{ "$comment": "User-facing AND internal changes, newest first. Every entry: title (one line) FIRST, then version (CalVer YYYY.M.N it ships in, authored as \"NEXT\"), date, area, type, userFacing, details (full), and optionally tickets (an array of bug-board numbers this entry closes, e.g. [83]). The title leads DELIBERATELY: it is the only line unique to an entry, and the release pipeline's stamp is a one-line change to the version right under it - without a unique line above it, rebasing a new entry over that stamp lands the number on the wrong entry (see docs/release-pipeline.md). The welcome screen shows userFacing entries released since the version last run; tickets are never shown to users - the release pipeline reads them to walk those tickets along the board as the release progresses (tools/release/move-tickets.py). See CLAUDE.md > Change log for the authoring rules. Ships with the product and is copied to the CDN on release.", "entries": [ { "title": "The Welcome window is frosted glass over the moving background, which is now a little brighter", "version": "2026.9.23", "date": "2026-10-06", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The Welcome window is no longer a solid panel: the moving picture behind the app now shows through its body and button strip, much more blurred and darkened, with a faint grain, like frosted glass, so the window feels part of the scene while its text stays easy to read. The moving background around the windows is also a little brighter: it had been dimmed to a quarter of its brightness, and now shows at 40%. In the light theme the glass is a light frosted panel. Help > About keeps its solid look." }, { "title": "The moving background is a little softer", "version": "2026.9.22", "date": "2026-10-06", "area": "UI", "type": "Improvement", "userFacing": true, "details": "Now that the moving picture sits behind the whole app, it is blurred a step more, so it reads as a soft backdrop around your windows rather than a picture competing with them. The shapes are still there; the cost of drawing it is unchanged." }, { "title": "The moving background now sits behind the whole app, not just the welcome screen", "version": "2026.9.21", "date": "2026-10-06", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The soft moving picture behind the welcome screen used to give way to a flat gradient as soon as you opened a song. It now stays for the whole session, so you see it around the sequencer when it isn't maximised, and between and around floating windows such as module panels. It costs nothing while windows cover it completely (as a maximised sequencer does): it simply stops drawing until some of it is visible again." }, { "title": "The welcome screen's moving background is smoother, lighter, and now shows on Linux too", "version": "2026.9.20", "date": "2026-10-05", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The soft, moving picture behind the welcome screen used to be a video, which only played on Windows (DirectX) and in the browser: on Linux, macOS and the OpenGL Windows build you got a flat gradient instead. It is now a set of pre-made frames, every frame of the original at its full frame rate, blended smoothly from one to the next. It looks the same everywhere, including Linux and macOS, and it costs less: it no longer keeps a video decoder running, it only redraws itself up to 15 times a second whatever your frame rate, it stops while the window is minimised, and it uses less memory. In the browser it loads in the background, so the page never waits for it." }, { "title": "Freyja's banner fades out smoothly, and the mouse pointer's edges are no longer too dark", "version": "2026.9.19", "date": "2026-10-05", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Freyja's banner art and the app's own mouse pointers were stored in a form that the drawing code darkened a second time. Wherever they turn see-through, they came out darker than drawn: the right end of Freyja's banner faded into the background too soon (the trees and wolf on that side all but vanished), and every pointer's soft outline and shadow came out darker and thinner than designed. Both are now stored the standard way, and they draw as authored. Pictures you add yourself (module art, panel images) were never affected.", "tickets": [ 697 ] }, { "title": "Lines, wires, waveforms and curves are now smooth", "version": "2026.9.18", "date": "2026-10-05", "area": "UI", "type": "Improvement", "userFacing": true, "details": "Lines in the app had no antialiasing at all: module wires, automation lanes, scope, envelope, LFO and spectrum curves, waveforms and knob pointers were drawn as hard-edged pieces snapped to whole pixels, so diagonals and curves came out as staircases with a notch at every joint, and a see-through wire went darker wherever two pieces overlapped. Every line is now one smooth, joined stroke with a soft edge one screen pixel wide, at any display scaling; waveforms (in both drawing modes) have soft edges, and so do knob pointers and modulation arcs, the Pixel knob skin's pointer included. Lines that run exactly flat or upright are kept on the pixel grid so they stay sharp. It costs a few hundredths of a millisecond a frame." }, { "title": "A beta you lose access to now offers to switch you back to the release", "version": "2026.9.17", "date": "2026-10-04", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "Since you can update from a release onto a beta, losing beta access (a pledge ending, or two weeks without reaching ezmuze central) left you on a beta whose supporter screen offered only Sign in, Get beta access and Quit, so the only way back was to find the release installer on the website. Every version of that screen now also offers Switch to the release, which names the release, asks once (songs a beta saved in a newer format may not open in it), then downloads it and restarts into it in place. Your songs, settings and sign-in are kept. Offline, the button waits for a connection and the screen links to the release downloads; a copy that was not installed by the installer gets Download the release instead; and on beta.ezmuze.studio it is Go to the release, which takes the tab to ezmuze.studio. Under the hood the switch is a Velopack downgrade, and the beta's own package is removed first so Velopack cannot quietly reinstall the beta on the next launch.", "tickets": [ 693 ] }, { "title": "Freyja's bottom row is one tidy strip, and its LFO waves are named in full", "version": "2026.9.16", "date": "2026-10-04", "area": "Modules", "type": "Improvement", "userFacing": true, "details": "Freyja's panel had an empty band between the Performance and Warp row and the LFO Shape section, left for Click and Noise controls Freyja does not have. Performance, Warp and LFO Shape now sit side by side in one row, so the band is gone and the panel is 236 points shorter (1450 x 675, was 911). The LFO's Wave is a dropdown, like Warp's Mode, so Triangle, Square and Sample & hold are no longer clipped to Tria..., Squa... and Sam..., and the Sync switch lines up with it. It reaches installed copies of Freyja with its next store version." }, { "title": "Freyja's panel wears its banner art across the top", "version": "2026.9.15", "date": "2026-10-04", "area": "Modules", "type": "Improvement", "userFacing": true, "details": "Freyja's text header (the name, a five-colour stripe and a subtitle) is replaced by the instrument's own banner art: the FREYJA wordmark, mountains, aurora and a rainbow waveform over a lake, across the panel's full width at its own proportions and fading out softly at the right. Everything below it moved down unchanged, so the panel is 161 points taller (1450 x 911). Like Bifrost's, the banner is built into the app, so the panel shows it whether or not the store package carries it. It reaches installed copies of Freyja with its next store version." }, { "title": "Wavetables keep their own level on every note, and their top end above middle C", "version": "2026.9.14", "date": "2026-10-04", "area": "Sound engine", "type": "Bug fix", "userFacing": true, "details": "Wavetables now play at the level they are stored at on every note, and keep their top end above middle C. From middle C up the oscillator used to swap a table for a copy cut to its first 14 harmonics (6 from F#5) and re-normalised every frame to a 0.95 peak, so an imported Vital, Serum or Phase Plant table that was quieter than that jumped in level at C4 (+13.6 dB for one peaking at 0.2), an FM modulator's depth jumped with it, and every bright sound lost everything above its 14th harmonic from C4 up. Anti-aliasing is now one band-limited copy of the table per octave, each keeping every harmonic below 0.45 of the sample rate at the top of its octave and fading into the next over the last quarter of it, at the table's own level; the spectral morph band-limits the same way, so turning a morph on no longer changes the level either. Factory sounds above middle C get brighter and more even. Measured old against new for every shipped bank, the Grrowl pack, the demos and a 703-preset Vital corpus (AC RMS through a 20 Hz high-pass, on a chord and on notes 48, 60 and 72): 63 presets that one output level holds within half a dB of their old render are re-levelled with it: Bifrost (Clav, Delay Lead, Delay Pluck, Fold Lead, Ping Delay Stab, Prism, Squashed, Tri Rhodes); Grrowl pack (Answer Back, Robot Voice, Screech, Talk Box Neuro, Two Speed Growl); Bank Manager (Bass And Keys, Choir Wash, Growl Below, Octave Lead, Press For Bell); Bueller (Baroque, Bent Pad, Bowed Air, Clavinet, Combo Organ, Glass Pad, Phased Brass, Square Vibrato, Sync Scream, Tension, Trumpets, Violins); Freyja (Bell Keys, Bright Strings, Chorus Wash, Clav, Electric Keys, Harp, Hollow Bass, Hollow Pad, Pluck Bass, Poly Brass, Pulse Poly, Soft Lead, Stepped, Toy Piano); Kenobi (Analog Organ, Choir Pad, Pulse Pluck, Rubber Bass, Sample and Hold, Square Lead); Laufeyson (Harpsi, High Pass Stab, Metal Ring, Sample and Hold); Rat 030 (Clicky, Detuned Acid, Distorted Square, Low Drone, Square Acid, Square Echo, Square Slides, Square Sweep, Sub Square). Laufeyson's own default Output Level goes from 0.13 to 0.125, because its eight-note chord now reached full scale (Init is 0.34 dB quieter; Two Saws and Square and Sub keep 0.13). 57 more change only above middle C, where they now have more top end, and are left as they are. Twelve change character and are deliberately left as the fix makes them: Bank Manager Vowel Lead, High Shimmer, Grand And Air, Choir And Glass, Three Zones and Bell Cloud; Bifrost Music Box, Vowel Cloud, Metal Pluck and Broken Radio; Grrowl Growly Vowel and Bell Screech. Their old sound above middle C came from the re-normalise lifting what the 14-harmonic cut left; Vowel Lead, for one, used to drop 10 dB at C4 and now rises through it smoothly. Imported vocal and speech tables whose energy sits above the 32nd harmonic get quieter high up, as they would in Vital: past C5 that content is above Nyquist. The demo songs are unchanged within 0.03 dB. Imported tables take about 3.4 times their own size for the mips (2 times before); the factory tables take less than before.", "tickets": [ 688 ] }, { "title": "Check for updates offers the newest beta too when you have beta access, clearly marked as a beta", "version": "2026.9.13", "date": "2026-10-03", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "The update check only ever looked at the channel the running build came from, so a release build never mentioned betas even to someone signed in with beta access. Now, when your ezmuze central account has beta access (checked with central, or a recent cached yes while offline), the check looks at both channels and offers whichever is newer, the latest beta or the latest release. A beta offer says so everywhere it appears: the dialog is titled Beta update available, the version reads e.g. 2026.9.12 (beta), a line says betas get new features first and may be less stable, and the button reads Update to beta or Download beta. The welcome-screen banner does the same. Signed out, or without beta access, you see releases exactly as before. Beta builds are now offered a newer release as well as the next beta. The same version on both channels is offered as the release, nothing older than what you run is ever offered, and the check runs again when you sign in or out. Which versions exist is now read from each channel's update feed, because no change-log entry carries a release's own number. Desktop only: the browser build always runs whatever its site serves, ezmuze.studio for releases and beta.ezmuze.studio for betas.", "tickets": [ 692 ] }, { "title": "A file dropped on a window goes to that window, not to the song behind it", "version": "2026.9.13", "date": "2026-10-03", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Dropping a sample on an open module panel, the preset browser or any other window over the sequencer used to put it on the timeline hidden underneath, as a new Audio channel and a clip. Now each window says which files it takes, and a drop goes to the window drawn on top where you let go. The sequencer places audio where it lands on the timeline and imports MIDI, SoundFonts, Vital banks, module bundles, packages and songs from other programs; the pattern explorer imports into the wave bank; the wave and sample pickers add audio to their bank; the Lua component editor takes any file; Module properties and Publish module take an image as the art; the UI editor takes images into the image library; the welcome screen opens songs from other programs. A window that cannot use the file says so: \"Chip - Chip doesn't know what to do with .wav files.\" Dropping several files at once names every type the window refused in one message. On macOS and Linux, which do not report where a drop lands, the mouse position at the drop is used, so a sample let go over the timeline is placed there too; the browser build sends the drop's own position. Built on a new GustUI drop-target registration (FileDropTrait and FileDropRouter).", "tickets": [ 683 ] }, { "title": "CI: the welcome video's content manifest regenerated, and a flaky voice-tap test fixed", "version": "2026.9.12", "date": "2026-10-03", "area": "Build", "type": "Internal", "userFacing": false, "details": "The welcome-video change hand-edited the 1.mp4 hash in place, but content.manifest is sorted by hash, so ContentFreshnessTests read it as stale on both the web and DX heads; it is now regenerated with tools/UpdateContentManifest. FollowNewestVoiceTests' component-tap test rendered a whole bar in one span after asking for the tap once. A tap nobody has read for a second stops being fed, and on a busy CI runner that render took longer than a second, so the feed stopped part way through and the 'latest' window came from wherever it stopped (with the test host starved onto one core it failed 6 runs of 6, the tap dropped before its first sample). The test now renders in 50 ms spans and asks for the tap between them, as an open display does every frame: 8 of 8 pass under the same starvation. The app was never affected; a display reads its tap every frame." }, { "title": "The welcome screen's background video is a fraction of the size, with the same soft look", "version": "2026.9.12", "date": "2026-10-03", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The out-of-focus concert video behind the welcome screen was a 3.5 MB 720p clip. It is now a 256x144 clip of 148 KB, so the download is smaller and each frame takes far less memory and work to decode and put on screen. Stretched over a whole window, a clip that small would show its pixels as a blocky grid, so it is drawn through the same blur as the backdrop behind a waiting dialog: the frame is stretched into a picture a quarter of the window's size, blurred there, and that is what fills the window. It looks as soft as it did before. The blur runs only while the video is on screen, and on a 4K window the welcome screen draws no slower than it did." }, { "title": "The preset browser has an Up button and a search box, and its lists fill the window", "version": "2026.9.12", "date": "2026-10-03", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The \"..