, which reads as a rendering fault rather than as a mono signal. It now says MONO SOURCE across the face and leaves the axis empty. No measurement changed; a stereo source draws exactly as before.

• Selecting a capture device did nothing. Only the source chosen at launch ever opened; every change made afterwards — from the status bar or from Settings, to a microphone, an interface, a loopback device, Silence or back to the test tone — was discarded. The choice was saved and reappeared as selected the next time you looked, so the meters looked like the failure: they went on showing the previous source, with its label, its channel count and its elapsed clock still running, exactly as though no input were reaching the machine. Each attempt also opened the device and held it with nothing reading it, which is why the system’s recording indicator came on while the meters stayed on the test tone. Relaunching with the device already selected was the only way through, and is no longer needed.

• Every panel’s four corners were missing their border. The surface is a rounded clip over a square border, so the clip removed the corner of the hairline along with everything else outside the arc: the border ran the flat edges and stopped dead at each tangent, leaving four bare arcs of panel fill fading into the barrier. The border now carries the same radius as the clip and runs continuously around the corner. Settings, the preset browser, the calibration editor, the report, the remote display panel and the shortcuts sheet are all built from that one surface, so all six are corrected — as is the startup notice, which had the same defect at a smaller radius.

• The elapsed clock sat two pixels high in the status bar. It is painted rather than built, and it was drawn from the top-left corner of a box taller than the line it holds, so every spare pixel fell below the digits and the clock rode above the optical centre of every label beside it. It now centres its line in its box, which is what a Text does with the space it is given.

• Everything in the tab strip’s action row sat a pixel below the tab names — most visibly the + that adds a tab, which stands directly beside the last one. A tab reserves the height of the active-tab rule whether or not it is the active tab, and the buttons reserved nothing, so the two rows of text were centred in boxes of different heights. The + was low for a second reason as well: a lone plus is drawn on the font’s math axis, below the middle of the band the words around it fill, and it is now raised onto their line.

• The LUFS meter’s two bars, and the digital meter’s channels, shared one background. The gap between bars was painted in the trough colour, so it only showed where one bar’s fill had risen past the other’s: two channels at the same level read as a single wide bar, and two at different levels as one bar with a step in it. Each bar now has its own trough and the gap shows the module behind it. The scale, the target band and the integrated rule still cross it — they belong to the meter, not to either bar.

• The VU meter’s face had four drawing defects. The needle was drawn from behind its own pivot with a tail longer than the cap meant to hide it, so a second short needle stuck out of the bottom pointing the other way. At rest it lay along the −20 mark and struck that label through — and every reading on the lower half of the face had the needle across it, because the labels were inside the sweep. The six scale numbers were centred on a common radius rather than cleared from the arc by a common gap, so they scattered, with 0 hanging below its own tick. And the dial sat high and left in the tile with the VU badge adrift in the empty corner. The needle now runs outwards only, the labels sit outside the arc where nothing can cross them, each one clears the arc by the same gap whatever its angle, and the badge sits under the pivot where the needle cannot reach. The ballistics and the scale are unchanged — the needle points at the same mark it did before.

• Four modules drew their contents in a corner instead of in the box. A Number Box put its reading against the left edge, which on an unavailable reading left a single em dash alone in the corner of a four-cell tile, looking like a rendering fault rather than “not measured yet”. An Alert Meter hung its block from the top edge of a module more than twice its height. The Super Meter centred itself as though it were a full circle when it opens 120° at the bottom, leaving a dead band a fifth of the module deep beneath it. The Validator stopped its rows at 34 px each and left the rest of the tile blank.

• A one-row Number Box drew a title bar and nothing else. A single grid row is about 55 px, and the title bar and the module’s inset account for all of it, so the reading had no room and the painter returned without drawing — blank, rather than the “too small” a module says when it cannot be read. The minimum is two rows now, and a stored layout holding a one-row box is clamped up rather than rejected.

