Skip to content

CHANGELOG.md

Latest commit

 

History

History
295 lines (216 loc) · 18.3 KB

CHANGELOG.md

File metadata and controls

295 lines (216 loc) · 18.3 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

0.8.1 - 2026-05-06

Fixed

  • 5h46m ghost recording / 0-byte M4A when stop arrives during AudioRecorder.start(). Swift actor reentrancy let stop() flip a single stopped flag while start() was suspended on SCShareableContent (~12 s); the resumed start created a live recorder no caller held, dropped every buffer, and never finalized. AudioRecorder is now a four-phase machine (notStarted -> starting -> running -> stopped) with a cancelRequested flag checked after every await, and a take-and-nil cleanup that's safe under reentrancy. No failure path leaves an orphan directory or zero-byte file under ~/Library/Application Support/Blackbox/Recordings/.
  • RecordingPipeline.start() self-cleans on partial-init throws (metadata save, AVAssetWriter init, startWriting): the recording directory is removed before the error propagates, instead of being left behind for pipeline.stop() to ignore.
  • Stop arriving during startup no longer surfaces an error toast. AudioMonitor pattern-matches RecorderError.cancelled and lost-race conditions silently; permission-denied and other real errors still surface.
  • After a user-initiated stop (manual stop, force-stop on auto-recording), the same bundle no longer immediately re-triggers auto-record on the next 3 s poll. The resolved parent bundle ID is suppressed until it disappears from the active-caller set; other apps (e.g. Zoom while Chrome is suppressed) still trigger auto-record. Grace-expiry stops do not suppress, so a re-detected call still records.

Added

  • RecorderError.cancelled for stop-during-start cancellation.
  • StartCheckpoint test seam in AudioRecorder and a StartGate actor for deterministic recorder-race tests.
  • Docs: docs/specification.md D4 amended with the always-live WebRTC caveat; new D13 covers the recorder lifecycle invariant, monitor cooperation, and suppression semantics.

0.8.0 - 2026-04-24

Added

  • Layered mic recovery (D12). AVAudioEngineConfigurationChange (existing) is now supplemented by a CoreAudio kAudioHardwarePropertyDefaultInputDevice listener and a 1 Hz buffer-arrival watchdog (2 s stall threshold). All three sources funnel through requestMicReinstall(source:) → debounced mic-tap reinstall. Catches same-format default-input swaps that the AVAudioEngine notification silently misses.

Changed

  • Reverted system audio capture from CoreAudio Process Tap (CATap) back to display-wide SCStream (ScreenCaptureKit). v0.6.0's SCStream approach had an empirical production track record with zero silent-recording reports; CATap produced three distinct silent-recording bugs in 5 days (Bluetooth HFP 24kHz pin, IO-proc stop when nothing plays, Chrome Meet routed to non-default idle output). Root cause: aggregate-device IO proc fires on a hardware output-device clock that can be idle, pinned, or stalled. SCStream's clock comes from the OS-composited mix, decoupled from any specific device.
  • System-audio track is now 2ch stereo 48 kHz 128 kbps AAC (was 1ch mono 64 kbps). Mic track remains 1ch mono 48 kHz 64 kbps. File size for a 30 min call: system track ~29 MB (was ~14 MB).
  • Gap fill (D8), leading silence, and tail padding apply to the mic track only. System track has no synthesised silence - matches v0.6.0 which shipped without system-track gap fill for weeks.
  • Onboarding now actively requests Screen Recording permission on completion via CGRequestScreenCaptureAccess().
  • Settings > Permissions row relabelled "Screen & System Audio Recording"; deep-links to the Screen Recording pane; uses CGPreflightScreenCaptureAccess() for live TCC state instead of a "recorded once" cache.