\" row at the top of a folder was easy to miss, so it has gone. To go back up, use the Up button at the start of the toolbar, as in the file browser, or press Alt+Up. Up is greyed out when you are already at the top. A new search box in the toolbar filters as you type and lists every match from every folder at once. It matches a preset's name, the folder it is filed in and who made it, and each result shows its folder beside its name. While you search, the folder list is dimmed. Press Escape or clear the box to return to the folder you were in. Select all, Rename and Delete act on the results you can see. The folder and preset lists now reach the bottom and right edges of the window rather than stopping short. You can also drag the browser to any size and maximise it, and the lists grow with it. Long lists and searches across big libraries stay quick because only the rows on screen are built.", "tickets": [ 689 ] }, { "title": "Control API: author_presets and delete_preset, for republishing a store module's presets", "version": "2026.9.12", "date": "2026-10-03", "area": "Modules", "type": "Internal", "userFacing": false, "details": "A module that left the app for the store is never seeded again, so a preset re-voiced in code (Bifrost's Twin Peak, Rat 030's re-levelled and re-voiced squares) had no way into the author's library, which is what the publish wizard ships as the Factory bank. author_presets {module, dryRun} writes the repo's bank (StoreModuleSources now carries each store module's preset builder) into the library and reports what it adds, what it changes, what differs only in a selector's stored position count (the warp selectors grew from 5 to 11, and applying a preset writes that count, so a bank left at 5 shrinks the dropdown) and what the library has that the repo does not; it never deletes. delete_preset {id} removes one library preset, because a stray test preset for a store module ships in its next version. StoreModuleSourcesTests holds every bank to its own module's id." }, { "title": "Bend, Sync, Fold and Mirror warps are now centred too", "version": "2026.9.11", "date": "2026-10-03", "area": "Sound engine", "type": "Bug fix", "userFacing": true, "details": "The four older warp modes could leave an oscillator's wave sitting above or below zero, just as the six new ones did before the last beta: Bend, for instance, dwells at the start of the table. That offset takes up headroom and thumps when notes start and stop. Every warp mode now has it taken back out, and the level kept within the table's own is worked out from the wave actually played, so a centred wave is not turned down for no reason. A warp at amount 0 sounds exactly as before. A few presets were adjusted so they sound as they did: Bifrost's Twin Peak was re-voiced, and Grrowl's Dominator and Triplet Growl and Rat 030's Bent Square were re-levelled. Your own sounds that use Bend, Sync, Fold or Mirror may come out slightly quieter (or, where the offset was pushing a drive, slightly louder), usually by well under half a decibel.", "tickets": [ 686 ] }, { "title": "The six new warp modes no longer push the sound off centre", "version": "2026.9.10", "date": "2026-10-03", "area": "Sound engine", "type": "Bug fix", "userFacing": true, "details": "Squeeze, Formant, Quantize, Pulse, Asym and Flip could leave an oscillator's wave sitting above or below zero: Pulse, for instance, holds the start of the table for most of each cycle, and Flip turns part of it over. That offset is inaudible on its own but takes up headroom, so a chord could clip sooner, and it made a thump when notes started and stopped. The oscillator now takes the offset back out and keeps the result within the table's own level, and the wavetable display draws the centred wave. Bend, Sync, Fold and Mirror are unchanged for now, so existing sounds that use them play exactly as before.", "tickets": [ 686 ] }, { "title": "Vital, Phase Plant and Serum imports now carry FM, ring mod, the new warp modes, spectral morph and filter routing onto Bifrost", "version": "2026.9.10", "date": "2026-10-03", "area": "Modules", "type": "Improvement", "userFacing": true, "details": "Importing a Vital bank, or a REAPER project with Vital, Phase Plant or Serum on a channel, now uses what Bifrost v8 added instead of leaving it out. Vital: oscillators that FM or ring-modulate each other land on Bifrost's FM and RM dials, with the oscillators reordered when the modulation runs backwards (Bifrost's B hears A, and C hears A or B); its Sync, Formant, Quantize, Bend, Squeeze and Pulse distortions become the matching warp modes; ten of its seventeen spectral morphs become Bifrost's morph, amount included; each oscillator and the sample go to Filter 1, Filter 2, both or around the filters as they did in Vital, and the two filters run in series or in parallel as Vital had them. Modulation routed onto a morph, warp or FM amount now carries too. Phase Plant: a generator modulating another at audio rate becomes FM (phase, frequency or pitch) or ring mod (level), and a silent modulator gets its own oscillator at level 0; generators, noise and the sampler reaching only a second filter go to Filter 2, in parallel when that filter sits on its own path. Serum: FM, RM and AM from the other oscillator become Bifrost's links (swapping A and B when A is the one modulated), and Asym, Flip, PWM, Quantize and the sync-window modes map to their warp modes. FM depth was measured against Vital and Phase Plant, and Bifrost's FM dial now reaches Vital's deepest FM at full scale. The import notes say plainly what is still approximated or left out: Vital's Vocode, Spectral Time Skew and other morphs Bifrost lacks, FM or ring mod from Vital's sample, Direct Out, and Serum's remap modes." }, { "title": "LFO, envelope and modulation displays follow the note you just played, including from the keyboard", "version": "2026.9.10", "date": "2026-10-03", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "When you played an instrument live from the computer keyboard or a MIDI keyboard, with the song stopped, the LFO cards on its modulation rail showed no moving playhead, and knobs an LFO or envelope was routed to showed no movement, although you could hear the modulation: live playing runs the instrument on its own, and those displays only ever asked the song's playback. They now follow whatever is sounding. And with several notes sounding, every one of those displays, and a scope on a part inside the instrument, now shows the note you played most recently. Before, they showed whichever note happened to be on the first voice, often an older or already released one. Letting go of a key does not make them jump back to an older note. The envelope graph's live dot also works on instruments whose envelope is set by knobs, such as Bifrost, where it never appeared.", "tickets": [ 685 ] }, { "title": "Bifrost v8: oscillators that modulate each other, six new warp modes, spectral morph, parallel filters and a live scope", "version": "2026.9.10", "date": "2026-10-03", "area": "Modules", "type": "New feature", "userFacing": true, "details": "Bifrost's three oscillators can now play off each other: Osc B takes FM, ring mod and hard sync from Osc A, and Osc C takes them from A or B. Each oscillator gains a spectral Morph with nine types (Low Pass, High Pass, Harmonic Stretch, Inharmonic Stretch, Formant Shift, Smear, Random Amp, Phase Disperse, Shepard) and its own amount, which you can modulate like any knob, and the wavetable display draws the result. Warp is now a menu of eleven modes, six of them new: Squeeze, Formant, Quantize, Pulse, Asym and Flip. Every source (A, B, C, Sub, Noise, Sample) can go to Filter 1, Filter 2 or straight out, and Filter Routing runs Filter 2 after Filter 1 (Serial) or beside it (Parallel). A new Output section shows a scope and a spectrum of everything Bifrost is playing, the whole chord, live whenever the panel is open. All of it rests at off, so every existing Bifrost sound is unchanged. The warp modes, ring mod, sync and morph live in the shared wavetable oscillator, so other synths built on it can offer them too. Also fixed: a panel switch bound to a value that is not itself an on/off switch read as on in both positions and wrote a huge number when clicked." }, { "title": "Chip, Kenobi, Bueller, Freyja, Rat 030 and LadderOne get six new warp modes", "version": "2026.9.10", "date": "2026-10-03", "area": "Modules", "type": "Improvement", "userFacing": true, "details": "The Warp section on these six synths now offers all eleven of the oscillator's warp modes: Squeeze, Formant, Quantize, Pulse, Asym and Flip join Off, Bend, Sync, Fold and Mirror. Its Mode control is now a dropdown in the same place, because eleven names do not fit a row of buttons. Existing songs and presets sound exactly as before. Chip has the change as soon as you update the app. Kenobi, Bueller, Freyja, Rat 030 and LadderOne come from the store, so their dropdown arrives with their next store version; until then their old row of buttons already reaches all eleven modes, squeezed into narrow buttons. Laufeyson has no Warp section (its Sync dial owns the oscillator's warp) and is unchanged." }, { "title": "The wave picker takes MP3, Ogg, FLAC and AIFF, has a Load file button, and no longer freezes on a long file", "version": "2026.9.9", "date": "2026-10-03", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The wave picker now accepts MP3, Ogg, FLAC and AIFF files as well as WAV, dropped onto it or chosen with a new Load file button. A file dropped onto an open picker now goes into the picker; before, it fell through to the song behind it and made a new audio channel. Long files are decoded in the background on desktop, so a four-minute MP3 no longer freezes the app for two seconds (on the web the decode still runs on the page's only thread). Long files dropped on the timeline decode in the background too. Also fixed: writing a preset pack dropped every preset's attached sample and imported wavetable, so a pack of converted Vital patches came out without their samples." }, { "title": "Bifrost v8 panel, banner and sample layer built in the app, ahead of its store release", "version": "2026.9.9", "date": "2026-10-03", "area": "Modules", "type": "Internal", "userFacing": false, "details": "Built in the repo, not yet published to the store (v8 will be a breaking version and is getting more before it goes out): the three oscillators stacked down the left with the filters, envelopes, tail, macros and Performance strip beside them; the product banner across the top (BifrostBanner.png, compiled in as a panel image); a Sample section under Osc C that is always shown, picking from the wave bank and copying the sample into the Sample Player so presets and packs carry it; the Sample Player in the code graph for every Bifrost, with older devices upgraded on pick, song load or a preset that carries a sample. New remote-control action author_graph puts the code graph onto the installed module for publishing. The panel needs an app that has the sample picker, which is why this lands in a beta first. The user-facing entry, closing #660, comes with the v8 publish." }, { "title": "The store says why Get did not work, and offers to sign you in again when that is the reason", "version": "2026.9.8", "date": "2026-10-02", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "When ezmuze central stopped accepting your sign-in (it expired, or newer sign-ins elsewhere pushed it out), pressing Get in the store looked like nothing: a loading line flashed, the same Get button came back, and the reason had faded away in the status bar behind the store window. That is what made Keyzoo 20/20 seem impossible to install. A failed Get now stays on the shelf in red, with a toast that waits to be dismissed. When the cause is the sign-in, a Sign in again button sits under the message and reloads the store once you are back in. The shelf says the same as soon as it opens, because an expired sign-in also makes the modules you own show Get. Every failure is logged with its cause.", "tickets": [ 680 ] }, { "title": "Long tooltips wrap to a readable width and always stay on screen", "version": "2026.9.8", "date": "2026-10-02", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Tooltips drew every line at whatever width it came to, so a store module tile's tooltip, which carries the whole description, ran as one line right across a 4K screen and off the edge. Every tooltip now wraps at about 400 px (never more than 40% of the window), keeps its own line breaks, flips to the left of or above the pointer at the window's edges and stays wholly on screen; anything taller than half the window ends in '...'. Short tooltips look exactly as before. Also fixed: moving the pointer straight from one element with a tooltip to another (adjacent tiles or toolbar buttons) no longer leaves the second one without its tooltip. The rules are in GustUI's TooltipLayout and design-guide.md.", "tickets": [ 681 ] }, { "title": "Re-importing your own store module's source now shows and publishes its new panel", "version": "2026.9.8", "date": "2026-10-02", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "Importing an .ezmodsrc stored its panel as not edited here, and an installed store package's panel wins over any panel that is not a person's edit. So when the author of a store module re-imported its source with a new panel, both the panel shown and the one the publish wizard bundled were still the installed package's, which is how Keyzoo 20/20 v3 went out with its old panel. An import by the module's author (the store's signed author grant on an installed copy, InstalledModules.IsAuthorOf) now marks the panel as their edit, so it beats the package's. Anyone else's import of a bundle still arrives as a plain seed that the installed package's panel outranks.", "tickets": [ 675 ] }, { "title": "Applying a preset no longer wipes a module's own imported wavetable or Lua file", "version": "2026.9.8", "date": "2026-10-02", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "A preset that carried no component state for a slot (one that only sets values, written outside the app or saved before a table was imported) cleared that slot's imported wavetable and Lua file. For a user module whose table is part of its own definition, every such preset turned the oscillator back into the built-in sine. Applying a preset now restores a slot the preset says nothing about to the module definition's own table and script (PresetMath.DefinitionComponentStates / RestoreDefinitionState), in the preset browser, the module editor, the store preview render and a store device's saved preset. A device with no definition behind it (a factory module) is still cleared, as before, and a preset that carries its own state still wins. Undo restores exactly what the device had before.", "tickets": [ 674 ] }, { "title": "The publish wizard's preview no longer comes out silent for a module that nests another one", "version": "2026.9.8", "date": "2026-10-02", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "Render preview built its list of module definitions from the module itself and the open song. A module that keeps its sound in a nested definition (Keyzoo 20/20's oscillators are in its nested Waves module) therefore rendered silence unless a channel holding it happened to be in the open song, and the wizard said there was nothing to upload. The preview now gets the module's nested definitions as well: the ones gathered for the bundle it is about to publish, or for an installed module, the closure from the library and installed packages (ModulePreviewRender.Render's nested list). Keyzoo 20/20 in an empty song now previews at the same level it does from a channel (raw peak 0.49).", "tickets": [ 673 ] }, { "title": "PXL4 from the store plays its 8 voices again, instead of arriving as a monosynth", "version": "2026.9.8", "date": "2026-10-02", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "PXL4's first version was published from an instance before a module's definition carried its voice count, so every version since says 1 voice in its definition and 8 only in its package manifest. Opening an installed package took the definition's number, so every PXL4 device made from it, the author's and a buyer's, had 1 voice, and the publish wizard's Graph check read '1 voice' and could call a panel-only version breaking. Opening an installed package now reconciles the two the way reading a bundle file always did (ModuleBundle.TakeManifestVoices): a definition that says mono takes the manifest's count when the manifest says more. The publish wizard's voice count also ranks an instance that was made from the installed package below the package itself, since it can only echo the package's definition (ModulePublish.PickVoices). An instance made from your own copy of a module still wins, as before.", "tickets": [ 671 ] }, { "title": "The publish wizard refuses a module whose own sample files are missing from the package", "version": "2026.9.8", "date": "2026-10-02", "area": "Services", "type": "Bug fix", "userFacing": true, "details": "A module that carries its own samples could be published without them: Keyzoo 20/20 v2 went to the store as 6.5 KB, its twelve sample players pointing at files the package did not have, and every note was silent. Neither the wizard nor the store noticed, although the app itself refused to import those same bytes. The wizard now checks the exact package it is about to upload: every embedded file a module uses must be its own, every file it lists must be in the package, and