• The delivery-target menu did not show which target was selected. Both arms of the ternary that picks the row colour returned the same value, so every entry was drawn identically and the active target was indistinguishable from the five it was listed with.

• The right-hand end of the status bar drifted away from the right edge. The elapsed clock, the calibration and the four buttons sat progressively further from the window’s right side the wider the window got, leaving a growing empty stretch of bar beside them. They are now flush with the edge at every width.

• The remote display panel could not be opened. Pressing REMOTE in the status bar threw “No OaaTheme in scope” and left the panel unbuilt, in release as well as debug — it was pushed with showDialog, so the Navigator built it above the palette the application provides. It is now pushed with showOaaPanel like the other five, and a test opens it through the button.

• Open Audio Analyzer could not be built for iPadOS at all. The engine was compiled as C on every platform, but miniaudio’s Core Audio backend is Objective-C on iOS — it configures an AVAudioSession, which has no C interface — so the build ended in several hundred errors inside Apple’s Foundation headers, none of which named a file in Open Audio Analyzer. The iOS build now compiles as Objective-C and links the frameworks miniaudio needs there. The tablet build is the remote display, and it now runs on an iPad.

• The iPad build would have been terminated by iOS the first time an input was chosen. NSMicrophoneUsageDescription was missing from the app’s Info.plist, which is not an error the app can catch — the system kills a process that touches the microphone without one.

• An iPad remembered nothing between launches and opened with “no configuration directory”. Open Audio Analyzer resolved its configuration from HOME, which iOS does not set — so every layout, skin and connection was lost when the app was closed. iPadOS now keeps its configuration in Library/Application Support/Open Audio Analyzer inside the app’s own container, which Open Audio Analyzer locates through the temporary directory rather than the environment. An Android tablet still persists nothing, and still says so at launch.

• A module’s resize grip was drawn outside the module. Both ticks ran to the corner of the module’s slot, which is past the frame’s rounded border — so they crossed the border and finished in the gutter between modules, and on a selected module they cut through the selection outline. They now sit inside the panel in both states; the area you can grab is unchanged.

• Selecting a module broke its own outline in two places. The rule under the title bar ran the full width of the frame and was painted after the border, so it printed the dim hairline colour over the bright selection border at each end. The rule now stops at the border’s inner edge.

• Nothing in a panel could be reached from the keyboard. Every control Open Audio Analyzer paints itself — buttons, toggles, segmented controls, list rows, icon targets — was a bare gesture detector, so none of them took focus, none responded to Enter or Space, none drew a focus ring, and none was visible to a screen reader. Settings, presets and the delivery-target editor were mouse-only with nothing on screen to say so. All of them are now focusable, keyboard operable, and announced.

• Opening the host picker a second time searched a browse that had already been shut down. A tablet’s search runs over one channel, and a channel keeps one subscription: opening a second picker while the first was still on screen — which happens for a frame every time one replaces another — ended the first one’s browse, and the first one’s teardown then ended the second’s. The panel left on screen said “Looking for hosts on this network…” over nothing that was running, found no host however long it was left, and logged “No active stream to cancel” when it was closed. The search is now shared: one browse for the application, handed to whoever is looking, and stopped when the last of them closes.

• Changing the signal source could replace every meter on the canvas with a red error box reading “A MeterClock was used after being disposed”. The old clock was torn down before the new one was installed, so any painter still mounted for that frame — which is all of them — tried to attach to a disposed notifier. Disposal now happens after the frame that replaces it.

• Tabs were clipped off the end of the strip while it was visibly half empty. The tab area and the gap before the action buttons were two flex children of equal weight, so the strip gave half its free width to the gap. Three tabs in an 800 px window were enough to lose one.

• Renaming a tab from its context menu opened a field that would not accept typing. The menu’s route restored focus to the canvas after the field had claimed it. Double-clicking a tab was never affected, which is what made this hard to see: the two ways in behaved differently.

