ten31-transcripts/Ten31Transcripts/Visual/VisualObserver.swift

import Foundation
import ScreenCaptureKit
import CoreMedia
import QuartzCore
import AppKit

/// Window-scoped visual capture: streams the call window's own rendered content
/// at ~`fps`, hands each frame to the app adapter, and **releases it immediately
/// — frames are never written to disk**. Builds the speaker timeline and records
/// `visual_gap`s when the window is minimized (SCK delivers non-live frames).
///
/// Window visibility/focus is NOT required — SCK captures a window even when it's
/// occluded or on another Space; only minimization freezes the backing buffer.
@available(macOS 13.0, *)
final class VisualObserver: NSObject, SCStreamDelegate, SCStreamOutput {
    private let bundleID: String
    private let adapter: any AppAdapter
    private let t0Host: Double
    private let fps: Int
    private let queue = DispatchQueue(label: "xyz.ten31.visual")

    private var stream: SCStream?
    private let builder = TimelineBuilder()
    private var gaps: [VisualTimeline.Gap] = []
    private var gapStart: Double?

    /// Optional live hook (e.g. for a debug HUD). Observations only; no frame.
    var onObservations: (([SpeakerObservation], TimeInterval) -> Void)?

    init(bundleID: String, adapter: any AppAdapter, t0Host: Double, fps: Int = 3) {
        self.bundleID = bundleID
        self.adapter = adapter
        self.t0Host = t0Host
        self.fps = max(1, fps)
    }

    func start() async throws {
        let content = try await SCShareableContent.excludingDesktopWindows(false, onScreenWindowsOnly: false)
        // The call window: the largest window owned by the target app.
        let candidates = content.windows.filter { $0.owningApplication?.bundleIdentifier == bundleID }
        guard let window = candidates.max(by: { $0.frame.width * $0.frame.height < $1.frame.width * $1.frame.height }) else {
            throw NSError(domain: "Ten31", code: 2,
                          userInfo: [NSLocalizedDescriptionKey: "No \(bundleID) window to capture."])
        }

        let filter = SCContentFilter(desktopIndependentWindow: window)
        let config = SCStreamConfiguration()
        config.minimumFrameInterval = CMTime(value: 1, timescale: CMTimeScale(fps))
        config.queueDepth = 3
        config.showsCursor = false
        config.pixelFormat = kCVPixelFormatType_32BGRA
        // window.frame is in points; capture at native pixels so OCR can read small
        // initials/names (a half-res Retina capture badly hurts recognition).
        let scale = NSScreen.main?.backingScaleFactor ?? 2
        config.width = max(2, Int(window.frame.width * scale))
        config.height = max(2, Int(window.frame.height * scale))

        let stream = SCStream(filter: filter, configuration: config, delegate: self)
        try stream.addStreamOutput(self, type: .screen, sampleHandlerQueue: queue)
        try await stream.startCapture()
        self.stream = stream
    }

    func stop() async -> (segments: [VisualTimeline.Segment], gaps: [VisualTimeline.Gap]) {
        if let stream { try? await stream.stopCapture() }
        stream = nil
        return queue.sync {
            if let gs = gapStart {
                gaps.append(.init(start: gs, end: CACurrentMediaTime() - t0Host, reason: "minimized"))
                gapStart = nil
            }
            builder.finish()
            return (builder.segments, gaps)
        }
    }

    /// Merge mic-VAD self spans into the visual timeline (call before `stop`'s read,
    /// or fold in afterwards in the packager).
    func addSelfSpans(_ spans: [VADSpan], selfName: String) {
        queue.sync { builder.mergeSelfSpans(spans, selfName: selfName) }
    }

    // MARK: - SCStreamOutput (on `queue`)

    func stream(_ stream: SCStream, didOutputSampleBuffer sampleBuffer: CMSampleBuffer,
                of type: SCStreamOutputType) {
        guard type == .screen, CMSampleBufferDataIsReady(sampleBuffer) else { return }
        let now = CACurrentMediaTime() - t0Host

        switch frameKind(sampleBuffer) {
        case .idle:
            // Window is live but static (no pixel change) — no new info, not a gap.
            return
        case .gap:
            // Minimized/blanked: the backing buffer is frozen. Open a gap once and
            // close any open speaker segments so none is carried across it.
            if gapStart == nil {
                gapStart = now
                builder.closeOpenSegments(at: now)
            }
            return
        case .live:
            if let gs = gapStart {
                gaps.append(.init(start: gs, end: now, reason: "minimized"))
                gapStart = nil
            }
            guard let pixelBuffer = CMSampleBufferGetImageBuffer(sampleBuffer) else { return }
            let observations = adapter.analyze(frame: pixelBuffer, at: now)  // frame released after this scope
            builder.ingest(observations, at: now)
            onObservations?(observations, now)
        }
    }

    func stream(_ stream: SCStream, didStopWithError error: Error) {}

    private enum FrameKind { case live, idle, gap }

    /// SCK delivers `.complete` only when content changes, `.idle` for a static
    /// (but visible) window, and `.blank`/`.suspended`/`.stopped` when frozen.
    private func frameKind(_ sampleBuffer: CMSampleBuffer) -> FrameKind {
        guard let attachments = CMSampleBufferGetSampleAttachmentsArray(sampleBuffer, createIfNecessary: false)
                as? [[SCStreamFrameInfo: Any]],
              let raw = attachments.first?[.status] as? Int,
              let status = SCFrameStatus(rawValue: raw) else { return .live }
        switch status {
        case .complete: return .live
        case .idle:     return .idle
        default:        return .gap        // .blank / .suspended / .stopped
        }
    }
}