every file it has must have its data. Anything missing is a refusal on a new 'Embedded files' row, naming the file. The store makes the same check. A new version that ships no presets where the version before shipped some now says so plainly on the Factory bank row, as a warning.", "tickets": [ 672 ] }, { "title": "Rule: a song-format bump needs a row in the store's format table", "version": "2026.9.8", "date": "2026-10-02", "area": "Services", "type": "Internal", "userFacing": false, "details": "The store works out which studio a module needs (the 'needs ezmuze studio X' badge) from ModuleFormat.FirstReadBy in ezmuze-central, a table of each song format and the first beta that read it. It stopped at v37 while the studio wrote v40, so every module version uploaded in v38-v40 got no floor. Central now has the rows (v38 2026.8.281, v39 2026.8.360, v40 2026.8.364), a backfill for the versions stored without one, and a deploy-time test that fails while the current format has no row. Song.CurrentFormatVersion and docs/release-pipeline.md now state the rule.", "tickets": [ 670 ] }, { "title": "Keyboard jamming has an on/off switch (Ctrl+J), so letter shortcuts and played notes no longer collide", "version": "2026.9.7", "date": "2026-10-01", "area": "UI", "type": "New feature", "userFacing": true, "details": "A new keyboard button on the toolbar, after Record, switches keyboard jamming (playing notes on the letter keys) on and off; Ctrl+J and Edit > Turn keyboard jamming off/on do the same, and the status bar says which. It is on to begin with and remembered. While it is on, the letters play notes from an open instrument panel, the piano roll or the module builder as before, and the letter shortcuts that share those keys (S split, M marker, Shift+M section, L loop the section, M add a cue in the Clip editor, S snap in the drum lanes) stand aside: pressing one plays its note and the status bar says how to turn jamming off, instead of the note and the edit both happening. Their menu items and buttons always work. While it is off, no letter plays a note and those shortcuts work. MIDI keyboards are not affected.", "tickets": [ 679 ] }, { "title": "Record is now Ctrl+R, so R plays a note when you jam on the keyboard", "version": "2026.9.7", "date": "2026-10-01", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "Record was a bare R, and R is also a note (F) on the QWERTY musical-typing keyboard. An open instrument panel plays that keyboard even while the sequencer has the keys, which it does as soon as you click a channel's arm dot, so playing a riff with an F in it started or punched out a recording. Record is now Ctrl+R (REAPER's key) in the sequencer, its tooltip and the help screen; holding Ctrl keeps the R from playing a note. In the browser, Ctrl+R no longer reloads the page.", "tickets": [ 678 ] }, { "title": "Sequence melody: turn a melodic loop into a played part in the Clip editor, seeing and editing its notes first", "version": "2026.9.7", "date": "2026-10-01", "area": "UI", "type": "New feature", "userFacing": true, "details": "The wave bank's Convert to instrument has moved into the Clip editor, beside Drum chop, as Sequence melody (the keyboard button, the Sequence melody menu, or 'Sequence the melody…' on a sample clip's menu). It shows the notes it would play the loop as, over the waveform: each note's start, its end and its name. Drag a start or an end, click the lane to add a note, drag one off the lane to remove it, and right-click a note to move it up or down a semitone or an octave. Sliders tune the detection (Sensitivity, Confidence and Off-grid for struck notes; Min note and Pitch hold for held ones; and which reader to trust), Play melody plays what you will get, and Make melody makes one sample on a key-tracking voice and a pattern that plays it where the clip did, on a new channel below or into a channel an earlier conversion made, replacing the clip in one undo step. The notes and settings last only while the mode is on. Convert to instrument in the wave bank is now a shortcut into it. Under the hood the melodic decomposer's swept constants became settings (the defaults unchanged), detection runs off the game thread with every answer cached by its settings, and the notes are laid out by the clip's own pieces, so loop passes, trims and stretches are followed.", "tickets": [ 677 ] }, { "title": "Ten store instruments' panels: values and labels no longer run into each other", "version": "2026.9.6", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "Modyssey, Bank Manager, Laufeyson, Kenobi, Cloudform, Freyja, PXL4, Bueller, Rat 030 and Rhodine each have a new compatible version in the store whose panel fits its words: envelope readouts such as '600 ms' sat in knobs narrower than the text and ran into the next one, Bank Manager's key ranges ('F#-1+35c') and captions ('HIGH KEY') overflowed, selector rows drew their buttons over their captions, and PXL4's SID knobs could not hold '2.58 kHz'. Knobs were widened and rows rebalanced (Bank Manager, Laufeyson and PXL4 grew a little wider); no text got smaller. Songs on 'latest compatible' pick the new panels up on their own. Every store panel is now held to PanelTextFitTests.", "tickets": [ 669 ] }, { "title": "Bifrost's panel breathes: macros beside the Tail, and the envelope and Performance knobs no longer overlap", "version": "2026.9.5", "date": "2026-10-01", "area": "Modules", "type": "Improvement", "userFacing": true, "details": "Bifrost v7 in the store (a compatible version, so songs pick it up on their own). The Modulation list beside the Tail is gone: the mod rail along the bottom lists every routing, with its depth and a delete, so it showed nothing the rail does not. The four macros took its place, as large knobs, which freed their strip for Performance, Click, Noise and LFO shape, now spread across the full width; and the envelope column grew, with A/D/S/R knobs as tall as their graphs. Before, the envelope knobs (60 wide on a 50 px pitch) and the strip's knobs (92 wide on 84) overlapped, so readouts ran together ('0.17 ms767 ms'). Bifrost is back in the panel geometry test. In the app, panels gained 'shown while a component exists' (an empty ShownWhenParam), and Bifrost's panel has a Sample section built on it for patches imported with a sample layer (#660); it is held back from the published panel (BifrostUi.SampleLayerShips) until an app that hides it on a plain Bifrost has shipped, because today's apps would show it as five dead controls.", "tickets": [ 661 ] }, { "title": "LadderOne's Osc 2 wave selector shows Saw, Pulse and Sub", "version": "2026.9.5", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "The Osc 2 wave selector on LadderOne's panel drew a single '-' cell, so the second oscillator's waveform could not be picked from the panel. The selector drives a three-way switch whose input is a plain number, so it had no names to show; it now carries its own (Saw, Pulse, Sub, in the switch's order). Arrives with LadderOne v4 in the store. For the store version, author_face now also gives a store module's selectors the position names its code graph has and the installed copy lacks (StoreModuleSources.CopyChoiceNames).", "tickets": [ 665 ] }, { "title": "LadderOne's envelope values and Warp mode no longer draw over each other", "version": "2026.9.5", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "LadderOne's envelope readouts ran together ('1 ms350 ms30%300 ms') because each value was wider than its knob, and the Warp MODE caption sat on its Off/Bend/Sync/Fold/Mirror buttons. The envelope column is 50 px wider with knobs on a 52 px pitch; the strip's segmented selectors (Warp Mode, Noise Type, LFO Wave, shared by the synth family) no longer reserve a value row, so their cells clear their captions; 'Osc2 oct/st', which was cut short, reads 'Osc2 tune'. Arrives with LadderOne v4. A new test, PanelTextFitTests, measures the text a panel draws (every value readout across its knob's travel, every caption) in the panel's own font against its control; store panels are held to it, and the ones that do not pass yet are listed against #669.", "tickets": [ 668 ] }, { "title": "Modules with an imported SoundFont can be published: give the font a licence from its SoundFont player", "version": "2026.9.5", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "A module whose SoundFont player played an imported .sf2 could never be published: the store said the font carried no data, and refused its credit whatever its licence was. Two causes. The bundle writer moves a font's bytes out of the module's definition into a separate file, and the store's check only looked in the definition; and the credit the app wrote for a font was a usage-only one the store does not accept. Now an imported font has a licence of its own: in the module editor, right-click the SoundFont player and choose Credit and licence of \"\"..., the same dialog embedded files use (CC0, CC BY, CC BY-SA or your own recording publish; Not set does not). The bundle's credit is built from it, and the publish wizard's Embedded audio check (the store's own rule, run over the package it is about to upload) passes for a licensed font, refuses an unlicensed one by name, and now says where to set the licence. Needs the matching store update, which ships first.", "tickets": [ 666 ] }, { "title": "The remote-control API answers at localhost on Linux and Mac, not only at 127.0.0.1", "version": "2026.9.5", "date": "2026-10-01", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "On the Linux and Mac builds every request to http://localhost:/ answered 404 while http://127.0.0.1:/ worked. Off Windows, .NET's HttpListener matches a request to a registered prefix by the host name in its Host header, and only http://127.0.0.1:/ was registered, so the request never reached the app. Those builds now also register http://localhost:/. The listener binds a named prefix to the first address the system resolves it to (::1 on a stock /etc/hosts, which is also where a client resolving localhost goes first), so it stays loopback only; a client that forces IPv4 while sending Host: localhost still gets the 404, and 127.0.0.1 keeps working everywhere. If the IPv6 bind fails the app falls back to 127.0.0.1 alone. Windows is unchanged. Verified on the Linux build under Docker: 127.0.0.1 and localhost both answer, and a request with any other Host is still refused.", "tickets": [ 667 ] }, { "title": "Choosing a window from View that a pinned window covers now says why it stayed behind", "version": "2026.9.4", "date": "2026-10-01", "area": "UI", "type": "Improvement", "userFacing": true, "details": "With the sequencer pinned to front, choosing a module panel or the preset browser from View left it behind the sequencer, and nothing said why: it looked as if View did nothing. A pinned window staying on top is the rule (the owner's call: pinned stays on top), so the window still stays behind it, but the status bar now says so, for example 'Chip - Chip is behind the pinned Sequencer - ...: unpin it to bring Chip - Chip forward', and the pinned window's pin button flashes three times so it can be found. It is said when a front-pinned window covers at least half of the chosen one; a pinned palette clipping a corner says nothing. In GustUI: ModalWindowElement.BehindPinnedNotice, CoveringPinned and FlashPin.", "tickets": [ 647 ] }, { "title": "Chip's Stack card shows its main sound controls instead of about 76 knobs", "version": "2026.9.4", "date": "2026-10-01", "area": "Modules", "type": "Improvement", "userFacing": true, "details": "Chip marked none of its controls In Panel, so its Stack card showed every input nothing drove: about 76 knobs, most of them internal plumbing such as table constants, envelope taps and unison spreads. Chip now has a face like the store synths: the two oscillator waves, pulse width, Osc2 detune and level, ring mod, noise level and crush, the filter's cutoff, resonance and envelope amount, and the amp envelope (15 controls, plus the card's own Voices). Everything else is still on its panel and in Advanced view. A Chip added from now on gets it; a Chip already in a song picks it up only if its card was never customised. The face is display only: no sound changes, and the demo parity tests still pass.", "tickets": [ 663 ] }, { "title": "Opening another song closes the wave picker, the preset browser and the other windows of the old one", "version": "2026.9.4", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Opening a different song closed the old song's editors but not every window that belonged to it. A wave picker opened from a module's Choose sample... stayed in the app behind the new sequencer, still bound to the old song, and took every file dropped on the app: a .mid dropped on the Pattern explorer failed with 'Not a RIFF/WAVE file' (#662). A module's preset browser stayed open, and listed under View, for a device the new song did not have (#664). Opening a song and closing the sequencer now share one list of the song's windows (EzmuzeStudioGame.CloseSongWindows): the editors as before, plus the wave picker, the content pack picker, the preset browser, a module's Properties, the wavetable designer, the Lua component editor, the loudness report, the sample recorder, Song settings and the Quantize/Swing/Humanize dialog, each of which holds the old song, one of its devices or one of its patterns.", "tickets": [ 662, 664 ] }, { "title": "Undoing Convert to instrument into an existing channel puts that channel back exactly, in one Ctrl+Z", "version": "2026.9.4", "date": "2026-10-01", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "Converting a melodic loop into a channel an earlier conversion made, then pressing Ctrl+Z, took two presses and still left the channel broken: its voice pointed at a wave bank slot that no longer existed and had lost its old root note, so it went silent. The conversion changed the voice's sample, root, fit and one-shot settings directly, outside the undo history, and recorded the new wave, pattern and clips as separate steps. Planning now changes nothing: the voice's new settings come back as ordinary undoable parameter edits, and the whole conversion (the new wave, the new channel or those settings, the pattern and every clip) is one edit, so one Ctrl+Z restores the song exactly and Ctrl+Y puts the conversion back. MelodicLoopConversion.Plan.ToEdit; a new test converts, undoes and compares the saved song with the original.", "tickets": [ 659 ] }, { "title": "Notes you play sound sooner on the Windows OpenGL, Mac and Linux versions", "version": "2026.9.3", "date": "2026-10-01", "area": "Sound engine", "type": "Improvement", "userFacing": true, "details": "A key on a MIDI keyboard or the QWERTY layout, or a click in the piano roll, now sounds about 20-30 ms later on the OpenGL builds instead of about 125 ms, as it already did on the DirectX build: the live audition plays 10 ms chunks there too, and a session that ever runs dry still queues one more chunk rather than clicking. It needed a fix in the audio backend, whose count of queued buffers only caught up once a frame: our KNI fork now counts them as they finish. To get it, the Windows OpenGL and Linux builds moved onto the KNI fork the DirectX and Mac builds already used, instead of the stock KNI package; the same SDL2 and OpenAL libraries ship, from the fork, and drawing is unchanged. Measured on Windows OpenGL: 134 ms before, about 20 ms after, no dropouts in 720 notes; Linux (headless): 124 ms before, 27 ms after.", "tickets": [ 613 ] }, { "title": "Kenobi, Bueller and LadderOne panels show the LFO and filter controls they were missing, and Plucktopus's card knobs work", "version": "2026.9.2", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "Store updates, not an app change: Kenobi v3 (LFO Sync and LFO Division on the Modulation section), Bueller v3 (the vibrato LFO's Wave, Sync and Division), LadderOne v3 (Filter Slope under Kind) and Plucktopus v4 (its Stack card's pick knobs are now Click Level and Click Tone, the dials that actually drive the pick; the old Pick Level/Pick Tone pair did nothing). All four are compatible versions, so songs on 'latest compatible' pick them up without asking. In the repo, the panels are authored in KenobiUi/BuellerUi/LadderOneUi (LadderOne joined StoreModuleSources so author_panel can reach it), and a new StoreModuleSourcesTests pair fails any store panel that leaves a wired control off or shows one wired to nothing; the store panels are also held to FactoryPanelLayoutTests' geometry now. Cloudform's and Laufeyson's published panels were checked against the source and already carried every control. Bifrost's sample layer (#660) and its panel overlaps (#661) are known gaps the tests list.", "tickets": [ 631 ] }, { "title": "The Grrowl preset pack is made for the current Bifrost, so its store page no longer says it was made for