• Every keyboard shortcut stopped working after a text field closed, until something was clicked. Focus fell back to the navigator’s scope, which sits above the bindings, and a key event travels up from the focused node — so there was nothing below it to reach.

• Closing the host picker on a tablet while it was still searching threw. Ending the Bonjour browse suspends on the channel, and the rest of the stop — which publishes an empty list of hosts — ran after the picker had already disposed the notifier it publishes to. It surfaced as “A ValueNotifier<List<DiscoveredHost>> was used after being disposed” in the log, most reliably on a hot restart. The search is now published as ended before anything is awaited, and teardown releases the browse without publishing at all.

• A display leaving mid-frame threw in the host it was leaving. Closing a socket that is still flushing a measurement is a StateError in Dart’s own I/O — “StreamSink is bound to a stream” — and it was raised inside the socket’s own close notification, where nothing could catch it. Every tablet that dropped off at the wrong moment logged what reads as a crash and left the connection it was there to give back.

• Changing the skin, the layout or the delivery target could disconnect a display. The new setting goes out the instant it changes, and a socket still flushing a measurement refuses to be written to — which the host could only read as “this display has gone”, and act on. It was likelier the slower the display was. A setting now waits for the frame in front of it and then goes, in order.

• Attaching to another machine flashed the host picker on the way in, and started a second search behind it. The link said it was connecting one turn after it began, and a display shows the picker until it does — so a screen that had already been told where to attach spent a frame drawing a panel asking where to attach, and a picker searches the network as soon as it is built.

• Closing the host picker while it was still opening its socket threw. Binding is real I/O, and the search could be shut down while it was in flight: what came back then was written to a notifier that had been disposed, and the socket it arrived with was never closed. On macOS the window is as long as the Local Network prompt is on screen.

• A panel row’s explanation no longer runs into the control above it. The caption sat two pixels under the row, and a row is as tall as whatever sits on its right — so under a bordered control it collided with the bottom edge rather than reading as a line of its own. The delivery target showed it worst: its note ended directly beneath the target menu and looked like part of it. It now clears the control, and still sits half as far from its own row as from the next one.

• The documentation site shows the current mark. Its header logo and favicon were a hand-copied version of the icon compiled into the site generator, so when the mark was redrawn they kept publishing the previous one on every page. The generator now reads assets/brand/oaa-mark.svg and fails when it is missing, so the site cannot hold a mark of its own again.

• The oaa CLI is published again. It was built with dart compile exe, which refuses a package whose dependencies have build hooks — oaa_engine has one — so no CLI reached a release at all. It is built with dart build cli now and ships as an archive of the executable and the engine beside it rather than as a single file; keep the two together. CI builds it the same way and runs the result, which nothing did before.

• The Windows package builds. Its manifest named the application Open Audio Analyzer where a package identifier is required, which the rename from Bel introduced and which fails validation with a line number and no mention of the name; the identifier is Oaa and the display name is unchanged.

• The Linux flatpak builds. Two things stopped it. Installing the runtime used a partial reference, which makes flatpak ask which of the matching ones is meant — a question -y does not answer and a CI runner cannot — so the job waited instead of failing. And the scalable icon was read through gdk-pixbuf by the AppStream step, which cannot decode an SVG unless librsvg has registered a loader; where it has not, a valid file is reported as an unrecognised format and the build fails with the package already assembled. The flatpak now ships the seven PNG sizes and no scalable icon, which is what a launcher uses either way.

🚧 Internal

• The install pages name the files a release actually publishes. Every installer is built as Open Audio Analyzer-<version>-… and GitHub replaces the spaces with periods when it attaches it, so the documented names — and the two shell commands printed for a reader to paste — matched nothing on the releases page. They are dot-separated now, and the difference is explained where it appears.

