feat(web): voice input — record, check the recognised text, send

Microphone button in the chat composer and in free-text steps. The browser
records 16 kHz mono WAV (no server-side ffmpeg), /me/{token}/stt transcribes,
the text lands in the field for the patient to check and edit — never sent
automatically, so safety and scrub() apply on send. Hidden when the server has
voice input off or the browser lacks a microphone API. nginx: 4 MB body limit.

Refs #117

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.env.example

@@ -11,6 +11,10 @@ CHAT_AGENT=auto
 OPENROUTER_API_KEY=
 OPENROUTER_MODEL=google/gemini-2.5-flash-lite
 
+# --- Voice input (speech-to-text): auto | off | mock | openrouter. OpenRouter = synthetic data only (ADR-014).
+STT_PROVIDER=auto
+STT_MODEL=google/gemini-2.5-flash
+
 # --- Pseudonymisation key: 32+ random bytes, hex. Generate: python -c "import secrets;print(secrets.token_hex(32))"
 PSEUDO_KEY=
 
@@ -23,6 +27,8 @@ CLINICAL_DB_URL=postgresql+asyncpg://clinical_app:clinical_app@localhost:5432/cl
 
 # --- Telegram: prototype on synthetic data only; prod channel is MAX / Gosuslugi. Empty = bot disabled.
 TG_BOT_TOKEN=
+# http://user:pass@host:port or socks5://… — needed on the RF server (docker has no route to Telegram over IPv4)
+TG_PROXY_URL=
 DOCTOR_NOTIFY_CHAT_ID=
 
 # --- Demo virtual clock (ADR-005)

.gitea/workflows/deploy.yml

@@ -0,0 +1,52 @@
+name: CD
+
+# Every merge into dev → lint + tests → deploy to the shared server (dev.meditron.tenvex.ru / meditron.tenvex.ru).
+# Secrets: DEPLOY_SSH_KEY (private key of the `ptsr-deploy` user), DEPLOY_HOST (user@host).
+on:
+  push:
+    branches: [dev]
+  workflow_dispatch:
+
+concurrency:
+  group: deploy-dev
+  cancel-in-progress: false
+
+jobs:
+  test:
+    name: ruff · pytest
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: astral-sh/setup-uv@v3
+        with:
+          version: "0.8.4"
+      - run: uv sync --frozen
+      - run: uv run ruff check .
+      - run: uv run pytest -q
+
+  deploy:
+    name: deploy dev
+    needs: test
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - name: ssh key
+        run: |
+          install -d -m 700 ~/.ssh
+          printf '%s\n' "${{ secrets.DEPLOY_SSH_KEY }}" > ~/.ssh/deploy_key
+          chmod 600 ~/.ssh/deploy_key
+      - name: deploy
+        env:
+          DEPLOY_HOST: ${{ secrets.DEPLOY_HOST }}
+          DEPLOY_SSH_OPTS: "-i /root/.ssh/deploy_key -o IdentitiesOnly=yes -o StrictHostKeyChecking=accept-new -o ConnectTimeout=20"
+        run: |
+          export DEPLOY_SSH_OPTS="${DEPLOY_SSH_OPTS//\/root\/.ssh/$HOME/.ssh}"
+          infra/deploy/deploy.sh HEAD
+      - name: smoke
+        run: |
+          sleep 10
+          for u in https://dev.meditron.tenvex.ru/ https://dev.meditron.tenvex.ru/api/v1/demo/personas https://dev.meditron.tenvex.ru/doctor/; do
+            code=$(curl -s -o /dev/null -w '%{http_code}' -m 20 "$u"); echo "$u -> $code"; [ "$code" = 200 ]
+          done

.streamlit/config.toml

@@ -1,15 +1,15 @@
 # Dashboard-only Streamlit settings (doctor dashboard + demo console).
 [theme]
 base = "light"
-primaryColor = "#2A78D6"
-backgroundColor = "#F6F7F9"
+primaryColor = "#285E4E"
+backgroundColor = "#F6F7F3"
 secondaryBackgroundColor = "#FFFFFF"
-textColor = "#1B1F24"
-borderColor = "#E3E6EA"
+textColor = "#213B38"
+borderColor = "#E1E7E3"
 baseRadius = "medium"
 # Falls back to the system sans-serif when offline (ADR-004).
-font = "Inter:https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&display=swap, sans-serif"
-headingFontWeights = 700
+font = "sans-serif"
+headingFontWeights = 600
 
 [client]
 toolbarMode = "minimal"

Makefile

@@ -6,7 +6,7 @@ install:
 up:
 	@test -f .env || cp .env.example .env
 	docker compose up -d --build
-	@echo "API: http://localhost:8000/docs  Dashboard: http://localhost:8501"
+	@echo "Patient app: http://localhost:8080  API: http://localhost:8000/docs  Doctor/demo console: http://localhost:8501"
 
 down:
 	docker compose down

README.md

@@ -1,14 +1,117 @@
 # PTSRaiAgent — ИИ-помощник сопровождения при ПТСР и депрессии
 
-> Хакатон MedITron · кейс SberMed AI «ИИ-помощник пациента». Прототип на **синтетических данных**.
-> Telegram используется только для прототипа на синтетических данных; прод-канал — MAX / Госуслуги.
+> Хакатон MedITron · кейс SberMed AI «ИИ-помощник пациента». Прототип работает **только на синтетических данных**.
+> Демо-стенд: **https://meditron.tenvex.ru** (кабинет врача: https://meditron.tenvex.ru/doctor/, API: https://meditron.tenvex.ru/docs).
 
-🚧 README заполняется к сдаче (issue «README по требованиям оформления»). Разделы по требованиям жюри:
+---
 
-1. Цели и задачи решения
-2. Что реализовано
-3. Развёртывание, запуск, использование
-4. Входные и выходные данные, API
-5. Ограничения и особенности медицинской тематики
+## 1. Цели и задачи
 
-Пока смотрите [`CLAUDE.md`](CLAUDE.md) и [`docs/`](docs/).
+**Проблема.** Ветераны СВО и люди с депрессией до помощи доходят с трудом, а начав лечение, часто его бросают: в первые недели побочные эффекты уже есть, а эффекта ещё нет. Врач между визитами ничего не знает о пациенте.
+
+**Решение.** Цифровой помощник сопровождает пациента между визитами к врачу. Из сценариев кейса закрыты два: **мониторинг после обращения** и **управление приёмом лекарств**.
+- Пациент общается с помощником в чате: веб-приложение или Telegram. В проде канал — MAX / Госуслуги.
+- Помощник собирает валидированные шкалы, ежедневный чек-ин и отметки о приёме препарата.
+- Он отвечает на вопросы о лечении по клиническим рекомендациям Минздрава, по инструкции ГРЛС и по карте пациента в ЕМИАС.
+- Врач видит «красные зоны» и получает сигналы при ухудшении.
+
+**Принципы.**
+- ИИ не ставит диагнозы и не назначает лечение.
+- LLM только формулирует ответы и извлекает данные.
+- **Риск, алерты и скоринг считают детерминированные правила** по валидированным шкалам (PHQ-9, PCL-5, GAD-7, AUDIT-C, C-SSRS, DAR-5).
+- Клинические решения принимает врач.
+
+## 2. Что реализовано
+
+| Блок | Что есть |
+|---|---|
+| **Чат-агент** | LLM с инструментами только для чтения: поиск по базе клинических рекомендаций (RAG, депрессия 2024, ПТСР 2023), карта пациента из ЕМИАС (мок), инструкция ГРЛС, детерминированная проверка препарата под пациента (сопутствующие болезни, непереносимость вспомогательных веществ, взаимодействия). Ответ печатается в реальном времени (SSE), видны шаги агента. Провайдеры: GigaChat (функции), OpenRouter (временно, только синтетика), офлайн-mock. |
+| **Диалоговый движок** | Обезличивание → проверка безопасности → FSM. Согласие (152-ФЗ, отзыв в один шаг), проверка возраста, онбординг ветерана (PC-PTSD-5 → PCL-5 → PHQ-9 → AUDIT-C), опросник intake (C-SSRS, PHQ-2→9, TSQ→PCL-5, RUS-AUDIT-S), ежедневный чек-ин, напоминания о дозах, побочные эффекты. На вопросы отвечает в любом состоянии. |
+| **Безопасность** | Лексикон риска (100% recall на наборе кризисных фраз), фиксированный кризисный скрипт без LLM, алерт врачу и дежурному, план безопасности. В демо кризисное окно показывается в конце теста и не блокирует диалог (`CRISIS_BLOCKING`). |
+| **Правила мониторинга** | 17 правил в `config/rules.yaml`: п.9 PHQ-9, рост баллов, побочки 3 дня подряд, намерение бросить, пропуски 3 из 7, сон < 4 ч, алкоголь, молчание > 3 дней и другие. |
+| **Кабинет врача** | Список «красных зон» с индексом риска прерывания терапии, карточка (графики шкал, сон, настроение, дозы, алерты, план безопасности, ИИ-саммари), живые алерты, подключение пациента с кодом и QR, аудит доступа. |
+| **Веб-приложение пациента** | Вход через ЕМИАС (демо: код от врача), анонимный вход или демо-вход в один клик (у каждого посетителя своя копия пациента). Экраны: сегодня, лекарства, чат, врач, профиль. |
+| **Демо-консоль** | Виртуальные часы (13 дней терапии за 3 минуты), проигрывание сценария, сплит-экран, тумблер начального теста. |
+| **Инфраструктура** | Docker compose, CI (ruff + pytest, около 1060 тестов), CD: мерж в `dev` выкатывает стенд. |
+
+## 3. Развёртывание и запуск
+
+**Требования:** только Docker с Compose v2 — всё остальное собирается в контейнерах. Для разработки — Python 3.12 и [uv](https://docs.astral.sh/uv/), Node 20.
+
+```bash
+cp .env.example .env          # по умолчанию офлайн: LLM_PROVIDER=mock, агент выключен
+make up                       # postgres + API (:8000) + кабинет врача/демо-консоль (:8501) + приложение пациента (:8080)
+make seed                     # синтетический кабинет: врач, ветеран Андрей и 7 пациентов
+```
+
+- API и Swagger: http://localhost:8000/docs
+- Кабинет врача и демо-консоль: http://localhost:8501
+- Веб-приложение пациента: http://localhost:8080 (собирается внутри Docker, Node на машине не нужен; для разработки — `cd web && npm run dev`, :5173)
+
+**Включить чат-агента** в `.env`:
+- `CHAT_AGENT=gigachat` + `GIGACHAT_CREDENTIALS=…` — целевой вариант для РФ;
+- или `CHAT_AGENT=openrouter` + `OPENROUTER_API_KEY=…` (+ `OPENROUTER_PROXY_URL` с сервера в РФ) — только для прототипа;
+- проверить из терминала: `uv run python -m app.llm.agent "Можно мне ибупрофен?" --persona andrey`.
+
+**Демо-переключатели:**
+- `SCREENING_ENABLED` — начальный тест (дублируется тумблером в демо-консоли);
+- `CRISIS_BLOCKING` — клинический протокол: кризис блокирует диалог до подтверждения врача;
+- `DEMO_CLOCK` — виртуальные часы;
+- `INTAKE_DEMO` — опросники intake на синтетике.
+
+**Сервер:** `infra/deploy/` — override compose, nginx-vhost, `deploy.sh`. Выкатка на стенд — `.gitea/workflows/deploy.yml`.
+
+**Разработка:** `uv sync`, `make lint test`.
+
+## 4. Входные и выходные данные, API
+
+**Входные данные:**
+- сообщения и ответы кнопками пациента;
+- назначения (мок МИС, `data/meds/mis_prescriptions.json`);
+- карта ЕМИАС (мок, `data/emias/records.yaml`);
+- выдержки ГРЛС (`data/meds/grls/`);
+- клинические рекомендации (`config/rag/corpus/`).
+
+**Выходные данные:** ответы пациенту, баллы шкал, алерты врачу, карточка пациента, журнал аудита.
+
+Все клинические конфиги (шкалы, пороги, правила, тексты, лексикон) лежат в `config/` и правятся без кода.
+
+Основные эндпоинты (`/api/v1`, полная спецификация — `/docs`):
+
+| Метод и путь | Назначение |
+|---|---|
+| `POST /dialog/{token}/events` | событие диалога: текст, кнопка, start → ответы помощника |
+| `POST /dialog/{token}/events/stream` | то же в виде SSE: шаги агента, текст по мере генерации, финальные ответы |
+| `GET /dialog/{token}/outbox` | системные сообщения: напоминания, чек-ины |
+| `POST /auth/emias`, `/auth/emias/demo`, `/auth/demo/fresh`, `/auth/anonymous` | входы веб-приложения |
+| `GET /me/{token}/today` | план на сегодня: дозы, отметки за неделю, визит |
+| `GET /doctors/{id}/patients` | список пациентов врача с цветом и индексом риска |
+| `GET /patients/{token}?doctor_id=` | карточка пациента (пишется в аудит) |
+| `GET /doctors/{id}/alerts`, `POST /alerts/{id}/ack` | алерты и их подтверждение |
+| `POST /patients/connect` | подключение пациента: назначение и одноразовый код |
+| `GET/POST /demo/clock…`, `/demo/reset`, `/demo/settings` | демо-управление |
+
+## 5. Ограничения и особенности медицинской тематики
+
+- **Только синтетические данные.** Персоны, карты ЕМИАС, назначения и выдержки ГРЛС вымышлены или пересказаны для прототипа и подлежат сверке.
+- **Не медицинское изделие в текущем виде.** Сервис не ставит диагнозы и не меняет терапию. Пороги шкал, тексты и лексикон — стартовые значения из литературы, их утверждает психиатр (`docs/clinical/`).
+- **152-ФЗ:** отдельное согласие на обработку данных о здоровье, две раздельные БД (identity и clinical) со связью через HMAC-токен, в LLM уходит только обезличенный текст, каждое обращение врача пишется в аудит.
+- **Зарубежные сервисы.** OpenRouter используется временно и только на синтетике. Для реальных пациентов нужен GigaChat в РФ-контуре.
+- **Безопасность.** Кризисный ответ фиксированный и выдаётся без LLM. Пропуск кризиса недопустим, ложные срабатывания допустимы. Демо-режим (кризис без блокировки) не равен клиническому протоколу: в проде `CRISIS_BLOCKING=true`.
+- **Ответы LLM** могут слегка различаться от раза к разу. Фактура берётся только из инструментов, а запрещённые формулировки (диагнозы, дозы, «отмените») фильтруются.
+
+## Структура репозитория
+
+```
+app/            бэкенд: core (privacy, safety, шкалы, правила, часы), dialog (движок и сценарии), llm (шлюз, RAG,
+                чат-агент), db, api, scheduler, integrations (моки МИС/ЕМИАС/ГРЛС, уведомления), services
+dashboard/      кабинет врача и демо-консоль (Streamlit)
+web/            веб-приложение пациента (React + Vite)
+config/         клинические конфиги: шкалы, правила, модули, тексты, лексикон, RAG-корпус
+data/           синтетика: сценарий демо, карты ЕМИАС, назначения, выдержки ГРЛС
+infra/          postgres init, деплой на сервер
+tests/          pytest (около 1060 тестов)
+docs/           спецификация, ADR (DECISIONS.md), клинические материалы, концепция и CJM, регламент
+```
+
+Подробнее: [`docs/info/SPEC.md`](docs/info/SPEC.md) · [`docs/DECISIONS.md`](docs/DECISIONS.md) · [`docs/info/CLINICAL_MODULES.md`](docs/info/CLINICAL_MODULES.md) · [`docs/source/`](docs/source/) (кейс, концепция, CJM).

