iOS "voice chat" had echo and no noise suppression: real iOS AEC/NS/AGC
come only from Apple's Voice-Processing I/O unit (VPIO), but the core
plays/captures via miniaudio's plain RemoteIO units, so .voiceChat mode
alone never engaged AEC.
Core (ABI PATCH 1->2):
- vc_set_mixed_output_sink + vc_set_external_playback. In external mode the
AudioEngine opens no hardware playback device; a mixer-timer thread drives
on_playback (decode+mix) on a ~20ms cadence and ships the final mix to the
sink. start() also skips the hardware capture device when the MIC stream is
external_feed (AudioParams.external_capture).
- New white-box test test_external_playback (drives the timer with no hw).
iOS/Swift:
- StreamDescriptor.externalFeed; VoiceCatClient.setMixedOutputSink /
setExternalPlayback wrappers.
- IOSVoiceProcessingEngine: AVAudioEngine + setVoiceProcessingEnabled; mic
tap -> feedPcm, mixed-sink lock-free ring -> AVAudioSourceNode (both share
the VPIO unit so AEC has its reference signal).
- IOSAudioRouter.currentConfigUsesVoiceProcessing scopes VPIO to the AEC
presets; SessionState join/leave + reconcileVoicePath() switch paths;
Voice Chat defaults to speaker; Settings surfaces AEC/NS state.
Known: pending on-device verification; a few bugs to fix afterward.
Multi-stage Dockerfile (builder → export → runtime) producing a 149 MB
Ubuntu 24.04 image, verified booting end-to-end on Docker Desktop. vcpkg
fetched via shallow git fetch at the pinned baseline, release-only overlay
triplets (x64-linux, arm64-linux) to halve intermediate disk usage, and
buildtrees deleted within the RUN layer so they never land in the image or
the BuildKit cache. Binary cache mount (VCPKG_BINARY_SOURCES) makes
subsequent rebuilds restore pre-built packages instead of recompiling.
Also adds:
- docker-compose.yml for one-command local deploy
- .dockerignore (excludes clients/, build/, .git/)
- .github/workflows/build-linux.yml — CI cross-build for amd64 + arm64
with downloadable artifacts (primary path for building from Windows)
- scripts/build-linux-binaries.sh — local Docker binary extraction fallback
- deploy/linux/voicecat.service — hardened systemd unit for bare-metal
- cmake/voicecat-toolchain.cmake now auto-wires VCPKG_OVERLAY_TRIPLETS
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
A bad UDP packet on a flaky link could permanently wedge the voice path,
unrecoverable even across app restarts. Three defects:
1. Anti-replay window was advanced from the UNAUTHENTICATED header seq
before the AEAD tag was checked, and not rolled back on failure. One
corrupted/forged frame shoved recv_highest_ far ahead, after which every
legitimate frame was rejected as "too old" forever. Reorder to
replay-check -> authenticate -> update (RFC 3711 3.3); the window now
moves only after a successful tag check.
2. The wire seq was only the low 16 bits of the nonce counter (zero-extended
on receive). After 65,536 frames the nonce desynced and all frames failed
auth. Widen the voice frame seq u16 -> u64 (header 14 -> 20 bytes). The
core owns all UDP framing, so Swift/C# clients need only a rebuild. This
is a versioned wire change: VOICECAT_PROTOCOL_VERSION 1 -> 2, handshake
rejects on mismatch.
3. Server leaked per-session UDP state on disconnect; unregister_session now
frees udp_endpoints_/udp_tokens_/ssrc_to_session_.
Also add rate-limited dropped-frame logging to MediaRelay so a wedged media
path is observable. New regression tests in test_media_aead.cpp cover the
poison (fails on old code) and the 16-bit wrap. ctest --preset dev
-E external_pcm: 22/22 pass (external_pcm aborts on a pre-existing CoreAudio
shutdown race, unrelated).
Let users choose what the SCREEN_AUDIO stream captures before sharing:
share everything, only selected apps, or all except selected apps, plus a
first-class "Exclude screen reader (VoiceOver) audio" toggle.
ScreenCaptureKit filters audio per application, so ScreenAudioCapture now
takes a ScreenAudioSelection and builds the matching SCContentFilter
(including:/excludingApplications:). New ScreenSharePickerSheet lists
running apps from SCShareableContent. iOS left untouched -- ReplayKit only
delivers the mixed system stream, so per-app filtering is impossible there.
Channel create/edit UIs only surfaced a subset of the core's vc_audio_config,
and DRED was exposed nowhere. While adding it, found a latent ABI mismatch:
both Swift AudioConfig and the C# VcAudioConfigNative blittable struct were one
int short of the native vc_audio_config (missing the trailing `dred`), so native
read past the managed struct in vc_create_channel/vc_edit_channel.
- core marshaling: thread `dred` through Swift (Models/Marshaling/toNative) and
C# (Structs/Models/Marshaling/VoiceCatClient) -- fixes the ABI gap + enables it
- windows: add the one missing DRED checkbox to ChannelEditDialog
- macos: ChannelEditSheet now exposes application, sample rate, packet loss,
complexity, and DRED (was stereo/bitrate/frame/FEC/DTX only)
- ios: rebuild ChannelEditView into a full create+edit form (all params); add
SessionState.editChannel + an admin Edit swipe action (iOS had no edit UI)
- guest nickname: add a dedicated `nickname` to SavedServer on macOS+iOS
(backward-compatible Codable), shown in Guest mode, wired into the guest auth
path -- guests could not set a display name on either before (only Windows)
Verified: macOS + iOS (sim, arm64) xcodebuild BUILD SUCCEEDED; core ctest 22/23
(only external_pcm aborts on a pre-existing shutdown mutex race; no C++ changed).