• Every identifier follows the name. The public C ABI is oaa_* and OAA_* in engine/include/oaa/oaa.h, the packages are oaa_core, oaa_ui, oaa_engine and oaa_wire, and the plugin’s classes are Oaa*. OAA_ABI_VERSION stays at 4: the header’s shape did not move, only its spelling.

• The release workflow reads OAA_SIGNING_IDENTITY, OAA_NOTARY_PROFILE, OAA_WINDOWS_CERT, OAA_WINDOWS_CERT_PASS and OAA_WINDOWS_PUBLISHER. The repository secrets have to be renamed to match, or every signed artefact quietly stops being signed.

• The repository, the documentation site and every download link moved to open_audio_analyzer.

• The cross-implementation wire golden is now plugin/test/golden/wire_v2.bin, regenerated from oaa_wire_fixture.

• The released sections below describe the application under its former name. They were renamed with the rest of the repository rather than left to contradict it; nothing else about them was rewritten.

• Stopping the engine decides whether to join on the flag that tracks the thread. It tested should_run — the flag stop itself clears — so anything that ever cleared it elsewhere would have turned destroying an engine into a free underneath a live analysis thread. It now tests whether a thread was started and not yet joined.

• PersistenceLayer is gone, and with it the last toImageSync in the application. PointBuckets replaces it: marks sorted by the colour they are drawn in, so a display of thirty thousand of them is a few dozen calls rather than thirty thousand. test/history_modules_test.dart fails if any of the three modules creates an image between frames again.

• packages/oaa_wire — the wire protocol, pure Dart and MIT, specified byte for byte in docs/WIRE.md. Three implementations speak it and none of them was written against another: the app’s host, the app’s display, and the plugin’s C++ sender. plugin/test/golden/wire_v2.bin holds the Dart codec against bytes the C++ actually produced, which is the only test that would catch the two drifting apart — and the drift is silent, because every frame is a fixed length, so a field written into the wrong slot still parses.

• MeterSource — the interface a meter module reads a measurement out of, in oaa_core. OaaEngine implements it and so does the remote display’s decoder, which is what lets the twelve modules run unchanged on a tablet with no engine in it. oaa_engine now depends on oaa_core for that one interface; the arrow still points away from dart:ffi.

• MeterClock decides what is new by comparing generations rather than by trusting what refresh() returned. With the remote host refreshing on its own timer there are two callers, and a one-shot “is this new” answer is consumed by whichever asks first — leaving the other to stop repainting, silently, only on the machines where somebody was using both screens at once.

• plugin/ is AGPL-3.0-or-later, not GPL-3.0. JUCE 7 and 8 are AGPLv3-or-commercial; only JUCE 6 offered GPLv3, which is what the plan was written against. Open Audio Analyzer takes the AGPLv3 option, which changes the licence of the plugin binary alone: the engine stays MIT, and the app stays GPL-3.0-or-later because it never links JUCE — it talks to the plugin over a socket. GPLv3 section 13 expressly permits the combination. One piece of good news the plan did not anticipate: Steinberg has relicensed the VST3 SDK to MIT, so JUCE is the only copyleft dependency and no separate SDK checkout is needed.

• engine/CMakeLists.txt builds liboaa as a static library for consumers that are not Dart. There are now two descriptions of the same compile, because a plugin CI runner has no Flutter SDK and a build hook cannot be handed to JUCE; plugin/test/sources_match.sh fails the build if they drift apart.

• The C++ and Dart implementations of the wire protocol are held against a committed golden that the plugin’s own serialiser generates, rather than each end round-tripping against itself. A field transcribed into the wrong slot still parses — every frame is a fixed length — so the app would draw a spectrum out of the scope buffer and look entirely plausible doing it. The golden asserts NaN and negative infinity survive unchanged, both being bit patterns a careless serialiser normalises.

• JUCE 8.0.15 is pinned and fetched rather than vendored, so the repository does not grow by 110 MB to record a version number.