app/api/doctor_views.py

@@ -34,6 +34,7 @@ from app.db import repo
 from app.db.clinical import Alert, DailyCheckin, Dose, PatientProfile, Prescription, SafetyPlan, ScaleResult
 from app.db.identity import Patient
 from app.integrations.base import Prescription as PrescriptionOut
+from app.services.anonymous_care import display_name as anonymous_name
 
 COLOR_ORDER: dict[str, int] = {"red": 0, "yellow": 1, "grey": 2, "green": 3}
 _MODULES = frozenset(get_args(Module))
@@ -96,6 +97,7 @@ class PatientData:
     doses: Sequence[Dose]
     prescriptions: Sequence[Prescription]
     alerts: Sequence[Alert]  # newest first
+    contact_requested_at: datetime | None = None
 
 
 async def load_patient_data(session: AsyncSession, profile: PatientProfile) -> PatientData:
@@ -104,6 +106,9 @@ async def load_patient_data(session: AsyncSession, profile: PatientProfile) -> P
     rx = (
         await session.scalars(select(Prescription).where(Prescription.token == token).order_by(Prescription.id))
     ).all()
+    dialog = await repo.get_dialog_state(session, token)
+    requests = (dialog.data.get("contact_requests") or []) if dialog else []
+    contact_at = datetime.fromisoformat(requests[-1]["ts"]) if requests else None
     return PatientData(
         profile=profile,
         checkins=await repo.list_checkins(session, token),
@@ -111,6 +116,7 @@ async def load_patient_data(session: AsyncSession, profile: PatientProfile) -> P
         doses=doses,
         prescriptions=rx,
         alerts=await repo.list_alerts(session, token=token),
+        contact_requested_at=contact_at,
     )
 
 
@@ -228,6 +234,10 @@ def _last_event(d: PatientData, now: datetime) -> str:
     alert = d.alerts[0] if d.alerts else None
     scales = [r for r in d.results if r.complete and r.scale in _SCALE_LABELS]
     scale = scales[-1] if scales else None
+    if d.contact_requested_at is not None and all(
+        d.contact_requested_at >= ts for ts in [x.ts for x in (alert, scale) if x is not None]
+    ):
+        return "Пациент просит врача связаться"
     if scale is not None and (alert is None or scale.ts > alert.ts):
         day = _therapy_day(d.profile, _local(scale.ts, now))
         return f"{_SCALE_LABELS[scale.scale]} = {scale.total} (день {day})"
@@ -249,7 +259,7 @@ def list_item(d: PatientData, patient: Patient | None, now: datetime) -> Patient
     rx = _active_rx(d, today)
     return PatientListItem(
         token=p.token,
-        display_name=patient.display_name if patient else "Пациент",
+        display_name=patient.display_name if patient else anonymous_name(p.token),
         age=_age(patient, p, today),
         modules=[cast(Module, m) for m in p.modules or [] if m in _MODULES],
         drug=_drug_label(rx) if rx else None,

app/api/fixtures.py

@@ -16,6 +16,7 @@ from app.api.schemas import (
     PatientListItem,
     SeriesPoint,
 )
+from app.core.safety import fill_phones
 from app.integrations.base import Prescription
 
 TZ = ZoneInfo("Europe/Moscow")
@@ -24,7 +25,8 @@ DOCTOR_ID = "doc-smirnova"
 
 # Synthetic personas for the demo console and the demo ЕМИАС sign-in of the patient web app.
 DEMO_PERSONAS: list[DemoPersona] = [
-    DemoPersona(token="tok-andrey", display_name="Андрей С.", therapy_day=13, scenario="data/synthetic/andrey.jsonl"),
+    # ЕМИАС demo sign-in shows this persona as «Никита В.» (the one-click demo on the welcome screen is «Андрей С.»).
+    DemoPersona(token="tok-andrey", display_name="Никита В.", therapy_day=13, scenario="data/synthetic/andrey.jsonl"),
 ]
 
 
@@ -262,7 +264,7 @@ def andrey_card() -> PatientCard:
             "Что помогает: прогулка с собакой, душ",
             "Люди и места: брат Олег, спортзал",
             "Кому звонить: жена, сослуживец Миша",
-            "Специалисты: д-р Смирнова, 8-800-2000-122, 112",
+            fill_phones("Специалисты: д-р Смирнова, {phone:helpline}, 112"),
             "Безопасная среда: ружьё хранится у брата",
         ],
         summary="За 13 дней: приверженность 12/13, эпизод намерения бросить на 5-й день (снят после звонка врача). "

app/api/patient_web.py

@@ -7,9 +7,9 @@ exactly like the web console, and consent, onboarding, scales and check-ins come
 Two ways in:
 - **ЕМИАС** (demo): the doctor's one-time connect code via `EmiasMock`, or a synthetic demo persona. The
   patient is linked to the attending doctor, who sees the data on the dashboard.
-- **Anonymous**: a random pseudonymous token and nothing else — no identity record, no doctor. The engine
-  supports tokens without an identity record (consent is kept in the dialog state), so no personal data is
-  stored at all. Crisis safety still works: the fixed crisis script does not depend on the doctor.
+- **Anonymous**: a random pseudonymous token, no identity record, and access to the prototype care team.
+  Clinical answers are stored on the server and visible to the assigned doctor under a pseudonym.
+  Consent is kept in the dialog state. Existing doctor-less profiles require explicit care opt-in.
 
 The token returned here is the session credential of the prototype. Production needs short-lived sessions
 issued by the server instead of handing out the pseudonymous token itself.
@@ -17,17 +17,19 @@ issued by the server instead of handing out the pseudonymous token itself.
 
 import logging
 import secrets
+import time as monotonic_time
 from collections import defaultdict
 from datetime import date, datetime, time, timedelta
 from typing import Any, Literal
 from zoneinfo import ZoneInfo
 
-from fastapi import APIRouter, HTTPException
+from fastapi import APIRouter, HTTPException, Request
 from pydantic import BaseModel, ConfigDict, Field
 from sqlalchemy import select
 
 from app.api import fixtures
 from app.api.schemas import DemoPersona
+from app.config import get_settings
 from app.core import clock
 from app.core.instrument_catalog import Catalog, catalog
 from app.core.intake_questions import DIARY_KINDS
@@ -35,12 +37,16 @@ from app.core.intake_runtime import intake_enabled, load_runtime
 from app.core.interpretation import texts as results_texts
 from app.core.meds_reference import find_drug, indication_for, load_reference
 from app.core.safety import CrisisResource, crisis_reply
-from app.db import repo
+from app.db import identity_repo, repo
+from app.db.audit import audit
 from app.db.clinical import DailyCheckin, Dose, Prescription
 from app.db.identity import TOKEN_LEN
 from app.db.intake import QuestionnaireRecord
-from app.db.session import get_clinical_session
+from app.db.session import get_clinical_session, get_identity_session
 from app.integrations.emias_mock import CODE_MAX_LEN, EmiasMock
+from app.llm.stt import InvalidAudio, SttError, SttNoSpeech, get_stt, wav_seconds
+from app.services import anonymous_care
+from app.services.garden import Garden, garden_for
 
 log = logging.getLogger(__name__)
 
@@ -79,14 +85,14 @@ async def _attach_doctor(token: str, doctor_id: str) -> None:
             await repo.upsert_profile(s, token=token, doctor_id=doctor_id)
 
 
-async def _seed_demo_persona(token: str) -> None:
+async def _seed_demo_persona(token: str, source: str | None = None) -> None:
     """Synthetic clinical start for a demo persona (no `make seed` yet): profile, prescription, today's doses.
 
     Idempotent: only missing pieces are created. Data mirror the dashboard fixtures, so the patient app and the
     doctor's card tell the same story.
     """
-    card = fixtures.card(token)
-    item = next((p for p in fixtures.PATIENTS if p.token == token), None)
+    card = fixtures.card(source or token)
+    item = next((p for p in fixtures.PATIENTS if p.token == (source or token)), None)
     if card is None or item is None:
         return
     now = clock.now()
@@ -146,26 +152,142 @@ async def auth_emias_demo(body: EmiasDemoIn) -> AuthOut:
     persona = next((p for p in fixtures.DEMO_PERSONAS if p.token == body.persona_token), None)
     if persona is None:
         raise HTTPException(404, "persona not found")
+    if persona.token == DEMO_SOURCE:
+        # Every visitor gets a private copy (never the shared account); this entry starts from consent and the
+        # initial questionnaires (subject to the screening switch), unlike the one-click chat demo.
+        return await _fresh_demo_copy(start_in_chat=False)
     await _seed_demo_persona(persona.token)
     log.info("web auth: emias demo tok=%s", persona.token[:6])
     return AuthOut(token=persona.token, mode="emias", display_name=persona.display_name)
 
 
+DEMO_SOURCE = "tok-andrey"
+ONE_CLICK_NAME = "Андрей С."  # «Демо: войти как Андрей»; the ЕМИАС demo persona is named in fixtures (Никита)
+
+
+@router.post("/auth/demo/fresh", response_model=AuthOut)
+async def auth_demo_fresh() -> AuthOut:
+    return await _fresh_demo_copy(start_in_chat=True)
+
+
+async def _fresh_demo_copy(*, start_in_chat: bool) -> AuthOut:
+    """Presentation entry: a fresh private copy of veteran Andrey for every sign-in — consent given, prescription
+    and ЕМИАС record in place, dialog resting in the chat. Nobody's crisis or half-done questionnaire leaks into
+    another viewer's demo. Synthetic data only."""
+    token = f"{DEMO_SOURCE}-{secrets.token_hex(4)}"
+    now = clock.now()
+    persona = next(p for p in fixtures.DEMO_PERSONAS if p.token == DEMO_SOURCE)
+    name = persona.display_name if not start_in_chat else ONE_CLICK_NAME
+    snils = f"000-{secrets.randbelow(10**3):03d}-{secrets.randbelow(10**3):03d} {secrets.randbelow(100):02d}"
+    async with get_identity_session() as s:
+        await identity_repo.create_patient(
+            s,
+            token=token,
+            display_name=f"{name} (демо)",
+            snils=snils,
+            clinic_id="demo",
+            doctor_id=fixtures.DOCTOR_ID,
+            now=now,
+            birth_year=now.year - 34,
+            sex="m",
+        )
+        if start_in_chat:
+            await identity_repo.record_consent(s, token=token, version="demo", now=now)
+    await _seed_demo_persona(token, source=DEMO_SOURCE)
+    async with get_clinical_session() as s:
+        await repo.upsert_profile(s, token=token, sex="m", birth_year=now.year - 34)
+        if start_in_chat:
+            consent = {"consent": True, "consent_ver": "demo", "consent_at": now.isoformat(), "revoked_at": None}
+            await repo.set_dialog_state(s, token=token, state="idle", data={**consent, "adult": True}, now=now)
+    log.info("web auth: demo fresh tok=%s", token[:14])
+    return AuthOut(token=token, mode="emias", display_name=name)
+
+
 @router.post("/auth/anonymous", response_model=AuthOut)
 async def auth_anonymous(body: AnonymousIn | None = None) -> AuthOut:
-    """Anonymous sign-in: a fresh random token; no identity record, no doctor.
+    """Anonymous sign-in: a fresh random token; care without an identity record.
 
-    The clinical side gets a doctor-less pseudonymous profile (the dialog API serves known profiles only). It holds
-    no personal data — only the module of the chosen `focus`; the stage 1 screening then sets the modules itself.
+    The clinical side gets a pseudonymous profile assigned to the prototype care team. It holds
+    no identity fields. Legacy clients may supply `focus`; the common stage 1 screening sets modules itself.
     """
     token = secrets.token_hex(TOKEN_LEN // 2)
     focus = body.focus if body is not None else None
     async with get_clinical_session() as s:
-        await repo.upsert_profile(s, token=token, doctor_id="", modules=[_FOCUS_MODULE[focus]] if focus else [])
+        await repo.upsert_profile(
+            s, token=token, doctor_id=anonymous_care.doctor_id(), modules=[_FOCUS_MODULE[focus]] if focus else []
+        )
     log.info("web auth: anonymous tok=%s", token[:6])
     return AuthOut(token=token, mode="anonymous")
 
 
+@router.get("/patient/capabilities")
+async def patient_capabilities() -> dict[str, bool]:
+    """Expose availability without leaking clinical policy or server configuration."""
+    return {"primary_screening": intake_enabled(), "voice_input": get_stt(get_settings()) is not None}
+
+
+class SttOut(BaseModel):
+    text: str
+    provider: str
+
+
+# 16 kHz mono 16-bit PCM = 32 000 bytes/s, + header slack.
+_WAV_BYTES_PER_SECOND = 32_000
+
+
+@router.post("/me/{token}/stt", response_model=SttOut)
+async def transcribe_voice(token: str, request: Request) -> SttOut:
+    """Voice → text for the patient to check and edit (#116). Body: raw `audio/wav`, 16 kHz mono.
+
+    Only transcribes: nothing is written to the dialog and the agent is not called — the patient sends the
+    checked text as an ordinary message, where safety and `scrub()` apply. The audio is neither stored nor logged.
+    """
+    settings = get_settings()
+    stt = get_stt(settings)
+    if stt is None:
+        raise HTTPException(404, "voice input is off")
+    async with get_clinical_session() as s:
+        if await repo.get_profile(s, token) is None:
+            raise HTTPException(404, "patient not found")
+    max_bytes = (settings.stt_max_seconds + 2) * _WAV_BYTES_PER_SECOND + 4096
+    declared = request.headers.get("content-length")
+    if declared and declared.isdigit() and int(declared) > max_bytes:
+        raise HTTPException(413, "recording too long")
+    audio = await request.body()
+    if len(audio) > max_bytes:
+        raise HTTPException(413, "recording too long")
+    try:
+        seconds = wav_seconds(audio)
+    except InvalidAudio as exc:
+        raise HTTPException(415, "expected audio/wav") from exc
+    if seconds > settings.stt_max_seconds + 1:
+        raise HTTPException(413, "recording too long")
+    started = monotonic_time.monotonic()  # latency only, not domain time (that is app.core.clock)
+    try:
+        text = await stt.transcribe(audio, "wav")
+    except SttNoSpeech as exc:
+        raise HTTPException(422, "no speech recognised") from exc
+    except SttError as exc:
+        log.warning("stt failed tok=%s provider=%s: %s", token[:6], stt.name, exc)
+        raise HTTPException(503, "speech recognition unavailable") from exc
+    ms = int((monotonic_time.monotonic() - started) * 1000)
+    log.info("stt tok=%s provider=%s secs=%.1f ms=%d chars=%d", token[:6], stt.name, seconds, ms, len(text))
+    return SttOut(text=text, provider=stt.name)
+
+
+@router.post("/me/{token}/doctor/connect")
+async def connect_anonymous_doctor(token: str) -> dict[str, bool]:
+    """Explicit opt-in for old doctor-less profiles; never reassign an existing doctor."""
+    async with get_clinical_session() as s:
+        profile = await repo.get_profile(s, token)
+        if profile is None:
+            raise HTTPException(404, "patient not found")
+        if not profile.doctor_id:
+            profile.doctor_id = anonymous_care.doctor_id()
+            await audit(s, actor=token, action="anonymous_care_opt_in", token=token, ts=clock.now())
+    return {"has_doctor": True}
+
+
 class DoseItem(BaseModel):
     id: int
     drug: str
@@ -305,6 +427,18 @@ async def my_today(token: str) -> TodayOut:
     )
 
