Technical Manual · v1.9

How Metis Meet works under the hood.

Reference guide covering single-window capture, twin-track audio clocks, 3-second echo hiding, AI provider configuration, Cursor CLI Seatbelt sandboxing, and local knowledge packs.

Single-Window HEVC + Twin-Track Audio Clock

RecorderEngine uses SCContentSharingPicker to select a single application window and configures SCStream at the window's Retina point size (capped at 1920×1080, 30 fps, H.265/HEVC). System audio (48 kHz stereo AAC, labeled Them) and microphone capture viaAVCaptureSession (48 kHz mono AAC, labeled Me) are synchronized onto one presentation timeline and flushed every 10 seconds using movie fragments.

~/Movies/MeetingRecorder/
  2026-11-19_1400_Architecture-Review.mov             # 1080p30 HEVC + Track 1 (Them) + Track 2 (Me)
  2026-11-19_1400_Architecture-Review_mixed.m4a       # Combined mono AAC for quick sharing
  2026-11-19_1400_Architecture-Review_transcript.txt  # Timestamped plain-text transcript
  2026-11-19_1400_Architecture-Review_transcript.srt  # Subtitle captions for Final Cut / QuickTime
  2026-11-19_1400_Architecture-Review_transcript.json # Structured sidecar + bookmarks + echo drops
  2026-11-19_1400_Architecture-Review_summary.md      # Structured AI summary & action items

On-Device SpeechAnalyzer + 3-Second Speaker-Bleed Hiding

LiveTranscriber streams PCM buffers from both audio tracks into two concurrent on-device recognizers (SpeechAnalyzer + SpeechTranscriber on macOS 26, falling back toSFSpeechRecognizer with requiresOnDeviceRecognition = true on macOS 15). WhenHide echo from speakers is enabled, microphone segments are held for 3.0 seconds and dropped if word-set overlap or word-pair similarity against recent system audio is ≥ 0.60.

Hybrid SQLite FTS5 + 512-dim NLContextualEmbedding Retrieval

Each knowledge pack lives under ~/Library/Application Support/MeetingRecorder/packs/<pack-id>/index.sqlite.KnowledgePackManager parses Markdown, source code (.swift, .py,.ts, .js, .go, .rs, .sql), PDFs viaPDFKit, and Word documents (.docx via /usr/bin/unzip +XMLParser) into 200-token chunks with 40-token overlap.

At query time, QuestionDetector scores completed utterances on the Them track (interrogative starters +2, question mark +2, prompt phrases +3, technical pack nouns +1; threshold ≥ 3). Hybrid search blends BM25 and cosine similarity over 512-dim NLContextualEmbedding mean-pooled vectors using Reciprocal Rank Fusion (vector weight 0.7), returning the top snippets in 16 ms.

Keyboard Shortcuts Reference

ShortcutAction
⌃⌥⌘RStart or stop recording the selected window or display
⌃⌥⌘MDrop a timestamped QuickTime chapter bookmark with optional note
⌃⌥⌘AAnswer the last 45 seconds of speech with Assistant
⌃⌥⌘HShow or hide the non-activating Assistant & Live Coaching panel
⌃⌥⌘CCopy the active Assistant talking points
⌃⌥⌘PPin the active hint card to the session tray
⌘FFind text inside the Live Transcript window
⌘,Open Recording, AI Provider, Knowledge Pack, and Coaching Settings