Repository navigation
Description
Milestone B verified triage — 2026-11-17:
P0/mixer/defect/needs-repro. Classification ≠ completion of product feature or physical FAT.Scope / release boundary: Core first-launch offline logical edit UX; current-build proof needed.
Exit criteria:
- On fresh alpha.26 with no audio device connected, create/rename tracks, configure FX/WIRE/session, Save/Open .smix without a fake locked reference GUI.
- After device binding/restart, reconcile exactly one logical→physical graph with truthful ACK/metering and no implicit hardware IN/ARM gate.
Canonical issue ledger. Preserve historical evidence below; no speculative merging or rewriting approved audio.
Owner-observed P0 regression (9 Nov 2026)
- On first opening the installed Windows application, an apparently complete fake/reference mixer appears but cannot be operated. Users assume the installed product is broken.
- Even after Audio I/O setup, FX insert requires hardware IN patched to at least one channel. Users cannot design FX/WIRE/channel sessions without a physical input. They must also be able to work with only the internal player.
Source-grounded audit baseline
main@381c140200d2f4ecb965183a2da1ba1e46d5d9c2; no code fix in this issue yet.
- Native GUI assembly derives from the interactive reference Pages model in
site/preview-v2/index.html(see demo state and initialization around lines 184,211,2838) and injectsweb/approved-native-bridge.jsat the end of body viascripts/build_approved_native_gui.py:250-260. Any visible pre-native reference state must not masquerade as operative Windows controls. Note: legacyweb/app.jsis another demo source but not the approved native GUI assembled by this script. src/Main.cpp:225-229acceptsarMixGuiReadyand marks WebView as loaded;web/approved-native-bridge.js:3148-3161sendsvisualReady:true,audioReady:falsebefore native device/graph admission.Main.cpp:4619-4645starts audio only after this handshake. GUI-ready is not session-edit-ready.web/approved-native-bridge.js:1457-1474definesdeviceReady()requiring at least 1 enabled capture input, 2 outputs and accepted graph or M4.:927-939blocks Add Track otherwise.:2934-2976makes insert require admitted ChannelId graph or verified legacy M4 FX lease;:3120-3132hides standalone Add FX.web/approved-native-contract.js:333-335rejects editing intents without a PCM-admitted graph.src/LiveDryRoutingLab.h:79-97rejects a device with zero inputs,:141-148rejects a silent session without verified physical masks,:1260-1277requires active ChannelId PCM for insert, and:1607-1621requires active PCM for Add Channel.src/LiveAudioEngine.h:846-855requires an attached authorized DIRECT device for legacy VST3 edit.tests/approved_native_cold_start_smoke.py:197-215expects Add Track disabled in NEEDS_REMAP, so test coverage currently locks in the wrong product behavior for offline editing.- Related: dynamic channel issue P0 dynamic mixer — qualify add/delete/reorder, native FX/routing and Show persistence #192, approved GUI/native issue P0 approved /preview/ — qualify native GUI parity and authoritative PCM readback #389, draft Player PR feat(player): native virtual PCM source, first-sound routing and lock-free stereo ring #486. Preserve other draft features / release branch independently. Do not merge feat(player): native virtual PCM source, first-sound routing and lock-free stereo ring #486 as an alleged fix of device-independent authoring.
Product contract (inspired by Gig Performer virtual channels + Chameleon, Mixxx independent decks/effect chains)
Separate logical session editing from device binding and PCM playback.
- WorkspaceReady: UI is the real editable native-backed session. From first launch, no fake FX/stored scene, no fake live meter. Channel/FX/WIRE/bus editing, Save/Open, FX catalog, internal virtual sources must work with 0 hardware inputs and even no audio device. Pending operations have real model ACK and persisted logical revision, not invented audio DSP ACK.
- DeviceBound: physical I/O identity, available pins, sample rate, buffer, permissions are attached independently. Input assignment is optional per channel. Empty input means SILENCE, not a disabled mixer.
- RenderReady: the single JUCE callback has compiled and installed the same logical session into an immutable native graph with prepared FX and verified channel/WIRE state. Only this level can claim real PCM, meters, monitoring and sound. Compilation can fail without discarding the user's editable session.