 
+@router.get("/me/{token}/garden", response_model=Garden)
+async def my_garden(token: str) -> Garden:
+    """The patient's garden (#118): grows from medication marks, check-ins/diary and finished questionnaires.
+
+    Never shrinks and never depends on answers. Seen only by the patient — not on the dashboard, not in summaries.
+    """
+    async with get_clinical_session() as s:
+        if await repo.get_profile(s, token) is None:
+            raise HTTPException(404, "patient not found")
+        return await garden_for(s, token)
+
+
 class ResultFinding(BaseModel):
     code: str
     title: str

app/api/routes.py

@@ -8,9 +8,14 @@ disclosed — the dashboard polls every 2–3 s), `ack_alert` (in `repo.ack_aler
 `app/api/fixtures.py` is kept for tests and as the demo cabinet reference (doctor id).
 """
 
+import asyncio
+import json
+import logging
 import secrets
+from typing import Any
 
 from fastapi import APIRouter, HTTPException
+from fastapi.responses import StreamingResponse
 from sqlalchemy import select
 
 from app.api import demo as console
@@ -42,9 +47,12 @@ from app.db.identity import Patient
 from app.db.session import get_clinical_session, get_identity_session
 from app.dialog.contracts import Event, get_engine
 from app.integrations.base import integrations
+from app.llm.agent import stream as agent_stream
 from app.scheduler import jobs as scheduler
+from app.services.anonymous_care import display_name as anonymous_name
 from app.services.monitoring import refresh_status
 
+log = logging.getLogger(__name__)
 router = APIRouter(prefix="/api/v1")
 router.include_router(max_notifications.router)
 router.include_router(questionnaire_router)
@@ -66,7 +74,7 @@ async def _patients(tokens: list[str]) -> dict[str, Patient]:
 
 def _name(patients: dict[str, Patient], token: str) -> str:
     p = patients.get(token)
-    return p.display_name if p else "Пациент"
+    return p.display_name if p else anonymous_name(token)
 
 
 # ------------------------------------------------------------------ doctor dashboard
@@ -218,6 +226,49 @@ async def dialog_event(token: str, body: DialogIn) -> DialogOut:
     return DialogOut(replies=await get_engine().handle(token, body))
 
 
+@router.post("/dialog/{token}/events/stream", tags=["dialog"])
+async def dialog_event_stream(token: str, body: DialogIn) -> StreamingResponse:
+    """Same as /events, as Server-Sent Events: `status` (agent steps), `delta` (answer as it is written),
+    `reset`, then `done` with the final replies (identical to /events). `error` if the engine failed."""
+    await _require_patient(token)
+    queue: asyncio.Queue[tuple[str, Any]] = asyncio.Queue()
+    marker = agent_stream.STREAM.set(queue)  # copied into the task's context below
+    try:
+
+        async def run() -> None:
+            try:
+                replies = await get_engine().handle(token, body)
+                queue.put_nowait(("done", [r.model_dump() for r in replies]))
+            except Exception:  # noqa: BLE001 — reported to the client, which falls back to /events
+                log.exception("dialog stream failed")
+                queue.put_nowait(("error", None))
+
+        task = asyncio.create_task(run())
+    finally:
+        agent_stream.STREAM.reset(marker)
+
+    async def events():
+        try:
+            while True:
+                try:
+                    kind, payload = await asyncio.wait_for(queue.get(), timeout=15)
+                except TimeoutError:
+                    yield ": keepalive\n\n"
+                    continue
+                yield f"event: {kind}\ndata: {json.dumps(payload, ensure_ascii=False)}\n\n"
+                if kind in ("done", "error"):
+                    return
+        finally:
+            if not task.done():
+                await task
+
+    return StreamingResponse(
+        events(),
+        media_type="text/event-stream",
+        headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"},  # nginx: do not buffer SSE
+    )
+
+
 @router.get("/dialog/{token}/outbox", response_model=DialogOut, tags=["dialog"])
 async def dialog_outbox(token: str) -> DialogOut:
     """System-initiated replies (reminders, check-ins) waiting for the web console."""
@@ -287,3 +338,22 @@ async def demo_personas() -> list[DemoPersona]:
             )
         )
     return out
+
+
+@router.get("/demo/settings", tags=["demo"])
+async def get_demo_settings() -> dict[str, bool]:
+    """Presentation switches: `screening_enabled` — initial questionnaires after consent (off → straight to chat)."""
+    from app.core import demo_settings
+
+    async with get_clinical_session() as s:
+        return await demo_settings.get_all(s)
+
+
+@router.post("/demo/settings", tags=["demo"])
+async def set_demo_settings(body: dict[str, bool]) -> dict[str, bool]:
+    from app.core import demo_settings
+
+    async with get_clinical_session() as s:
+        if demo_settings.SCREENING in body:
+            await demo_settings.set_screening(s, bool(body[demo_settings.SCREENING]))
+        return await demo_settings.get_all(s)

app/channels/MAX_SETUP.md

@@ -0,0 +1,150 @@
+# Подключение будущего бота MAX
+
+Серверная часть готова к подключению. Сам бот ещё не создан, секреты не заполнены,
+реальная отправка не включена. Изменений интерфейса сайта в этой поставке нет.
+Интеграция уведомляет **самого пациента**; врачебные оповещения остаются в существующем канале проекта.
+
+## Что реализовано
+
+- Одноразовая ссылка для привязки аккаунта MAX к текущему пациенту сайта, срок по умолчанию 10 минут.
+- Подтверждение подключения, `/help`, `/stop`, отключение при остановке или удалении бота в MAX.
+- Нейтральные напоминания из существующего планировщика сайта: тесты, дневник, приём лекарств и другие события outbox.
+- Уведомление о завершении теста: сервер каждые примерно 30 секунд находит новые завершённые
+  `QuestionnaireRecord` и `ScaleResult`, начиная с момента подключения MAX. Незавершённые и старые тесты не отправляются.
+- Постоянная очередь, восстановление после перезапуска, ограниченные повторы при сетевых ошибках и ответах 429/5xx.
+- В сообщении только короткий текст и кнопка на сайт. Ответы, оценки, диагнозы, назначения и ключ входа на сайт
+  в MAX не передаются. Переписка с ботом не сохраняется и не отправляется в LLM.
+
+## 1. Создать бота — будущему владельцу
+
+1. Открыть [MAX для партнёров](https://business.max.ru/) и пройти регистрацию/проверку профиля.
+   Условия допуска зависят от типа владельца; актуальный порядок — в
+   [официальной инструкции подключения](https://dev.max.ru/help/platform_connection).
+2. Создать чат-бота, оформить описание и пройти требуемую MAX модерацию.
+3. В настройках бота получить токен и скопировать имя из его публичной ссылки `https://max.ru/ИМЯ`.
+4. Передать администратору сервера имя бота, а токен — через защищённое хранилище секретов.
+   Токен не нужен в переписке, MR или репозитории.
+
+## 2. Заполнить настройки сервера
+
+Нужны доступная identity/clinical БД проекта, публичный HTTPS-адрес сайта и HTTPS-адрес API на порту 443.
+Входящие POST от MAX должны доходить до API без браузерной CAPTCHA или страницы авторизации.
+Прокси должен передавать заголовок `X-Max-Bot-Api-Secret` и не журналировать тело webhook/Authorization.
+Дополнительные таблицы identity создаются существующим `init_db()` при запуске приложения.
+
+В `.env` сервера (вне Git) заполнить:
+
+```dotenv
+MAX_ENABLED=false
+MAX_BOT_TOKEN=
+MAX_BOT_USERNAME=
+MAX_WEBHOOK_SECRET=
+MAX_WEBHOOK_URL=https://api.example.org/api/v1/webhooks/max
+MAX_SITE_URL=https://patient.example.org/
+MAX_LINK_TTL_SECONDS=600
+# При необходимости — PEM-файл доверенного CA, доступный внутри контейнера:
+MAX_CA_BUNDLE=
+DEMO_CLOCK=false
+```
+
+- `MAX_BOT_USERNAME` — только имя, без `@`, домена или параметров ссылки.
+- `MAX_WEBHOOK_SECRET` — отдельная случайная строка, не токен бота: 32–256 символов из `A–Z a–z 0–9 _ -`.
+  Сгенерировать локально, например `python -c 'import secrets; print(secrets.token_urlsafe(32))'`, сохранить в секреты.
+- `MAX_SITE_URL` ведёт в обычный личный кабинет; не добавлять в адрес ключи доступа или данные пациента.
+- Исходящие запросы идут на `https://platform-api2.max.ru`, токен передаётся заголовком `Authorization`.
+  Верификация TLS всегда включена. Если системе нужен корневой сертификат Минцифры, установить его по
+  официальной инструкции MAX и указать PEM-путь в `MAX_CA_BUNDLE`; не отключать проверку TLS.
+- Переменные должны попасть именно в окружение API-контейнера/процесса, а не только панели сайта.
+  В текущем Compose сервис `app` уже получает `.env` через `env_file`.
+  После изменения окружения пересоздать процесс/контейнер API. Новые зависимости не требуются.
+
+## 3. Проверить и включить
+
+В каталоге проекта, в том же окружении, где работает API:
+
+```sh
+uv run python -m app.channels.max_setup check
+```
+
+Команда проверит токен и совпадение имени бота. Сообщений пользователям она не отправляет.
+Затем установить `MAX_ENABLED=true`, перезапустить API и выполнить:
+
+```sh
+uv run python -m app.channels.max_setup subscribe
+uv run python -m app.channels.max_setup status
+```
+
+`subscribe` регистрирует только заданный webhook с секретом и событиями
+`bot_started`, `message_created`, `bot_stopped`, `dialog_removed`. Старые чужие подписки автоматически
+не удаляются: при переносе сервера владелец должен проверить их отдельно через MAX API.
+`status` проверяет регистрацию адреса; это ещё не проверка доступности сервера из MAX.
+Команды не печатают токены, секрет webhook или сырые ответы платформы.
+
+При запуске через Compose вместо `uv run python ...` использовать
+`docker compose exec app python -m app.channels.max_setup check` (также `subscribe` и `status`).
+После изменения `.env` пересоздать API: `docker compose up -d --build --force-recreate app`.
+
+Запускать один процесс API с MAX на один токен бота: его worker отправляет до одного сообщения в секунду.
+БД защищает очередь от параллельного захвата, но общий лимит скорости нескольких процессов не координируется.
+Планировщик и MAX worker запускаются вместе с FastAPI, отдельный процесс бота не нужен.
+При `MAX_ENABLED=false` сетевых вызовов MAX и фонового worker нет; сайт продолжает обычную работу.
+
+## API для разработчика сайта
+
+В проекте используется ключ текущего пациента, полученный при входе. Для этих ручек передавать его
+как `Authorization: Bearer <patient_token>`; это существующая авторизация прототипа, не новая система SSO.
+Не брать token другого пациента и не публиковать демонстрационные учётные записи для реального использования.
+
+| Метод и путь | Запрос | Ответ |
+|---|---|---|
+| `GET /api/v1/me/notifications/max` | Authorization | `available`, `connected`, `last_delivery` |
+| `POST /api/v1/me/notifications/max/link` | Authorization; JSON `{"consent": true}` | `url`, `expires_at` |
+| `DELETE /api/v1/me/notifications/max` | Authorization | состояние с `connected: false` |
+| `POST /api/v1/webhooks/max` | MAX JSON + `X-Max-Bot-Api-Secret` | `{"success": true}` после записи события |
+
+Поток подключения: пациент явно соглашается на уведомления → сайт получает `url` → открывает ссылку MAX →
+пользователь запускает бота → webhook подтверждает привязку → повторный GET показывает `connected: true`.
+Ссылка одноразовая; новая ссылка отменяет предыдущую. Изначальный ключ пациента не входит в URL MAX.
+Для привязки и доставки дополнительно нужны действующее согласие и подтверждение совершеннолетия в состоянии пациента.
+Активную привязку нельзя молча переназначить другому аккаунту: сначала отключить её.
+
+Коды ошибок: 401 — неизвестный/отсутствующий ключ пациента; 403 — нет согласия или неверный секрет webhook;
+422 — неверный запрос; 413 — webhook больше 64 КиБ; 503 — MAX выключен или хранилище события недоступно.
+Повтор webhook с тем же идентификатором события не создаёт новых отправок. Групповые чаты игнорируются.
+
+`last_delivery`: `queued`, `sending`, `sent`, `failed`, `cancelled` или `null`.
+`sent` означает принятие сообщения API MAX, а не прочтение человеком.
+
+## Пробное подключение после создания бота
+
+1. Войти на сайте синтетическим пациентом и дать необходимые согласия.
+2. Получить ссылку через POST выше (Swagger `/docs` или будущий интерфейс сайта), перейти в MAX и запустить бота.
+3. Убедиться, что GET показывает `connected: true` и приходит подтверждение.
+4. Завершить тест на сайте; примерно через 30 секунд плюс очередь должно прийти сообщение без результатов.
+5. Дождаться события планировщика и проверить нейтральное напоминание с кнопкой на сайт.
+6. Отправить `/stop`: GET должен показать `connected: false`, ожидающие уведомления отменяются.
+7. Повторить подключение новой ссылкой, проверить DELETE через API и остановку бота через интерфейс MAX.
+
+При недоставке проверить настройки и подписку, доступность HTTPS, состояние очереди и коды ошибок в логах.
+`http_401`/`http_403` требуют проверки токена/прав; `http_429` — ожидания; `transport` — сети/TLS;
+`failed` после шести попыток требует вмешательства администратора. Детали ответов и идентификаторы в логи не выводятся.
+Отключение отменяет очередь; уже принятый MAX запрос отозвать невозможно.
+
+## Гарантии и ограничения
+
+Очередь и подтверждения событий сохраняются в identity БД. Клинические записи остаются в clinical БД.
+Для событий завершения тестов БД повторно просматривается после перезапуска; ключи уведомлений предотвращают
+повторную постановку. Напоминания используют существующий механизм outbox/повторов планировщика.
+При временных ошибках — до 6 попыток с увеличением задержки и учётом `Retry-After`; зависшая отправка получает
+новую аренду через 60 секунд. Потеря процесса между принятием сообщения MAX и записью успеха в БД может
+привести к повторному сообщению: MAX не предоставляет ключ идемпотентности отправки.
+Не удалять строки дедупликации/очереди без согласованной политики хранения. При большом количестве пациентов
+нужны пакетное сканирование завершённых тестов и общий ограничитель скорости; текущая реализация рассчитана на MVP.
+
+Проверки разработчика: `make lint test`; тесты MAX работают с синтетикой и подменённым HTTP-транспортом.
+Реальную доставку подтвердит только пробное подключение после выдачи токена и настройки публичного сервера.
+
+Официальные контракты: [отправка сообщений](https://dev.max.ru/docs-api/methods/POST/messages),
+[подписка webhook](https://dev.max.ru/docs-api/methods/POST/subscriptions),
+[события](https://dev.max.ru/docs-api/objects/Update),
+[ссылки запуска бота](https://dev.max.ru/docs/chatbots/bots-coding/prepare).

