Seiten · Health Debug

M6 — Cycle (Fertility) — Session Card

Planning artifact. This card is opened by a future build session so it knows exactly what to build for module M6 Cycle (Fertility). It contains no production code. It is faithful to docs/plan/13-execution-modules-and-sessions.md (the session framework), 03-protocol-engines.md §6 (the engine spec), 04-backend-togo.md (backend conventions), 02-platform-divergences.md §2.1 (this module closes Divergence #1 on iOS), and 12-legacy-reference-map.md (legacy precedents).

1. Purpose

The Cycle (Fertility) engine tracks user-logged menstrual cycles and offers predictions only once enough consistent data exists, and otherwise fails safe to "no prediction." It is a backend state machine in ToGO (the Single Source of Truth); every client — web, mcp, apple, android, desktop (there is no chrome surface for this module) — only renders the engine state the API returns and submits log intents. This module closes platform Divergence #1 (Meds & Cycle Tracking, historically missing on iOS): in the greenfield build the 3-cycle prediction gate and irregularity fail-safe live once, server-side, so the iOS/Android/web screens are pure presentation against an already-live contract and can never re-derive the rule. Because cycle/fertility data is highly sensitive, cross-user isolation (404-not-403) and per-user SSE/event routing (no cross-user leakage) are hard requirements, not niceties — the legacy build had to fix an event broker that leaked across users, and this module must not reintroduce that class of bug.

2. Engine parameters relevant to this module

Spec-fixed constants (from `03-protocol-engines.md §6.2` and shared context — do NOT invent others)

Parameter	Value	Notes
Calibration requirement	3 consecutive logged cycles before any prediction is made	Hard gate. No prediction, ever, below this.
Irregularity rule	On irregularity the engine fails safe: no prediction	Predictions are withdrawn, never "widened."
Prediction sentinel	`unavailable` (+ reason code)	NOT NULL rule: an explicit sentinel value, never SQL NULL.
States	`Insufficient` / `Calibrated` / `Suspended`	Categorical only — the user never sees a confidence percentage (Core Opinion 1).
"Consecutive"	no unlogged cycle between logged ones	Engine never infers a skipped cycle happened.

State transitions (03 §6.3): Insufficient → Calibrated on the 3rd consecutive cycle; Calibrated → Suspended when irregularity is detected; a logging gap resets the consecutive counter to Insufficient; Suspended → Insufficient on recalibration (a fresh 3-consecutive run — see PENDING items).

Legacy precedent (PENDING RATIFICATION — precedents from `cycle.go`, NOT spec-fixed; carry as fail-safe categorical/temporal, no medical certainty)

These are the concrete numbers the legacy engine used. They are candidate answers, not authority — the greenfield spec (03 §6.2, and Open Question #6 in 03 §Open Questions) leaves the quantitative definition of "irregularity" undefined. Use them as the default implementation to be ratified during the backend session, and encode whatever is ratified as literal constants inside the engine core/ package.

Precedent	Legacy value (`cycle.go`)
Minimum cycles for prediction	`MinCyclesForPrediction = 3` (matches the spec gate)
Irregularity threshold	`IrregularThresholdDays = 9`-day spread → refuse to predict
Averaging window	average of the last 6 cycles
Ovulation estimate	next cycle start −14 days
Fertile window	−5 / +1 day around the ovulation estimate
BBT confirmation	3 readings ≥ +0.2 °C over the prior 6-day mean
Pregnancy dating	Naegele: +280 days from cycle start
Safety data	`SafetyProvider` interface with `NoSafetyData{}` default — no drug DB wired, by design

All predictions are behavioral-data extrapolations, explicitly labelled non-medical, and every AI-adjacent / fertility surface ends in "Consult your doctor" (Core Opinion 2). The engine never smooths, averages away, or best-guesses an irregular history.

Data model (this module owns)

Tables: cycles and cycle_days (per-user). Every column NOT NULL with a defined default; UUID ids; timestamptz timestamps (04 §4/§6).
Endpoints (names from 03 §6.6, under the /api/v1/ prefix per 04 §API): POST /api/v1/engines/cycle/logs (log a cycle event; accepts Idempotency-Key), GET /api/v1/engines/cycle/state (calibration state + prediction-or-unavailable + reason), GET /api/v1/engines/cycle/history.
Engine output shape: { state, consecutive_cycles_logged, prediction: value|unavailable, reason }.

3. Session table

One row per session. Surface order per 13 §3: backend → web → mcp → (no chrome) → apple → android → desktop. The backend session must land before any client session (13 §4). Replace <date> (YYYY-MM-DD) and NN (per-cell counter) when the session actually runs.