an earlier version", "version": "2026.9.2", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "Grrowl v2 in the store: the fifty basses checked against Bifrost v3 and later (whose oscillator routing changed the interface in v3) and published against that interface, with the levels retuned when Bifrost's missing wavetables came back. In the app, publish_preset_pack takes 'flc' to publish a new VERSION of a pack rather than a second product (the wizard keeps the listing's name and description, which a version would otherwise overwrite), an author may republish the presets their own installed pack lent (anybody else's are still refused), apply_preset applies a preset by name or id from the control API, and module_ui opens an installed store module on a fresh profile instead of calling it not installed.", "tickets": [ 634 ] }, { "title": "Set the licence of a sample inside your module, so the store can accept it", "version": "2026.9.2", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "A file added to a module with Add a file... had no licence, and there was no way to give it one, so a module built on samples could never be published: the store refuses any sound it cannot redistribute. Right-click the file's row in the module editor and choose Credit and licence... to set its licence (CC0, CC BY, CC BY-SA or My own recording), author and source URL; the row now shows the licence, and the dialog says straight away whether the store will accept it. The publish wizard's Embedded audio check now runs the store's own rule over the exact package it is about to upload, so it refuses what the upload would refuse, in the same words, before anything is sent.", "tickets": [ 630 ] }, { "title": "The Compressor can duck to another channel: turn on Side-chain and pick the Key on the Stack", "version": "2026.9.2", "date": "2026-10-01", "area": "Sound engine", "type": "New feature", "userFacing": true, "details": "The suite Compressor now has a Side-chain switch. Set the channel's Key on the Stack's top strip (the kick, say), turn Side-chain on, and the Compressor turns this channel down when the key is loud instead of when the channel itself is: the classic pad pumping to the kick, without reaching for the legacy Sidechain or Compressor. The SC high-pass filters the key, so a kick's sub can be ignored. Side-chain is Off by default, so songs made before it sound exactly the same. Measured on a rendered song: a pad keyed by a kick drops about 21 dB under each hit and is untouched between them.", "tickets": [ 640 ] }, { "title": "File > Export > Song as ezmuze project saves the song as a real .ezproj file (a download on the web)", "version": "2026.9.2", "date": "2026-10-01", "area": "Web", "type": "Bug fix", "userFacing": true, "details": "In the web app, the Browser storage service said \"Use Download to get a real file\", but there was no Download anywhere. File > Export now has Song as ezmuze project (.ezproj)..., which downloads the open song as a project file in the browser and offers a save dialog on the desktop. It is a copy: where Ctrl+S saves does not change. The Browser storage description now points at it.", "tickets": [ 655 ] }, { "title": "Pressing Okay in Preferences no longer switches off a --control-port session's remote control", "version": "2026.9.2", "date": "2026-10-01", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "A session started with --control-port lost its remote control API the moment anyone pressed Okay (or Apply) in Preferences, because Okay re-applied the stored Preferences > Advanced setting, which is Off unless you turned it on there, and stopped the server until a restart. Apply and Okay now only touch the remote control server when you actually change the enabled switch or the port; a real change still takes effect immediately.", "tickets": [ 651 ] }, { "title": "The Stack-face docs agree with the app: a Dial does not narrow a card", "version": "2026.9.2", "date": "2026-10-01", "area": "Modules", "type": "Internal", "userFacing": false, "details": "With a Dial in a module and nothing marked In Panel, the Stack card shows the Dial plus every other unwired control. That is the locked rule (design-guide 2026-08-28: a face is its controls PLUS anything still unreachable), but docs/stack-design.md section 2 still described the older controls-only rule. The doc now carries the amendment, ExposedParamPorts says so, and a test pins it: the Dial joins the card, the port it drives leaves it, and marking is what narrows it. The docs portal pages (make-synth-deeper, make-components) say the same.", "tickets": [ 650 ] }, { "title": "Chip's Osc2 Detune reads in semitones", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Osc2 Detune read as a note name, C-1 at its centre and C-1-1200c at its minimum, on the panel, the UI editor and the Stack card. The dial is added to the oscillator's pitch alongside the note, so it is a transpose: it now reads -12 st to +12 st, and typing \"+7 st\" sets it. The same holds for any dial wired into a pitch input that something else also feeds. A note name off the end of the keyboard also stopped growing a thousand cents: it reads as the end note.", "tickets": [ 649 ] }, { "title": "Envelope cards on the mod rail show their knobs", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "With the mod rail open, Chip's Noise Env, Amp Env and Click Env cards (and any envelope source) were wide empty boxes: the rail made room for the envelope's own graph and skipped its knobs, but it can only draw an LFO's editor, so nothing appeared. A source now gets an editor on the rail only when the rail can draw that editor; every other source gets its row of knobs (Attack, Decay, Release, Curve, Delay, Hold for the ones nothing else drives). Amp Env shows only Curve, Delay and Hold, because Chip's own dials drive the rest.", "tickets": [ 648 ] }, { "title": "Chip's Osc1 Wave and Osc2 Wave read Pulse, Saw and Triangle", "version": "2026.9.2", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "Chip's wave selectors read as bare numbers on the Stack card (0 and 1) and as unlabelled segments on its panel, because they feed a select into a table number rather than a choice port they could borrow names from. They now carry their own names, Pulse, Saw and Triangle, on the card, the panel and the dropdown. Along the way: any selector that names its own positions and drives no choice port now reads its value as the index it is, so its names line up with the sound (Freyja's Range at 1 read 16' instead of 8', for one). Chip devices already in a song keep their numbers until the device is added again.", "tickets": [ 658 ] }, { "title": "The Stack closes when an undo removes the channel it shows", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Undoing the add of a channel the Stack was showing (a new return channel, say) took the channel out of the song but left the Stack open on it, still titled with its name and showing its cards. Channel-list edits now rebuild the rows in place rather than the whole view, so nothing checked the Stack any more. It now closes when its channel has left the song, and follows the channel when an edit replaces the channel object with one of the same id. A Stack showing any other channel is untouched.", "tickets": [ 657 ] }, { "title": "Convert to instrument no longer offers a drum kit as the channel to play into", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "The wave picker's Convert to instrument... listed every channel with a wave-bank voice as a place to put the new instrument, a Capacitor drum kit and ordinary wave-bank channels included, and choosing one rewrote that channel's first sampler: the kit's pads stopped working as before, and a wave-bank channel, which picks its sample by note, would have played a different sample for every note. It now lists New channel, plus only the channels that are already a played sampler (the ones an earlier conversion made), so nothing is silently overwritten.", "tickets": [ 653 ] }, { "title": "The Samples tab's source list no longer runs one source into the next", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "In the pattern explorer's Samples tab, a source with a status under its name (\"not signed in\", \"offline\") was drawn in a row only one line tall, so the status spilled out of the row and ran into the next source's name, as \"ezmuze central / not signed in\" did into \"Freesound\". Those rows are now two lines tall, with a small gap between sources.", "tickets": [ 656 ] }, { "title": "Control API: explorer_tab brings the pattern explorer's tab to the front", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": false, "details": "With the pattern explorer sharing a window with another tab (a channel's Stack, the Clip editor), POST /action explorer_tab answered ok while the other tab stayed in front, so the explorer's tiles were not in the element tree and clicks aimed at them missed. Opening the explorer when it is already open now shows its tab in whichever window hosts it, which is what explorer_tab (and the app's own calls after an import or a Stack action) relied on. docs/remote-control-api.md says so.", "tickets": [ 622 ] }, { "title": "Pattern tiles draw their waveform and preview on every channel, imported MIDI included", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "In the pattern explorer, the tiles of a channel whose generator has automation lanes (The Dark's Plink, for one) drew no waveform and did nothing when clicked, and so did a pattern just imported from a MIDI file. A tile showed the waveform the timeline had rendered, and the timeline renders each placement separately when automation makes them differ, so no picture of the pattern on its own was ever made; an imported pattern is not placed anywhere yet, so nothing asked for one at all. The explorer now asks for each tile's pattern itself, with its audio, so the tile draws and the click previews it. Playback is unchanged.", "tickets": [ 654 ] }, { "title": "The floating wave picker draws every row in full after scrolling", "version": "2026.9.2", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Scrolling the floating wave picker (Choose sample... on Capacitor, for one) could leave some rows drawn with only their number, name and waveform: no play button, no R button, no detail line. It was GustUI's cache of which parts of an element are on screen, which did not notice when the element itself moved, so a row that had been half outside the list kept drawing only the half it could see then. The cache now notices. Any scrolled list could show it; the wave picker was where it was seen.", "tickets": [ 652 ] }, { "title": "The store tells you where an installed loop pack's loops are", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "An installed loop pack's store page said its loops were \"in the pattern explorer, under Loop packs\", and the install message pointed at the pattern explorer too; there is no Loop packs label anywhere. A pack's loops arrive as the patterns of a channel added from it, so both now say so: add a channel with + > From Pack, and the loops are in that channel's Patterns tab.", "tickets": [ 635 ] }, { "title": "Plucktopus's Stack card source points its pick knobs at the dials that drive the pick", "version": "2026.9.1", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": false, "details": "The Stack card face for Plucktopus listed \"Pick Level\" and \"Pick Tone\", two dials its graph carries but never wired, so those card knobs did nothing; the pick is the shared click layer, which the panel already drives through \"Click Level\" and \"Click Tone\". The face now names those. It reaches the store's Plucktopus only when its author re-applies the face (author_face) and publishes a new version; until then the store package still carries the old marks. A new test fails any store module whose card has a control wired to nothing." }, { "title": "The conversion report for an ezmuze 3 song lists what could not be brought over", "version": "2026.9.1", "date": "2026-10-01", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "Importing an ezmuze 3 song could leave things behind without saying what: channels left out because their loop pack is not licensed to you (only a count on the status line), loops no installed pack has (silent), and TriggerFX that were skipped (not mentioned). The conversion report now has a line for each, under the channel it belongs to, and TriggerFX under their own heading. When anything was lost, a \"Song imported\" notice offers the report as soon as the song opens; it is also in Song settings > Conversion report at any time.", "tickets": [ 629 ] }, { "title": "The control API's glue question uses the clips you name", "version": "2026.9.1", "date": "2026-10-01", "area": "Services", "type": "Bug fix", "userFacing": false, "details": "POST /action glue with ask:true and a clips list asked the mixed-selection question of whatever was already selected, so naming a channel's pattern and sample clips was answered \"the selection has no channel with both pattern and sample clips\" although Ctrl+G on the same clips asked it. The ask path now selects the named clips first, exactly as the glue itself does.", "tickets": [ 639 ] }, { "title": "Dropping a store package (.ezmod) onto the window installs it", "version": "2026.9.1", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "A sealed store package dragged onto the window was turned away (on Windows, with the no-entry cursor) or treated as audio and refused, although the store has always said a dropped package installs. Dropping one now runs the very install the store's Get runs, with the same checks and no others: the package has to be signed by ezmuze central and open with the signed-in account's key, so only a module licensed to you installs. Signed out, a package bought on another account, or a listen-only package is refused with the reason and nothing is installed.", "tickets": [ 627 ] }, { "title": "Saving a song opened from an ezmuze 3 file asks where to save it, and never overwrites the original", "version": "2026.9.1", "date": "2026-10-01", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "After File > Open of an ezmuze 3 .eez or .eee file, a plain Save (Ctrl+S) wrote the new ezmuze studio format straight over the original file, which ezmuze 3 then could not read; the same went for a legacy song opened from ezmuze central. The imported song now starts with no file of its own, so Save asks where to put it, exactly as Save as does, suggesting a .ezproj. The original legacy file is never touched, and it stays in your recent projects, where opening it imports it afresh.", "tickets": [ 626 ] }, { "title": "Synths added from the store play chords again instead of one note at a time", "version": "2026.9.1", "date": "2026-10-01", "area": "Modules", "type": "Bug fix", "userFacing": true, "details": "A store module added to a song (Kenobi, Plucktopus, Rhodine and the rest) arrived with Voices set to 1 whatever its author published, so chords played one note at a time. The voice count lives on each device in the song, and the store's add path never copied the published one across the way the factory and your own modules always have. New store devices now start with the polyphony their package declares. A device already in a song keeps its saved Voices setting; raise it on the Stack card.", "tickets": [ 633 ] }, { "title": "Starting a new song from a pack in File > New project asks about unsaved changes first", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "File > New project > (a pack) replaced the open song straight away, so any unsaved work in it was lost without a word. It now asks the same Save / Discard / Cancel question as Empty project, Open and opening a demo. The welcome screen's pack templates go through the same guard; there it never shows, because no song is on screen.", "tickets": [ 628 ] }, { "title": "Warp Mode on Chip, Kenobi and the other synths is a named choice: Off, Bend, Sync, Fold, Mirror", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "The Stack showed Warp Mode as a number (0 for Off). A selector borrows the option names of the choice it drives, but only when it drove exactly one, and Warp Mode drives every oscillator's warp at once. A selector whose destinations all offer the same options now reads those names; one driving choices that differ still shows a number. Saved songs are unaffected: the mode is stored as its index, as before.", "tickets": [ 636 ] }, { "title": "The Sends card shows a new return's send straight away", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "After Sends + > New return channel (reverb), the return channel and the send were both made, but the Sends card kept saying No sends until you selected the channel again. Adding the return is one undo step that adds a channel and a send together, and the sequencer rebuilt its rows for the new