Fixed

  • Silent system-audio track on FaceTime (and any app routing through communication audio paths on macOS 26). SCStream delivers non-interleaved stereo Float32 CMSampleBuffers; the PCM round-trip helper inherited from the CATap era mis-copied non-interleaved payloads, producing system tracks with mean_volume near -66 dB. SCStream buffers are now appended directly to a stereo 128 kbps AAC writer input (v0.6.0 parity). Post-fix: FaceTime system track mean_volume -37.8 dB (from -66.4 dB); Chrome -27.1 dB.
  • Full-hour mic-only recordings when default output was idle while Chrome Meet routed call audio to a non-default output via its in-page device picker.
  • AudioRecorder.stop() now tears down the D12 default-input CoreAudio listener and watchdog timer (previously only removed in deinit's defensive path). Closes a listener leak on kAudioObjectSystemObject that accumulated across recording sessions.
  • Onboarding gate now treats CGPreflightScreenCaptureAccess() as the source of truth on every launch. First-run denial and post-install permission revocation both re-surface onboarding; prior behaviour sticky-marked onboarding complete even when the TCC prompt was denied.
  • mappedStartError now routes SCStream -3802/-3821 codes to RecorderError.systemStopped (new case), matching the in-flight handleStreamStopped routing.

Removed

  • CATap tap/aggregate-device/IO-proc pipeline, output-device change listener, IO proc buffer pool, and mic recovery-after-device-change paths.
  • AudioRecorder.pcmBuffer(from:) (interleaved-only PCM helper - root cause of the FaceTime silence on macOS 26), systemFormat cache, systemBuffersConversionFailed counter.

0.7.0 - 2026-04-17

Added

  • CoreAudio Process Tap (CATap) system audio capture, replacing SCStream
  • Hardware smoke test suite with file-based IPC test mode (--ui-test-mode) for automated real-audio validation
  • RecordingPipeline type for AVAssetWriter management, gap filling, tail padding, and audio level metering (extracted from AudioRecorder for testability)
  • AudioMonitorDependencies dependency injection with TestClock / TestRecorderFactory for deterministic call-detection tests
  • Silence gap filling (D8) across both pipelines with clean LPCM format descriptions
  • Device latency offset compensation (D9) for mic-system alignment
  • Drift compensation via CATap kAudioSubTapDriftCompensationKey
  • Output device change handling with aggregate device rebuild and silence gap filling
  • Pre-allocated AVAudioPCMBuffer pool for CATap IO proc callback (RT-safe per spec D5)
  • AudioHardwareSystem Swift wrappers (switf-macos) for typed CoreAudio access

Changed

  • Output is now 2-track M4A: system audio (1ch mono 48kHz) + mic (1ch mono 48kHz). Legacy 3-track recordings remain playable.
  • AudioRecorder converted from class to actor with custom DispatchSerialQueue executor
  • Call detection simplified to polling-only, removing ~140 lines of CoreAudio listener management
  • macOS 26.1+, Swift 6.2, warnings-as-errors

Fixed

  • Chrome/WebRTC silent system-audio bug (eliminated by CATap architecture)
  • Mic stall recovery after output device changes (mic is recreated if stalled post-rebuild)
  • Track alignment: session starts on max(first_sys_pts, first_mic_pts) so both tracks begin with real content

Known Issues

  • Tail padding may report a short residual (up to several seconds) when macOS stalls output device routing mid-call. Tracked in #8.

0.6.0 - 2026-04-04

Added

  • Dual-SCStream recording: display-wide stream (guaranteed completeness) runs alongside per-app stream (cleaner AEC reference) simultaneously
  • Per-app audio track in recordings when single caller detected (3-track M4A: display-wide + per-app + mic)
  • Display and per-app audio buffer stats logging (received, appended, dropped, peak level)
  • Track selector shows App option for 3-track recordings

Changed

  • AEC post-processing uses per-app track as reference when available (cleaner than display-wide)
  • Transcription skips display-wide track in 3-track files to avoid doubling call audio
  • Removed mid-recording restart logic (no more interrupted recordings from caller detection churn)

Fixed

  • Chrome/WebRTC calls producing silent system audio track (per-app SCStream of Chrome captures silence, display-wide now always present as fallback)

0.5.1 - 2026-03-16

Fixed

  • Mic audio recorded at 2x speed when device switches mid-recording (e.g. AirPods connecting during a call) - tap now resamples to 48kHz instead of passing native device rate to AVAssetWriter

0.5.0 - 2026-03-13

Added

  • Per-app audio capture: when a single calling app is detected, captures only that app's audio via SCContentFilter. Falls back to display-wide capture for multiple callers or unresolved apps
  • Crash recovery watchdog helper (BlackboxWatchdog)
  • AEC regression test suite and golden reference validation scripts
  • App icon and screenshot in README