Session ID	Surface	Scope (what this one session builds)	Entry deps	Done criteria (template from `13 §5`)	Plan docs to read
`S-<date>-M6-backend-NN`	backend	The cycle ToGO plugin: pure `core/` state machine (Insufficient/Calibrated/Suspended, the 3-consecutive gate, irregularity → no-prediction fail-safe, `unavailable` sentinel + reason) + `plugin.go` registration; `cycles` + `cycle_days` expand-contract migration; the three endpoints; engine unit tests as executable spec; integration test on real Postgres. Ratify the PENDING legacy numbers (§2) and encode them literally in `core/`. Verify cross-user isolation (404-not-403) and that any SSE/event emitted for cycle carries no cross-user leakage.	M0.1–M0.4 (foundation backend: kernel, TOGO auth + isolation test, plugin framework, expand-contract + error-envelope harness); M1 backend as the reference-engine pattern to copy	Backend-engine template: plugin registered on the kernel; migration is expand-contract and every column `NOT NULL` with a default; endpoints return computed state; engine unit tests encode every spec/ratified constant literally and pass (no prediction until 3 consecutive cycles; fail-safe = no prediction on irregularity); integration test on real Postgres passes (idempotency replay; contradictory-edit → `Suspended`; cross-user isolation 404-not-403); no protocol constant appears outside the engine's `core/` package	`13` (§3–§6), `03 §6` (engine spec), `04` (§2 layout, §3 sqlc+Atlas, §4 NOT NULL rule, §5 expand-contract, §8 plugin/purity, §9 testing), `12 §2` (legacy `cycle.go` constants)
`S-<date>-M6-web-NN`	web	The cycle screen (`/engines/cycle`): renders server state — shows calibration progress / "needs 3 consecutive logged cycles" when `Insufficient` or `Suspended`, and shows a prediction only when the server state is `Calibrated`. Cycle-log entry control submits intents only. Both themes; Arabic RTL.	`S-<date>-M6-backend-NN`; M0.5–M0.7 (web shell, theme engine, i18n engine); M1 web as the reference card pattern	Client-engine template: renders the cycle endpoint state and submits log intents only; thin-client/SSOT audit passes — no gate/irregularity/prediction logic in client code, grep for the protocol constants finds nothing, prediction shown only when server says `Calibrated`; both themes render per the strict color spec; Arabic RTL mirrors correctly (incl. progress/calibration bars); strings are ICU, none concatenated; "Consult your doctor" deferral rendered verbatim	`13` (§5 client template, §6), `03 §6`, `05` (§3 routing `/engines/cycle`, §4 query keys, §5b RTL strategy, component inventory), `02 §2.1` (Divergence #1)
`S-<date>-M6-mcp-NN`	mcp	Cycle status read-only MCP tool(s): `cycle_status` → `GET /api/v1/engines/cycle/state`, returned verbatim. No prediction tool beyond what the backend exposes — the engine's fail-safe rules apply unchanged; no client-side inference.	`S-<date>-M6-backend-NN`; M0.8 (MCP server + CLI skeleton)	MCP template: each tool calls the real API with personal-access-token auth, returns server state verbatim, holds no protocol logic and no AI keys; the excluded tools stay excluded (no cycle/fertility prediction tool beyond the backend `state`/`history` reads; the general `complete_break` exclusion still holds)	`13` (§5 MCP template, §6), `11` (§thin-layer rule, planned tools, deliberately-excluded tools incl. "no cycle/fertility prediction tool"), `03 §6`
`S-<date>-M6-apple-NN`	apple	Native Swift Meds & Cycle screen (iOS; + macOS/widget surfaces as they apply) rendering cycle engine state against the live API — this session closes Divergence #1 on iOS. No local re-implementation of the 3-cycle gate or irregularity rule.	`S-<date>-M6-backend-NN`; M1 apple (Apple project scaffold + reference engine pattern)	Client-engine template: renders the cycle endpoint state and submits log intents only; thin-client/SSOT audit passes (no gate/prediction/irregularity logic in Swift — grep finds no protocol constants; iOS never re-derives the rule); both themes per the strict color spec; Arabic RTL mirrors correctly; strings are ICU/String-Catalog, none concatenated; "Consult your doctor" deferral present	`13` (§5 client template, §6), `03 §6`, `07` (§principles, project layout, iOS app row — "Meds & Cycle tracking (closes Divergence #1)"), `02 §2.1`
`S-<date>-M6-android-NN`	android	Native Kotlin/Compose cycle screen (phone; + Wear/Glance as they apply) rendering cycle engine state; cycle-log intents only.	`S-<date>-M6-backend-NN`; M1 android (Android scaffold + reference engine pattern)	Client-engine template: renders the cycle endpoint state and submits log intents only; thin-client/SSOT audit passes (no protocol logic in Kotlin, grep finds no constants); both themes per the strict color spec; Arabic RTL mirrors correctly; strings are ICU, none concatenated; "Consult your doctor" deferral present	`13` (§5 client template, §6), `03 §6`, `08` (project layout, engine screen pattern, en/ar parity), `02 §2.1`
`S-<date>-M6-desktop-NN`	desktop	Electron cycle surface, reusing the M6 web components (`13 §3` step 7) — thin renderer of server cycle state.	`S-<date>-M6-backend-NN`; `S-<date>-M6-web-NN` (component reuse); M1 desktop (Electron scaffold pattern)	Client-engine template: renders the cycle endpoint state and submits log intents only; thin-client/SSOT audit passes (reused web components carry no protocol logic; PAT auth via safeStorage, no keys embedded); both themes per the strict color spec; Arabic RTL mirrors correctly; strings are ICU, none concatenated	`13` (§5 client template, §6), `03 §6`, `09` (Electron layout, web-component reuse), `05` (reused components)