app/channels/max_bot.py

@@ -0,0 +1,224 @@
+"""Notifications-only MAX adapter; no patient text enters the dialog engine or an LLM.
+
+Webhook transactions persist receipts and replies together, then return immediately.
+The worker delivers from identity DB with leases and bounded retries. MAX has no send
+idempotency key: a crash after API acceptance and before commit can duplicate a message.
+"""
+
+import asyncio
+import contextlib
+import json
+import logging
+from functools import lru_cache
+
+import yaml
+from pydantic import BaseModel, Field
+from sqlalchemy import select
+from sqlalchemy.exc import IntegrityError
+
+from app.channels.max_client import MaxApiError, MaxClient
+from app.config import get_settings
+from app.core import clock
+from app.db.max_notifications import MaxBinding, MaxReceipt
+from app.db.session import get_identity_session
+from app.dialog.contracts import Reply
+from app.integrations import max_store as store
+from app.integrations.base import Outbox, integrations
+
+log = logging.getLogger(__name__)
+
+
+class Texts(BaseModel):
+    connected: str = Field(max_length=4000)
+    not_connected: str = Field(max_length=4000)
+    invalid: str = Field(max_length=4000)
+    stopped: str = Field(max_length=4000)
+    help: str = Field(max_length=4000)
+    notification: str = Field(max_length=4000)
+    completed: str = Field(max_length=4000)
+    open_site: str = Field(max_length=128)
+
+
+@lru_cache
+def texts() -> Texts:
+    path = get_settings().config_dir / "texts" / "max.yaml"
+    return Texts.model_validate(yaml.safe_load(path.read_text(encoding="utf-8")))
+
+
+def _event(event: dict) -> tuple[str, int, str, str] | None:
+    """Normalize only documented private-user events; unknown/malformed events are ignored."""
+    kind = event.get("update_type")
+    stamp = event.get("timestamp")
+    if type(stamp) is not int or stamp < 0:
+        return None
+    message_id = ""
+    if kind == "message_created":
+        message = event.get("message")
+        if not isinstance(message, dict):
+            return None
+        recipient, body = message.get("recipient"), message.get("body")
+        if not isinstance(recipient, dict) or recipient.get("chat_type") != "dialog" or not isinstance(body, dict):
+            return None
+        user, text, message_id = message.get("sender"), body.get("text"), body.get("mid")
+        if not isinstance(message_id, str) or not message_id or not isinstance(text, str):
+            return None
+    elif kind in ("bot_started", "bot_stopped", "dialog_removed"):
+        user, text = event.get("user"), event.get("payload") or ""
+    else:
+        return None
+    if not isinstance(user, dict) or user.get("is_bot") is not False:
+        return None
+    uid = user.get("user_id")
+    if type(uid) is not int or not 0 < uid < 2**63 or not isinstance(text, str) or len(text) > 4000:
+        return None
+    # Names, profile details and raw text are neither persisted nor logged.
+    receipt = store.digest(f"{kind}:{uid}:{message_id or stamp}")
+    return kind, uid, text.strip(), receipt
+
+
+async def handle_update(event: dict) -> None:
+    if not get_settings().max_enabled or (parsed := _event(event)) is None:
+        return
+    kind, user_id, text, receipt = parsed
+    now = clock.now()
+    try:
+        async with get_identity_session() as session:
+            if await session.get(MaxReceipt, receipt):
+                return
+            session.add(MaxReceipt(digest=receipt, received_at=now))
+            await session.flush()
+            binding = await session.scalar(select(MaxBinding).where(MaxBinding.user_id == user_id))
+            reply_token = None
+            command, _, payload = text.partition(" ")
+            if kind in ("bot_stopped", "dialog_removed") or command.lower() == "/stop":
+                if binding:
+                    await store.disconnect(session, binding.token)
+                # When MAX itself stops/deletes the bot, do not attempt a reply.
+                if kind != "message_created":
+                    return
+                reply = "stopped"
+            elif kind == "bot_started" or command.lower() == "/start":
+                code = text if kind == "bot_started" else payload.strip()
+                if code:
+                    token = await store.code_token(session, code, now) if len(code) <= 128 else None
+                    linked = (
+                        token and await store.active_patient(token) and await store.bind(session, code, user_id, now)
+                    )
+                    reply = "connected" if linked else "invalid"
+                    reply_token = token if linked else None
+                else:
+                    reply = "connected" if binding and binding.enabled else "not_connected"
+                    reply_token = binding.token if binding and binding.enabled else None
+            else:
+                reply = "help"
+            await store.enqueue(session, key=receipt, user_id=user_id, token=reply_token, kind=reply, now=now)
+    except IntegrityError:
+        # A concurrent delivery of the same webhook can race on its receipt primary key.
+        async with get_identity_session() as session:
+            if await session.get(MaxReceipt, receipt) is None:
+                raise
+
+
+class MaxOutbox:
+    def __init__(self, fallback: Outbox):
+        self.fallback = fallback
+
+    async def push(self, token: str, replies: list[Reply]) -> None:
+        if replies and get_settings().max_enabled and await store.active_patient(token):
+            async with get_identity_session() as session:
+                binding = await session.get(MaxBinding, token)
+                if binding and binding.enabled:
+                    # Hash the event for retry deduplication; never store/transmit its clinical text.
+                    event = json.dumps([r.model_dump() for r in replies], sort_keys=True, ensure_ascii=False)
+                    key = store.digest(f"{token}:{binding.connected_at.isoformat()}:{clock.today()}:{event}")
+                    await store.enqueue(
+                        session, key=key, user_id=binding.user_id, token=token, kind="notification", now=clock.now()
+                    )
+        await self.fallback.push(token, replies)
+
+
+async def deliver_one(client: MaxClient) -> bool:
+    async with get_identity_session() as session:
+        row = await store.claim(session, clock.now())
+    if row is None:
+        return False
+    if row.token is not None:
+        async with get_identity_session() as session:
+            binding = await session.get(MaxBinding, row.token)
+            valid = (
+                binding
+                and binding.enabled
+                and binding.user_id == row.user_id
+                and binding.connected_at <= row.created_at
+            )
+        if not valid or not await store.active_patient(row.token):
+            async with get_identity_session() as session:
+                await store.finish(session, row, state="cancelled", now=clock.now())
+            return True
+    settings, messages = get_settings(), texts()
+    try:
+        await client.send(
+            row.user_id, getattr(messages, row.kind), site_url=settings.max_site_url, button=messages.open_site
+        )
+    except MaxApiError as exc:
+        retry = exc.retryable and row.attempts < store.MAX_ATTEMPTS
+        delay = max(exc.retry_after, min(300, 5 * 2 ** (row.attempts - 1)))
+        async with get_identity_session() as session:
+            await store.finish(
+                session,
+                row,
+                state="queued" if retry else "failed",
+                now=clock.now(),
+                error=exc.code,
+                retry_seconds=delay if retry else 0,
+            )
+        log.warning("MAX delivery deferred=%s code=%s", retry, exc.code)
+    else:
+        async with get_identity_session() as session:
+            await store.finish(session, row, state="sent", now=clock.now())
+    return True
+
+
+async def _run(client: MaxClient) -> None:
+    tick = 0
+    while True:
+        try:
+            if tick % 30 == 0:
+                async with get_identity_session() as session:
+                    await store.collect_completed(session, clock.now())
+            await deliver_one(client)
+        except Exception as exc:
+            # No exception text: SQL/HTTP errors may contain identifiers or parameters.
+            log.error("MAX worker failed type=%s", type(exc).__name__)
+        tick += 1
+        # One send per second per process, below MAX's two-per-second per-dialog limit.
+        await asyncio.sleep(1)
+
+
+_client: MaxClient | None = None
+_task: asyncio.Task | None = None
+_wrapper: MaxOutbox | None = None
+
+
+async def start() -> None:
+    global _client, _task, _wrapper
+    if not get_settings().max_enabled or _task is not None:
+        return
+    texts()  # Validate templates before replacing a working outbox.
+    _client = MaxClient(get_settings())
+    _wrapper = MaxOutbox(integrations.outbox)
+    integrations.outbox = _wrapper
+    _task = asyncio.create_task(_run(_client), name="max-notifications")
+
+
+async def stop() -> None:
+    global _client, _task, _wrapper
+    if _task:
+        _task.cancel()
+        with contextlib.suppress(asyncio.CancelledError):
+            await _task
+    if _wrapper and integrations.outbox is _wrapper:
+        integrations.outbox = _wrapper.fallback
+    if _client:
+        await _client.close()
+    _client = _task = _wrapper = None

app/channels/max_client.py

@@ -0,0 +1,101 @@
+"""Small MAX Bot API client. Secrets stay in headers; failures never expose response bodies."""
+
+import logging
+import ssl
+
+import httpx
+
+from app.config import Settings
+
+API_URL = "https://platform-api2.max.ru"
+UPDATE_TYPES = ["bot_started", "message_created", "bot_stopped", "dialog_removed"]
+
+
+class _HideRecipient(logging.Filter):
+    def filter(self, record: logging.LogRecord) -> bool:
+        # httpx INFO normally logs query strings, which include a MAX user_id.
+        return "platform-api2.max.ru" not in record.getMessage()
+
+
+logging.getLogger("httpx").addFilter(_HideRecipient())
+
+
+class MaxApiError(Exception):
+    def __init__(self, code: str, *, retryable: bool = False, retry_after: float = 0):
+        self.code, self.retryable, self.retry_after = code, retryable, retry_after
+        super().__init__(code)
+
+
+class MaxClient:
+    def __init__(self, settings: Settings, *, transport: httpx.AsyncBaseTransport | None = None):
+        context = ssl.create_default_context()
+        if settings.max_ca_bundle:
+            context.load_verify_locations(cafile=str(settings.max_ca_bundle))
+        self.http = httpx.AsyncClient(
+            base_url=API_URL,
+            headers={"Authorization": settings.max_bot_token.get_secret_value()},
+            timeout=httpx.Timeout(5),
+            verify=context,
+            follow_redirects=False,
+            transport=transport,
+        )
+
+    async def close(self) -> None:
+        await self.http.aclose()
+
+    async def request(self, method: str, path: str, **kwargs) -> dict:
+        try:
+            response = await self.http.request(method, path, **kwargs)
+        except httpx.RequestError:
+            raise MaxApiError("transport", retryable=True) from None
+        if not response.is_success:
+            try:
+                delay = min(3600, max(0, float(response.headers.get("Retry-After", "0"))))
+            except ValueError:
+                delay = 0
+            raise MaxApiError(
+                f"http_{response.status_code}",
+                retryable=response.status_code in (408, 429) or response.status_code >= 500,
+                retry_after=delay,
+            )
+        try:
+            data = response.json()
+        except ValueError:
+            raise MaxApiError("invalid_response", retryable=True) from None
+        if not isinstance(data, dict):
+            raise MaxApiError("invalid_response", retryable=True)
+        if data.get("success") is False:
+            raise MaxApiError("rejected")
+        return data
+
+    async def send(self, user_id: int, text: str, *, site_url: str, button: str) -> None:
+        data = await self.request(
+            "POST",
+            "/messages",
+            params={"user_id": user_id, "disable_link_preview": "true"},
+            json={
+                "text": text,
+                "notify": True,
+                "attachments": [
+                    {
+                        "type": "inline_keyboard",
+                        "payload": {"buttons": [[{"type": "link", "text": button, "url": site_url}]]},
+                    }
+                ],
+            },
+        )
+        if not isinstance(data.get("message"), dict):
+            raise MaxApiError("missing_message", retryable=True)
+
+    async def subscribe(self, settings: Settings) -> None:
+        data = await self.request(
+            "POST",
+            "/subscriptions",
+            json={
+                "url": settings.max_webhook_url,
+                "update_types": UPDATE_TYPES,
+                "secret": settings.max_webhook_secret.get_secret_value(),
+            },
+        )
+        if data.get("success") is not True:
+            raise MaxApiError("subscription_not_confirmed")

app/channels/max_setup.py

@@ -0,0 +1,56 @@
+"""Operator commands for a future bot owner. Never runs as part of application startup."""
+
+import argparse
+import asyncio
+import sys
+
+from pydantic import ValidationError
+
+from app.channels.max_client import MaxApiError, MaxClient
+from app.config import Settings
+
+
+async def run(command: str) -> None:
+    # Validate the same configuration even while the deployed app is still disabled.
+    settings = Settings(max_enabled=True)
+    client = MaxClient(settings)
+    try:
+        bot = await client.request("GET", "/me")
+        if bot.get("is_bot") is not True:
+            raise MaxApiError("not_a_bot")
+        if bot.get("username") != settings.max_bot_username:
+            raise MaxApiError("bot_username_mismatch")
+        if command == "subscribe":
+            await client.subscribe(settings)
+            print("Webhook MAX зарегистрирован. Проверьте командой status, затем выполните пробное подключение.")
+        elif command == "status":
+            result = await client.request("GET", "/subscriptions")
+            subscriptions = result.get("subscriptions", [])
+            configured = any(isinstance(s, dict) and s.get("url") == settings.max_webhook_url for s in subscriptions)
+            print("Нужный webhook зарегистрирован." if configured else "Нужный webhook не найден.")
+            if not configured:
+                raise MaxApiError("webhook_missing")
+        else:
+            print("Токен принят MAX, имя бота совпадает с настройкой. Подписка не изменялась.")
+    finally:
+        await client.close()
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Подготовка MAX: параметры берутся только из .env/окружения.")
+    parser.add_argument("command", choices=("check", "subscribe", "status"))
+    args = parser.parse_args()
+    try:
+        asyncio.run(run(args.command))
+    except ValidationError:
+        print("Проверьте параметры MAX в .env по app/channels/MAX_SETUP.md.", file=sys.stderr)
+        return 1
+    except (MaxApiError, OSError) as exc:
+        code = exc.code if isinstance(exc, MaxApiError) else "tls_or_file_error"
+        print(f"Не удалось выполнить настройку MAX: {code}. Секреты и ответ сервера скрыты.", file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

app/channels/telegram.py

@@ -19,6 +19,7 @@ from typing import Protocol
 
 import yaml
 from aiogram import Bot, Dispatcher, F, Router
+from aiogram.client.session.aiohttp import AiohttpSession
 from aiogram.filters import CommandObject, CommandStart
 from aiogram.types import CallbackQuery, InlineKeyboardButton, InlineKeyboardMarkup, Message
 from pydantic import BaseModel
@@ -193,11 +194,28 @@ class TelegramChannel:
             await self._dispatch(chat_id, person.token, Event(type="start", params={"code": code}))
             return
         token = await self.directory.token_for(chat_id)
+        if token is None and get_settings().demo_clock:
+            token = await self._demo_copy(chat_id)
         if token is None:
             await self._say(chat_id, texts.not_connected)
             return
         await self._dispatch(chat_id, token, Event(type="start"))
 