channel without telling the Stack about the send. The Stack now refreshes after any such combined edit, as long as the channel it shows is still in the song.", "tickets": [ 643 ] }, { "title": "Dropping several audio files makes Audio 1, Audio 2, Audio 3 instead of three Audio 2s", "version": "2026.9.1", "date": "2026-10-01", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "A new channel made by dropping an audio file was numbered by counting the channels that have devices. An audio channel has none, so the number never moved and every dropped file landed on another channel called Audio 2. New Audio and Waves channels are now numbered one past the highest number already in use under that name.", "tickets": [ 642 ] }, { "title": "Moving a renamed channel names it by your name for it", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "After you renamed a channel, the status bar still used its original name when you moved it into a group (Moved \"Audio 2\" into ... for a channel you had called Kick). Status messages about a channel (moving it, placing a clip on it, and the no-instrument hint) now use the name you gave it, and the channel's own name only when it has none.", "tickets": [ 638 ] }, { "title": "Measure loudness says Close once the report is ready", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "File > Measure loudness kept its Cancel button after the measurement had finished and the report was showing. The button now reads Cancel while the song is being rendered and Close once the result (or an error) is up. Escape presses it either way.", "tickets": [ 646 ] }, { "title": "The wave picker shows a short sample's real length, in milliseconds", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "The wave picker printed 0.0s for every sample under a second, and because the length was worked out with whole-number division every other length was cut down to whole seconds too (a 1.5 second sample read 1.0s). Lengths under a second now read in milliseconds (250 ms) and longer ones in seconds to one decimal (1.5 s), the same way in the wave picker, the Wave bank and Samples tiles and the message after adding a sample.", "tickets": [ 645 ] }, { "title": "Capacitor's Choose sample... button is readable", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "The caption was dark grey on Capacitor's dark button. The panel uses the vintage theme, whose dark text is meant for a cream background, but the button has its own dark fill. A sample or pack picker's caption now keeps the theme's text colour only when it can be read on the button's own fill, and otherwise switches to black or white, whichever reads.", "tickets": [ 644 ] }, { "title": "Help's Quick Tips match the app again, and View > Show wave bank says what it does", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Three Quick Tips in F1 Help were out of date. Clips said M switches a clip between repeat and stretch: M now adds a marker, every clip repeats, and dragging an edge by its bottom half changes its speed. Saving and exporting pointed at File > Publish, which no longer exists: audio is File > Share or File > Export, and File > Export also writes MIDI. Blue and red lines described filters nothing in the app shows any more, so that tip is gone. View > Toggle wave bank only ever showed the wave bank, so it is now called Show wave bank.", "tickets": [ 632 ] }, { "title": "Shortcuts show ] [ and 0 instead of OemCloseBrackets, OemOpenBrackets and D0", "version": "2026.9.1", "date": "2026-10-01", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "The Markers menu listed its next/previous marker keys as OemCloseBrackets and OemOpenBrackets, and F1 Help showed the module editor's zoom reset as Ctrl+D0. Menus printed the key's internal enum name and the help screen only knew a handful of keys. Both now use one key-name table in GustUI (KeyNames.Display), so every menu row and every help listing prints the key's own legend: ], [, 0, Esc, Page Down and so on.", "tickets": [ 637, 641 ] }, { "title": "The channel cost test waits for the cost window instead of guessing its timing", "version": "2026.8.376", "date": "2026-10-01", "area": "Build", "type": "Internal", "userFacing": false, "details": "ChannelCost_DoesNotThrow_WhenAChannelIsAddedMidWindow rendered for a fixed 1.1 s and expected the 1 s cost window to have rolled; with the release dry run sharing the CI box that 100 ms margin failed a beta build. It now renders until the reading is back, up to 10 s." }, { "title": "The release workflow builds the macOS head too", "version": "2026.8.376", "date": "2026-09-30", "area": "Build", "type": "Internal", "userFacing": false, "details": "promote.yml's head matrix never gained osx-arm64-gl when build.yml did, so a release dry run failed in finalize on the missing .dmg. It now builds all five heads, and PromoteWorkflowTests fails if the two lists drift apart again." }, { "title": "If a song fails to open, you are told why instead of being left with an empty window", "version": "2026.8.375", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "When something went wrong while the sequencer was being built for a song you opened (a demo, a file, a recent song), the welcome screen closed, no sequencer appeared, and nothing was said or logged: the only way out was to restart. The failure is now shown as \"Could not open : \" on the status line and in a message that stays until you close it, the welcome screen comes back so you can try again, and the app log records what failed and where. Other song opens and imports that ran in the background now report their failures the same way.", "tickets": [ 625 ] }, { "title": "The Samples tab previews like the other tabs, has a tile view and a right-click menu, and every tab can search", "version": "2026.8.374", "date": "2026-09-30", "area": "UI", "type": "New feature", "userFacing": true, "details": "The Pattern explorer's three tabs now share one preview player. Whatever you are previewing - a pattern, a wave bank sample, a sample in the Samples tab, the wave picker or Add a sample - shows a playhead moving across its waveform with the part still to come dimmed, its play button turns into a stop, a playing tile shows a small stop button, and the status bar says what is playing. Clicking the item that is playing stops it. The Samples tab can now show its samples as tiles like the other two tabs as well as a list (the two buttons at the right of its path line; your choice is remembered), and both stay fast with thousands of files. Samples from your own folders draw their waveforms as they scroll into view, the detail line says how many bars a sample is, and the waveform no longer stretches across the whole width of a wide window. Right-click a sample to preview it, add it to the wave bank without placing it, place it on the selected channel, or show it in its folder or in the store. The Patterns and Wave bank tabs have a search box that filters as you type.", "tickets": [ 624 ] }, { "title": "GustUI's windowed list can lay its items out as a grid", "version": "2026.8.374", "date": "2026-09-30", "area": "UI", "type": "Internal", "userFacing": false, "details": "VirtualListElement gained a multi-column mode (SetLayout: line pitch, items per line, column pitch). Items fill a line left to right, the window is whole lines, and ScrollToIndex brings an item's line into view, so a tile grid of thousands builds only the tiles on screen, as the list does. The arithmetic (Lines, ColumnsFor, GridRange, CellOrigin) is pure and unit-tested. Previews are now tracked by their preview key as well as a Guid (LoopPreview.PlayingKey), which is what lets the shared preview player (PreviewFeedback) follow items that have no Guid, and Add a sample and the Samples tab's Add to wave bank share one add path (WaveBankAdd)." }, { "title": "Stretched loops in older songs no longer change key", "version": "2026.8.373", "date": "2026-09-30", "area": "Sound engine", "type": "Bug fix", "userFacing": true, "details": "Songs using the Phase vocoder stretch (older and imported songs, and anything that names RubberBand) fitted pack loops to the song's tempo with Preserve pitch on, but the stretch corrected the pitch the wrong way: instead of keeping the loop's key it moved it further, so a loop made at 100 BPM played in a 124 BPM song came out about seven semitones sharp. It now keeps the key, like every other algorithm, and a test checks every algorithm at four stretch lengths so this cannot come back unnoticed. Cached renders made with the old stretch are thrown away and rebuilt. The shipped demos sound exactly as before: none of them stretches a loop this way.", "tickets": [ 623 ] }, { "title": "Play in time previews start straight away in step with the song, and keep their key", "version": "2026.8.372", "date": "2026-09-30", "area": "UI", "type": "Improvement", "userFacing": true, "details": "With Play in time on, a preview no longer waits for the next bar. It starts as soon as it is ready, part-way into its loop, exactly where it would be had it started on the bar the song is in, so it is in step from the first beat. Previews also always keep their key now, at 0.5x, 1x and 2x: they used to follow the song's own stretch, which changes the pitch when Preserve pitch is off and on songs using the Phase vocoder algorithm (every demo and imported song), so a sped-up loop could sound higher.", "tickets": [ 621 ] }, { "title": "Play in time: explorer previews can start on the next bar, fitted to the song's tempo, and loop", "version": "2026.8.371", "date": "2026-09-30", "area": "UI", "type": "New feature", "userFacing": true, "details": "The Pattern explorer (patterns, wave bank and samples) and the sample browser have a Play in time switch, with a length choice of 0.5x, 1x or 2x beside it. With it on, a preview whose length is known in bars (pack loops and patterns, or a sample whose source, file name or WAV tags give its tempo) starts on the song's next bar while the song plays, or at once when it is stopped; is stretched to the song's tempo with the song's own loop stretch, keeping its key when Preserve pitch is on (0.5x plays it at double the tempo, 2x at half); and loops until you click it again, preview something else, or switch Play in time off. Anything with no known tempo, such as a one-shot, still plays once as it is. The switch and length are remembered and shared by every explorer.", "tickets": [ 621 ] }, { "title": "Clicking a tile in the Pattern explorer previews it while the song is playing", "version": "2026.8.371", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "A click on a pattern or wave bank tile did nothing at all while the song played: no sound, no Previewing line, and a pack pattern that had never been heard was not even fetched, so its tile stayed blank. The explorer skipped every preview during playback. It now previews whether the song is playing or not, through the preview's own sound, so the song carries on undisturbed and you can audition patterns against it.", "tickets": [ 620 ] }, { "title": "Add channel lists From Pack before Pack group", "version": "2026.8.370", "date": "2026-09-30", "area": "UI", "type": "Improvement", "userFacing": true, "details": "In the + Add channel menu, From Pack (start a channel from one of your packs) now comes before Pack group (add a pack's standard folders to the song, a tool for pack authors), so the everyday choice is the nearer one." }, { "title": "The sample editor can paste over a selection without changing its length, repeat a copy to fill it, and clear it", "version": "2026.8.369", "date": "2026-09-30", "area": "UI", "type": "New feature", "userFacing": true, "details": "Three new edits for the selected area in the Clip editor's sample view, on the right-click menu and the Edit menu. Paste over writes the copied audio into the selection from its start, replacing samples in place: a shorter copy stops when it runs out and leaves the rest of the selection alone, a longer one stops at the selection end, and the sample never changes length. Paste over repeating does the same but repeats a shorter copy back to back until the selection is full, cutting the last repeat at the end. Clear area (the item that was called Silence) sets the selection to silence. They are greyed out with no selection, and the pastes with nothing copied. Each is one undo step. Menu only: Ctrl+Shift+V already means the opposite in the sequencer.", "tickets": [ 619 ] }, { "title": "A pack pattern dragged onto the timeline without previewing it first now plays and shows its waveform", "version": "2026.8.368", "date": "2026-09-30", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "A pack loop's place in the wave bank holds no audio of its own; it names the loop, and the loop's audio is fetched separately. Clicking a pattern in the pattern explorer fetched it, but dropping one on the timeline never did, so a pack pattern dragged straight from a channel added from a pack (or a sample-pack loop dropped before its row had loaded) landed as a blank clip that played silence. The drop still places the clip at once, and now fetches the loop's audio behind it; the clip fills in with its waveform and plays as soon as it arrives.", "tickets": [ 618 ] }, { "title": "The line between channels runs the full width of a widened channel header", "version": "2026.8.367", "date": "2026-09-30", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "The resize bar under each channel header, which is also the line that splits one channel from the next, was sized once when the row was built. Widening the header column moved it but never widened it, so the split line stopped at the old width. It now follows the column, and so does the one under each automation lane band.", "tickets": [ 617 ] }, { "title": "The desktop app opens in seconds again, instead of sitting for minutes with no window", "version": "2026.8.366", "date": "2026-09-30", "area": "Desktop", "type": "Performance", "userFacing": true, "details": "Startup seeds every factory preset and module on every launch, and each seed rewrote the whole library to disk: hundreds of preset seeds times about 900 preset files, and about 50 module seeds times about 70 definitions. As the library grew with the effects suite this reached minutes with no window. An unchanged preset is now not written at all, a changed one writes only itself and the index, and the module library writes only the entries that changed. Measured on the DX build: from several minutes, then 35 s, to about 6.6 s to the window.", "tickets": [ 616 ] }, { "title": "Web: the pattern that crashed store modules (#581) is removed from 13 more places and guarded by a test", "version": "2026.8.365", "date": "2026-09-30", "area": "Web", "type": "Internal", "userFacing": false, "details": "#581 was a conditional expression (a ? b : c) inside a try whose catch has a 'when' filter: under .NET WASM AOT the reference it produced could be lost to the garbage collector, and the web runtime died. Measured in Edge on Bank Manager's real data, that exact method stored a dangling list in about a third of versions and killed the runtime every run, while the same code with a plain catch, with no conditional, or moved into a method the try only calls was clean in 600 versions each; building objects inside the filtered try was not the problem on its own. Every filtered catch in the projects the web build compiles was checked (50 sites). The 13 that still had a conditional, ?? or ?. inside the filtered try now call a separate method instead: loading a song's store modules (TryMaterialize, which did not reproduce but was split anyway), module upgrades, listen grants, the module About panel, embedding a store module, two publish-wizard checks, the package-key and licence-claim calls, review writes, and Serum import. WebAotExceptionFilterTests fails on any new one. docs/web-aot-exception-filters.md has the rule, the measurements and the full site list.", "tickets": [ 614 ] }, { "title": "Automation segments can curve, hold or S-bend", "version": "2026.8.364", "date": "2026-09-30", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "Every segment of an automation lane now has a small square at its middle. Drag it up or down to bend the segment (or Alt+drag anywhere along it), double-click it to straighten it again, or right-click it for Straight, Exponential, Logarithmic, S-curve and Hold (step). With points selected, the lane menu's Shape the selected segments sets many at once. Playback, the lane drawing, the row overlay, copy, paste, undo and saving all follow the shape, and moving a point keeps its segment's shape. Straight segments play exactly as before. REAPER import now brings envelope point shapes across: square, slow start/end, fast start and fast end land as the matching shapes, and a bezier as a curve of the same tension. Songs saved with this version need this version or later to open (song format 