Session count: 6 (backend, web, mcp, apple, android, desktop — no chrome).

4. Notes & gotchas

Legacy precedents worth reusing

internal/protocols/cycle.go and its *_test.go are the executable protocol specification — validate the new core/ engine against those cases first (12 §2). Constants: MinCyclesForPrediction=3, IrregularThresholdDays=9, last-6-cycle average, ovulation −14d, fertile window −5/+1d, BBT +0.2 °C ×3 over prior 6-day mean, Naegele +280d.
The legacy SafetyProvider / NoSafetyData{} seam (no drug DB wired, by design) is the correct posture — keep the safety hook empty rather than inventing safety data.
Legacy per-user CRUD already modeled cycles and cycle_days (and pregnancies with a partial-unique "one ongoing per user") — a schema precedent, but re-evaluate the legacy anti-patterns (text ids → use UUIDs; absent dates as empty text → use proper defaults; no FK to the auth users table; CSV-text fields).

Liabilities to avoid

#1 liability — duplicated protocol logic on a client. The legacy iOS app re-implemented engines locally and never called the API; that is exactly the drift this module closes. Every client session's done-criteria include the thin-client/SSOT grep audit for precisely this reason.
Sensitive-data routing. Fertility data is sensitive: the legacy fix replaced ToGO's event broker (which leaked events across users) with a per-user SSE hub carrying resource names only, no state (12 §4). Any cycle event/SSE emitted here must be per-user with zero cross-user leakage; foreign rows return 404 not 403. Add an isolation assertion to the backend integration test.
Never emit a prediction outside Calibrated. The prediction field must carry the unavailable sentinel (not SQL NULL) plus a reason in Insufficient/Suspended. Clients must show a prediction only when the server state is Calibrated — no client-side "we probably have enough data" heuristic.
No medical certainty. Predictions are behavioral extrapolations, never fertility guarantees or contraceptive advice; all such copy defers to "Consult your doctor" (Core Opinion 2).

PENDING DESIGN / PENDING RATIFICATION

PENDING RATIFICATION: the quantitative definition of "irregularity" is undefined in spec (03 §6.2, Open Question #6). Ratify the legacy 9-day-spread / last-6-average / BBT numbers (or a replacement) during the backend session and encode them literally in core/.
PENDING (spec open, 03 §6.7): whether recalibration after Suspended requires a fresh full 3-consecutive-cycle run (planning default: yes, fail-safe); and whether backfilled historical cycles count toward "consecutive" (planning default: stored but not counted — only forward-logged cycles calibrate).
PENDING PRODUCT/CLINICAL (03 §7): whether an active contraceptive method (M7) should suspend cycle predictions — a product/clinical decision, not an engine guess. Coordinate at M7, do not hardcode an interaction now.
PENDING SCOPE: pregnancy dating (Naegele +280d) and a pregnancies table — decide in the backend session whether it is in M6 scope or deferred; it is a legacy precedent, not a spec-fixed M6 deliverable.
PENDING DESIGN: the calibration-progress UI (Insufficient state) and the cycle/prediction screen layouts across web/apple/android come from the separate design track (02 §2.1, 03 §6.3).

5. Handoff reminder

At the end of every M6 session, per 13 §6: write one session-handoff memory to the healthdebug brain (retain, then memory/edit to attach metadata — retain cannot set metadata) using the exact handoff content structure, with source_ref: handoff/M6-cycle-fertility/<surface>/<n> (unique per session; never reuse) and metadata: {project:"healthdebug", category:"session-handoff", era:"greenfield", status:"current", surface:"<surface>", module:"M6", title:"Handoff <session-id>", next:"<module+surface>"}. Then update the single build-ledger memory (source_ref: build-ledger, edited in place) so one recall shows the whole frontier and the recommended next session. Any architectural decision or new open question also gets its own classified memory and, if it changes the plan, an edit to the relevant plan doc.