domain/practice_insights — the practice insights rollup

backend/src/domain/practice_insights.py (207 lines). Pure-Python rollup over a user's recent PracticeSession rows: the router fetches the last 60 days in a single query and hands them to build_insights. "Keeping the aggregator DB-free keeps it cheap to test (no fixtures), and forces the SQL layer to stay a thin 'select by user and date window'" (practice_insights.py:1-6). All week math runs in the user's timezone so a Pacific user doesn't see a week-boundary jump at 5 PM Sunday (practice_insights.py:8-12).

Constants (practice_insights.py:25-36)

Constant	Value	Meaning
WEEKLY_TARGET_SESSIONS	4	The spec's "meeting the practice goal" cadence (≥4 sessions per local calendar week)
WEEKLY_HISTORY_WEEKS	8	Rolling window for the weekly bar chart — "enough to spot a 6-week streak with a soft on-ramp"
ROLLING_30D_WINDOW_DAYS	30	Window for total/average/per-mode rollups; named so the router's SQL window stays in lock-step

Output

PracticeInsights(weekly_counts, streak_weeks, total_minutes_30d, avg_duration_minutes_30d, per_mode_counts, last_insight) — mirrors schemas.practice.PracticeInsightsResponse "so the router can re-shape with a single model_validate call" (practice_insights.py:47-61).

Rules

• Week bucketing — ISO weeks (Monday-start, _monday_of), matching the practice-cadence rule elsewhere in the app (practice_insights.py:64-71); the 8 week-start dates are ordered oldest-first so the chart reads left-to-right (practice_insights.py:74-81).
• Zero-duration guard — sessions with duration_minutes <= 0 are skipped everywhere: "partial sessions count toward weekly totals iff duration > 0. Zero-duration aborts don't move the cadence needle" (practice_insights.py:92-95), mirrored in the 30-day stats so "a quick-cancel session never inflates per_mode_counts or drags the average toward zero" (practice_insights.py:138-140).
• Streak weeks — consecutive weeks ending now with count >= 4, scanning the weekly buckets from the newest backwards. "The current week counts even if it's still in progress — the spec's '4 x/week for 3 weeks running' UX shows users their momentum as it accrues" (practice_insights.py:101-118).
• 30-day window — strict lower bound today - 30 < local_day <= today "so the span is exactly 30 distinct calendar days (today-29 .. today); an inclusive lower bound would count 31" (practice_insights.py:121-128).
• Last insight — the most recent non-null insight across the full 60-day fetch window, deliberately wider than the 30-day rollup: "a user who took a multi-week pause should still see their last takeaway when they return rather than a blank card" (practice_insights.py:156-174).
• avg_duration_minutes_30d is None (not 0) with no qualifying sessions (practice_insights.py:151-153).

Worked example

Today = Thu 2026-07-30 (Monday 07-27); sessions this week: 4×15 min → current week bucket = 4; previous week = 5; two weeks ago = 2. Then streak_weeks = 2 (this week and last hit 4; the 2-count week breaks the scan). total_minutes_30d sums all positive-duration sessions in 07-01..07-30; a 0-minute abort on 07-29 affects nothing.

Endpoint: GET practice insights — see api/practice-sessions.

---

Grounded in adepthood@fbc529d, 2026-07-31.