40).", "tickets": [ 592 ] }, { "title": "Adding a store module no longer crashes the web version of the studio", "version": "2026.8.363", "date": "2026-09-30", "area": "Web", "type": "Bug fix", "userFacing": true, "details": "In the browser, swapping a channel's instrument for a store module (Bank Manager, Rat 030 and other large ones) could freeze the studio for good, with the console full of \".NET runtime already exited\" errors. It happened whether or not the song was playing. The first time a session opened one of those packages, reading its versions went wrong: the code that reads a version sat inside an error handler of a kind that the browser build's compiled code does not handle safely, and a memory clean-up in the middle of it threw away a module's definition while the version still pointed at it. That memory was then reused, and the next write through it damaged the app's memory until the runtime stopped. The reading now happens outside that handler, so every version keeps its own definitions. Tested by opening every store module and swapping instruments during playback 70 times in Edge, where the unfixed build crashed within the first few swaps in 7 of 8 runs. The desktop apps were never affected.", "tickets": [ 581 ] }, { "title": "Notes you play answer much faster on the DirectX build: about 25 ms instead of about 125 ms", "version": "2026.8.362", "date": "2026-09-30", "area": "Sound engine", "type": "Improvement", "userFacing": true, "details": "Playing a MIDI keyboard, typing notes on the QWERTY layout, or hitting notes in the piano roll now sounds about 25 ms after the key, down from about 125 ms (plus, in both cases, your audio device's own delay). That is fast enough to play in time. The live audition now renders 10 ms of audio at a time and keeps about 30 ms queued instead of 150 ms. While notes are being played it asks Windows for a 1 ms timer, so it is never late refilling that queue. If an instrument is ever too heavy to keep up, it queues a little more rather than clicking. Measured with the synth demo playing and half the CPU busy: no dropouts, in the audition or the song. This applies to the Windows DirectX build. The OpenGL builds keep the old queue for now, because their audio backend only counts what it has played once a frame. The browser still plays pre-rendered notes, and its lack of bend and sustain is unchanged and not part of this.", "tickets": [ 610 ] }, { "title": "MIDI keyboards and controllers work on Linux, through the ALSA sequencer", "version": "2026.8.362", "date": "2026-09-30", "area": "Desktop", "type": "New feature", "userFacing": true, "details": "The Linux build now hears MIDI input as the Windows build and the browser do: notes with velocity, the sustain pedal, pitch bend and MIDI learn, from every MIDI source the ALSA sequencer knows about (USB keyboards, virtual ports, other programs). It connects to them by itself, shows them in Preferences > MIDI with the same on/off switches, and follows devices being plugged in and out. On a machine without ALSA's sequencer (no sound driver, or libasound missing) MIDI is simply absent, and Preferences says why. This was built and tested without a real MIDI device, so the first reports from Linux users are welcome.", "tickets": [ 608 ] }, { "title": "MIDI keyboards on macOS (built for the Mac, not yet tried on one)", "version": "2026.8.362", "date": "2026-09-30", "area": "Desktop", "type": "New feature", "userFacing": true, "details": "The macOS app now has a MIDI input backend, CoreMIDI, so MIDI keyboards and controllers can play instruments and be MIDI-learned there as they already can on Windows and in the browser: every connected input is listed in Preferences > MIDI by name, and plugging one in or out is picked up by itself. Please note it was written and built without a Mac to run it on, so it has never been tried on real hardware. It is built to fail safely: if CoreMIDI cannot be started, Preferences > MIDI says why and lists no devices, and nothing else is affected. Reports from Mac users are very welcome.", "tickets": [ 609 ] }, { "title": "Record onto the timeline: arm channels and record audio from your microphone or notes from a MIDI keyboard, in time with the song", "version": "2026.8.361", "date": "2026-09-30", "area": "Sound engine", "type": "New feature", "userFacing": true, "details": "Arm a channel with the new red dot beside its Mute button, then press Record (the red dot after the metronome in the toolbar, or R). From a stop it counts in if a count-in is set, then records; while playing it punches in, and pressing it again punches out. With Loop selection on, only the selection is recorded and each pass becomes a take: the last one plays, and Ctrl+Z steps back through the earlier ones, each take being one undo step. Audio takes are placed as sample clips, embedded in the song. Notes played on a MIDI keyboard during a recording become a pattern on each armed channel that has an instrument. Takes land in time: the app works out when each sample was heard from the audio devices' own timing, and Preferences > Audio has a Recording latency setting with a Measure button that plays the metronome out of your speakers and records it back to find the rest. Also in Preferences > Audio: 'Metronome only while recording'. Not yet: comping between takes, choosing the input device, stereo input, and quantize on record.", "tickets": [ 594 ] }, { "title": "Loop selection works again once a song has been fully cached for playback", "version": "2026.8.361", "date": "2026-09-30", "area": "Sound engine", "type": "Bug fix", "userFacing": true, "details": "On desktop, once the whole song had been rendered to the background cache, playback came off disk and ignored the loop mode: Loop selection played straight through the selected bars, and No loop wrapped to the top instead of stopping. A session now plays from the cache only while the whole song is looping, and live in the other two modes, where the loop rules are followed.", "tickets": [ 612 ] }, { "title": "Record automation live: move a control while the song plays and it is written into its lane", "version": "2026.8.361", "date": "2026-09-30", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "Each channel has an automation mode, set from its right-click menu (Automation): Read plays the lanes and records nothing, as before; Touch records a control while you hold it and hands back to the lane when you let go; Latch keeps recording its last value after you let go, until playback stops; Write records every lane on the channel from play to stop. Armed channels show T, L or W in their row. Moving a knob on a device panel or Stack card while playing then writes points where you heard the move, even though the audio is rendered ahead, creating the lane if it has none. The points are thinned so a sweep is a handful of points and a held value is two, and each recording pass is one undo step. Re-recording a lane plays your new value as you move it. Send levels and MIDI-learned controllers record too. The mixer's own volume and pan cannot be automated yet.", "tickets": [ 596 ] }, { "title": "Menu shortcuts spelled as a key name, like Delete and Escape, no longer wrap onto two lines", "version": "2026.8.361", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "A menu row's shortcut had a fixed 22-pixel box for its key, which fits a letter but not a name: the piano roll's and drum lanes' Pattern menus showed \"Del\" / \"ete\" and \"Esc\" / \"ape\" spilling onto the next row. The key's width is now measured (in GustUI's menu row), the menu grows to fit it, and a long label gives way to the shortcut rather than running under it.", "tickets": [ 605 ] }, { "title": "On the OpenGL builds, held notes now sustain and chords sound when you play the keyboard, a MIDI keyboard or the piano roll", "version": "2026.8.360", "date": "2026-09-30", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "The live audition stream (what makes a held note keep sounding until you let go, lets several notes sound together and follows pitch bend) was only built into the DirectX build. The OpenGL builds for Windows, Linux and macOS fell back to a short pre-rendered one-shot of about 400 ms for every note from the computer keyboard, the piano roll or a MIDI keyboard. They now use the same streaming audition as DirectX, through KNI's DynamicSoundEffectInstance on OpenAL.", "tickets": [ 611 ] }, { "title": "Control API: select_lane_points without a range only reads the lane", "version": "2026.8.360", "date": "2026-09-30", "area": "Services", "type": "Bug fix", "userFacing": false, "details": "POST /action select_lane_points with no startBeats/endBeats defaulted to 0..end and selected every point of the lane, although docs/remote-control-api.md says omitting the range selects nothing and just reads the points. It now reads only: the lane's selection is left as it was and the lane need not be expanded. Either end alone runs from the lane's start or to its end. The range is read by ControlArgs.LanePointRange, unit-tested.", "tickets": [ 607 ] }, { "title": "Sends and return channels: share one reverb or delay between channels", "version": "2026.8.360", "date": "2026-09-30", "area": "Sound engine", "type": "New feature", "userFacing": true, "details": "Add channel > Return Channel makes a return: a channel with no clips holding a reverb at 100% wet. Any channel can then send to it from the new Sends card at the end of its Stack, with a level knob per send (right-click it for pre- or post-fader, its automation lane, or to remove it). A post-fader send follows the channel's fader and pan; a pre-fader one does not. Send levels can be automated like any knob. Returns sum into the master (or their group) like any channel, can send on to another return, and say which channels feed them. Soloing a channel keeps its returns audible, and soloing a return plays everything sent to it. REAPER projects now bring their sends across as sends to return channels, with send volume envelopes as lanes, where they used to be dropped. A lookahead device on a return runs uncompensated for now, and its card says so. Songs with sends need this version (song format 39); songs without them sound exactly as before.", "tickets": [ 595 ] }, { "title": "Play instruments from a MIDI keyboard, and MIDI-learn any knob to a controller", "version": "2026.8.360", "date": "2026-09-30", "area": "UI", "type": "New feature", "userFacing": true, "details": "MIDI keyboards and controllers now work on Windows (both builds) and in Chrome, Edge and Firefox. Notes play the selected channel's instrument live, through the same sound QWERTY musical typing uses, with velocity, the sustain pedal and the pitch-bend wheel. Right-click any knob in the Stack and choose MIDI learn, then move a knob or slider on your controller: it follows from then on, as an undoable edit like a drag (the mod wheel is learnable the same way). Preferences has a new MIDI page listing your devices, each with an on/off switch and a light that blinks when it sends anything, plus the bend range and your learned controllers with a Forget button. Devices plugged in later appear by themselves. In the browser MIDI is off until you switch it on, because the browser asks permission. Not yet on Linux or macOS.", "tickets": [ 593 ] }, { "title": "Markers and named sections on the timeline: jump to them, loop a section in one step", "version": "2026.8.359", "date": "2026-09-30", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "The sequencer's ruler has a marker lane under the bar numbers. Press M to drop a marker at the playhead (on the beat while playing), double-click the lane to add one there, or select some bars and press Shift+M to make them a named section such as Intro, Verse or Drop. Click a marker to put the playhead on it; drag it to move it (bar lines, Shift for free); drag a section's end to resize it; double-click to rename; right-click for Loop this section, Select this section, colour and delete. ] and [ jump to the next and previous marker, L loops the section the playhead is in, and the new Markers menu lists them all to jump to. Selecting a section makes it the area Copy, Cut and Paste act on, and a cut or paste across every channel carries the markers after it along. Markers are undoable, saved in the song (an older build opens such a song and simply leaves them off), and REAPER imports bring the project's markers and regions in as markers and sections.", "tickets": [ 590 ] }, { "title": "A Loudness meter (LUFS and true peak), and File > Measure loudness for the whole song", "version": "2026.8.359", "date": "2026-09-30", "area": "Sound engine", "type": "New feature", "userFacing": true, "details": "A new Loudness meter effect (Utility) measures whatever passes through it, which it leaves untouched, to ITU-R BS.1770-4 and EBU R 128: momentary and short-term loudness as bars, integrated loudness since Play written large against a target you pick (-14 LUFS streaming, -16 podcast, -23 broadcast or -9 club), the loudness range, the highest true peak (warned above -1 dBTP) and the loudest readings. The readings stay on screen when the song stops, so playing the song through gives its number. Put it on the master's Stack to meter the mix, or on any channel. File > Measure loudness... renders the whole song as Share does and reports the same readings without playing it, with what a streaming service's normalisation will do to it. The meter is checked against the EBU Tech 3341 and 3342 reference cases. It sits beside the existing peak meters rather than replacing them: they answer a different question (is anything about to clip, right now).", "tickets": [ 591 ] }, { "title": "Keyboard shortcuts work again straight after renaming something", "version": "2026.8.359", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "After renaming a channel (or anything else through the small Rename dialog), every keyboard shortcut stopped working until you clicked into another text box: the closed dialog's text field kept the keyboard. A text field that is no longer on screen now lets go of it.", "tickets": [ 606 ] }, { "title": "Export a song or a pattern as a MIDI file, from File > Export or a pattern's right-click menu", "version": "2026.8.359", "date": "2026-09-30", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "File > Export > Song as MIDI file writes the whole song as a Standard MIDI File (format 1) that any DAW or hardware sequencer opens: the tempo, the time signature, and one track per channel that plays notes, named after the channel, with every note where it actually plays (clip trims, repeats and stretches included) and its velocity. Right-clicking a pattern in the pattern explorer offers Export as MIDI file for that pattern alone. Notes between keys or with a drawn bend are written as MPE on that track only (each note on its own channel, bends at a 48-semitone range); tracks without them stay plain MIDI on channel 1. Loops and samples are audio and are not included, nor are automation and instruments. A file exported this way reads back through the app's own .mid drop with its notes unchanged.", "tickets": [ 588 ] }, { "title": "Save a song as MP3 (192, 256 or 320 kbps) or FLAC, on the desktop and on the web", "version": "2026.8.359", "date": "2026-09-30", "area": "Sound engine", "type": "New feature", "userFacing": true, "details": "Save as audio (File > Share, and now also File > Export > Song as audio) offers MP3 and FLAC beside Ogg Vorbis and WAV. MP3 has a bitrate row of its own: 192, 256 (the default) or 320 kbps; FLAC is lossless, about half the size of WAV, and carries the song's title and author. Both encoders are written in ezmuze itself, so they work the same in the browser as on the desktop and add no licence to the app. The dialog now shows how far the encoding has got. The MP3 encoder has no psychoacoustic model yet, which is why it does not offer bitrates below 192, and like any MP3 without a gapless header it starts with 24 ms of silence.", "tickets": [ 589 ] }, { "title": "Control API: save_to answers the next save dialog, so exports can be scripted", "version": "2026.8.359", "date": "2026-09-30", "area": "Desktop", "type": "Internal", "userFacing": false, "details": "POST /action {\"name\":\"save_to\",\"path\":...