What & why

The meeting detector runs scan_process — a depth-25 accessibility-tree walk (every node a synchronous cross-process AX IPC) — on each running meeting-capable app every 5–10s, gated only by screen-lock and app-presence, never by audio.

The worst case is the common case: with no meeting there's no signal to early-break on, so it walks every window to full depth, forever, finding nothing. For an always-open browser with a huge web AX tree that's the dominant steady-state non-model CPU cost (it showed up as a top continuous non-model consumer in live sampling).

screenpipe already computes per-chunk audio activity (RMS-thresholded) and calls MeetingDetector::on_audio_activity — but that hook was a no-op, and is_in_audio_session() was a circular wrapper around the v2 flag the detector itself sets. This wires the existing-but-discarded signal into a real recency stamp and gates the idle scan cadence on it.

Inspired by fastrepl/anarlog (hyprnote), which keys meeting detection off audio-device activity rather than UI walking.

The change (idle scan cadence)

Before — every 10s, always (apps open, no call):
0s    10s   20s   30s   40s   50s   60s
●─────●─────●─────●─────●─────●─────●     7 depth-25 AX walks / min

After — 30s when silent; instant wake on audio:
0s                  30s                 60s
●───────────────────●───────────────────●   ~2 walks / min
                ⚡ audio onset → immediate scan (no detection latency)

Why detection accuracy cannot regress

• The state machine (advance_state) is untouched → all existing detector tests pass unchanged.
• A real meeting produces audio → the onset trigger scans immediately → detection is at least as fast as today.
• The only behavior delta: an app open with zero audio and no call is scanned every 30s instead of 10s. The 30s safety-net scan still catches even a fully-silent meeting; the instant anyone speaks, onset fires.
• With no detector (tests / detector disabled) the cadence is byte-identical to before.

Tests

• screenpipe-audio: 6/6 (4 new recency tests, incl. one pinning the signal is independent of the v2 flag — no circular dep).
• screenpipe-engine: 102 passed / 0 failed / 2 pre-existing ignored (4 new gate tests + the existing advance_state suite, all green).
• Standalone logic proof (gate decisions + recency + rising-edge arithmetic): 11/11.

Validation status

• ✅ Compiles clean (both crates), no warnings; all unit tests green.
• ⏳ End-to-end before/after CPU on a running build needs a full app build (CI/release). Methodology: count AX-walk frequency + meeting-detector thread CPU over a fixed idle window (apps open, muted) — expect ~3× fewer idle walks with no missed detection.

Files

• crates/screenpipe-audio/src/meeting_detector.rs — real on_audio_activity recency stamp + audio_onset Notify + audio_active_within() / wait_for_audio_onset().
• crates/screenpipe-engine/src/meeting_detector.rs — IDLE_QUIET_SCAN_INTERVAL + pure apps_present_scan_interval() + audio-onset select! arm.

🤖 Generated with Claude Code

A rendered before/after diagram is available (HTML→PNG) — gist hosting rejected the binary, so the ASCII timeline above stands in. Happy to attach the image inline if you'd prefer it.