Three coordinated fixes for the bug where enabling stereo mic capture
caused all audio output (A2DP, speaker, wired) to go silent:
1. audio_engine.cpp — open playback before capture
On iOS, starting the stereo capture AudioUnit can trigger an audio
route reconfiguration that drops A2DP before the playback device has
a chance to claim the route. Opening and starting the playback device
first commits the output route (A2DP), so iOS is less likely to drop
it when stereo capture activates afterward.
2. client.cpp — decouple set_capture_channels from engine restart
Previously vc_set_capture_channels() stopped and restarted the engine
immediately, which opened capture first (old ordering) and raced
against the settling AVAudioSession route. Now it only stores the
channel count; the caller (Swift via vc_audio_restart) controls when
the engine restarts, after the route has settled.
3. IOSAudioRouter.swift — call audioRestart() after channel config
selectCaptureChannels() and applyPreset() now call audioRestart()
after applyConfiguration() + setCaptureChannels(). This is the
vc_audio_restart() path that was added to the ABI in fdcc84f but
never wired up in the Swift layer. The restart sees the stored
channel count and reopens devices in the correct order (playback
first, capture second).
The doStartMicStream path is unaffected: setCaptureChannels is called
before the server acknowledges the stream (engine not yet running), so
ensure_audio_running() picks up capture_channels=2 directly when the
stream is confirmed and opens with the right count from the start.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Apple client (macOS + iOS)
Built in M4 (see docs/roadmap.md). One shared Swift core
(VoiceCatCore package) wrapping the C ABI (core/include/voicecat.h),
with platform-specific UIs: AppKit for macOS (best VoiceOver accessibility), SwiftUI
for iOS. See docs/architecture.md §4 and
docs/tech-stack.md §2.
What's here now
VoiceCatCore Swift Package — ✓ complete (2026-06-18)
The shared Swift core that both the macOS AppKit app and the iOS SwiftUI app will consume.
Mirrors the Windows client's VoiceCat.Interop layer (clients/windows/)
using Swift-native C interop instead of P/Invoke.
clients/apple/
├── Package.swift # SPM: binary target (XCFramework) + VoiceCatCore library + tests
├── VoiceCatCore.xcframework/ # BUILT ARTIFACT — produced by scripts/build-xcframework.sh (gitignored)
├── scripts/
│ └── build-xcframework.sh # builds libvoicecat + vcpkg deps → fat .a → XCFramework + module map
├── Sources/VoiceCatCore/
│ ├── Enums.swift # Swift-idiomatic mirrors of the 9 voicecat.h C enums
│ ├── Config.swift # VoiceCatConfig (wraps vc_config)
│ ├── Event.swift # VoiceCatEvent — copies ev.text inside the callback (the #1 lifetime rule)
│ ├── Models.swift # Channel, User, Stream, Device, Permissions, Account, AudioConfig, …
│ ├── Marshaling.swift # C arrays → Swift arrays + immediate vc_free_* (callers never manage native lifetime)
│ ├── Callbacks.swift # @convention(c) on_event/on_level + Unmanaged.passUnretained context bridging
│ └── VoiceCatClient.swift # the public Swift surface — owns vc_client*, all 38 C functions, event delivery on @MainActor
└── Tests/VoiceCatCoreTests/
└── VoiceCatClientSmokeTests.swift # 6 XCTest smoke tests against a real voicecat-server (6/6 green)
Key patterns (carried over from the proven C# VoiceCat.Interop — see
docs/architecture.md §4 per-platform binding notes):
- C interop via module map:
import VoiceCatC— Swift sees all C enums/structs/functions directly. No manual struct/function redeclaration (unlike C# P/Invoke). The module map (module VoiceCatC { header "voicecat.h" }) is staged into the XCFramework headers bybuild-xcframework.sh. @convention(c)callbacks: plain C function pointers (not ARC-managed closures) +Unmanaged.passUnretained(self)as theusercontext — the Swift analog of C#'s[UnmanagedCallersOnly]+GCHandle.deinitcallsvc_client_destroy(joins all threads) before the object's memory is freed, so no callback can fire with a dangling pointer.- Config string lifetimes: the core stores raw pointers from
vc_config(doesn't copy). Native CString storage (strdup) is held for the client's entire lifetime, freed indeinitaftervc_client_destroy. - Event delivery: events buffered in a lock-protected array + coalesced
DispatchQueue.maindrain (one async block at a time) — the Swift analog of C#'sChannel<VoiceCatEvent>+ 30ms WinForms Timer pump.ev.textis copied toStringinside the callback before enqueueing (dangling-pointer rule). - Level meters: coalesced to latest-per-stream-id (intermediate values are visually
irrelevant, same as C#'s
ConcurrentDictionary<uint,float>). - Immediate
vc_free_*on list reads — callers never manage native list lifetime.
