What problem does this solve?
CoachAnomalyKind.restingHRDrift has existed since the proactive-alert path landed, with a comment explaining why nothing fires it:
Reserved for a future baseline-aware detector (needs multi-day history).
That's accurate — CoachAnomalyDetector reads a 12-hour NotificationContextPacket, and drift is only meaningful against a multi-day baseline. But the baseline already exists: RestingHRBaselineService learns and persists the 10th percentile of 30 days on UserProfile.hrRestingBaseline. It just never reaches the packet.
An elevated resting heart rate is the signal that moves first under infection, alcohol, heat and under-recovery — typically about a day before you notice anything — so this is a cheap, high-value detector sitting one field away from working.
Proposed solution
Carry the learned baseline and last night's resting figure in the notification packet, and implement the detector.
Measure both sides the same way — the interpolated 10th percentile, sharing RestingHRBaselineService's own helper. Comparing a night's mean against a 30-day percentile would produce a difference that is mostly an artefact of the two formulas.
Bound the night by the sleep session rather than a clock window, so a late night or a shift schedule is measured over the hours actually slept.
Area
AI Coach (tools, prompts, on-device LLM)
What problem does this solve?
CoachAnomalyKind.restingHRDrifthas existed since the proactive-alert path landed, with a comment explaining why nothing fires it:That's accurate —
CoachAnomalyDetectorreads a 12-hourNotificationContextPacket, and drift is only meaningful against a multi-day baseline. But the baseline already exists:RestingHRBaselineServicelearns and persists the 10th percentile of 30 days onUserProfile.hrRestingBaseline. It just never reaches the packet.An elevated resting heart rate is the signal that moves first under infection, alcohol, heat and under-recovery — typically about a day before you notice anything — so this is a cheap, high-value detector sitting one field away from working.
Proposed solution
Carry the learned baseline and last night's resting figure in the notification packet, and implement the detector.
Measure both sides the same way — the interpolated 10th percentile, sharing
RestingHRBaselineService's own helper. Comparing a night's mean against a 30-day percentile would produce a difference that is mostly an artefact of the two formulas.Bound the night by the sleep session rather than a clock window, so a late night or a shift schedule is measured over the hours actually slept.
Area
AI Coach (tools, prompts, on-device LLM)