- Channels tab is now a drill-down on iPhone: ChannelBrowserView lists top-level
channels; ChannelDetailView shows the people in a channel, its sub-channels, and
an explicit Join button (with password prompt). iPad split view unchanged.
- Extract self-contained UserRow (context menu + sheets) from UserListView so admin
actions are reused in the drill-down.
- Fix off-screen chat compose box: pin VoiceControlsView via per-tab
.safeAreaInset(edge: .bottom) instead of a floating overlay, so it reserves layout
space above the tab bar (keeping the compose box visible, cooperating with keyboard
avoidance) without covering the tab bar buttons.
- Collapse Activity into Chat like macOS/Windows: ChatView renders a merged,
time-sorted timeline of messages + activity (activity rows in gray); remove the
Activity tab and ActivityLogView.
- Label the RPSystemBroadcastPickerView inner UIButton for VoiceOver
("Share/Stop sharing screen audio") instead of relying on an outer SwiftUI label.
Implement system/desktop audio sharing on the Apple clients, feeding the
existing SCREEN_AUDIO Opus -> AEAD -> UDP path via vc_stream_feed_pcm. No
C++/protocol/codec changes -- the core was already ready (the Windows-only
loopback is #ifdef VOICECAT_HAS_LOOPBACK; off Windows the stream just waits
for fed PCM). Audio only; video is dropped.
macOS (in-process):
- ScreenAudioCapture.swift drives an audio-only SCStream
(excludesCurrentProcessAudio), converts Float32 -> int16 in the channel's
mono/stereo mode, and calls feedPcm. Capture starts on the self
.streamStarted event (effective config known then). Wired into
MainWindowController.screenAudioClicked().
iOS (forward-to-host, single session):
- VoiceCatBroadcast: a ReplayKit Broadcast Upload Extension consumes
.audioApp only, resamples to 48kHz int16 stereo (AVAudioConverter), and
writes a shared App Group SPSC ring (BroadcastAudioRing.swift). It does
not link libvoicecat.
- Host BroadcastAudioPump drains the ring (reacting to the extension's
Darwin notifications) and feeds the SCREEN_AUDIO stream it owns, downmixing
to mono when the channel is mono. Screen audio appears as a second stream
of the same user; no credentials persisted. UI is RPSystemBroadcastPicker
View in VoiceControlsView. Removes the speculative BroadcastCredentials.
Docs: voice.md s9, CLAUDE.md status, PROGRESS.md.
Mirrors the Windows client's UI overhaul (commit 97fa659 + 540ec13) adapted to
Mac-native conventions. The main window is now just toolbar + channels + users
+ chat; audio device settings moved to a modeless Settings window.
- NSToolbar: Join Voice, Share Screen Audio, Mute, Deafen (SF Symbol toggle
buttons) + Output Volume slider (NSSlider 0-100, default 80). Voice actions,
mute/deafen, and output volume moved out of the bottom panel into the toolbar
- Audio device settings (input mode, VAD sensitivity, PTT key, device picker,
level meter) moved to a new SettingsWindowController -- a modeless window
opened via the app menu's "Settings..." (Cmd+,) item. Source-of-truth for
audio state lives in MainWindowController so voice start applies settings even
before the window has been opened; SettingsWindowController reads from /
writes back to those properties and applies changes live when voice is active.
Level meter forwarded from handleLevel -> updateLevel(rms:)
- Unified log: chat NSTextView + activity NSTableView collapsed into a single
NSTextView -- activity events in secondaryLabelColor (gray), chat in default
- Private messaging: scope dropdown removed; compose always sends to the
current channel. Each PM conversation opens in its own modeless
PrivateMessageWindowController. Incoming .textMessage with .private scope
routed to the right window; outgoing PMs echoed by server arrive through the
same path. "Send Private Message..." added to user context menu.
- Messages menu: "New Private Message..." (Cmd+Shift+N) opens a UserPickerSheet
listing all server users so you can PM anyone on the server
- Channel tree now shows live user counts, e.g. "General (3)"; refreshChannelTree
called on .userJoined/.userLeft (was missing)
- Voice menu: Join Voice (Cmd+Shift+V), Share Screen Audio (Cmd+Shift+S),
Mute (Cmd+Shift+M), Deafen (Cmd+Shift+D) -- NSMenuItem key equivalents with
[.command, .shift] mask, dispatched by the responder chain
- setOutputVolume(_:) wrapper added to VoiceCatClient.swift (was missing -- the
C ABI + C# wrapper shipped in commit 97fa659 but the Swift wrapper was never
added); wired end-to-end: toolbar slider -> client.setOutputVolume(gain)
Part A -- fixed and verified the previously-uncompiled Swift from the external
PCM feed/tap commit (615d2a8):
- Rebuilt the macOS xcframework slice (regenerated the module map from current
voicecat.h, exposing vc_pcm_sink_cb / vc_stream_feed_pcm / vc_set_pcm_sink)
- Fixed feedPcm type bug: size_t imports as Int in Swift not UInt; the original
UInt(samplesPerChannel) was wrong
- Added VoiceCatPcmSinkCallback typealias -- a Swift-idiomatic public alias for
the C vc_pcm_sink_cb so consumers (tests, the macOS app) can declare a sink
callback without directly importing the VoiceCatC C module. Mirrors the C#
VcPcmSinkCallback delegate
- keyCodeName helper deduplicated (was in PttKeyCaptureSheet.swift +
MainWindowController.swift -- now shared)
Platform-specific adaptations (vs. Windows): NSToolbar instead of ToolStrip;
global menu bar + NSMenuItem key equivalents (Cmd not Ctrl, responder-chain
dispatched, no custom key monitor needed); PM windows as modeless NSWindows;
picker as Mac sheet; gray = secondaryLabelColor; SF Symbols for toolbar icons.