- OutputAudible: requires an actual safe output path, no feedback loop, and genuinely accepted renderer; do not auto-redirect devices or unmute unsafe routes.
Implementation boundaries
- Create/control one authoritative
MixerSessionDocument/ChannelIdmodel from cold launch and persist its changes, independent ofdeviceEpoch. Initial real channel/MAIN graph andSOURCE_NONEinputs are valid without capture pins. - Input and output device binding becomes a separate transition, never a prerequisite for opening plugin browser, inserting/editing FX, BUS or WIRE. Config-only transaction ACK separate from PCM-applied ACK; show precise, quiet status on actual device/FX failures.
- When device is bound or changed, control-worker prepares VST3 instance with negotiated rate/buffer, routes, latency compensation, native lease and one atomic publisher handoff. Audio callback must never instantiate, scan, allocate, lock, disk I/O or run browser audio. Missing vendor plugin preserves placeholder + intended state/route, never a fake processing claim.
- Native first paint must show genuine non-mock editing state, not preview-demo state. Distinguish GUI rendered, session loaded/editable, PCM admitted, output audible without a blocking wizard/lab mode/approval checkbox. Remove demo event listeners and demo meter motion from native entrypoint.
- Player (once feat(player): native virtual PCM source, first-sound routing and lock-free stereo ring #486 qualified) is a typed virtual source into the same ChannelId graph, not another renderer; Play must not require a physical capture pin or hidden ARM/MONITOR step.
Minimum acceptance
- Fresh install, no device and zero capture pins: immediately add INPUT, GROUP, FX, set sends/WIRE, rename, adjust controls, Save/Open .smix, restart; exact logical state persists. No fake moving meters or dead clickable controls. Editor stays fully usable after unplug.
- Output-only setup (no capture): add FX, load internal music player, Play → FX → BUS/MAIN → playback; verify real native PCM and meter evidence. No mandatory mic patch.
- Connect ASIO later: bind real physical IN to any existing channel without replacing channel ID/FX/WIRE/presets; same ASIO output safety; replace device/period/rate and preserve session.
- Reject missing/bad VST3 or invalid routing with a real actionable error and unchanged good state; lazy background scan, no UI/main/audio callback stall.
- Cold launch browser smoke + native Windows process FAT for first GUI frame, no-device, output-only, live input and restart; assert edit ACK vs PCM ACK separately and verify actual native plugin effect with PCM. Keep CI selective (fast preflight, then one exact-SHA Windows qualification).
Guardrail
Do not merely set .disabled=false or rewrite errors while the underlying session owner still refuses operations. No second PCM engine, WebAudio fallback, dummy physical capture pin, forced input selection, or mandatory Monitor/ARM/START ritual.
Activity
masarray commentedon Nov 9, 2026
Deep-audit supplement — 2026-11-09, implementation #491
Current P0 milestone: draft #491, exact SHA
75e99c63a8ef46d6c361113b33f58e325848361e(fast CI green, exact-SHA Windows CTest dispatched; not certified yet).Additional liveness / deadlock risks discovered
src/Main.cpp:225-229acceptsarMixGuiReadybefore PCM;timerCallback():4633-4645callsengine_.start()synchronously on JUCE message thread. Driver scan/open can block WebView events. ItsaudioEngineStarted_one-shot is set before successful start, so a failed startup never proactively retries.ConsoleAppdismisses splash on stage 2 even whenstartupError_is nonempty, revealing reference mixer when underlying device/native graph is unavailable. Fix by separating startup state and true control-plane session, not by lying about visualReady.src/PluginCatalog.h::~PluginCataloginvokesworker_.stopThread(-1)aftersignalThreadShouldExit(), but scanner uses third-party VST3 metadata discovery which can hang/non-cooperate. Simply adding a finite join timeout and destroyingPluginCatalogwould risk use-after-free; the robust P0 repair is process-isolated VST scan + bounded process lifetime, keep owner state only on message/control thread, validate metadata and never load arbitrary vendor code on RT.src/Main.cpp::handleApprovedNativeControlandLiveAudioEngine::addVst3()can