+    async def _demo_copy(self, chat_id: int) -> str | None:
+        """Demo stand (DEMO_CLOCK): `/start` without a code gets a private copy of veteran Andrey (synthetic),
+        starting from consent and the intake — the same entry as the ЕМИАС demo button of the web app."""
+        from app.api.patient_web import _fresh_demo_copy
+
+        try:
+            token = (await _fresh_demo_copy(start_in_chat=False)).token
+        except Exception:
+            log.exception("start: demo copy failed")
+            return None
+        if not await self.directory.bind(chat_id, token):
+            return None
+        log.info("start: demo copy bound tok=%s", _short(token))
+        return token
+
     async def on_text(self, chat_id: int, text: str | None) -> None:
         texts = load_texts()
         token = await self.directory.token_for(chat_id)
@@ -272,6 +290,12 @@ _task: asyncio.Task[None] | None = None
 _prev_outbox: Outbox | None = None
 
 
+def make_bot(token: str) -> Bot:
+    """aiogram `Bot`, through `TG_PROXY_URL` when it is set."""
+    proxy = get_settings().tg_proxy_url
+    return Bot(token, session=AiohttpSession(proxy=proxy)) if proxy else Bot(token)
+
+
 async def start() -> None:
     """Start polling if TG_BOT_TOKEN is set; no-op otherwise."""
     global _bot, _dp, _task, _prev_outbox
@@ -280,7 +304,7 @@ async def start() -> None:
         if not token:
             log.info("telegram disabled: TG_BOT_TOKEN is empty")
         return
-    bot = Bot(token)
+    bot = make_bot(token)
     channel = TelegramChannel(bot)
     dp = Dispatcher()
     dp.include_router(build_router(channel))

app/config.py

@@ -14,7 +14,7 @@ ROOT_DIR = Path(__file__).resolve().parent.parent
 
 class Settings(BaseSettings):
     # Empty values in .env (e.g. `TG_BOT_TOKEN=`) fall back to defaults instead of failing validation.
-    model_config = SettingsConfigDict(env_file=".env", extra="ignore", env_ignore_empty=True)
+    model_config = SettingsConfigDict(env_file=".env", extra="ignore", env_ignore_empty=True, hide_input_in_errors=True)
 
     # Two separate databases (ADR-003): identity holds direct identifiers, clinical holds pseudonymised data.
     identity_db_url: str = "postgresql+asyncpg://identity_app:identity_app@localhost:5432/identity"
@@ -37,12 +37,22 @@ class Settings(BaseSettings):
     openrouter_api_key: str = ""
     openrouter_model: str = "google/gemini-2.5-flash-lite"
     openrouter_base_url: str = "https://openrouter.ai/api/v1"
+    openrouter_proxy_url: str = ""  # OpenRouter blocks RF IPs: the server goes through an HTTP proxy (as vobot does)
     gigachat_agent_model: str = "GigaChat-2-Pro"  # function calling quality matters more than price here
     chat_agent_timeout_s: float = 25.0
     chat_agent_max_steps: int = 5
 
+    # Voice input (#116): speech-to-text for the patient's messages. `auto` = OpenRouter if a key is set, else mock.
+    # OpenRouter STT is synthetic data only (ADR-014): the audio leaves the RF and cannot be scrubbed.
+    stt_provider: Literal["auto", "off", "mock", "openrouter"] = "auto"
+    stt_model: str = "google/gemini-2.5-flash"  # needs audio input
+    stt_timeout_s: float = 30.0
+    stt_max_seconds: int = Field(default=60, ge=5, le=300)
+
     # Telegram is used only for the prototype on synthetic data; prod channel is MAX / Gosuslugi.
     tg_bot_token: str = ""
+    # The RF server reaches api.telegram.org over IPv6 only; docker has IPv4 only → go through a proxy (http/socks).
+    tg_proxy_url: str = ""
     doctor_notify_chat_id: int | None = None
 
     # MAX is opt-in. Its adapter is delivered in the related backend/channel changes.
@@ -77,6 +87,11 @@ class Settings(BaseSettings):
 
     # Virtual clock for the demo (ADR-005).
     demo_clock: bool = True
+    # Initial screening (onboarding questionnaires) on/off — the default; the demo console can flip it at runtime.
+    screening_enabled: bool = True
+    # Crisis blocks the dialog until the doctor's ack (clinical protocol) — off for the demo: the crisis message
+    # and the doctor's alert still happen, the patient keeps using everything.
+    crisis_blocking: bool = False
     tz: str = "Europe/Moscow"
 
     api_base_url: str = "http://localhost:8000"

app/core/demo_settings.py

@@ -0,0 +1,34 @@
+"""Presentation switches (demo): stored in `demo_settings`, default from settings/env.
+
+    SCREENING_ENABLED=false                 # env default
+    POST /api/v1/demo/settings {"screening_enabled": false}   # one button in the demo console
+
+Screening off: after consent the patient goes straight to the chat — no onboarding questionnaires.
+"""
+
+from typing import Any
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.config import get_settings
+from app.db.clinical import DemoSetting
+
+SCREENING = "screening_enabled"
+
+
+async def get_all(session: AsyncSession) -> dict[str, Any]:
+    row = await session.get(DemoSetting, SCREENING)
+    return {SCREENING: bool(row.value) if row is not None else get_settings().screening_enabled}
+
+
+async def screening_enabled(session: AsyncSession) -> bool:
+    return (await get_all(session))[SCREENING]
+
+
+async def set_screening(session: AsyncSession, enabled: bool) -> None:
+    row = await session.get(DemoSetting, SCREENING)
+    if row is None:
+        session.add(DemoSetting(key=SCREENING, value=enabled))
+    else:
+        row.value = enabled
+    await session.flush()

app/core/instrument_catalog.py

@@ -13,7 +13,7 @@ from pydantic import BaseModel
 
 from app.config import get_settings
 from app.core.cssrs_content import load_cssrs_russian
-from app.core.intake_questions import DIARY_PARTS, content, diary_config, texts
+from app.core.intake_questions import DIARY_PARTS, content, diary_config, red_flags, texts
 from app.core.scales import load_scale
 
 Section = Literal["safety", "mood", "trauma", "alcohol", "about", "diary", "scale"]
@@ -92,6 +92,9 @@ def catalog() -> Catalog:
                 ),
             )
 
+    flags = red_flags()
+    add(FormInfo(id="red_flags", title="Вопросы о безопасности", section="safety"), [i["id"] for i in flags["items"]])
+
     cssrs = load_cssrs_russian()
     add(FormInfo(id="cssrs", title=cssrs.title, section="safety"), ["cssrs_" + i.id for i in cssrs.items])
 

app/core/intake_questions.py

@@ -53,11 +53,30 @@ def cssrs_answers(answers: dict) -> dict[str, int]:
     return {k.removeprefix("cssrs_"): v for k, v in answers.items() if k.startswith("cssrs_")}
 
 
+def red_flags() -> dict:
+    return yaml.safe_load((get_settings().config_dir / "instruments/red_flags.yaml").read_text(encoding="utf-8"))
+
+
+def red_flag_category(item_id: str) -> str | None:
+    return next((i["category"] for i in red_flags()["items"] if i["id"] == item_id), None)
+
+
+def red_flag_triggers(item_id: str) -> bool:
+    """False for questions that are asked and stored but never stop/alert (rf_weapons, team decision 2026-10-04)."""
+    return next((i.get("trigger", True) for i in red_flags()["items"] if i["id"] == item_id), True)
+
+
 def next_intake(answers: dict, policy: IntakePolicy, sex: str | None) -> Question | None:
-    """Stage 1 in the fixed order of «Опросники для MVP»: PHQ-2, trauma filter → TSQ, RUS-AUDIT-S, C-SSRS screener,
-    then the context questions; stage 2 (rest of PHQ-9, PCL-5, full RUS-AUDIT) only where stage 1 is positive.
-    Items inside a scale keep the order of the source. The risk stop rules run after every answer (flows/intake.py),
-    so a positive answer interrupts the questionnaire wherever it comes."""
+    """Stage 1: the red-flag block first (config/instruments/red_flags.yaml), then the fixed order of «Опросники
+    для MVP»: PHQ-2, trauma filter → TSQ, RUS-AUDIT-S, C-SSRS screener, the context questions; stage 2 (rest of
+    PHQ-9, PCL-5, full RUS-AUDIT) only where stage 1 is positive. Items inside a scale keep the order of the source.
+    The risk stop rules run after every answer (flows/intake.py), so a positive answer interrupts the questionnaire
+    wherever it comes."""
+    flags = red_flags()
+    for n, item in enumerate(flags["items"]):
+        if item["id"] not in answers:
+            text = (flags["intro"] + "\n\n" if n == 0 else "") + item["text"]
+            return Question(item["id"], text, tuple((o["value"], o["label"]) for o in flags["options"]), "now")
     for item in ("phq9_1", "phq9_2"):
         if item not in answers:
             return scale_question("phq9", item)

app/core/interpretation.py

@@ -12,6 +12,7 @@ import yaml
 from app.config import get_settings
 from app.core.intake_questions import content
 from app.core.questionnaire import IntakeDecision, IntakePolicy
+from app.core.safety import fill_phones
 
 # Stage 1 finding → diary module of stage 2 (PTSD is the «veteran» module of the profile).
 FINDING_MODULE = {"depression_suspected": "depression", "ptsd_suspected": "veteran"}
@@ -23,7 +24,7 @@ PCL5_IMPROVED_BELOW = 28
 
 
 def texts() -> dict[str, Any]:
-    return yaml.safe_load((get_settings().config_dir / "texts/results.yaml").read_text(encoding="utf-8"))
+    return yaml.safe_load(fill_phones((get_settings().config_dir / "texts/results.yaml").read_text(encoding="utf-8")))
 
 
 def render(interpretation: dict[str, Any]) -> str:

app/core/questionnaire.py

@@ -50,7 +50,6 @@ class IntakePolicy(BaseModel):
     phq9_specialist_at: int = Field(ge=0, le=27, strict=True)
     phq9_priority_at: int = Field(ge=0, le=27, strict=True)
     pcl5_specialist_at: int = Field(ge=0, le=80, strict=True)
-    pcl5_priority_at: int = Field(ge=0, le=80, strict=True)
     pcl5_behavior_flag_at: int = Field(ge=0, le=4, strict=True)
 
     @model_validator(mode="after")
@@ -65,8 +64,6 @@ class IntakePolicy(BaseModel):
             raise ValueError("choose a total threshold OR a zone for alcohol expansion")
         if not self.phq9_psychologist_at < self.phq9_specialist_at < self.phq9_priority_at:
             raise ValueError("PHQ-9 thresholds must be ordered")
-        if self.pcl5_specialist_at >= self.pcl5_priority_at:
-            raise ValueError("PCL-5 thresholds must be ordered")
         return self
 
 
@@ -276,7 +273,9 @@ def evaluate_intake(
     if c.positive:
         flags.append("cssrs_positive")
     flags.extend(c.conflicts)
-    if item9_positive or c.risk_lower_bound == "high":
+    # One crisis rule (ADR-013): a high C-SSRS risk. Item 9 alone is a crisis only while the screener is not complete;
+    # with a negative / low screener it is the «discrepancy» flag of the routing table, not the crisis screen.
+    if c.risk_lower_bound == "high" or (item9_positive and not c.complete):
         return IntakeDecision(policy.version, "crisis", "crisis", (), tuple(flags), c.missing)
     if c.positive and policy.approval and policy.cssrs_positive_policy == "stop_any":
         return IntakeDecision(policy.version, "crisis", "crisis", (), tuple(flags), c.missing)
@@ -327,9 +326,11 @@ def evaluate_intake(
     # PHQ-2-negative does not invent a PHQ-9 total; it simply does not trigger expansion.
     phq_total = phq.total if phq.complete else 0
     pcl_total = pcl.total if pcl.complete else 0
+    # Risk and flags first, scales second (clinical review 2026-10-04). A high PCL-5 alone is not a reason for a
+    # priority appointment: it goes to the planned specialist route.
     if c.risk == "moderate":
         primary = "urgent_psychiatrist"
-    elif phq_total >= policy.phq9_priority_at or pcl_total >= policy.pcl5_priority_at:
+    elif phq_total >= policy.phq9_priority_at:
         primary = "priority_psychiatrist"
     elif phq_total >= policy.phq9_specialist_at or pcl_total >= policy.pcl5_specialist_at:
         primary = "psychiatrist_or_psychotherapist"

app/core/safety.py

@@ -395,6 +395,25 @@ def _crisis_cfg(path: Path = CRISIS_PATH) -> _CrisisCfg:
     return _CrisisCfg.model_validate(yaml.safe_load(path.read_text(encoding="utf-8")))
 
 
+_PHONE_REF = re.compile(r"\{phone:([a-z_]+)\}")
+
+
+def fill_phones(text: str, path: Path = CRISIS_PATH) -> str:
+    """Replace `{phone:<resource>}` with the number from the single table in config/texts/crisis.yaml.
+
+    Every patient text and prompt refers to the table instead of repeating a number, so a number is changed in one
+    place. An unknown resource fails loudly: a wrong number on the crisis screen must never pass silently.
+    """
+    resources = _crisis_cfg(path).resources
+
+    def number(m: re.Match[str]) -> str:
+        if m.group(1) not in resources:
+            raise ValueError(f"unknown crisis resource {m.group(1)!r}")
+        return resources[m.group(1)].phone
+
+    return _PHONE_REF.sub(number, text)
+
+
 def _render(body: str, resources: list[CrisisResource], doctor_notified: str, grounding: str) -> str:
     lines = [body.strip(), ""]
     for r in resources:

app/db/clinical.py

@@ -192,3 +192,12 @@ class ClockState(ClinicalBase):
 
     id: Mapped[int] = mapped_column(Integer, primary_key=True)
     offset_seconds: Mapped[float] = mapped_column(Float, default=0.0)
+
+
+class DemoSetting(ClinicalBase):
+    """Presentation switches changed at runtime from the demo console (e.g. the initial screening on/off)."""
+
+    __tablename__ = "demo_settings"
+
+    key: Mapped[str] = mapped_column(String(64), primary_key=True)
+    value: Mapped[Any] = mapped_column(JSONType)

app/dialog/crisis.py

@@ -14,6 +14,7 @@ Decisions:
 
 from typing import Any
 
+from app.config import get_settings
 from app.core.safety import CrisisResource, CrisisScript, crisis_reply
 from app.dialog.contracts import Reply
 from app.dialog.fsm import Ctx, Step
@@ -37,6 +38,9 @@ def open_crisis(ctx: Ctx, categories: list[str], source: str) -> Step:
         text = "\n".join([ctx.texts.crisis.repeat, *_resource_lines(script.resources)])
         return Step(replies=[Reply(text=text, kind="crisis")])
     crisis = {"open": True, "opened_at": ctx.now.isoformat(), "source": source, "variant": script.variant}
+    if not get_settings().crisis_blocking:
+        # Non-blocking (demo default): the script is shown, the doctor is alerted, the patient keeps going.
+        return Step(replies=[Reply(text=script.text, kind="crisis")], data={"crisis": crisis})
     return Step(
         replies=[Reply(text=script.text, kind="crisis")],
         next_state="crisis",
@@ -44,6 +48,10 @@ def open_crisis(ctx: Ctx, categories: list[str], source: str) -> Step:
     )
 