} queues a path for the next native save dialog, which then writes there without showing, the save-side twin of pick_file. Added to drive the MP3, FLAC and MIDI exports end to end in a running app (#588, #589)." }, { "title": "A metronome: a click on every beat, from a new button beside Loop or the K key, with an optional count-in", "version": "2026.8.358", "date": "2026-09-30", "area": "Sound engine", "type": "New feature", "userFacing": true, "details": "The transport has a metronome. Turn it on with the new button beside Loop in the sequencer's toolbar, or K; it is remembered between sessions. It clicks on every beat of the song's time signature, with the first beat of each bar higher and louder, and it follows loops and jumps exactly, because each click is placed against the audio actually being heard. Preferences > Audio sets its volume, its sound (Click, Beep or Woodblock) and a count-in of one or two bars before playback starts from a stop, off by default; the count-in plays whether the metronome is on or not, and moving the playhead while playing never counts in again. The click is only ever heard: it is never in a bounce, an export or the bounce cache. There is no 'only while recording' option yet, because the song does not record yet.", "tickets": [ 585 ] }, { "title": "Quantize, swing and humanize the notes of a pattern, in the piano roll and the drum lanes", "version": "2026.8.358", "date": "2026-09-30", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "The Pattern menu of the piano roll and of the drum lanes has Quantize (also Ctrl+Q), Swing and Humanize. Each opens a small dialog and acts on the selected notes, or on every note when nothing is selected, as one undo step. Quantize pulls note starts to a grid from 1/4 to 1/32, triplets included, with a strength so it can tighten without flattening, and can pull note ends too. Swing delays the off-beat eighths or sixteenths, from straight to a full triplet shuffle; applying it again replaces the swing rather than adding to it. Humanize moves note starts by up to 40 ms and velocities by up to half either way, at random but repeatably, and undo puts back exactly what was there.", "tickets": [ 587 ] }, { "title": "Fixed memory use growing with every song opened, which took the app past 6 GB in an hour", "version": "2026.8.358", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Every song you opened stayed in memory after you moved on to the next one. GustUI keeps a list of every window so that Escape and menu shortcuts reach the right one; a window joined it when it was made and left only when it was closed. The pattern explorer makes its window whenever a song opens but shows it only when asked, so a window that was never shown was never closed and never left the list, and through it the whole previous song, its sequencer and its patterns stayed reachable. In a scripted soak on a Debug build, each round of four demo opens added about 37 MB of managed heap that no garbage collection could free: 205 MB to 660 MB over 13 rounds, with the working set going from 450 MB to 1.75 GB. The list now holds its windows weakly, so it no longer keeps a window alive by itself. Over 17 rounds the heap now stays at 172 to 182 MB and the working set levels off at 800 to 950 MB. Playing, the piano roll, module panels, the Clip editor, Drum chop previews, theme switches and scrolling were soaked the same way and did not grow.", "tickets": [ 604 ] }, { "title": "Recording works on the OpenGL builds with microphones they could not open before, such as a LifeCam webcam", "version": "2026.8.358", "date": "2026-09-30", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "On the OpenGL builds (Windows GL, Linux, macOS), Record a sample failed with 'Failed to open capture device. - OpenAL Error: 40964' on some microphones, a Microsoft LifeCam VX-2000 among them. The audio library those builds ship could only open a device in the exact format it mixes at (48 kHz stereo for the LifeCam), and the app always asked for 44.1 kHz mono. When that fails the app now opens the device itself, trying mono then stereo at 44.1, 48, 96, 32 and 16 kHz, records at the rate the device accepts and folds stereo down to mono. Microphones that already worked open exactly as before. The DirectX build was not affected.", "tickets": [ 582 ] }, { "title": "The compiled GL visualiser shaders no longer depend on the line endings of the checkout that built them", "version": "2026.8.358", "date": "2026-09-30", "area": "Build", "type": "Internal", "userFacing": false, "details": "VisEffectsGL.xnb embeds its hand-written GLSL as text, line endings included. The committed DesktopGL and web copies had been built from a tree with one CRLF source among three LF ones, which no checkout reproduces, so the file showed as modified after every local content build. The .frag, .vert, .glsl and .glslfx sources are now eol=lf in .gitattributes, the KNI RawGlslEffectImporter normalises them to LF when it reads them (so a CRLF copy compiles to the same bytes), and both .xnb were rebuilt once from LF sources. The shaders themselves are unchanged. ContentFreshnessTests now checks that the committed DesktopGL .xnb embeds the LF form of every source.", "tickets": [ 580 ] }, { "title": "Opening the piano roll no longer closes an open module panel", "version": "2026.8.358", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Opening the piano roll, the drum lanes or the wave picker from a clip closed any device panel you had open, and its UI editor with it; opening a panel likewise closed the piano roll. That rule dated from when both were full-screen sheets and only one could be on screen. They are ordinary windows now, so they stay open together: the pattern editor comes to the front, and the panel is one click away behind it, or in the View menu's window list. A piano roll opened with a panel up also used to be left unlaid-out and frozen, because only one of the two got its per-frame update; both run now. The module builder still replaces a panel or a piano roll, and only one pattern editor is open at a time.", "tickets": [ 603 ] }, { "title": "Piano roll note waveforms now show release tails past the end of a note", "version": "2026.8.357", "date": "2026-09-30", "area": "Sequencer", "type": "Improvement", "userFacing": true, "details": "A note's waveform no longer stops at the end of the note: when its sound keeps ringing (a release, a long decay), the rest is drawn past the note in a dimmer version of the note colour, so you can see how far each note really carries. It keeps the note's height, and anywhere silent is still left empty. A note that starts while another is ringing sits on top of that tail.", "tickets": [ 602 ] }, { "title": "The piano roll shows each note's own waveform inside it", "version": "2026.8.356", "date": "2026-09-30", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "Each note in the piano roll now draws the waveform it actually produced: its attack, how it decays and how loud it came out. The pattern is rendered in the background through the channel's instrument (without effects), and every voice is heard separately before they are mixed, so each note of a chord shows its own sound. All notes share one scale, so a quiet note looks quiet; the waveform is not shrunk again by the note's velocity. It is drawn only inside the note. After an edit the waveforms update a moment later. Notes on a hosted plugin instrument, and bent notes, show no waveform yet.", "tickets": [ 602 ] }, { "title": "The wave picker's Convert to drum kit now opens Drum chop, which fixes a kit landing at bar 1 on top of the loop", "version": "2026.8.356", "date": "2026-09-30", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Convert to drum kit… on a wave picker row is now a shortcut to the Clip editor's Drum chop for that wave, so you see and adjust the cuts before anything is made. It opens on the first clip in the song that plays the wave; when nothing plays it yet, the wave is placed first, on a new channel at bar 1 as a drop would place it (one undo step), and chopped there. The old one-step conversion is gone, and with it three problems: the kit was always placed once at bar 1, on top of a loop that kept playing, because it only looked for the loop in pattern clips (#599); its hits were timed at the song's tempo rather than the loop's, so a 125 BPM loop in a 122 BPM song drifted against the arrangement (#599); it took about fifteen presses of Ctrl+Z to undo (#600); and it analysed the loop on the game thread, freezing the app for over half a second (#601). Making a kit onto an existing Capacitor channel is not offered any more.", "tickets": [ 599, 600, 601 ] }, { "title": "Drum chop in the Clip editor: see where a loop will be cut into drums, adjust it with sliders, hear it and make the kit", "version": "2026.8.356", "date": "2026-09-30", "area": "UI", "type": "New feature", "userFacing": true, "details": "Open a drum loop in the Clip editor and press the Drum chop button (or Drum chop > Chop into a drum kit…, or right-click the clip on the timeline > Chop into a drum kit…). Markers appear where the loop would be cut into hits, and a coloured band under the waveform shows which drum pad each slice will play on: slices in one colour are the same drum. Click the marker lane to add a marker, drag one to move it, drag it off the lane to remove it; markers you place or move are pinned (solid flags) and the sliders leave them alone, while detected ones (hollow ticks) follow the sliders. Sensitivity and Min gap decide where the hits are; Match and Layers decide which hits share a pad; Keep one-offs gives a sound heard only once a pad of its own. Click a slice to hear it, Play kit to hear the loop rebuilt from the pads, Play loop to compare. Make kit puts a Capacitor kit on a new channel under the clip with a pattern that plays it where the clip did, in one undo step; Save as cues keeps the markers on the clip. The loop is analysed at its own tempo and nothing heavy runs on the game thread, so the sliders never freeze the app. Ctrl+Z inside Drum chop undoes marker changes; Escape leaves it.", "tickets": [ 598 ] }, { "title": "The Clip editor's sample view can draw and edit drum chop markers", "version": "2026.8.356", "date": "2026-09-30", "area": "UI", "type": "Internal", "userFacing": false, "details": "The sample view half of the Drum chop mode (#598), not yet reachable. In chop mode the marker lane holds the chop's slice markers instead of the loop and cues: pinned markers are the solid amber cue flag, detected ones a hollow amber tick with a fainter line, and a 3 px band along the foot of the wave shows which pad each slice plays on, in one of sixteen pad colours, with the pad number beside its marker. A click in the lane adds a marker, a drag moves one, dragging one well above or below the lane removes it, a double-click hands a pinned one back to detection, and a click on the wave reports the slice to audition." }, { "title": "A drum kit can be made from chop markers, with pads shared by sound and hits at their own level", "version": "2026.8.356", "date": "2026-09-30", "area": "Sound engine", "type": "Internal", "userFacing": false, "details": "The kit-making half of the Clip editor's Drum chop mode (#598). Send cues to drum pads' machinery (a render of the clip as it plays, pads cut from it, the kit channel below the clip with its effects and mixer, one undo step) now also takes a list of markers to cut at and a pad map: pieces the chop-and-cancel grouping says are the same drum share one pad even though each slice's audio differs, and each hit keeps its level relative to the pad's first as its velocity. The clip's own cues are left alone. Send cues to drum pads itself is unchanged." }, { "title": "Drum chop markers: detected, pinned and vetoed, with their own undo", "version": "2026.8.356", "date": "2026-09-30", "area": "Sound engine", "type": "Internal", "userFacing": false, "details": "The model behind the Clip editor's Drum chop mode (#598), with no UI yet. DrumChopSession holds a loop's slice markers: detected ones follow the settings, markers the user adds or drags are pinned and never moved by them, and a detected marker the user removes leaves a veto so no setting brings it back. It works out the loop's own tempo from the clip (its beats over its seconds), runs the grouping on a captured input so it can go off the game thread and be cancelled, decides which sounds get pads (every sound by default, most used first past sixteen, numbered by first heard; one-off sounds left out when Keep one-off sounds is off), renders what the kit will sound like, and keeps an undo history of its own that leaves the song alone." }, { "title": "The drum loop analysis is split into stages and its four swept constants become settings", "version": "2026.8.356", "date": "2026-09-30", "area": "Sound engine", "type": "Internal", "userFacing": false, "details": "For the Clip editor's Drum chop mode (#598). DrumLoopDecomposer is now Analyse (the mono mix and spectral flux, computed once), PickOnsets (cheap: re-run on every slider move), SnapToAttack (for a marker placed by hand) and Group (the chop-and-cancel grouping over ANY list of onsets, cancellable, and saying which chop each hit came from). The onset threshold, minimum gap, match tolerance and layer ceiling are a Settings record whose defaults are the swept values; the attack window and the 2 ms cancellation slide stay private on purpose. With the defaults the output is bit-identical to before (the golden test)." }, { "title": "Drum loop analysis is pinned by a golden test before it is opened up for the Clip editor", "version": "2026.8.356", "date": "2026-09-30", "area": "Sound engine", "type": "Internal", "userFacing": false, "details": "Groundwork for the Clip editor's Drum chop mode (#598). DrumLoopDecomposerGoldenTests fingerprints today's chop-and-cancel decomposition bit for bit (every sound's samples, every hit's position and gain, the residual) on synthetic breaks at several tempos, captured before the method is split into stages and its swept constants become settings. The defaults must keep reproducing these fingerprints exactly." }, { "title": "Tabbed windows: a calmer title bar with a grip to drag by, and inactive tabs that match the active one", "version": "2026.8.355", "date": "2026-09-30", "area": "UI", "type": "Improvement", "userFacing": true, "details": "Follow-up to the tabbed-window rework. The empty part of a tabbed window's title bar was either a black recess or the title bar's own accent colouring, and with the accent line running under the window buttons they looked like another tab. It is now one plain surface in the theme's colours behind both the empty space and the buttons, with a small dot grip beside the buttons and the move cursor to show where to drag the window. The active tab is the only accented one. Inactive tabs now use the same gradient as the active tab, dimmed, and no longer draw a bottom edge.", "tickets": [ 597 ] }, { "title": "Tabbed windows get their close and maximise buttons back, and tabs drag to reorder or tear off", "version": "2026.8.354", "date": "2026-09-30", "area": "UI", "type": "Improvement", "userFacing": true, "details": "When windows were tabbed together they lost their window buttons: every tab carried its own pin, pop-out and maximise, nothing sat at the far right, and dragging a tab only moved the whole window. Now a tabbed window keeps pin, maximise/restore and close at the far right like any window, and that close closes every tab (anything that asks before closing still asks). Each tab carries only its own close. Drag a tab sideways to reorder the tabs; drag it up or down about twice the tab's height to tear it off into its own window, which keeps following the pointer so you can drop it, dock it or merge it straight into another window. That replaces the pop-out button. Drag the empty strip beside the window buttons to move, dock or merge the whole tabbed window.", "tickets": [ 597 ] }, { "title": "Betas build faster: the heads build alongside the tests, and the preset checks sample", "version": "2026.8.353", "date": "2026-09-30", "area": "Build", "type": "Internal", "userFacing": false, "details": "The beta pipeline went from about 25 minutes to 32-47. Three changes. (1) build.yml: the five heads no longer wait for the test job; they start with it, and deploy-beta waits for both, taking 9-12 minutes of tests off the critical path in front of the ~21-minute web AOT. (2) test.sh runs the suite with --filter Category!