Tests — 6/6 green
swift test
# ✓ testVersionStringIsNonEmpty
# ✓ testResultStringRoundTrips
# ✓ testConnectTofuAuthListChannelsRoundTrips (connect → TOFU → confirm → guest auth → channels → permissions → guest ListAccounts rejected)
# ✓ testAdminChannelCrudAccountCrudRoundTrips (admin auth → channel create/edit/delete → account create/list/reset/delete)
# ✓ testScreenAudioStreamStartsAndStops (screen-audio stream start/stop through Swift interop)
# ✓ testPerStreamRecvControlsRoundTrip (two clients, per-stream gain/mute/NR round-trip)
Prerequisites for tests: cmake --preset dev && cmake --build --preset dev (builds
voicecat-server + voicecat-admin into build/dev/bin/).
What's NOT here yet (next steps)
- macOS AppKit app (
clients/apple/macOS/) — the M4 UI: connect dialog, saved-server list (Keychain for passwords), TOFU identity dialog, main window (NSOutlineView channel tree, NSTableView user list, NSTextView chat, activity log), voice controls, per-user tuning, full VoiceOver accessibility. Mirrors the WindowsVoiceCat.Appfeature set. - iOS SwiftUI app — AVAudioSession, mic permission, foreground voice.
vc_audio_suspend/vc_audio_resumeABI hooks — deferred until the iOS client milestone (keep ABI stable).- ReplayKit Broadcast Upload Extension for iOS
SCREEN_AUDIO(docs/voice.md§9). - macOS
SCREEN_AUDIOvia ScreenCaptureKit (currently stub returnsfalse). - iOS XCFramework slices —
apple-ios/apple-ios-simpresets are scaffolding; runscripts/build-xcframework.sh --allonce the iOS vcpkg triplets are validated.
Building the XCFramework
The XCFramework is a local build artifact (gitignored, like the Windows client's
build/windows-client/bin/voicecat.dll). Run the build script before swift build /
swift test:
# Prerequisites: VCPKG_ROOT set, Xcode installed
export VCPKG_ROOT=/path/to/vcpkg
# Build the macOS slice + fat static lib + XCFramework (validated)
scripts/build-xcframework.sh
# → clients/apple/VoiceCatCore.xcframework/ (macOS-arm64 slice)
# Build all 3 slices (macOS + iOS device + iOS sim) — iOS still scaffolding
scripts/build-xcframework.sh --all
Fat static library
The apple-dev CMake preset produces a 1.9 MB libvoicecat.a containing only voicecat's
own object files — vcpkg's static dependencies (protobuf, mbedtls, libsodium, opus, sqlite3,
spdlog, asio, abseil, …) are 107 separate .a files under vcpkg_installed/arm64-osx/lib/.
A Swift Package binary target can only link ONE .a per XCFramework slice, so
build-xcframework.sh merges them all into a single self-contained libvoicecat-fat.a
(~30 MB) using libtool -static. This is the Apple equivalent of how the Windows client
ships a single voicecat.dll with all deps statically linked (via MinGW's -static flags
in core/CMakeLists.txt).
Swift Package
swift build # builds VoiceCatCore library
swift test # runs 6 smoke tests against a real voicecat-server
The Package.swift declares:
- A binary target (
VoiceCatCoreXCF) pointing at the localVoiceCatCore.xcframework. - A library target (
VoiceCatCore) that depends on the binary target and provides the Swift wrapper. - A test target (
VoiceCatCoreTests) withlinkerSettings: [.linkedLibrary("c++")]— the fat static lib is C++20, so the final executable must link libc++ (the LLVM C++ standard library on macOS). vcpkg's static deps are already in the.a; macOS system frameworks (CoreAudio/CoreFoundation) are auto-discovered by the linker.