swift test 10/10 (4 ExternalPcmTests + 6 VoiceCatClientSmokeTests against a
live server); xcodebuild Debug + Release BUILD SUCCEEDED with 0 Swift warnings.
Adds Opus 1.6 DRED support end-to-end: encoder embeds 20 ms of ML
redundancy in every packet when enabled; decoder recovers lost frames
from the next buffered packet's DRED extension rather than falling back
to PLC comfort noise.
Protocol: bool dred = 11 added to AudioConfig (backward-compatible,
defaults false). C ABI: int dred added to vc_audio_config. Encoder:
OPUS_SET_DRED_DURATION(2) when dred=true. Decoder: OpusDREDDecoder +
per-stream OpusDRED scratch pre-allocated off the RT thread;
JitterBuffer::try_copy_front_payload peeks at the next packet without
popping on every PLC step; opus_decoder_dred_decode reconstructs the
lost frame if DRED data is present, otherwise falls back to PLC.
New test: test_dred_toggle (22/22 ctest green).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
1398 → ~340 lines. All "Done" journal entries replaced by a compact
"Recent completed work" summary; M0–M4 sections condensed to scan-able
paragraphs. Pending items (iOS on-device verification, external PCM
feed/tap API plan) kept at full length. M5 task list updated to mark
macOS/iOS UI done and list remaining open items.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The real root cause of "selecting Stereo Mic kills headphone/A2DP output on Join
Voice." Every prior fix worked on the Swift IOSAudioRouter under the false premise
that "miniaudio does NOT touch AVAudioSession on iOS." It does: the core opened
devices via ma_device_init(nullptr, ...), and with a NULL context miniaudio 0.11.25
runs an iOS "hack" that sets the session category by device type, then
ma_context_init__coreaudio calls setCategory()+setActive() on every device open --
capture -> AVAudioSessionCategoryRecord with zero options. That wipes the
.playAndRecord category, the mode, and .allowBluetoothA2DP / .mixWithOthers /
.allowAirPlay that IOSAudioRouter had just configured, killing headphone/A2DP (and
even wired) output. Stereo presets break worst because they rely on the A2DP output
route the wipe removes. TeamTalk avoids this by opening RemoteIO/VPIO AudioUnits
directly and leaving the session entirely to the app.
Fix (core, cross-platform safe): AudioEngine now owns a ma_context built by
make_context_config() with coreaudio.sessionCategory = ma_ios_session_category_none
and noAudioSessionActivate/Deactivate = MA_TRUE, and routes all ma_device_init calls
(playback, capture, loopback) plus enumerate_devices through it. miniaudio no longer
touches AVAudioSession; IOSAudioRouter is the sole owner (the session is already
activated on connect in AppState before any device opens). Context is lazily inited
in start(), reused across restarts, uninited in ~AudioEngine.
Adds TEMP AudioSessionManager.logSessionState() diagnostics (after activate, on route
change, on .streamStarted) to verify on-device that the category stays
PlayAndRecord+allowBluetoothA2DP instead of flipping to Record. Remove once confirmed.
Windows: cmake --build --preset dev clean; ctest --preset dev 21/21.
iOS build + on-device verification pending on Mac.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Diagnosed by comparing against TeamTalk5 (Client/iTeamTalk), which
achieves stereo mic + A2DP output. Five fixes:
1. configureStereoCapture now calls setPreferredInput +
setInputDataSource (mirroring TeamTalk5's SoundDevicesModel).
Previously omitted based on incorrect diagnosis that
setPreferredInput collapsed A2DP — the real culprit was
setPreferredInputNumberOfChannels(2), which neither project uses.
2. New C ABI: vc_audio_restart (full stop + re-init, unlike
suspend/resume which only stop/start). Swift wrapper added.
The withAudioSuspend wrapper that used it was removed after
on-device testing showed it killed all audio (including
VoiceOver) when switching presets — the core's
set_capture_channels handles engine restart internally.
3. Bluetooth options: Voice Chat preset now includes BOTH
.allowBluetoothHFP AND .allowBluetoothA2DP (matching TeamTalk5's
UtilSound.swift:228). Previously HFP-only blocked A2DP headphones.
4. Capture channels now reset when switching stereo→mono via
selectCaptureChannels/applyPreset. AudioSessionManager tracks
activeMicStreamId (set by SessionState on join/leave voice).
5. Docs synced: voice.md, tech-stack.md, architecture.md,
PROGRESS.md. Removed stale setPreferredInputNumberOfChannels(2)
references.
Verified: ctest --preset dev 21/21 green, iOS client builds.
Stereo mic + A2DP output still needs on-device debugging — the
core recipe is correct but iOS 26 route behavior requires
hands-on testing with a debugger.