 
+def crisis_resource_lines(resources: list[CrisisResource]) -> list[str]:
+    return _resource_lines(resources)
+
+
 def no_consent_crisis_text(script: CrisisScript) -> str:
     """Crisis script without the "doctor notified" line — for people we may not (or did not) notify about."""
     return "\n".join([script.body, "", *_resource_lines(script.resources), "", script.grounding])

app/dialog/engine.py

@@ -28,6 +28,7 @@ Logs carry only the token, states and event codes (guardrail #8).
 
 import asyncio
 import logging
+import re
 from dataclasses import dataclass, field
 from datetime import datetime, timedelta
 from typing import Any
@@ -36,6 +37,7 @@ from weakref import WeakValueDictionary
 from sqlalchemy.exc import InterfaceError, OperationalError
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from app.config import get_settings
 from app.core import clock, safety
 from app.core.intake_runtime import intake_enabled
 from app.core.privacy import scrub
@@ -46,7 +48,7 @@ from app.db.session import get_clinical_session
 from app.dialog import scale_session
 from app.dialog.actions import execute
 from app.dialog.contracts import Event, Reply, register_engine
-from app.dialog.crisis import crisis_open, no_consent_crisis_text, open_crisis
+from app.dialog.crisis import crisis_open, crisis_resource_lines, no_consent_crisis_text, open_crisis
 from app.dialog.flows import load_flows
 from app.dialog.flows.core import close_crisis, revoke
 from app.dialog.fsm import (
@@ -66,6 +68,16 @@ from app.services.contracts import MonitoringService, get_monitoring
 
 log = logging.getLogger(__name__)
 
+LAST_PROMPT = "last_prompt"
+DEFERRED_CRISIS = "deferred_crisis"
+QUESTIONNAIRE_PREFIXES = ("scale.", "veteran.onboarding.", "intake.")
+DEFERRABLE_KINDS = frozenset({"crisis", "safety_plan"})
+MEDICINE_WORDS = re.compile(
+    r"(лекарств|таблет|препарат|сертралин|пароксетин|флуоксетин|эсциталопрам|антидепрессант|побочк)", re.I
+)
+# States with their own free-text handling (idle → flows' `text:idle`; explicit question states).
+SIDE_ANSWER_SKIP = frozenset({"idle", "kb.question", "meds.question"})
+FREE_TEXT_PREFIXES = ("daily.", "meds.", "kb.")  # steps that expect free text (check-in, side effects, questions)
 # Enter hooks may chain (scale done → next onboarding state → next scale); a cycle must not hang a request.
 MAX_ENTER_CHAIN = 8
 _patient_locks = WeakValueDictionary()
@@ -137,12 +149,18 @@ class Engine:
 
         await self._apply(ctx, await self._start_pending(ctx), run)
 
+        self._defer_crisis_messages(ctx, event, run)
+
         if _consented(ctx):
             for reply in run.replies:
                 # Outgoing texts may quote the safety plan (names of relatives) — scrub before storing too.
                 await repo.add_message(
                     session, token=token, direction="out", text=scrub(reply.text), ts=ctx.now, fsm_state=ctx.state
                 )
+        prompt = next((r for r in reversed(run.replies) if r.buttons), None)
+        if prompt is not None and ctx.state not in RESTING_STATES:
+            # Re-shown after an off-script question so the patient (and the web screen) gets the buttons back.
+            ctx.data[LAST_PROMPT] = prompt.model_dump()
         if event.type != "system" or ctx.state != state_before:
             # System events the step ignores (reminders inside a scale) do not count as the patient being there.
             ctx.data[STEP_AT] = ctx.now.isoformat()
@@ -165,6 +183,21 @@ class Engine:
         )
         return run.replies
 
+    def _defer_crisis_messages(self, ctx: Ctx, event: Event, run: _Run) -> None:
+        """Crisis messages triggered by test answers wait until the questionnaire is over (team decision
+        2026-10-04: let the patient finish the test, then show the crisis message; it blocks nothing). The doctor's
+        alert is not delayed — only the patient-facing message. A typed acute message is never deferred."""
+        if get_settings().crisis_blocking:
+            return  # clinical protocol: the crisis takes over immediately
+        if event.type == "button" and ctx.state.startswith(QUESTIONNAIRE_PREFIXES):
+            held = [r for r in run.replies if r.kind in DEFERRABLE_KINDS]
+            if held:
+                run.replies[:] = [r for r in run.replies if r.kind not in DEFERRABLE_KINDS]
+                ctx.data[DEFERRED_CRISIS] = [*(ctx.data.get(DEFERRED_CRISIS) or []), *(r.model_dump() for r in held)]
+            return
+        if ctx.data.get(DEFERRED_CRISIS) and not ctx.state.startswith(QUESTIONNAIRE_PREFIXES):
+            run.replies.extend(Reply.model_validate(r) for r in ctx.data.pop(DEFERRED_CRISIS))
+
     def _abandoned(self, ctx: Ctx) -> bool:
         """The current question step was left unanswered longer than `abandon_after_hours` (app.core.clock)."""
         if intake_enabled() or not _consented(ctx) or ctx.state in RESTING_STATES or crisis_open(ctx.data):
@@ -289,8 +322,47 @@ class Engine:
             step.notify_monitoring = True
             return step
 
+        side = await self._side_answer(ctx)
+        if side is not None:
+            return side
         return await self._dispatch(ctx, event)
 
+    async def _side_answer(self, ctx: Ctx) -> Step | None:
+        """A question asked in the middle of a questionnaire, a check-in or a crisis is answered right away
+        (agent → guidelines/instruction), then the pending question is shown again with its buttons — the
+        chat never ignores a question. Idle and the explicit question states have their own handlers.
+        In a crisis the answer always carries the emergency contacts; the crisis itself stays open until the
+        doctor acknowledges it, and acute messages never get here (handled above)."""
+        from app.dialog.flows.knowledge import knowledge_answer, looks_like_question
+
+        in_crisis = (crisis_open(ctx.data) or ctx.state == "crisis") and await ctx.monitoring.is_in_crisis(ctx.token)
+        free_text_step = ctx.state in SIDE_ANSWER_SKIP or ctx.state.startswith(FREE_TEXT_PREFIXES)
+        if ctx.text is None or (free_text_step and not in_crisis):
+            return None  # idle / check-in / question states parse free text themselves
+        asks = looks_like_question(ctx.text) or bool(MEDICINE_WORDS.search(ctx.text))
+        if asks:
+            answer = await knowledge_answer(ctx, explicit=True)
+        elif in_crisis:
+            # Not a question, but in a crisis every message gets a human reply (with contacts below), never silence.
+            answer = await ctx.gateway.phrase("support")
+        else:
+            return None  # e.g. a free-text answer to the current scale item — the step handles it
+        if not answer:
+            return None
+        log.info("dialog side answer token=%s state=%s crisis=%s", ctx.token, ctx.state, in_crisis)
+        replies = [Reply(text=answer)]
+        if in_crisis:
+            script = crisis_reply([])
+            replies.append(
+                Reply(
+                    text="\n".join([ctx.texts.crisis.repeat, *crisis_resource_lines(script.resources)]), kind="crisis"
+                )
+            )
+        elif ctx.data.get(LAST_PROMPT):
+            last = Reply.model_validate(ctx.data[LAST_PROMPT])
+            replies.append(last.model_copy(update={"text": f"{ctx.texts.side_answer_return}\n\n{last.text}"}))
+        return Step(replies=replies)
+
     def _can_screen(self, ctx: Ctx, screen: str) -> bool:
         if crisis_open(ctx.data) or scale_session.active_scale(ctx.state) == screen:
             return False

app/dialog/flows/core.py

@@ -66,9 +66,13 @@ def menu_reply(ctx: Ctx, text: str | None = None) -> Reply:
 async def onboarding_entry(ctx: Ctx, after: str | None = None) -> str:
     """Entry state of the next module onboarding: `"{module}.onboarding.start"` for the first module of the
     profile (after `after`, if given) that has such a state registered; `"idle"` when none is left."""
+    from app.core.demo_settings import screening_enabled
+    from app.core.intake_runtime import intake_enabled
+
+    if after is None and not await screening_enabled(ctx.session):
+        return "idle"  # demo switch: no initial questionnaires, straight to the chat
     profile = await ctx.profile()
     modules = list(profile.modules) if profile is not None and profile.modules else []
-    from app.core.intake_runtime import intake_enabled
 
     # Stage 1 is the same for everyone (ADR-012): it does not depend on modules, which it sets itself — an anonymous
     # person has no profile at all before it.

app/dialog/flows/intake.py

@@ -4,6 +4,7 @@ import math
 from dataclasses import asdict, replace
 from zoneinfo import ZoneInfo
 
+from app.config import get_settings
 from app.core import interpretation
 from app.core.intake_questions import (
     DIARY_KINDS,
@@ -14,6 +15,8 @@ from app.core.intake_questions import (
     diary_questions,
     group_answers,
     next_intake,
+    red_flag_category,
+    red_flag_triggers,
     scale_question,
     texts,
 )
@@ -27,6 +30,7 @@ from app.dialog.contracts import Button, Event, Reply
 from app.dialog.crisis import crisis_open, open_crisis
 from app.dialog.fsm import Step, on_enter, state, trigger
 from app.services import questionnaire as service
+from app.services.followup import schedule_post_crisis_call
 
 
 def allowed(ctx):
@@ -105,6 +109,7 @@ async def urgent(ctx, row=None, *, code="help_now", categories=None):
         level=AlertLevel.ACUTE,
         params={"record_id": row.id if row else None, "flags": flags},
     )
+    await schedule_post_crisis_call(ctx.session, ctx.token, ctx.now, source="questionnaire:" + code)
     step = open_crisis(ctx, categories or [], source="questionnaire:" + code)
     step.data["questionnaire_record"] = None
     step.data["intake_review_required"] = True
@@ -113,6 +118,74 @@ async def urgent(ctx, row=None, *, code="help_now", categories=None):
     return step
 
 
+def _defer_urgent(row, code, categories=None):
+    """Remember the first crisis signal of this questionnaire; the questionnaire goes on (team decision
+    2026-10-04: the crisis message comes at the end and blocks nothing)."""
+    if not row.details.get("pending_urgent"):
+        row.details = {**row.details, "pending_urgent": {"code": code, "categories": categories or []}}
+
+
+async def _flush_urgent(ctx, row, step):
+    """Questionnaire over: acute alert + callback for the doctor, then the (non-blocking) crisis message."""
+    pending = (row.details or {}).get("pending_urgent")
+    if not pending or step.next_state == "intake.collect":
+        return step
+    row.details = {**row.details, "pending_urgent": None}
+    code, categories = pending["code"], pending.get("categories") or []
+    profile = await ctx.profile()
+    if profile:
+        profile.enhanced_monitoring = True
+    await service.alert(
+        ctx.session,
+        ctx.token,
+        f"{row.id}:{code}",
+        code,
+        ctx.now,
+        level=AlertLevel.ACUTE,
+        params={"record_id": row.id, "deferred": True},
+    )
+    await schedule_post_crisis_call(ctx.session, ctx.token, ctx.now, source="questionnaire:" + code)
+    crisis = open_crisis(ctx, categories, source="questionnaire:" + code)
+    for reply in crisis.replies:
+        reply.buttons.append([Button(label=ctx.texts.safety_plan.title.rstrip(":"), data="safety_plan")])
+    return step.then(crisis)
+
+
+async def item9_discrepancy(ctx, row):
+    profile = await ctx.profile()
+    if profile:
+        profile.enhanced_monitoring = True
+    await service.alert(
+        ctx.session,
+        ctx.token,
+        row.id + ":phq9_cssrs_discrepancy",
+        "phq9_cssrs_discrepancy",
+        ctx.now,
+        level=AlertLevel.HIGH,
+        params={"record_id": row.id},
+    )
+    row.details = {**row.details, "item9_flagged": True}
+
+
+async def item9_assessment(ctx, row):
+    """Thoughts of death in PHQ-9 item 9 → the C-SSRS screener at once; the screener's level decides (graded)."""
+    await service.alert(
+        ctx.session,
+        ctx.token,
+        row.id + ":item9_assessment",
+        "phq9_item9_cssrs",
+        ctx.now,
+        level=AlertLevel.HIGH,
+        params={"record_id": row.id},
+    )
+    await service.finish(ctx, row, "complete", extra={"stop_reason": "phq9_item9_cssrs"})
+    profile = await ctx.profile()
+    if profile:
+        profile.enhanced_monitoring = True
+    task = await service.queue_task(ctx.session, ctx.token, "cssrs_since_contact", row.id + ":cssrs", ctx.now)
+    return await start(ctx, "cssrs_since_contact", task)
+
+
 async def manual_review(ctx, row, code):
     await service.finish(ctx, row, "review_required", extra={"stop_reason": code})
     await service.alert(
@@ -132,12 +205,36 @@ async def manual_review(ctx, row, code):
 
 
 async def advance(ctx, row):
+    return await _flush_urgent(ctx, row, await _advance(ctx, row))
+
+
+async def _advance(ctx, row):
     a = row.answers
     c = score_cssrs_recent(cssrs_answers(a))
     policy = load_policy()
-    # PHQ-9 item 9 ≥1 — in the full scale or in the depression module of the diary — stops at once (stage 1 rule).
-    if a.get("phq9_9", 0) > 0 or a.get("dphq_9", 0) > 0 or c.risk_lower_bound == "high":
-        return await urgent(ctx, row, code="suicide_signal")
+    # Red flags (asked first): any «Да» stops the questionnaire with the help screen of its category.
+    for item_id, value in a.items():
+        if item_id.startswith("rf_") and value == 1 and red_flag_triggers(item_id):
+            category = red_flag_category(item_id)
+            if get_settings().crisis_blocking:  # clinical protocol: stop at once
+                return await urgent(ctx, row, code="red_flag_" + item_id[3:], categories=[category] if category else [])
+            _defer_urgent(row, "red_flag_" + item_id[3:], [category] if category else [])
+            break
+    # One crisis rule for the whole app (ADR-013): the crisis screen opens at a HIGH C-SSRS risk only. Moderate →
+    # stop + psychiatrist (about 24 h); low → flag, safety plan, continue.
+    if c.risk_lower_bound == "high":
+        if get_settings().crisis_blocking:
+            return await urgent(ctx, row, code="suicide_signal")
+        _defer_urgent(row, "suicide_signal")
+    if a.get("dphq_9", 0) > 0 and row.kind in DIARY_KINDS:
+        # Diary depression module, item 9 ≥1 → the C-SSRS right away («Этап 2»); its risk level decides.
+        return await item9_assessment(ctx, row)
+    if a.get("phq9_9", 0) > 0 and not c.complete:
+        # Item 9 without a completed screener (only in a scheduled PHQ-9): assess the risk right away.
+        return await item9_assessment(ctx, row)
+    if a.get("phq9_9", 0) > 0 and not row.details.get("item9_flagged"):
+        # Item 9 ≥1 with a negative / low C-SSRS = discrepancy (routing table): flag to the doctor, safety plan.
+        await item9_discrepancy(ctx, row)
     if c.positive and policy.cssrs_positive_policy == "stop_any":
         return await urgent(ctx, row, code="cssrs_positive")
     if c.conflicts:
@@ -165,7 +262,7 @@ async def advance(ctx, row):
         return await manual_review(ctx, row, "answer_conflict")
     if q:
         replies = [question_reply(row, q)]
-        if c.complete and c.risk == "low" and not row.details.get("plan_shown"):
+        if ((c.complete and c.risk == "low") or row.details.get("item9_flagged")) and not row.details.get("plan_shown"):
             from app.dialog.actions import safety_plan_reply
 
             replies.insert(0, await safety_plan_reply(ctx))
@@ -225,11 +322,12 @@ async def advance(ctx, row):
         if row.kind in ("phq9", "pcl5"):
             sid = row.kind
             total = extra["totals"][sid]["total"]
-            priority = getattr(policy, sid + "_priority_at")
+            # Only PHQ-9 has a priority threshold; PCL-5 alone never makes an appointment a priority.
+            priority = getattr(policy, sid + "_priority_at", None)
             specialist = getattr(policy, sid + "_specialist_at")
             extra["route"] = (
                 "priority_psychiatrist"
-                if total >= priority
+                if priority is not None and total >= priority
                 else ("psychiatrist_or_psychotherapist" if total >= specialist else "review_with_treating_clinician")
             )
             earlier = [
@@ -284,7 +382,13 @@ async def advance(ctx, row):
         return step
     from app.dialog.flows.core import menu_reply
 
