domain/invitations — readiness-signal computation

backend/src/domain/invitations.py (185 lines). Pure computation of invitation coordinates: given a snapshot of cross-feature engagement (habit streaks, sustained practice weeks, active days in a rolling window) plus optional Creek Vault corpus-theme readings, decide which deeper depths the moment invites the user to consider. A later persistence pass turns the coordinates into InvitationSignal rows (invitations.py:1-8).

Ethics, stated in code (backend/src/domain/invitations.py:17-22):

This module encodes NO shaming and NO FOMO: it never counts what the user
*failed* to do, never compares them to anyone, and stays silent by default.
An absent signal is the norm, not a deficiency. The thresholds below are
deliberately conservative so an invitation only appears once a rhythm is
genuinely established — the "you choose your depth" principle honoured before a
single row is written.

Pure by design: no session, no I/O, no clock (invitations.py:24-28).

Thresholds

Constant	Value	Rationale
SUSTAINED_HABIT_STREAK_DAYS	21	"Three unbroken weeks … an integrated habit rather than a burst of early enthusiasm" (invitations.py:36-39)
SUSTAINED_PRACTICE_WEEKS	4	Four consecutive weeks meeting the ≥4-sessions/week cadence = "depth already reached" (invitations.py:41-44)
HIGH_ENGAGEMENT_ACTIVE_DAYS	25	Active 25 of the last 30 days before the embodied-community invitation — "never a nudge to do more" (invitations.py:46-49)
ENGAGEMENT_WINDOW_DAYS	30	The rolling window (invitations.py:51-54)
CORPUS_THEME_FULLNESS_THRESHOLD	0.75	A corpus theme this full is "a genuinely lived Aspect, not a passing mention" (invitations.py:56-59)

Target/kind strings are referenced from the model enums' values (not literals) "so the candidate coordinates can never drift from the persisted vocabulary — the drift-guard test asserts exactly this coupling" (invitations.py:61-70).

Inputs and output

ReadinessAggregates(habits: list[HabitSignal], practices: list[PracticeSignal], active_days_in_window: int, corpus_themes: list[CorpusThemeSignal] = []) — corpus_themes was appended last with a default "so every existing positional construction (behavioral-only) stays valid" (invitations.py:110-122). Output: InvitationCandidate(target_type, target_id, kind); target_id is None for ring-level invitations (invitations.py:73-83).

compute_invitation_candidates(aggregates)

Each source contributes independently (invitations.py:169-185):

Source	Rule	Emitted candidate
Habits	each habit with streak_days >= 21	(habit, habit_id, consistency) (invitations.py:125-131)
Practices	each practice with sustained_weeks >= 4	(practice, practice_id, mastery) (invitations.py:134-140)
Engagement	active_days_in_window >= 25	at most one (embodied_community, None, readiness) (invitations.py:143-149)
Corpus themes	filter to fullness >= 0.75; strongest wins — highest fullness, lowest stage_number breaking ties	at most one (course, stage_number, readiness) (invitations.py:152-166)

Worked example

habits=[(id=1, streak=25), (id=2, streak=20)], practices=[(id=7, weeks=4)], active_days=26, corpus_themes=[(stage=3, 0.8), (stage=6, 0.8), (stage=2, 0.5)] → four candidates: (habit, 1, consistency), (practice, 7, mastery), (embodied_community, None, readiness), and (course, 3, readiness) — stage 3 beats stage 6 on the tie-break. Habit 2 (20 < 21) and stage 2 (0.5 < 0.75) stay silent.

Persistence dedupe (a declined invitation is never recreated) is DB-level — see data-model/preferences-invitations; endpoints in api/invitations; the vault reading in creek-vault.

---

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