Three iOS client problems fixed plus a new core stereo-mic capture ABI:
1. Channel-id sync bug (mic button permanently dimmed): SessionState never
synced currentChannelId from the self user's channelId on connect, so the
mic button (gated on currentChannelId == 0) stayed dimmed. Added
syncSelfChannel() (mirrors macOS MainWindowController.swift:461,491,522);
called from init/.channelList/.userJoined/.userLeft/.userUpdated/.joinResult.
Added applyServerMuteState() + serverMuted/serverDeafened to VoiceState.
2. Join/Leave Voice button: replaced icon-only mic toggle with explicit
text button (parity with macOS). Mute/deafen disable when not in voice.
3. IOSAudioRouter.swift (new): full AVAudioSession routing layer — input
port selection, built-in mic orientation/polar patterns, Bluetooth
HFP/A2DP/Off modes, Standard/Raw mic processing, stereo capture, AirPlay,
UserDefaults persistence. AudioSessionManager delegates to it.
4. Core stereo-mic capture (append-only ABI): vc_set_capture_channels()
lets the core open the mic device in stereo (2-ch interleaved). LocalStream
gains capture_channels; ensure_audio_running reads it; audio_engine.cpp
capture_accum_ + on_capture updated to channel-aware accumulation. Test
test_stereo_mic_capture (headless, L!=R stereo round-trip). Swift wrapper
VoiceCatClient.setCaptureChannels.
5. Settings UI rework: AVAudioSession-derived input/output tree replaces
miniaudio device picker.
6. iOS deployment target raised to 18.0 (Package.swift + project.pbxproj).
swift-tools-version 6.0 with swiftLanguageModes .v5.
Docs: tech-stack.md, architecture.md, voice.md, roadmap.md, building.md
updated; stale 'vc_audio_suspend/resume deferred' claims corrected.
Verified: ctest --preset dev 21/21 green; swift test 6/6 green;
xcodebuild -target VoiceCatiOS -sdk iphonesimulator BUILD SUCCEEDED.
Full SwiftUI app at clients/apple/iOS/VoiceCatiOS.xcodeproj:
- 24 Swift source files: AppState + SessionState (@Observable @MainActor),
AudioSessionManager (AVAudioSession owner + interruption/route handling),
ServerListStore/SavedServer (App Group container + Keychain sharing),
and 14 SwiftUI views covering the full feature set
- NavigationSplitView on iPad, TabView on iPhone (horizontalSizeClass)
- Channel tree via OutlineGroup, user list with context menu admin actions
- PTT via DragGesture(minimumDistance: 0) + @GestureState
- onEvent closures hop to MainActor via Task { @MainActor in ... }
- App Group: group.cat.voice.VoiceCat (shared with future ReplayKit extension)
C ABI: add vc_audio_suspend / vc_audio_resume (AudioEngine::suspend/resume)
called by AudioSessionManager on AVAudioSession interruption events.
XCFramework: add ios-arm64 and ios-arm64-simulator slices to build-xcframework.sh;
Package.swift gains .iOS(.v17) platform; CMakePresets.json adds apple-ios /
apple-ios-sim presets with arm64-ios / arm64-ios-simulator vcpkg triplets.
Verified: xcodebuild -target VoiceCatiOS -sdk iphonesimulator26.5 BUILD SUCCEEDED.
The previous "shipped" claim was false — xcodebuild had never been run and
the macOS app source was never committed. This commit adds the 17 Swift
source files + xcodeproj and fixes three real defect classes so Debug and
Release both build clean:
1. MainWindowController.swift compile errors:
- NSAccessibility.post arg order (element:notification:userInfo:)
- NSAccessibilityPriorityMedium -> NSAccessibilityPriorityLevel.medium
- StreamSummary.streamId -> .id (Identifiable conformance)
- drop redundant VoiceCatResult.description extension
2. Linker: add -lc++ to OTHER_LDFLAGS (libvoicecat-fat.a is C++20; pure-Swift
app target has no .cpp sources so libc++ wasn't pulled in — swift test
passed because Package.swift testTarget has linkerSettings: c++).
3. Release config: add ONLY_ACTIVE_ARCH=YES (XCFramework only has arm64).
Verified: clean Debug + Release builds, otool -L shows libc++.1.dylib,
nm shows _vc_client_create/_vc_version_string, app launches and runs.
Lays the groundwork for the macOS (AppKit) and iOS (SwiftUI) clients with a shared
Swift core wrapping the C ABI, mirroring the proven Windows VoiceCat.Interop layer.
Architecture decision: macOS UI = AppKit (not SwiftUI) for the most mature VoiceOver
accessibility story — same rationale as the Windows client's WinForms-over-WinUI-3
decision. iOS stays SwiftUI. Recorded in docs/roadmap.md §2.
Build infrastructure (Phase 0):
- clients/apple/scripts/build-xcframework.sh: runs cmake --preset apple-dev, merges
libvoicecat.a + 107 vcpkg static deps into a single ~30 MB fat static library
(libvoicecat-fat.a) via libtool -static (SPM binary targets link one .a per slice),
stages voicecat.h + a generated module.modulemap (module VoiceCatC) into the headers,
runs xcodebuild -create-xcframework -> clients/apple/VoiceCatCore.xcframework.
VoiceCatCore Swift Package (Phase 1):
- Package.swift: binary target (VoiceCatCoreXCF) + library (VoiceCatCore) + test target.