construct/prepare vendor processors on the message thread. A single slow plugin can freeze UI even though audio callback remains RT-safe. Prepare off-RT with validated lifetime and bounded ACK; do not migrate arbitrary vendor operations to the audio callback. When threading is incompatible, use a separate child process or plugin-specific safe fallback.src/NativeMixerSessionDocument.hmodels a saved hardware binding plus graph;web/approved-native-contract.js::intentrejects without installed PCM;LiveDryRoutingLabneeds an active device to publish. P0 first-launch: genuine SOURCE NONE ChannelId mixer, zero-capture output-only FX workspace #491's zero-capture output-only startup is not no-device edit support. Full P0 requires one logical ChannelId document that commits offline edits with a distinct model revision and safely compiles once device becomes available.site/preview-v2/index.html, where the embedded reference reducer seeds demo channels, catalog and meters. Native bridge injection near</body>stops demo meters/overrides interaction after its own startup; when no native state arrives, a convincing yet dead mixer remains. Startup must render real model-backed editable controls; until model has loaded, only honest lightweight initialization, no simulated FX/meters or fake authorizations.Workstream sequencing (not cosmetic)
SESSION_EDIT_ACK+ laterPCM_APPLIED_ACK(never conflate). Open/Save and 10→N dynamic ChannelIds must survive no hardware and restart.Merge fence: keep #491 Draft until actual Windows CTest + real GUI/PCM qualifications; do not claim #490 resolved by permitting zero physical capture alone. Never sacrifice existing approved GUI or current main release to rush this change.
masarray commentedon Nov 10, 2026
P0 owner field regression — 10 Nov 2026 — PUBLIC alpha.8 NOT hardware-qualified
Owner installed published
v0.5.0-alpha.8and reports an unusable/fake-looking reference mixer, plus ASIO/PCM pipeline regression. This supersedes previous CI-green release claims. No new tag until actual mixer startup and audio device FAT pass.Source root-cause audit:
site/preview-v2first runs browser reference model;arMixGuiReadywas visual-only andLiveAudioEngine::startruns after that. When saved ASIO does not reopen, there is no native ChannelId PCM graph; the reference model stays on screen.#491SOURCE NONE zero capture is already in main but cannot create a logical mixer without playback. Rebind paths also special-case ASIO4ALL and did not recover a pristine SOURCE NONE session after driver restart.Active focused correction: draft PR #511,
6d92bbcc77e595b55ad77dbb3049fc311d0c683a(read/write code, not a cosmetic-only issue):/preview/byte-for-byte. Rejects Pages-only GUI-ready event. Native demo labels hidden.Remain open / release blockers: device-independent logical mixer document editing/save/restore with no playback; native offline edit ACK separate from PCM ACK; successful fresh-install ASIO output; all-channel WIRE/BUS/FX PCM evidence; UI/driver test on affected physical Windows machine. These cannot be called done because CI compiled or reference DOM painted. Do not force hardware remapping or synthesize a fake source/meter/PCM pipeline. Avoid release controller / fat dispatch until the exact candidate passes focused checks.
masarray commentedon Nov 10, 2026
Forensic correction to prior P0 assessment: Work PR #491 HEAD
3015960bis an ancestor of published alpha.8ffdf279a; core ASIO reattach/pin policies are byte-identical. The confirmed first-paint regression is commit0b0e637bdeleting Work's native-only guard fromscripts/build_approved_native_gui.py, later shipped via #503/#505. Full analysis in PR #511 discussion. Do not reimplement Work or promote draft #511 without hardware qualification.masarray commentedon Nov 10, 2026
2026-11-10 — Inherited native GUI acceptance from #389 (consolidated)
This P0 issue is the authoritative owner for genuine first-frame/native session UX. The single approved public/native
/preview/renderer has been integrated; do not build a new GUI or revive fixed M4 reference startup.Acceptance inherited from #389:
Native audio/physical WIRE and FX evidence stays separately in #412; dynamic channel lifecycle stays in #192. Full cross-issue archive: reconciliation PR #524. This transfer does not mark #490 resolved.