Health fixes: no freeze on Email Support, cheaper idle call detection, lighter Speakers page

[r3dbars]

Requested by Justin · project thread

Why

Before: clicking Email Support or Send Diagnostic walked the whole model cache folder (lots of CoreML files) and read the full reliability log (up to ~10 MB) on the main thread, so the app could stall right when someone is reporting a problem. While the app sat idle, call detection's 5 s backstop read the bundle ID of every audio process on the Mac (30-60 of them), even ones doing no audio. And the Speakers page re-read up to 256 KB of every saved transcript after every rename, merge, discard, or meeting save (~250 MB per rename at 1,000 meetings).

After: the support disk reads run off the main thread. The idle scan only fetches a bundle ID for processes actually using the mic or speakers, and the mic and camera backstop timers get 500 ms leeway so macOS can batch their wakes. The Speakers scan skips transcripts it already knows have no unnamed voices until the file changes.

Three commits, one problem each, from an app health audit (thread linked above).

Product Impact

• Affects: meetings (call detection, speaker review), settings support actions
• Lane: meeting reliability
• Why this matters: the app idles for hours on ~150 Macs; fewer wakes and disk reads mean less battery, and support/settings flows shouldn't beachball.

What changed

• TranscriptedSupportActions: diagnosticsSnapshot is now async. The model-cache walk and reliability log read run in a Task.detached; everything else stays on the main actor. sendFeedback / sendDiagnosticEvent / feedbackEmailURL / diagnosticsText became async. An in-flight flag stops a double click from opening two drafts.
• TranscriptedSettingsActions.sendDiagnosticEvent is now () async -> String?. The settings view awaits it in a Task and has its own in-flight guard.
• MicActivityMonitor.currentProcessAudioState: reads isRunningInput/isRunningOutput first and the bundle ID only when one is true. All three consumers already ignore bundle IDs of processes with the relevant flag false, so detection results are unchanged. Mic and camera backstop timers: 500 ms leeway.
• SpeakerReviewQueueScanner: new NoPendingSpeakersCache (path → modification date + size, lock-guarded, capped at 20k entries). A transcript with no db_pending speaker can never produce a review row regardless of profiles, so it is skipped until its fingerprint changes. Files with pending speakers are still read every scan. New hasPendingSpeakers(in:) helper. Two new fast-test suites.

How I checked it

• bash scripts/dev/linux-checks.sh (48 passed)
• python3 scripts/dev/check-source-pins.py --changed-only (pass)
• bash build.sh --no-open / bash run-tests.sh / bash run-integration-smoke.sh: via CI

Checks I could not run, and why:

• No Swift toolchain in this session, so build and tests rely on CI.

Mac or hardware test still needed? If yes:

• Optional: start a Zoom/Meet call and check the "record this meeting?" prompt still shows (unchanged behavior). Click Email Support and check the draft opens once with diagnostics. Name a pending voice on the Speakers page and check it leaves the queue.

Risk Review

• Privacy / local-first behavior reviewed (no new data leaves the device)
• No new analytics properties or Sentry tags
• Checked the text-pin tests for every file I edited
• No storage path changes
• Agent PRs got an independent deep review of the full diff

Notes

Audio real-time code untouched. No new AVAudioEngine/inputNode touches.

🤖 Generated with Claude Code

https://claude.ai/code/session_0126tPTsXGHmWx6EmsR32sjX