- Sources/VoiceCatCore/: 7 files mirroring the C# VoiceCat.Interop patterns adapted to
Swift native C interop — Enums (9 Swift mirrors of C enums, UInt32-backed), Config,
Event (copies ev.text to String inside the callback — the #1 lifetime rule), Models
(10 Swift value types), Marshaling (C arrays -> Swift + immediate vc_free_*),
Callbacks (@convention(c) + Unmanaged.passUnretained, the Swift analog of C#'s
[UnmanagedCallersOnly] + GCHandle), VoiceCatClient (owns vc_client* as OpaquePointer,
all 38 C ABI functions, deinit -> vc_client_destroy then frees config CStrings, event
delivery on main queue via coalesced DispatchQueue.main drain).
Tests — 6/6 green (swift test against a real voicecat-server):
- testConnectTofuAuthListChannelsRoundTrips, testAdminChannelCrudAccountCrudRoundTrips,
testScreenAudioStreamStartsAndStops, testPerStreamRecvControlsRoundTrip, plus two
static smoke tests. Catches Swift-specific interop bugs (@convention(c) callback
lifetime, Unmanaged pointer resolution, CString memory management, enum raw-value
bridging, struct layout) that C++ ctest cannot. C++ suite still 21/21 green.
Docs updated (house rule): tech-stack.md §2, architecture.md §4, roadmap.md M4 + §2,
clients/apple/README.md (full rewrite), PROGRESS.md, .gitignore.
macOS port groundwork — core, server, tools, and tests now build and run on
macOS 26.5 / Apple Silicon. ctest --preset dev green 21/21 (2 consecutive runs).
apple-dev produces valid arm64 libvoicecat.a + XCFramework for the Swift Package.
Three real cross-platform bugs found and fixed (all latent on Windows/Linux):
1. test_m2_voice.cpp POSIX branch missing <netdb.h> — Linux glibc transitively
includes it, macOS doesn't. Would fail on any strict POSIX system.
2. SIGPIPE killing processes on macOS — writing to a closed TCP socket raises
SIGPIPE by default (doesn't exist on Windows, benign on Linux). Fixed by
ignoring SIGPIPE in both core client init and server startup (POSIX-only,
#ifndef _WIN32). Production fix, not just tests.
3. Use-after-free of Asio's kqueue reactor on server shutdown — the
deterministic test_tofu_flow segfault. TcpServerConn's tls_read_loop runs on
a blocking-I/O thread; when Server::run() returned, io_context was destroyed
while those threads were still running. On macOS kqueue the reactor pointer
is null'd immediately -> segfault in socket.close(). Latent on Windows IOCP
and Linux epoll. Fix: TcpAcceptor now tracks connections; new shutdown()
closes all + joins threads before io is destroyed; Server::stop() now closes
acceptor + media_relay too (was just io.stop()).
Verified: dev + apple-dev presets build green, 21/21 tests pass, server starts
+ two vccli text chat over TLS (M1 on Mac), vccli --voice starts MIC stream via
CoreAudio (M2 protocol-level), vccli --list-devices enumerates CoreAudio
devices, xcodebuild -create-xcframework produces valid VoiceCatCore.xcframework.
No ABI or proto changes. Docs updated: building.md, clients/apple/README.md,
PROGRESS.md, CLAUDE.md status line.
Rationalize the preset set to match the project's actual state (past M5):
- Rename dev->skeleton (no-deps stub smoke), m1-dev->dev (default dev preset)
- Drop m2-dev (cache-identical to m1-dev)
- Add release preset (optimized + tests on, symbols kept)
- Strip server-release binaries (-s linker flag)
- Add apple-dev/apple-ios/apple-ios-sim scaffolding presets for XCFramework
Add cmake/voicecat-toolchain.cmake wrapper that auto-resolves the vcpkg
triplet from the host platform (x64-mingw-static/x64-linux/arm64-osx) so
the main presets work on Windows/Linux/macOS without per-OS variants.
Update all docs (building.md, CLAUDE.md, README.md, AGENTS.md, deployment.md,
tech-stack.md, client READMEs) and stale preset-name references in code
comments. No C++ behavior changes — the core was already portable.
Three reported bugs traced to one root cause plus two missing designed features:
1. Stale users + eternal PLC hiss (root cause): ConnSession::close() silently
erased dropped users without broadcasting UserEvent::LEFT, so peers never
learned the user left and their audio engines never called remove_stream —
Opus PLC synthesized comfort noise forever. Fix: broadcast_left() helper
+ close() broadcasts LEFT before erasing.
2. PLC cap (defense-in-depth): on_playback now caps pure PLC at ~2s, then
emits digital silence so a stale stream can never hiss forever even if
remove_stream is skipped. Resets automatically on fresh packets.
3. No timeout / no ping: client never sent Ping, server had no last_seen /
reaper, so half-open connections (NAT timeout, wifi loss, sleep) left
ghost users forever. Fix: client Ping every 15s with RTT measurement,
ConnSession::last_seen bumped on every inbound TCP/UDP frame, steady_timer
reaper sweeps every 15s and drops sessions older than 45s (configurable
via server::Config).
4. UDP KEEPALIVE: client sends plaintext kFrameKeepalive every 5s; server
bumps last_seen + echoes back. Keeps NAT bindings alive and lets media
activity defer the reaper independently of TCP.