• miniaudio v0.11.25 vendored under engine/third_party/ (public domain / MIT-0), with everything but the device layer compiled out.

• dr_wav 0.14.6, dr_flac 0.13.4 and dr_mp3 0.7.4 vendored under engine/third_party/dr_libs/ (public domain / MIT-0), compiled in one translation unit separate from the device layer. MA_NO_DECODING in oaa_device.c is what stops miniaudio compiling its own bundled copies of the same three and colliding with them at link time — it was already load-bearing for measurement correctness and is now load-bearing for the build as well.

• MP3 is the last format tried when identifying a file, not the first. dr_mp3 recognises a file by scanning for something that parses as a frame, and arbitrary binary data contains such sequences often enough that, given first refusal, it will open a FLAC file and decode noise from it.

• OAA_ABI_VERSION is 4. The change is additive — oaa_snapshot is byte for byte what it was at 3 — and adds only the oaa_file_* decoding calls.

• File paths reach the decoder as UTF-8 and are widened to UTF-16 on Windows. dr_libs’ plain _init_file calls go through fopen, which reads the path in the process’s ANSI code page; an umlaut in a user name is enough to make a file unopenable on one platform only, with an error that says nothing about encoding.

• Cancelling an analysis goes through a flag in native memory rather than by killing the worker isolate, which would leak the engine and the open decoder it was holding. A message cannot reach an isolate busy in a decode loop that never yields to its event queue; a pointer both isolates can read can.

• The conformance suite generates its own signals rather than reading WAV fixtures, so it runs on a headless runner with no network and no decoder.

• Loudness is asserted to be independent of both sample rate and push block size — properties the standard does not state but which catch two classes of error that 48 kHz single-block tests cannot.

• pffft vendored under engine/third_party/ (FFTPACK licence, permissive), and the maths-constant feature macro added to the build hook’s POSIX defines — M_PI is XSI rather than ISO C, and asking glibc for a POSIX level hides it. Linux-only compile failure, the same shape as the Phase 0 clock_gettime one.

• One set of transforms feeds the analyser, the spectrogram and the stereo cloud. Three modules running their own FFT over the same audio would cost three times as much and could disagree about where a peak is.

• Three primitives came out of writing the modules rather than being guessed at in advance: a shared dB scale and graticule, a paragraph cache that re-lays out only when a formatted string actually changes, and a buffer of marks sorted by the colour they are drawn in, which is how a display made of tens of thousands of them stays a few dozen draw calls.

• The accumulating modules advance on the engine’s publish counter rather than on every paint, so a resize or a theme change cannot scroll a spectrogram through time that no audio passed through.

• The canvas placement rules — overlap, clamping, id allocation — are pure functions over TabSpec in oaa_core, so they are covered by tests that need no window, and so the remote display cannot come to a different conclusion about where a module goes than the app did.

• The canvas and workspace tests build their own sparse layouts instead of reading geometry off the default preset. What the app opens with is a product decision, and a drag test that measured itself against it failed the day the default improved.

• Dragging a module rebuilds nothing. The module stays where it is and one painter draws where it would land, so pointer movement cannot stall a canvas of live meters.

• Module painters now extend a MeterPainter base and the module frame’s chrome is painted rather than decorated. CustomPainter.hitTest and BoxDecoration both absorb pointer events by default, which left every meter’s face unclickable with nothing reported anywhere.

• .github/workflows/ci.yml now runs dart test packages/oaa_wire and plugin/test/sources_match.sh. Both were named as gates in CLAUDE.md and README.md for a phase before either was wired in, which is the worst state for a gate to be in: everybody believes it is running.

• .github/workflows/release.yml builds the four installers and the CLI on a tag and on demand. On demand matters — an installer built only at release time is one whose script has been broken for weeks by the time anyone finds out.

• .github/workflows/docs.yml builds the documentation site on every pull request and publishes it from main.