=Probe: the per-synth level probes, the two ReportEveryPresetsHeadroom reports and BarAheadBisectTests (a per-channel bisect of the ShippedDemoBarAheadParityTests gate) print tables for a person writing a bank or chasing a parity failure and cost about 4 CPU-minutes a build; run them with --filter Category=Probe. (3) The render-heavy factory-bank checks (audible and no clipping, chord survival, no two presets alike) and HALL 9000's two render loops hear 5 presets drawn at random each run (PresetSample) instead of every preset of every bank; EZMUZE_ALL_PRESETS=1 sweeps every preset, and the seed is printed and replayable with EZMUZE_PRESET_SEED." }, { "title": "Set note velocity by dragging a note's top or bottom edge, in the piano roll and the drum lanes", "version": "2026.8.352", "date": "2026-09-29", "area": "Sequencer", "type": "New feature", "userFacing": true, "details": "Notes always had a velocity (MIDI and REAPER imports carry it, and instruments play louder, brighter or with accents because of it) but there was no way to see it properly or change it: quieter notes were only drawn a little darker, and every new note was full velocity. Now a note's height is its velocity, in both the piano roll and the drum lanes: full height at 100%, a thin line at 0%, centred on its row. Drag a note's top or bottom edge (or just outside it, so even a very quiet note can be grabbed) to set it; a readout shows the value. With several notes selected, dragging one moves them all by the same amount. New notes start at 80% and then use the last velocity you set. One undo step per drag. In the piano roll, zoom in if the rows are too short to grab an edge.", "tickets": [ 586 ] }, { "title": "The sequencer's vertical scrollbar no longer ends up off the right edge after zooming rows or resizing", "version": "2026.8.351", "date": "2026-09-29", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "The sequencer hands the vertical scrollbar's strip back to the timeline when every channel fits, and takes it again when they overflow. It decided that from the sizes of the previous redraw and never checked again, so zooming the rows in (Ctrl+wheel) until they overflowed, or going straight from a maximised window to a smaller one, put the scrollbar just past the right edge, where it could not be seen or grabbed until the window was resized again. Going back left an empty strip where it had been. The sequencer now lays itself out again as soon as the scrollbar starts or stops being needed. tools/lua/sequencer-scrollbar-edge.lua checks all four cases in a running app.", "tickets": [ 583 ] }, { "title": "Dragging into the space under the last channel shows that it will make a new channel", "version": "2026.8.350", "date": "2026-09-29", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "Dragging a wave, sample, pattern or loop from the explorer into the empty strip under the last channel showed the drop landing on that last channel, with its plain name badge, but letting go made a new channel instead. The preview now tells the truth: while you hover in that strip it shows the new-channel ghost at bar 0, where the drop puts it, with the \"— new channel\" badge, and the existing rows dim. What the drop does is unchanged.", "tickets": [ 584 ] }, { "title": "On a song with many channels, the add-channel button can be scrolled into view again", "version": "2026.8.349", "date": "2026-09-29", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "The + button under the last channel was left out of the sequencer's scroll range. Scrolled all the way down, by wheel or scrollbar, it sat just below the edge of the list, and when the channels fitted the window it was cut in half. A song with more channels than the window could show therefore could not gain a channel. The scroll range and the scrollbar now run to the bottom of the + row, and channels that fit the window leave room for it, so it is always reachable. The same band is where you drop a sample or pattern to make a new channel, so that works on long songs too.", "tickets": [ 579 ] }, { "title": "The release pipeline keeps the last 30 betas on the CDN and refuses to run on a nearly full disk", "version": "2026.8.348", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "Every beta was kept on the CDN for ever; 205 of them at ~400-500 MB each filled the VPS disk to 100% on 2026-09-29, breaking bug-board uploads. deploy-beta.sh now keeps the newest EZMUZE_BETA_KEEP (default 30) version directories under beta/ after each successful deploy: version-sorted, only three-integer directory names (never feed/ or changelog.json), the just-published version protected, each removal logged, never fatal. Promotion is unaffected because promote.yml rebuilds from source and publish.sh copies its own stage. prepare.sh, deploy-beta.sh and publish.sh now fail early with a clear message when a volume they write to has under EZMUZE_MIN_FREE_GB (default 10) free, and prepare.sh prunes unused Docker build cache older than 72h box-wide (it had reached 36 GB, mostly from ezmuze-central's per-push image builds). tools/release/prune-betas.sh previews or runs retention by hand; test_prune_betas.sh tests it and runs in test.sh." }, { "title": "The macOS build leaves out the Windows and Linux natives the full KNI dependencies bring", "version": "2026.8.347", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "After #578 made KNI's ThirdParty/Dependencies the full MonoGame.Dependencies submodule, the platform project's content also carried x86/ and x64/ natives (SDL2.dll, soft_oal.dll, libopenal.so) and MonoGame.Framework.dll.config into the macOS publish, and the Contents/MacOS Mach-O-only check failed the build (run 36569856069). build-head.sh now drops them after asserting the two dylibs are there; none is ever loaded on macOS." }, { "title": "One KNI for local builds and CI: ezmuze/kni main", "version": "2026.8.347", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "Local builds compiled KNI from a fork branch (madninjaskillz/kni feature/raw-glsl-vis) while CI compiled a snapshot in ezmuze/kni, with no shared history, so the two could drift (#578). The fork's later work was replayed onto ezmuze/kni main: the Slang shader processor and its fixes, DX recording through Media Foundation, and the XAudio2 quit fix. ezmuze/kni now carries KNI's third-party code as submodules, as the fork did, so the content builder can be built from it. The local ../kni checkout tracks ezmuze/kni main, and the fork branch is an archive. The release build fetches those submodules itself (kni_submodules in tools/release/lib.sh, called from build-head.sh): the DX, macOS and web heads compile NVorbis and the Stb image code from them, and actions/checkout's deploy key cannot fetch other repos. Verified: all five heads build in Release from a fresh clone once the submodules are fetched; the merged content builder reproduces all 23 committed .xnb files byte for byte; the DX and DesktopGL apps play, show the welcome video and quit cleanly.", "tickets": [ 578 ] }, { "title": "Quitting the Windows app while audio plays no longer occasionally hangs", "version": "2026.8.347", "date": "2026-09-29", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "About one quit in eight with sound playing could leave the DirectX build stuck in the background instead of closing. The cause was in the audio library underneath the app (KNI's XAudio2 back end). The callback XAudio2 makes when a buffer finishes took a lock that the code tearing a voice down already held, while that teardown waited for the callback's thread to go idle, so the two waited on each other forever. The callback now waits at most 2 ms for the lock and otherwise hands the buffer end to the next frame's update, so no audio is lost and nothing can deadlock. The same deadlock could also freeze the app for good mid-session, when one sound was replaced while another played. The app side of orderly quitting shipped earlier (#549); this is the library half, now in the KNI build CI uses.", "tickets": [ 555 ] }, { "title": "Removing a module closes its panel and UI editor", "version": "2026.8.346", "date": "2026-09-29", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Deleting a device from the Stack (or undoing the step that added it, or deleting its whole channel) left its custom UI panel and its UI editor open on a device that no longer existed, along with any sample or content picker opened from that panel. They now close as the device goes. The module builder already did this.", "tickets": [ 577 ] }, { "title": "The mouse wheel over the piano roll's keyboard scrolls up and down", "version": "2026.8.346", "date": "2026-09-29", "area": "UI", "type": "Bug fix", "userFacing": true, "details": "Over the piano keys at the left of the piano roll, the mouse wheel did the grid's gesture, zooming time, which does nothing useful there and nothing at all once the view is zoomed right out, so the scrollbar was the only way up and down. It now works like the sequencer's track headers: the wheel over the keys scrolls up and down, and Ctrl+wheel there makes the rows taller or shorter. Over the grid nothing changed: the wheel zooms time, Shift+wheel scrolls and Ctrl+wheel zooms the rows.", "tickets": [ 576 ] }, { "title": "The drum lanes select, move and resize like the piano roll", "version": "2026.8.345", "date": "2026-09-29", "area": "UI", "type": "Improvement", "userFacing": true, "details": "The drum lane editor lagged behind the piano roll. Right-drag now draws a box that selects every hit it touches (Shift adds to the selection), and dragging a selected hit moves the whole selection together, keeping its shape at the edges. A hit's end now shows a resize pointer and drags to lengthen or shorten it, and dragging the end of a selected hit resizes them all by the same amount. That matters because Capacitor stops a sample when its hit ends, so a crash or a long 808 needs a long hit to ring. A clicked hit is now a 16th instead of a whole bar. Dragging no longer rebuilds the whole grid on every step, and the pattern's waveform re-renders once when you let go rather than on every step (see the entry on pattern editors).", "tickets": [ 574 ] }, { "title": "Dragging notes no longer re-renders the pattern on every step, in any pattern editor", "version": "2026.8.345", "date": "2026-09-29", "area": "Sequencer", "type": "Performance", "userFacing": true, "details": "Moving a note re-rendered the pattern's waveform and ring-out on every step of the drag: 52 renders for an eight-second drag in the drum lanes, and 83 for the same drag in the piano roll while the song played (the piano roll waited only while stopped; playback bypassed it, #575). Edits now park the render until the drag ends, or until 400 ms after the last edit for key nudges and undo, and queue it once; play, save and closing the editor queue it at once. The same drags now queue 2 to 4 renders. What you hear still follows the drag: an edited bar plays its instrument live until the new render lands. The pattern's old ring-out is dropped the moment it is edited, so it can no longer play the notes you just moved. The wave picker uses the same rule. /stats patternRenders.deferred shows what is parked.", "tickets": [ 575 ] }, { "title": "When the audio device runs dry, the log says whether a garbage collection was the cause", "version": "2026.8.344", "date": "2026-09-29", "area": "Desktop", "type": "Internal", "userFacing": false, "details": "A single 5 ms output-starved event on Elevation (#573) could not be reproduced in 96 further plays of the four demos straight after loading them, a third of them with the full test suite running as CPU load. Each output-starved event now logs one line: how long the device ran dry, the time since the last buffer against the audio the device held, and the garbage collections (by generation) and GC pause time in that gap. The counters are read at every submit, which costs nothing; only the line itself allocates, and only when the device ran dry (the first 20, then every 100th)." }, { "title": "No more dropout in the first moment of playback straight after opening a song", "version": "2026.8.344", "date": "2026-09-29", "area": "Desktop", "type": "Bug fix", "userFacing": true, "details": "Pressing play straight after opening a heavier song could drop out for about 20 ms right at the start, and /stats counted one audio underrun each time (three of the four demos on the release smoke test). The output device asks for its first three buffers the moment it starts, and the render-ahead thread had not rendered them yet; the fallback that renders on the spot could not get at the renderer while that thread was busy with the song's first, slowest render, and played silence instead. Play now waits until those three buffers are ready before starting the device: 1 to 28 ms on the demos straight after loading, capped at 250 ms, after which it starts anyway as before. After the fix, 24 plays of the four demos straight after loading counted no underruns. A repeatable check is in tools/lua/start-of-play-underruns.lua.", "tickets": [ 572 ] }, { "title": "Duplicating or inserting bars inside a long clip splits it instead of stacking the copy on top", "version": "2026.8.343", "date": "2026-09-29", "area": "Sequencer", "type": "Bug fix", "userFacing": true, "details": "Duplicate, Insert empty bars and Paste insert open a gap and push the rest of the selected channels right, but a clip crossing the insert point was left whole, so the inserted material landed on top of its second half: duplicating bars 4-8 of the Synth demo's Bass and Arp left their 4-14 clips in place with the copy at 8-12 on top (#569). A clip crossing the point is now split there, the way Split at playhead splits: the head stays, the tail (playing on from where the head stopped, a sample clip's fades on the outer ends) moves with everything else. Clips that end or start at the point are not split, and Cut-collapse never splits. The ripple edits moved into Ezmuze.Core (RippleEdits, with RippleMath.Straddlers) so the whole edit is tested; Split at playhead uses the same split. The decision is in design-guide.md.", "tickets": [ 569 ] }, { "title": "Build hygiene finished: nullable and the correctness analyzers are enforced everywhere", "version": "2026.8.343", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "The last step of #451: analyzer rule CA2211 (no public mutable static fields) is switched on and is a build error. Three injection seams became properties: LocalFolderService.OsSearch, DroppedFileWaveSource.AssetResolver and GameThreadSyncContext.ContinuationFailed. The audio engine's 24 diagnostic switches and counters stay fields, each suppressed with its reason: 10 are [ThreadStatic], two are updated through Interlocked, and all are read on the render path. With this, everything the ticket asked for is in place. Nullable warnings are errors in every project (the app heads since the rollout, and now the audio engine, the Splice plugin, the tests and every tool). The four correctness analyzers (CA1305, CA1310, CA1001, CA2211) are errors. The broad Recommended analyzer set stays off by decision (docs/build-hygiene.md section 3). .editorconfig and central package management were already in place. docs/build-hygiene.md records the counts and the method.", "tickets": [ 451 ] }, { "title": "Types holding disposable fields are checked at build time", "version": "2026.8.343", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "Part of #451: analyzer rule CA1001 (a type holding a disposable field should itself be disposable) is switched on and is a build error. None of its 23 findings was a leak. 19 are views holding GustUI elements, whose Dispose is empty because the element tree owns them. ExternalCallThread's worker keeps its wake event for the life of a thread that may be abandoned mid-call. ListingGate and the wave browser leave superseded CancellationTokenSources to the GC, because in-flight work still holds their tokens. The web page owns the game for the life of the tab. Each is suppressed on its type with that reason, so a new disposable field on any other type fails the build." }, { "title": "String prefix and suffix checks say how they compare", "version": "2026.8.343", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "Part of #451: analyzer rule CA1310 is switched on and is a build error. The five app checks it found were machine text compared culture-sensitively and were fixed under #568. The remaining 50 (39 in the tests, 11 in tools) compare ASCII literals against machine or test strings and now say StringComparison.Ordinal. A new StartsWith, EndsWith or IndexOf(string) without a comparison now fails the build." }, { "title": "Every number and date the code formats now says which culture it uses", "version": "2026.8.343", "date": "2026-09-29", "area": "Build", "type": "Internal", "userFacing": false, "details": "Part of #451: analyzer rule CA1305 is switched on and is a build error. Its 239 findings (134 in the app and libraries, 36 in the tests, 69 in tools) now pass a culture explicitly. Text a person reads keeps the user's culture, now written as CultureInfo.CurrentCulture, so nothing on screen changes. Machine text (hex, ids, cache keys, stored integers) says CultureInfo.InvariantCulture, chosen only where the output is identical in every culture. The six sites where the culture really did change a machine format (the crash reporters' time stamps) were fixed separately under #568. A new ToString, Parse or string.Format without a culture now fails the build." }, { "title": "Crash reports and a few path and command-line checks no longer depend on your language settings", "version": "2026.8.343", "date": "2026-09-29", "area": "Core", "type": "Bug fix", "userFacing": false, "details": "Found while switching on the culture analyzers (#451). The desktop, Linux and Mac crash reporters stamped the report and its crash-