5. Graceful client disconnect: vc_disconnect() sends Disconnect{code=0} via
a flag-based io-thread exit (no double-close race); server handles
client-sent Disconnect with immediate close() + LEFT broadcast.
3 new tests: disconnect_left, plc_cap, reaper_timeout. 21/21 ctest green.
Docs: protocol.md §6/§7, voice.md §6, architecture.md §5, PROGRESS.md.
start_loopback_capture hardcoded channels=1, forcing miniaudio to downmix the system's stereo mix to mono before the encoder saw it -- on_capture_frame then upmixed L=R to produce fake stereo. Now the loopback device opens in the channel's mode (stereo when the channel is stereo), CaptureCallback carries an explicit channels param so the encoder gets real interleaved L/R, and a mono fallback covers unusual render endpoints. New test_loopback_stereo_capture asserts L!=R end-to-end; 18/18 ctest green.
The core already supported SCREEN_AUDIO capture on Windows (post-M3 WASAPI
loopback via VOICECAT_HAS_LOOPBACK) and the C# Interop layer was complete
(VcStreamKind.ScreenAudio, StartStream/StopStream/SetRemoteStream). Only the
UI was missing -- no core, proto, or C ABI changes needed.
Adds a 'Share Screen Audio' toggle to the voice panel, independent of mic
voice (can share without joining voice). Disconnect/teardown now stops the
screen stream cleanly. New smoke test exercises the full StartStream ->
StreamStarted -> StopStream -> StreamStopped path through P/Invoke.
The media AEAD nonce is an implicit per-direction monotonic counter;
open() reconstructs it from the 14-byte header seq field (the AAD), so
the contract is header.seq == the counter seal() used. The SFU relay
decrypted inbound frames with the sender key, re-sealed with the
recipient send_crypto (its own counter), but forwarded the sender
header verbatim -- so seq carried the wrong counter and the recipient
rebuilt the wrong nonce, silently dropping every relayed frame. It only
worked for a single first-ever sender into a fresh recipient, which is
why reverse/3rd-party audio failed.
Rewrite the outgoing header seq to the recipient peek_send_counter()
before re-sealing so each server->client direction is one contiguous
monotonic counter and the nonce always matches. Safe: the jitter buffer
orders by timestamp, not seq. No wire-format/proto/ABI change.
Adds test_relay_interleaved_reseal regression coverage.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A connected Windows client would randomly snap from its joined channel
back to Lobby. Root cause was a state-sync inconsistency, not a drop:
the server delivered self-initiated state changes (channel join/leave,
stream announce/stop) only as a private *Result to the actor and
broadcast the authoritative UserEvent::UPDATED to everyone else. The
core never applied the result to its SessionModel, so vc_list_users()
kept self in the old channel; the Windows HandleUserUpdated rebuilds
_currentChannelId from vc_list_users() on any user's UPDATED event, so
the next unrelated event surfaced the stale self-channel.
Fix, per the response-vs-broadcast contract now documented in
docs/protocol.md §6: the *Result is pure ack/correlation/actor-private
payload; the resulting state change is broadcast to every client
INCLUDING the actor, and clients apply it to their local model rather
than re-deriving own state from a *Result.
- server: join/leave/stream announce+stop broadcast with exclude=0
- server: text fan-out includes the sender (channel + private echo)
- core: response handlers no longer mutate session_model_
- windows: drop optimistic text echo; render own message via the relay
- docs/protocol.md §6: document the response-vs-broadcast contract
Registry-level admin broadcasts (move/mute/kick/channel CRUD) already
used exclude=0 and were correct. ctest build/m1-dev 18/18 green;
VoiceCat.App builds 0 warnings.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
RemoteStream::playout_ts was seeded to 0 and only advanced inside the
decode loop (including on every PLC iteration), so it free-ran at ~1x
wall-clock regardless of whether the sender was transmitting. The
sender's frame timestamps only advance while it actually sends (the
VAD/PTT gate returns before ls.timestamp += samples). Across a late join
or any VAD/PTT silence gap the two clocks diverged without bound; once
past the jitter buffer's 500 ms late-drop window every real frame was
dropped-as-late (clock ahead) or never-due (clock behind) -> permanent
silence, while the talk indicator (driven by push_recv_frame, independent
of the jitter buffer) stayed lit.
Add JitterBuffer::peek_front_ts() (try-lock, RT-safe) and seed/re-sync
playout_ts to the earliest buffered frame on the first frame and whenever
it has drifted past +/-200/500 ms. This seeds startup and recovers after
every silence gap.
New regression test test_playout_resync free-runs the clock ~2 s past the
drop window, pushes a ts=0 frame, and asserts audible output: fails
(energy=0) without the fix, passes with it. ctest --preset m1-dev: 14/14.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
on_playback() was passing miniaudio's hardware playback-callback frame
count to opus_decode()'s max_samples, instead of the decoder's fixed
frame size (960 samples @ 20ms/48kHz). Since real packets decode to
more samples than the (often smaller, e.g. ~480 on default low-latency
WASAPI) hardware period, opus_decode returned OPUS_BUFFER_TOO_SMALL on
nearly every callback -- packets were received/decrypted/jitter-buffered
correctly but never decoded into audible PCM. Result: control-plane
events and VAD worked, but zero audio in headphones.
mix_for_test()'s white-box test masked this since it always called
on_playback with frames == frame_samples, the one case where the bug
is invisible.