• tool/docs.dart generates the site with no dependencies, so the docs job is a Dart SDK and no Flutter. The page list is written out rather than globbed, so docs/PLAN.md cannot be published to users by accident.

• packaging/icon/make_icons.dart describes the mark once as geometry and renders every size the installers want. Exported by hand, thirty-odd files across four containers drift.

• The project has a logo, in assets/brand/: the icon’s four bars with the tile taken away, and the name set beside them over two lines. It is the same mark as the application icon rather than a second one, and the wordmark is Inter SemiBold converted to outlines so the file needs no font installed. Three files — the lockup for dark backgrounds, the lockup for light ones, and the mark on its own. Nothing bundles them; they are for the README, the site and anywhere else the project is shown.

• make_icons.dart draws the mark in three shapes rather than one, because the desktops, iOS and Android each mask it differently, and writes iOS without an alpha channel — an icon that has one is refused by the App Store on upload rather than at build time.

• The canvas’s refusal toast is a provider rather than private widget state, so the shortcut layer above the canvas reports “no room for that” through the same channel a refused drop does instead of growing a second one.

• The README leads with the application icon, a badge row, a table of contents and the documentation site, and it now enumerates the thirteen modules and what each one shows — a count it stated and never listed. Presentation only: nothing it says about what Open Audio Analyzer measures changed.

0.1.0 — 2026-08-15

First release. The architecture is proven end to end and the app runs; most meters do not exist yet. See the roadmap.

📐 Measurement

• Peak, peak max, RMS, crest factor, inter-channel correlation and stereo balance are measured. Peak uses a 1.5 s hold and a 20 dB/s fall; RMS is smoothed with a 300 ms time constant.
• Every loudness quantity is unmeasured and reads as a dash — LUFS-M, LUFS-S, LUFS-I, LRA, true peak, TP max, DR-S, DR-I, PLR and PSR. They are NaN behind OAA_FLAG_LOUDNESS_UNAVAILABLE, never a zero that looks like a reading. K-weighting, R128 gating, LRA and true-peak oversampling arrive in the same release as the EBU conformance vectors that prove them.
• Open Audio Analyzer does not implement Decibel’s proprietary TrueDyn and will not approximate it. DR-S and DR-I are defined in docs/METRICS.md instead, reproducibly.
• All dB readings clamp to a −144.0 floor rather than negative infinity, so that differences between them stay finite.

✨ Added

• A C11 measurement engine with a lock-free snapshot, published by a dedicated analysis thread and read once per frame over FFI with no allocation.
• A built-in 1 kHz test signal, so the engine is measurable on a machine with no audio hardware.
• The Precision Instrument design system: one spacing scale, one border weight, no shadows, tabular figures on every number.
• A domain model with a 24-column grid layout, delivery-target calibrations for streaming, podcast, EBU R 128, ATSC A/85 and CD, and forward-compatible preset serialisation that skips module kinds a build does not have.
• Number Box module and the application shell.

🐛 Fixed

• The analysis loop no longer runs progressively slower than real time on a loaded machine. nanosleep only guarantees at least the delay requested, so every block overshoots slightly under contention; the loop discarded that error on each iteration instead of absorbing it, and on an oversubscribed host it settled at about a third of real speed. Lateness up to 250 ms is now made up one block at a time, and only a longer stall — a laptop waking from sleep, a debugger — resynchronises.

🚧 Internal

• Native code builds through a Dart build hook on all five platforms. There is no CMakeLists.txt, podspec or build.gradle for it anywhere.
• POSIX feature-test macros are declared in the build hook. Without them -std=c11 hides clock_gettime and nanosleep, which built cleanly on the development machine and failed on both POSIX CI runners.
• CI runs analysis, formatting, and the domain, widget and engine suites, with the engine built on Linux, macOS and Windows.
• Licensing is split: MIT for engine/, oaa_engine and oaa_core; GPL-3.0-or-later for the application, UI, CLI and plugin.