Changed

  • AEC post-processing streams chunk-by-chunk instead of loading entire tracks into memory (constant ~16KB vs ~3x recording size)

Fixed

  • Crash when Krisp switches audio devices during recording (ObjCTryBlock broken in release builds due to NS_NOESCAPE block optimization)
  • Uncaught exception handler crash on background thread (inherited @MainActor isolation)
  • Config change observer crash on CoreAudio I/O thread (inherited @MainActor isolation)
  • Race between stop() and config change handlers causing concurrent AVAudioEngine mutation
  • Zero-format (0Hz, 0ch) from inputNode during device transitions now rejected before installTap
  • Rapid device switching causing unnecessary mic audio gaps (300ms debounce)

0.4.3 - 2026-03-11

Added

  • Mic diagnostic logging: input device name, permission status, and peak audio level logged per recording for debugging silent mic issues

0.4.2 - 2026-03-11

Fixed

  • False positive "Previous session crashed unexpectedly" error shown on every launch (crash detection flag was read after being overwritten)
  • Mic track silent in recordings: AVAssetWriterInput was configured for 2-channel output but AVAudioEngine mic delivers 1-channel audio
  • Mic buffer diagnostics: logs received/appended/dropped counts on recording stop for easier debugging

0.4.1 - 2026-03-11

Fixed

  • Crash on echo cancellation: DTLN-aec CoreML model bundle was not included in app bundle (only present in dev builds)
  • Silent crashes now detected on next launch with error shown in menu bar
  • Uncaught Objective-C exceptions now logged before process exit

Changed

  • Echo cancellation is now manual (button in recording detail view) instead of automatic on recording stop

0.4.0 - 2026-03-11

Added

  • Echo cancellation post-processing (DTLN-aec CoreML, 256-unit model) for cleaner mic recordings
  • Original/Processed audio toggle in recording detail view (defaults to processed when available)
  • Echo cancellation indicator icon in recordings list
  • Waveform visualization in recording detail view (amplitude bars via Canvas, click/drag to seek)
  • Track selector (Both/System/Mic) for playback controls
  • System notification when Screen Recording permission is revoked during recording
  • Input+output check for call detection (filters out dictation, Siri, voice memos)

Changed

  • Playback and transcription prefer echo-cancelled audio (audio-processed.m4a) when available
  • Processed recordings stored at 16kHz mono AAC alongside originals for size savings and debugging
  • Mic capture uses AVAudioEngine instead of SCStream .microphone (independent pipeline, automatic device following)
  • Call detection uses polling-only (no CoreAudio property listeners)
  • Dual-track M4A is now the final output format (no post-recording mixing)
  • Display-wide audio capture no longer excludes any apps

Fixed

  • dispatch_assert_queue_fail crash in disk space monitor (MainActor isolation inherited by DispatchSource handler on audioQueue)
  • Config change data race: AVAudioEngine handler now dispatches to audioQueue instead of running on arbitrary CoreAudio thread
  • Format mismatch on mic device change: preserves original tap format so AVAssetWriterInput encoder doesn't fail mid-stream
  • Inaccurate mic timestamps: uses AVAudioTime from tap callback via CMClockMakeHostTimeFromSystemUnits for proper multi-track sync
  • Idle sleep could interrupt recording: changed ProcessInfo activity to .userInitiated (prevents sleep)
  • Auto-recording not retried when initial start fails during active call
  • Silent file loss on finishWriting timeout: now saves partial file (playable due to movieFragmentInterval)

Removed

  • Apple Voice Processing (setVoiceProcessingEnabled) - removed for reliability (VPIO aggregate device caused format issues)
  • Post-recording audio mixing (AVMutableComposition + AVAssetExportSession)
  • Virtual audio processor exclusion list (Krisp, SoundSource, Loopback)
  • CoreAudio device change listener for mic following
  • CoreAudio process list and per-process input listeners (~140 lines)
  • Dead code: Log.fault, unused TranscriptionService error cases, stale .blackbox path checks