Fix: RemoteStream gained a small ring buffer (init_ring/push_ring/
pop_ring) that decouples decode cadence from playback-callback cadence.
on_playback now tops the ring up by decoding whole Opus frames (always
decoder.frame_samples(), never the hardware frame count) and drains
exactly what the callback asks for, silence-padding (PLC) on underrun.
Side effect: also fixes playout_ts, which was advancing by the wrong
unit (hardware frames instead of decoded samples) -- it now tracks
correctly against jitter-buffer timestamps.
ctest --test-dir build/m1-dev: 12/12 green.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Closes the three items PROGRESS.md's M3 section explicitly carried forward as
out of scope:
- Device enumeration (vc_list_devices) + input device selection
(vc_set_input_device), backed by AudioEngine::enumerate_devices() via
miniaudio's ma_context_get_devices. Device ids are opaque hex-encoded
ma_device_id strings.
- VAD/PTT send-side input gate (vc_set_input_mode, vc_set_push_to_talk).
webrtc-audio-processing (the originally-planned APM) has no working
Windows/MSVC build upstream (GCC-only Meson, unfinished MinGW support, hard
abseil-cpp dependency), so VAD is a new lightweight, dependency-free
energy/RMS processor (EnergyVadProcessor) behind the existing ApmProcessor
interface. Gating is MIC-only; SCREEN_AUDIO/AUX_DEVICE always bypass it.
- True stereo playback: AudioEngine's mixer and output device now carry
stereo end-to-end (mono streams upmix L=R) instead of downmixing decoded
stereo streams to mono before mixing.
- Real WASAPI loopback capture for SCREEN_AUDIO (Windows-only, via
miniaudio's loopback device type), replacing test-only injection as the
production capture path.
Also: vccli gains --list-devices, --input-device, --input-mode, and
--share-screen-audio flags, plus a stdin command loop (ptt on/off, mode
vad/ptt) for manual verification. New test_vad_ptt_devices.cpp covers all
four items (ABI-level + a white-box AudioEngine stereo-mix check).
Docs updated to match: voice.md, roadmap.md (decision-log entry superseding
the original webrtc-audio-processing choice), tech-stack.md, README.md,
architecture.md, CLAUDE.md, PROGRESS.md.
Still explicitly out of scope, documented not silently dropped: real
webrtc-audio-processing/AEC (no AEC/NS/AGC exists at all yet), macOS/iOS
SCREEN_AUDIO capture, process-specific loopback, and a pre-existing
RT-thread rule violation in the capture path that predates this work.
Verified: ctest 12/12 green across 3 consecutive full-suite runs (both dev
and m1-dev presets build clean); test_vad_ptt_devices passed 5 consecutive
standalone runs; manually verified live (vccli --list-devices against real
hardware, vccli --voice --input-mode vad streaming without incident).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements docs/roadmap.md M3: multiple concurrent streams per user (MIC +
SCREEN_AUDIO + AUX_DEVICE), independent per-stream receiver gain/mute/noise-
reduction, talk indicators, and enforced per-channel Opus configurability
(mono/stereo, bitrate, frame size, FEC/DTX, application).
Bugs fixed along the way (found while implementing, not pre-existing scope):
- Server hard-coded stream_id=1 for every announce, so a second stream from
the same user silently overwrote the first in SessionRegistry::set_user_stream.
Now a per-session counter (ConnSession::next_stream_id_); handle_stream_stop
validates against announced_stream_ids_ before clearing.
- Client dropped mode/dtx/complexity/application from effective_audio even for
the single M2 stream -- only sample_rate/bitrate_bps/frame_ms/fec were ever
applied to OpusParams. Fixed on both the send (handle_stream_announce_result)
and receive (sync_remote_streams) paths via a shared
opus_params_from_audio_config() helper.
- OpusEncoder always used OPUS_APPLICATION_VOIP; added OpusParams::application
and wired it through.
- on_playback's per-stream decode passed the wrong frame_size to opus_decode
(total samples instead of samples-per-channel), which would have overflowed
the decode buffer for any stereo stream.
- teardown_voice() raced when called concurrently from run_io()'s own cleanup
and from disconnect() on a different thread -- both could see
udp_thread_/talk_timer_thread_ as joinable() at once and race to join() the
same std::thread (intermittent std::system_error under ctest). Fixed with a
teardown_mu_ guard instead of carrying the flake forward.
New:
- Per-channel AudioConfig: SessionRegistry now seeds Lobby (mono/24kbps/VOIP/
FEC+DTX) and a new "Music Room" channel (stereo/128kbps/AUDIO/no DTX);
handle_stream_announce enforces the channel's config, clamping (not
overriding) bitrate_bps to its ceiling.
- core/src/core/client.h/.cpp: local-stream state is now a
std::unordered_map<int, LocalStream> keyed by vc_stream_kind, with
request_id-correlated announce/result handling (request_id already
round-tripped on the wire; just wasn't read before). on_capture_frame is
kind-aware and upmixes mono capture to stereo when a stream's config calls
for it. set_self_mute's mic_muted now only gates the MIC kind. NS is wired
through set_remote_stream. New run_talk_timer() thread emits
VC_EVENT_TALK_STATE from both remote and local edge detection.
- core/src/audio/audio_engine.h/.cpp: kind-keyed injection taps
(inject_capture), stereo-to-mono downmix at the decode/mix boundary,
RemoteStream gains recv_ns (lazy ApmProcessor) + noise_reduction_enabled
and last_voice_ms/talking; new set_stream_noise_reduction() and
poll_talk_transitions().