-    return Step(replies=[menu_reply(ctx, shown)], next_state="idle", data={"questionnaire_record": None})
+    replies = [menu_reply(ctx, shown)]
+    if ((c.complete and c.risk == "low") or row.details.get("item9_flagged")) and not row.details.get("plan_shown"):
+        # The safety plan is never skipped because the risk signal came with the last question.
+        from app.dialog.actions import safety_plan_reply
+
+        replies.insert(0, await safety_plan_reply(ctx))
+    return Step(replies=replies, next_state="idle", data={"questionnaire_record": None})
 
 
 # Diary flags that bring a full scale forward («Этап 2», p.4).

app/dialog/flows/veteran_daily.py

@@ -62,6 +62,7 @@ from app.dialog.flows.meds import (
 )
 from app.dialog.fsm import Ctx, Step, abandonable, on_enter, state, trigger
 from app.dialog.scale_session import register_scale_intro
+from app.llm import agent
 
 PREFIX = "daily"
 STEPS = ("mood", "med", "sleep", "nightmares", "alcohol", "text")
@@ -429,6 +430,9 @@ async def _idle_text(ctx: Ctx, event: Event) -> Step:
     assert ctx.text is not None
     facts = await ctx.gateway.extract_checkin(ctx.text)
     if facts is None:
+        smart = await agent.answer_chat(ctx.token, ctx.text)
+        if smart is not None:
+            return Step(replies=[menu_reply(ctx, agent.render(smart))])
         text = await knowledge_answer(ctx, explicit=False) if looks_like_question(ctx.text) else None
         return Step(replies=[menu_reply(ctx, text or await ctx.gateway.phrase("support"))])
     fields: dict[str, Any] = {}
@@ -446,6 +450,14 @@ async def _idle_text(ctx: Ctx, event: Event) -> Step:
             text = (await side_effect_reply(ctx, codes)).text
     if fields:
         await save_checkin(ctx, **fields)
+    if not facts.quit_intent:
+        # Chat first: whatever the patient writes in idle gets the agent's answer; check-in signals above are
+        # recorded silently for the doctor. Quit intent keeps its rule reaction (explanation + doctor notified).
+        smart = await agent.answer_chat(ctx.token, ctx.text)
+        if smart is not None:
+            return Step(
+                replies=[menu_reply(ctx, agent.render(smart))], notify_monitoring=bool(fields or facts.side_effects)
+            )
     if text is None and facts.med_question and not facts.quit_intent:
         text = await answer_question(ctx, await current_prescription(ctx))
     # A question is answered even if it also mentioned alcohol/side effects (those are recorded above).

app/dialog/scale_session.py

@@ -33,6 +33,7 @@ from dataclasses import replace
 from datetime import timedelta
 from typing import Any
 
+from app.config import get_settings
 from app.core.privacy import scrub
 from app.core.scales import Item, ScaleDef, ScaleResult, load_scale, score
 from app.db import repo
@@ -237,7 +238,10 @@ async def _record(
     sex = profile.sex if profile is not None and profile.sex in ("m", "f") else None
     result = score(scale.id, answers, sex=sex)  # type: ignore[arg-type]
     crisis_flags = [f for f in result.flags if f in ctx.texts.settings.crisis_flags]
-    if crisis_flags or session["idx"] + 1 >= len(scale.items):
+    # Clinical protocol (crisis_blocking): a crisis flag ends the scale at once. Demo default: the scale is finished
+    # first and the crisis message comes at the end of the questionnaire (engine defers it), blocking nothing.
+    stop_now = bool(crisis_flags) and get_settings().crisis_blocking
+    if stop_now or session["idx"] + 1 >= len(scale.items):
         return await _finish(ctx, scale, session, answers, result, crisis_flags)
     updated = {**session, "answers": answers, "idx": session["idx"] + 1}
     sessions = {**(ctx.data.get("scales") or {}), scale.id: updated}
@@ -278,7 +282,7 @@ async def _finish(
     # Crisis path decided by the scale's own flags (deterministic), not by the LLM.
     risk_screen = ctx.texts.settings.risk_screen_scale
     if crisis_open(ctx.data):
-        crisis = Step(next_state="crisis")
+        crisis = Step(next_state="crisis") if get_settings().crisis_blocking else Step()
     else:
         crisis = open_crisis(ctx, categories=[], source=f"scale:{scale.id}")
     step = done.then(crisis)

app/dialog/texts.py

@@ -7,6 +7,7 @@ import yaml
 from pydantic import BaseModel, ConfigDict, Field
 
 from app.config import get_settings
+from app.core.safety import fill_phones
 
 
 class _Strict(BaseModel):
@@ -91,6 +92,7 @@ class ReachOutTexts(_Strict):
     silence: str
     missed_doses: str
     default: str
+    after_crisis: str
 
 
 class VisitTexts(_Strict):
@@ -111,6 +113,7 @@ class DialogTexts(_Strict):
     reach_out: ReachOutTexts
     visit: VisitTexts
     unavailable: str
+    side_answer_return: str = "Вернёмся к вопросу:"
 
 
 def dialog_texts_path() -> Path:
@@ -120,4 +123,4 @@ def dialog_texts_path() -> Path:
 @lru_cache
 def load_dialog_texts(path: Path | None = None) -> DialogTexts:
     path = path or dialog_texts_path()
-    return DialogTexts.model_validate(yaml.safe_load(path.read_text(encoding="utf-8")))
+    return DialogTexts.model_validate(yaml.safe_load(fill_phones(path.read_text(encoding="utf-8"))))

app/integrations/emias_record.py

@@ -63,7 +63,11 @@ def _load(path: Path) -> dict[str, PatientRecord]:
     raw = yaml.safe_load(path.read_text(encoding="utf-8")) or {}
     out: dict[str, PatientRecord] = {}
     for snils, rec in (raw.get("records") or {}).items():
-        out[pseudo_id(snils)] = PatientRecord.model_validate(rec)
+        aliases = rec.pop("tokens", []) if isinstance(rec, dict) else []
+        record = PatientRecord.model_validate(rec)
+        out[pseudo_id(snils)] = record
+        for token in aliases:  # demo tokens used by the patient web app (e.g. tok-andrey) share the record
+            out[token] = record
     return out
 
 
@@ -73,7 +77,11 @@ class EmiasRecordSource:
 
     async def get(self, token: str) -> PatientRecord | None:
         try:
-            base = _load(self._path or default_records_path()).get(token)
+            records = _load(self._path or default_records_path())
+            base = records.get(token)
+            if base is None and "-" in token:
+                # Fresh demo copies (POST /auth/demo/fresh): "tok-andrey-1a2b3c4d" → the "tok-andrey" record.
+                base = records.get(token.rsplit("-", 1)[0])
         except (OSError, yaml.YAMLError, ValueError) as exc:
             log.error("emias records unavailable: %s", type(exc).__name__)
             base = None

app/integrations/notify.py

@@ -92,9 +92,9 @@ class TelegramDoctorChat:
 
     def _get_sender(self) -> ChatSender:
         if self._sender is None:
-            from aiogram import Bot
+            from app.channels.telegram import make_bot
 
-            self._bot = Bot(self._bot_token)
+            self._bot = make_bot(self._bot_token)
             self._sender = self._bot.send_message
         return self._sender
 

app/llm/agent/__init__.py

@@ -14,12 +14,14 @@ Guarantees (on top of what the dialog engine already did — scrub, consent, saf
 import asyncio
 import json
 import logging
+import re
 from typing import Any
 
 from pydantic import BaseModel
 
 from app.config import Settings, get_settings
 from app.core.types import ScrubbedText
+from app.llm.agent import stream
 from app.llm.agent.prompts import SYSTEM_PROMPT
 from app.llm.agent.providers import (
     GigaChatToolProvider,
@@ -35,6 +37,19 @@ from app.llm.guard import assert_scrubbed
 log = logging.getLogger(__name__)
 
 MAX_ANSWER_CHARS = 1200
+MAX_TOOL_CALLS_PER_TURN = 4  # seen live: one «как дела?» turn issued 23 parallel guideline searches
+# Greetings / small talk: answered without the LLM (instant, free, never drags medical findings into «привет»).
+SMALL_TALK = re.compile(
+    r"^\W*(прив\w*|здравств\w*|здаров\w*|добр(ый|ое|ого)\s+\w+|хай|hello|hi|ку)\b"
+    r"|\b(как\s+(дела|ты|вы|жизнь|поживаешь)|что\s+делаешь|чем\s+занят\w*|спишь|кто\s+ты|ты\s+(бот|робот|кто))\b",
+    re.I,
+)
+SMALL_TALK_MAX_WORDS = 8
+SMALL_TALK_REPLY = (
+    "Здравствуйте! Я на связи. Могу рассказать о вашем лечении: как действует препарат и когда ждать эффекта, "
+    "какие бывают побочные эффекты, что можно и нельзя сочетать с вашими таблетками, что в составе. "
+    "Спросите — отвечу по вашей карте, инструкции и клиническим рекомендациям."
+)
 NUDGE = "Сформулируй ответ пациенту по результатам инструментов выше (3–6 предложений)."
 _ALLOWED_MSG_KEYS = {"role", "content", "tool_calls", "tool_call_id"}
 
@@ -66,6 +81,15 @@ def build_provider(settings: Settings | None = None) -> ToolChatProvider | None:
     return None
 
 
+def is_small_talk(text: str) -> bool:
+    """Short greeting / chit-chat without a medical question in it."""
+    words = re.findall(r"\w+", text)
+    medical = r"(лекарств|таблет|препарат|побоч|доз|лечени|можно|ибупрофен|нурофен|алкогол|пив|сертралин|сон|сплю)"
+    if len(words) > SMALL_TALK_MAX_WORDS or re.search(medical, text, re.I):
+        return False
+    return bool(SMALL_TALK.search(text))
+
+
 async def _chat_with_retry(provider: ToolChatProvider, messages: list[dict[str, Any]]):
     """One retry: OpenRouter occasionally fails a single call (seen live as intermittent ProviderError)."""
     wire = _wire(messages) if provider.name == "openrouter" else messages  # gigachat needs the tool `name`
@@ -102,6 +126,9 @@ async def answer_chat(
     except ValueError:
         log.warning("chat agent: unscrubbed input rejected")
         return None
+    if is_small_talk(question):
+        return AgentAnswer(text=SMALL_TALK_REPLY, sources=[], tools=[], provider="small_talk")
+    stream.emit("status", "Думаю над вопросом…")
     ctx = ToolContext(token=token)
     messages: list[dict[str, Any]] = [
         {"role": "system", "content": SYSTEM_PROMPT + await _patient_brief(ctx)},
@@ -123,17 +150,18 @@ async def answer_chat(
                     {
                         "role": "assistant",
                         "content": turn.content,
-                        "tool_calls": [
+                        "tool_calls": [  # the calls we actually run (capped) — history must match the results
                             {
                                 "id": c.id,
                                 "type": "function",
                                 "function": {"name": c.name, "arguments": json.dumps(c.arguments, ensure_ascii=False)},
                             }
-                            for c in turn.tool_calls
+                            for c in turn.tool_calls[:MAX_TOOL_CALLS_PER_TURN]
                         ],
                     }
                 )
-                for c in turn.tool_calls:
+                for c in turn.tool_calls[:MAX_TOOL_CALLS_PER_TURN]:
+                    stream.emit("status", stream.TOOL_STATUS.get(c.name, "Проверяю данные…"))
                     result = await run_tool(ctx, c.name, c.arguments)
                     messages.append({"role": "tool", "tool_call_id": c.id, "name": c.name, "content": result})
             log.warning("chat agent: step limit reached tools=%s", ",".join(ctx.used))
@@ -159,8 +187,14 @@ def _finish(text: str, ctx: ToolContext, provider: str) -> AgentAnswer | None:
     return AgentAnswer(text=text, sources=ctx.sources[:4], tools=ctx.used, provider=provider)
 
 
+_MD_BULLET = re.compile(r"^\s*[*-]\s+", re.M)
+_MD_MARKS = re.compile(r"\*\*|__|^#+\s*", re.M)
+SOURCES_MIN_CHARS = 160  # short replies (greetings, small talk) go without the sources footer
+
+
 def render(answer: AgentAnswer) -> str:
-    """Patient-facing text: answer + short sources footer."""
-    if not answer.sources:
-        return answer.text
-    return f"{answer.text}\n\nИсточники: " + "; ".join(answer.sources)
+    """Patient-facing text: answer (markdown stripped to plain text with «•» bullets) + short sources footer."""
+    text = _MD_MARKS.sub("", _MD_BULLET.sub("• ", answer.text)).strip()
+    if not answer.sources or len(text) < SOURCES_MIN_CHARS:
+        return text
+    return f"{text}\n\nИсточники: " + "; ".join(answer.sources)

app/llm/agent/prompts.py

@@ -1,6 +1,8 @@
 """System prompt of the chat agent (Russian). Prohibitions follow SPEC §7; grounding: answer only from tools."""
 
-SYSTEM_PROMPT = """Ты — помощник пациента в сервисе сопровождения лечения между визитами к врачу. Пациент лечится у
+from app.core.safety import fill_phones
+
+_SYSTEM_PROMPT = """Ты — помощник пациента в сервисе сопровождения лечения между визитами к врачу. Пациент лечится у
 психиатра; ты отвечаешь на его вопросы о состоянии, лечении и лекарствах.
 
 Как работать:
@@ -30,6 +32,17 @@ SYSTEM_PROMPT = """Ты — помощник пациента в сервисе
 - проводить психотерапию; обещать конфиденциальность сверх реальной;
 - выдумывать факты, которых нет в результатах инструментов.
 
