Skip to content

YC (S26) | Open Computer History | Continuously record your company computer work, map your workflows, help you find work worth automating, and power your agents' context

fix(meeting): end calls on silence, not just captured-audio existence (#4265)

Louis Beaumont6ab6b917fe85
diff --git a/crates/screenpipe-engine/src/meeting_detector.rs b/crates/screenpipe-engine/src/meeting_detector.rs
index 8495ab275..62003909e 100644
--- a/crates/screenpipe-engine/src/meeting_detector.rs
+++ b/crates/screenpipe-engine/src/meeting_detector.rs
@@ -1729,6 +1729,23 @@ fn is_browser_app(app_name: &str) -> bool {
.any(|b| contains_case_insensitive(app_name, b))
}
+/// Whether recent output audio should keep an `Ending` meeting alive.
+///
+/// `recent_output_chunk` comes from the DB (`has_recent_output_audio`), which
+/// matches *any* `(output)` audio chunk in the window — but the system-audio tap
+/// writes those continuously even while silent. On its own it therefore pins a
+/// meeting "live" forever once the call ended (controls gone, tap still running),
+/// so the meeting never auto-finalizes (it flaps Active<->Ending). We additionally
+/// require `recent_voice_activity` (RMS-gated, from the capture pipeline) so a
+/// quiet stretch lets the grace timer run out and the meeting end normally, while
+/// a genuinely audible call (tab-switched / minimized / screen-sharing) stays alive.
+fn output_audio_keeps_meeting_alive(
+ recent_output_chunk: bool,
+ recent_voice_activity: bool,
+) -> bool {
+ recent_output_chunk && recent_voice_activity
+}
+
/// Advance the state machine based on scan results.
///
/// Returns the new state plus an optional action to perform (DB insert/update).
@@ -3054,7 +3071,19 @@ pub async fn run_meeting_detection_loop(
// native meeting apps (e.g., Zoom). Audio activity is a strong signal
// that the user is still in the meeting even if UI controls are hidden.
let has_output_audio = if matches!(state, MeetingState::Ending { .. }) {
- db.has_recent_output_audio(30).await.unwrap_or(false)
+ // A recent (output) chunk alone isn't enough: the system-audio tap
+ // writes those continuously even during silence, which would keep an
+ // already-ended call "live" forever (controls gone, tap still running)
+ // and the meeting would never auto-finalize. Require RMS-gated voice
+ // activity too, so a quiet call ends on schedule but an audible one
+ // (tab-switched / minimized / screen-sharing) stays alive. No detector
+ // wired (tests / detector disabled) -> default true to preserve prior
+ // behavior.
+ let recent_output_chunk = db.has_recent_output_audio(30).await.unwrap_or(false);
+ let recent_voice_activity = detector.as_ref().map_or(true, |d| {
+ d.audio_active_within(AUDIO_GATE_WINDOW.as_millis() as u64)
+ });
+ output_audio_keeps_meeting_alive(recent_output_chunk, recent_voice_activity)
} else {
false
};
@@ -3535,6 +3564,20 @@ async fn insert_new_meeting(
mod tests {
use super::*;
+ #[test]
+ fn output_audio_keepalive_requires_real_voice_not_just_a_chunk() {
+ // The system-audio tap writes a recent (output) chunk continuously, even
+ // during silence — that alone must NOT keep an ended call alive, or a
+ // detected meeting never auto-finalizes (flaps Active<->Ending forever).
+ assert!(!output_audio_keeps_meeting_alive(true, false));
+ // A genuinely audible call (controls hidden: tab-switch / minimize /
+ // screen-share) stays alive.
+ assert!(output_audio_keeps_meeting_alive(true, true));
+ // No recent output chunk at all -> never kept alive by this signal.
+ assert!(!output_audio_keeps_meeting_alive(false, true));
+ assert!(!output_audio_keeps_meeting_alive(false, false));
+ }
+
// ── audio-gated scan cadence tests ─────────────────────────────────
// These pin the CPU optimisation: with apps open but no recent audio the
// Idle scan rate drops; a tracked meeting (any non-Idle state) is never