0.3.0 - 2026-03-10

Added

  • Per-process mic detection using macOS 14.2+ CoreAudio APIs (kAudioProcessPropertyIsRunningInput) - replaces system-wide DeviceIsRunningSomewhere listener
  • Microphone capture on auto-recordings - both system audio and mic are now recorded during calls
  • Call app name resolution from bundle ID (shown in HUD, notifications, and file names)
  • Transcription service with Soniox integration
  • Recordings detail view with built-in audio player and transcription UI
  • NavigationSplitView layout for recordings (sidebar + detail pane)
  • Soniox API key field in Settings
  • M4A export for recordings (single-track copy, auto-mixes if multi-track)
  • Real-time audio level metering with animated waveform icon in menu bar
  • HUD-based error notifications with configurable duration
  • Disk space pre-check (50 MB minimum) before starting a recording
  • Restart rate limiting for auto-recovery (max 3 restarts per 30-second window)
  • Virtual audio processor exclusion (Krisp, SoundSource, Loopback) to prevent voice duplication
  • "Report a Bug" menu item (opens GitHub issues)
  • Keyboard shortcuts for playback: Space (play/pause), Left/Right arrows (skip 15s)
  • Low disk space monitoring during recording (warning at 500 MB, auto-stop at 100 MB)
  • Polished DMG installer with proper icon layout via create-dmg
  • Swift Testing framework setup (make test, make check)
  • /release-dmg slash command for automated release pipeline

Changed

  • Recording uses dual-track capture (system audio + mic as separate AVAssetWriterInputs) with auto-mix to single-track M4A on save via AVMutableComposition
  • Recordings stored as plain directories instead of .blackbox macOS package bundles
  • Transcription uses Soniox stt-async-v4 model with mix-first approach (multi-track files mixed before upload)
  • Auto-recordings now include microphone audio (previously system audio only)
  • Recordings UI redesigned from flat table to split view with playback and transcription
  • "Record Microphone" setting now applies to all recordings, not just manual ones
  • make run now kills the previous Blackbox process before launching
  • Notifications switched from system banners (UNUserNotification) to in-app HUD toasts
  • "Recording Saved" HUD click now opens the main window instead of Finder
  • Soniox API key stored in macOS Keychain instead of UserDefaults (with one-time migration)
  • Transcription file upload uses streaming (64 KB chunks) instead of loading entire file into memory
  • Auto-recording now degrades gracefully on mic failure (continues without mic instead of stopping)

Removed

  • .blackbox bundle format and UTI declaration
  • One-time migration from flat .m4a to .blackbox format (no longer needed)

Fixed

  • Voice duplication when using Krisp or other virtual audio processors
  • Soniox speaker field parsing (v4 API returns String, not Int)
  • Transcription quality - proper speaker diarization with timestamps instead of single text blob
  • stop() no longer returns URL for corrupt files (guards on writer.status == .completed)
  • AVAssetWriter failure during recording now triggers auto-recovery instead of silently truncating
  • applicationShouldTerminate double-reply race prevented with hasReplied flag
  • Menu bar audio level icons use valid SF Symbols (speaker.wave.1/.2/.3)
  • Removed unnecessary .screen output registration from SCStream (wasted GPU resources)
  • Menu bar countdown/elapsed timer font changed to .monospacedDigit() for stable width
  • Auto-recording now starts after manual recording stops during an active call
  • HUD click handler fires only once (was possible to double-fire on rapid clicks)
  • Dropped audio buffers are now logged for diagnostic purposes
  • Recordings list debounced on window activation (prevents redundant disk scans)

0.2.0 - 2026-03-09

Added

  • Release pipeline with notarization, stapling, and Sparkle appcast
  • DMG creation with Applications symlink

0.1.0 - 2026-03-08

Added

  • Auto-recording triggered by microphone activity detection
  • Manual recording via menu bar
  • Dual-track M4A output (system audio + mic)
  • Recording HUD with start/save notifications
  • Grace period for call resumption detection
  • Auto-recovery on stream failures
  • Device following (seamless mic switching)
  • Crash-safe recordings via movie fragment intervals
  • Onboarding flow with permissions walkthrough
  • Settings: launch at login, grace period, save directory, notifications
  • Structured logging with os.Logger + file sink
  • Sparkle auto-update support
  • Developer ID code signing