- core/src/session/session.h/.cpp: Stream now carries the full AudioConfig,
not just sample_rate/frame_ms.
- New additive C ABI (core/include/voicecat.h): vc_audio_config +
vc_get_stream_audio_config (effective Opus config for any stream you own or
a peer's); vc_test_inject_capture (test-only synthetic PCM injection,
clearly marked, mirrors AudioEngine::inject_capture).
- tests/test_m3_multistream.cpp: the M3 exit criterion through the real ABI
(mirrors test_voice_client_abi.cpp's approach, not raw sockets) -- two
concurrent local streams, independent gain/mute/NS control, per-channel
config divergence via vc_get_stream_audio_config, talk indicators.
Explicitly out of scope for this pass (tracked in PROGRESS.md, not silently
dropped): VAD/PTT input gate + device enumeration; real WASAPI loopback
capture for SCREEN_AUDIO (synthetic injection only); true stereo playback
output (AudioEngine's mixer/output device stays mono -- Opus itself is fully
stereo-correct on the wire).
ctest --test-dir build/m1-dev: 11/11 green, verified across 3 consecutive
full-suite runs plus 8 standalone runs of the new test.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
test_m2_voice passed against raw BSD sockets, but vc_client::stream_start/stop,
UDP binding, and capture/recv were still VC_ERR_NOT_IMPLEMENTED stubs -- meaning
vccli and any GUI client still couldn't actually talk. Implements the real
client-side UDP-binding handshake, media key derivation, capture->encode->seal->
send and recv->open->decode->playback paths, plus server-side StreamInfo
broadcast so peers learn about each other's streams via sync_remote_streams().
Adds test_voice_client_abi (two real vc_client instances, not raw sockets) and
vccli --voice/--mute/--text flags, manually verified live between two instances.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds the full voice pipeline: 14-byte binary frame header, ChaCha20-Poly1305
AEAD keyed from the TLS exporter, libopus encode/decode with FEC/PLC/DTX,
an adaptive per-ssrc jitter buffer, a miniaudio capture/playback engine, an
APM passthrough stub, and the UdpBinding/StreamAnnounce signaling chain
wired through ConnSession/SessionRegistry into a new server-side SFU
(MediaRelay) that decrypts and re-encrypts frames per channel member.
Exit criterion verified: test_m2_voice — two headless clients relay 50
encrypted Opus frames through the server; ctest --preset m1-dev is 9/9
green. Also corrects protocol.md's UdpBinding diagram, which described the
UDP-side binding packet as AEAD-sealed when it is in fact a plaintext
bootstrap frame (separate from the TCP/TLS UdpBinding ack).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Implements the full M1 milestone. Two clients authenticate over TLS 1.3
(guest + Argon2id password) and exchange channel + private text messages
through a real server. All five ctest --preset m1-dev tests pass in ~1 s.
Key components added:
- vcpkg baseline + m1-dev preset (protobuf/mbedTLS/libsodium/asio/sqlite3)
- FrameCodec feed+emit, encode/decode_envelope, protobuf codegen
- TcpServerConn with blocking TLS handshake thread + tls_read_loop
- TlsContext (mbedTLS 1.3, ECDSA-P256 self-signed cert, TOFU on client)
- WorkerPool (3 threads, used for Argon2id)
- Database: SQLite + libsodium Argon2id, account lifecycle, bootstrap admin
- ServerIdentityManager: Ed25519 key + cert generate/persist/fingerprint
- ConnSession state machine: WaitingHello -> WaitingAuth -> Authenticated
- SessionRegistry: channel tree, user map, text routing, broadcast
- vc_client full M1 C ABI: connect/TLS/handshake/auth/text/disconnect
- voicecat-admin CLI: account add/reset/del/list
- test_m1_integration: M1 exit criterion, verified green
Bug fixed: double-framing in ConnSession::send_envelope -- encode_envelope
was adding the [4-byte len] prefix, then TcpServerConn::send_frame added
a second one, causing the client to parse [len][proto] as protobuf (silent
failure). Fixed by serializing raw protobuf bytes in send_envelope and
letting send_frame apply the single length prefix.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Turn the design into a buildable, dependency-free M0 skeleton plus the
onboarding layer so a new agent can pick up instantly.
Build system:
- CMake + CMakePresets (dev = no deps; server-release = vcpkg) + vcpkg.json
- Skeleton builds with just a C++20 compiler; deps stay off until needed
- .gitattributes (LF), .gitignore, .clang-format
Core (libvoicecat):
- core/include/voicecat.h: full C ABI (the client/server contract), stubbed
- core/proto/voicecat.proto: control-plane wire format, matches docs/protocol.md
- src/{net,crypto,codec,protocol,session,audio,core}: subsystem stubs that
return VC_ERR_NOT_IMPLEMENTED, each pointing to its design doc
- server/ (voicecat-server) and tools/vccli/ link the core
- tests/: CTest smoke test asserting the C ABI contract (behavior, not just build)
- clients/{apple,windows}: M4 placeholders
Onboarding for agents:
- CLAUDE.md: hub — build/test commands, architecture at a glance, doc map, rules
- AGENTS.md: working method (behavior-driven; clean compile is the floor not the goal)
- PROGRESS.md: living tracker — M0 done, M1 task checklist, "where we left off"
Verified: cmake --preset dev && cmake --build --preset dev && ctest --preset dev → green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