-Стиль: по-русски, на «вы», спокойно и по-человечески, 3–6 коротких предложений, без markdown-заголовков.
+На приветствия и болтовню («привет», «как дела?», «спишь?») отвечай 1–2 фразами, тепло, от лица помощника
+сервиса (не называй себя «языковой моделью» или «ИИ-моделью») и предложи, с чем можешь помочь: вопросы о
+лечении, лекарствах, побочных эффектах, самочувствии. Находки из карты в ответ на болтовню не упоминай.
+Никогда не переспрашивай, о каком препарате речь, если вопрос о «моём лекарстве»: это текущее назначение.
+Про аналоги и замену: назови группу текущего препарата по инструкции (например, «антидепрессант группы СИОЗС»),
+скажи, что в группе есть и другие препараты, а подбор и замену делает только лечащий врач.
+
+Стиль: по-русски, на «вы», спокойно и по-человечески, 3–6 коротких предложений. Без markdown: никаких «*», «#»,
+«**»; если нужен список — короткие строки с «•».
 Если в вопросе есть признаки угрозы жизни — коротко скажи, что важно сразу позвонить 112 или на телефон
-доверия 8-800-2000-122, и что врач будет уведомлён."""
+доверия {phone:helpline}, и что врач будет уведомлён."""
+
+# Numbers come from the single table config/texts/crisis.yaml.
+SYSTEM_PROMPT = fill_phones(_SYSTEM_PROMPT)

app/llm/agent/providers.py

@@ -16,6 +16,7 @@ import httpx
 from pydantic import BaseModel, Field
 
 from app.config import Settings
+from app.llm.agent import stream
 
 log = logging.getLogger(__name__)
 
@@ -49,9 +50,12 @@ class OpenRouterProvider:
         self._model = settings.openrouter_model
         self._url = settings.openrouter_base_url.rstrip("/") + "/chat/completions"
         self._timeout = settings.chat_agent_timeout_s
+        self._proxy = settings.openrouter_proxy_url or None
         self._client = client
 
     async def chat(self, messages: list[dict[str, Any]], tools: list[dict[str, Any]]) -> AssistantTurn:
+        if stream.active():
+            return await self._chat_stream(messages, tools)
         body: dict[str, Any] = {"model": self._model, "messages": messages, "tools": tools, "temperature": 0.1}
         body["max_tokens"] = 900
         if not any(m["role"] == "tool" for m in messages):
@@ -61,7 +65,7 @@ class OpenRouterProvider:
             if self._client is not None:
                 resp = await self._client.post(self._url, json=body, headers=headers, timeout=self._timeout)
             else:
-                async with httpx.AsyncClient(timeout=self._timeout) as client:
+                async with httpx.AsyncClient(timeout=self._timeout, proxy=self._proxy) as client:
                     resp = await client.post(self._url, json=body, headers=headers)
         except httpx.HTTPError as exc:
             raise ProviderError(type(exc).__name__) from exc
@@ -173,6 +177,78 @@ class GigaChatToolProvider:
         return AssistantTurn(content=msg.content or None, tool_calls=calls)
 
 
+# ------------------------------------------------------------------------------------------------ streaming
+
+
+async def _iter_sse(resp: httpx.Response):
+    async for line in resp.aiter_lines():
+        if not line.startswith("data:"):
+            continue  # comments («: OPENROUTER PROCESSING») and blank lines
+        data = line[5:].strip()
+        if data == "[DONE]":
+            return
+        try:
+            yield json.loads(data)
+        except json.JSONDecodeError:
+            continue
+
+
+async def _openrouter_stream(self: "OpenRouterProvider", messages: list[dict[str, Any]], tools: list[dict[str, Any]]):
+    """Same request with `stream: true`: content deltas go to the live channel, tool calls are assembled."""
+    body: dict[str, Any] = {
+        "model": self._model,
+        "messages": messages,
+        "tools": tools,
+        "temperature": 0.1,
+        "max_tokens": 900,
+        "stream": True,
+    }
+    if not any(m["role"] == "tool" for m in messages):
+        body["tool_choice"] = "required"
+    headers = {"Authorization": f"Bearer {self._key}", "X-Title": "PTSRaiAgent prototype (synthetic data)"}
+    content: list[str] = []
+    calls: dict[int, dict[str, Any]] = {}
+    try:
+        client = self._client or httpx.AsyncClient(timeout=self._timeout, proxy=self._proxy)
+        try:
+            async with client.stream("POST", self._url, json=body, headers=headers, timeout=self._timeout) as resp:
+                if resp.status_code != 200:
+                    await resp.aread()
+                    raise ProviderError(f"http {resp.status_code}")
+                async for chunk in _iter_sse(resp):
+                    choice = (chunk.get("choices") or [{}])[0]
+                    delta = choice.get("delta") or {}
+                    if delta.get("content"):
+                        content.append(delta["content"])
+                        stream.emit("delta", delta["content"])
+                    for tc in delta.get("tool_calls") or []:
+                        slot = calls.setdefault(tc.get("index", 0), {"id": "", "name": "", "args": ""})
+                        slot["id"] = tc.get("id") or slot["id"]
+                        fn = tc.get("function") or {}
+                        slot["name"] += fn.get("name") or ""
+                        slot["args"] += fn.get("arguments") or ""
+        finally:
+            if self._client is None:
+                await client.aclose()
+    except httpx.HTTPError as exc:
+        raise ProviderError(type(exc).__name__) from exc
+    out = []
+    for i in sorted(calls):
+        c = calls[i]
+        try:
+            args = json.loads(c["args"] or "{}")
+        except json.JSONDecodeError:
+            args = {}
+        out.append(ToolCall(id=c["id"] or c["name"] or f"call{i}", name=c["name"], arguments=args))
+    if out and content:
+        stream.emit("reset", None)  # a preface before tool calls is not the answer
+    log.info("openrouter stream ok calls=%d chars=%d", len(out), sum(map(len, content)))
+    return AssistantTurn(content="".join(content) or None, tool_calls=out)
+
+
+OpenRouterProvider._chat_stream = _openrouter_stream  # type: ignore[attr-defined]
+
+
 # ------------------------------------------------------------------------------------------------ offline mock
 
 _DRUG_WORDS = re.compile(

app/llm/agent/stream.py

@@ -0,0 +1,32 @@
+"""Live progress of the chat agent for streaming channels (web SSE).
+
+The SSE endpoint sets a queue in the context; the agent pushes events while it works:
+    status — a tool step in plain Russian («Смотрю вашу карту ЕМИАС…»)
+    delta  — a chunk of the answer as the model writes it
+    reset  — the streamed text so far was a preface before tool calls: clear it
+Channels without a queue (Telegram, tests) are unaffected. Deltas are the raw model text: the final reply
+(forbidden-output filter, sources footer) always comes in the closing `done` event and replaces it.
+"""
+
+import asyncio
+from contextvars import ContextVar
+from typing import Any
+
+STREAM: ContextVar[asyncio.Queue[tuple[str, Any]] | None] = ContextVar("agent_stream", default=None)
+
+TOOL_STATUS = {
+    "get_patient_record": "Смотрю вашу карту ЕМИАС…",
+    "check_drug_for_patient": "Проверяю препарат по вашей карте: болезни, непереносимости, другие лекарства…",
+    "get_drug_info": "Открываю инструкцию к препарату (ГРЛС)…",
+    "search_guidelines": "Ищу в клинических рекомендациях Минздрава…",
+}
+
+
+def active() -> bool:
+    return STREAM.get() is not None
+
+
+def emit(kind: str, payload: Any) -> None:
+    q = STREAM.get()
+    if q is not None:
+        q.put_nowait((kind, payload))

app/llm/gateway.py

@@ -43,7 +43,9 @@ FORBIDDEN_OUTPUT: tuple[re.Pattern[str], ...] = tuple(
     for p in (
         r"диагноз",
         r"у (вас|тебя) (\w+ )?(депресси|птср|тревожн\w* расстройств|расстройств|биполяр|шизофрен)",
-        r"(это|похоже на|признаки) (\w+ )?(депресси|птср|посттравматическ|расстройств)",
+        # Diagnosis-like statements about the patient. A plain definition («ПТСР — это расстройство, которое…»)
+        # is education, not a diagnosis, so bare «это …» is no longer blocked.
+        r"(похоже на|признаки|скорее всего,? у вас|у вас,? похоже) (\w+ )?(депресси|птср|посттравматическ|расстройств)",
         r"(увеличьте|уменьшите|снизьте|повысьте|удвойте|измените|увеличь|уменьши|снизь) (\w+ )?доз",
         r"(?<!не )отмен(ите|и) ",
         r"(?<!не )(прекратите|перестаньте|бросьте) (\w+ )?(приём|прием|принимать|пить)",

app/llm/stt.py

@@ -0,0 +1,162 @@
+"""Speech-to-text for the patient's voice input (#116, slide 11: record → check the text → send).
+
+The service only transcribes: the text goes back to the patient to check and edit, and reaches the dialog as an
+ordinary text message — safety, `scrub()` and the LLM gateway apply there. Audio is never stored or logged.
+
+* `OpenRouterSTT` — an audio-capable model via OpenRouter `/chat/completions` (`input_audio`). TEMPORARY, synthetic
+  data only (ADR-014): voice is personal data that `scrub()` cannot depersonalise, and OpenRouter is a foreign
+  service (152-FZ). Real patients need an RF provider (SaluteSpeech, #121) behind the same `SpeechToText` contract.
+* `MockSTT` — offline, returns a fixed synthetic phrase (ADR-004: the demo works without network).
+"""
+
+import asyncio
+import base64
+import logging
+import struct
+from functools import lru_cache
+from pathlib import Path
+from typing import Protocol
+
+import httpx
+import yaml
+from pydantic import BaseModel
+
+from app.config import Settings
+
+log = logging.getLogger(__name__)
+
+_RETRY_STATUSES = {429, 500, 502, 503, 504}
+_ATTEMPTS = 3
+
+
+class SttTexts(BaseModel):
+    system: str
+    user: str
+    silence_marker: str
+    mock_transcript: str
+
+
+@lru_cache
+def load_texts(path: Path | None = None) -> SttTexts:
+    from app.config import get_settings
+
+    path = path or get_settings().config_dir / "texts" / "stt.yaml"
+    return SttTexts.model_validate(yaml.safe_load(path.read_text(encoding="utf-8")))
+
+
+class SttError(RuntimeError):
+    """The provider failed (network, HTTP error, bad response)."""
+
+
+class SttNoSpeech(SttError):
+    """The recording holds no recognisable speech."""
+
+
+class InvalidAudio(ValueError):
+    pass
+
+
+def wav_seconds(audio: bytes) -> float:
+    """Duration of a PCM WAV file; raises `InvalidAudio` for anything else."""
+    if len(audio) < 44 or audio[:4] != b"RIFF" or audio[8:12] != b"WAVE":
+        raise InvalidAudio("not a WAV file")
+    pos, byte_rate = 12, 0
+    while pos + 8 <= len(audio):
+        chunk_id, size = audio[pos : pos + 4], struct.unpack_from("<I", audio, pos + 4)[0]
+        if chunk_id == b"fmt " and size >= 16:
+            byte_rate = struct.unpack_from("<I", audio, pos + 16)[0]
+        elif chunk_id == b"data":
+            if not byte_rate:
+                break
+            return min(size, len(audio) - pos - 8) / byte_rate
+        pos += 8 + size + (size & 1)
+    raise InvalidAudio("malformed WAV")
+
+
+class SpeechToText(Protocol):
+    name: str
+
+    async def transcribe(self, audio: bytes, fmt: str = "wav") -> str: ...
+
+
+class MockSTT:
+    name = "mock"
+
+    async def transcribe(self, audio: bytes, fmt: str = "wav") -> str:
+        return load_texts().mock_transcript
+
+
+class OpenRouterSTT:
+    name = "openrouter"
+
+    def __init__(self, settings: Settings, client: httpx.AsyncClient | None = None) -> None:
+        self._key = settings.openrouter_api_key
+        self._model = settings.stt_model
+        self._url = settings.openrouter_base_url.rstrip("/") + "/chat/completions"
+        self._timeout = settings.stt_timeout_s
+        self._proxy = settings.openrouter_proxy_url or None
+        self._client = client
+
+    async def transcribe(self, audio: bytes, fmt: str = "wav") -> str:
+        texts = load_texts()
+        content = [
+            {"type": "text", "text": texts.user},
+            {"type": "input_audio", "input_audio": {"data": base64.b64encode(audio).decode(), "format": fmt}},
+        ]
+        body = {
+            "model": self._model,
+            "messages": [{"role": "system", "content": texts.system}, {"role": "user", "content": content}],
+            "temperature": 0,
+            "max_tokens": 800,
+            # Transcription needs no thinking: with reasoning on, gemini-2.5-flash takes ~60 s instead of ~3 s.
+            "reasoning": {"enabled": False},
+        }
+        data = await self._post(body)
+        try:
+            text = (data["choices"][0]["message"]["content"] or "").strip()
+        except (KeyError, IndexError, TypeError) as exc:
+            raise SttError("bad response") from exc
+        text = text.strip("«»\"' \n")
+        if not text or text.lower() == texts.silence_marker.lower():
+            raise SttNoSpeech("no speech")
+        return text
+
+    async def _post(self, body: dict) -> dict:
+        headers = {"Authorization": f"Bearer {self._key}", "X-Title": "PTSRaiAgent prototype (synthetic data)"}
+        for attempt in range(_ATTEMPTS):
+            try:
+                if self._client is not None:
+                    resp = await self._client.post(self._url, json=body, headers=headers, timeout=self._timeout)
+                else:
+                    async with httpx.AsyncClient(timeout=self._timeout, proxy=self._proxy) as client:
+                        resp = await client.post(self._url, json=body, headers=headers)
+            except httpx.HTTPError as exc:
+                if attempt + 1 == _ATTEMPTS:
+                    raise SttError(type(exc).__name__) from exc
+                await asyncio.sleep(0.5 * 2**attempt)
+                continue
+            if resp.status_code in _RETRY_STATUSES and attempt + 1 < _ATTEMPTS:
+                log.warning("stt openrouter http %s, retry %d", resp.status_code, attempt + 1)
+                await asyncio.sleep(0.5 * 2**attempt)
+                continue
+            if resp.status_code != 200:
+                # Provider error bodies carry no patient data — log the message to make failures diagnosable.
+                log.error("stt openrouter http %s: %s", resp.status_code, resp.text[:200])
+                raise SttError(f"http {resp.status_code}")
+            try:
+                return resp.json()
+            except ValueError as exc:
+                raise SttError("bad json") from exc
+        raise SttError("retries exhausted")
+
+
+def get_stt(settings: Settings) -> SpeechToText | None:
+    """The configured provider; `None` = voice input is off. `auto` = OpenRouter if a key is set, else mock."""
+    mode = settings.stt_provider
+    if mode == "auto":
+        mode = "openrouter" if settings.openrouter_api_key else "mock"
+    if mode == "openrouter":
+        return OpenRouterSTT(settings) if settings.openrouter_api_key else None
+    if mode == "mock":
+        return MockSTT()
+    return None

app/scheduler/jobs.py

@@ -22,11 +22,13 @@ from app.db import repo
 from app.db.clinical import PatientProfile
 from app.db.intake import QuestionnaireTask
 from app.db.session import get_clinical_session
-from app.dialog.contracts import Event, Reply, get_engine
+from app.dialog.contracts import Button, Event, Reply, get_engine
+from app.dialog.texts import load_dialog_texts
 from app.integrations.base import integrations
 from app.scheduler import demo
 from app.services import monitoring
 from app.services import questionnaire as service
+from app.services.followup import next_day_check_due
 
 log = logging.getLogger(__name__)
 _worker = None
@@ -43,6 +45,15 @@ async def plan():
             state = await repo.get_dialog_state(session, profile.token)
             if not state or state.data.get("consent") is not True or state.data.get("adult") is not True:
                 continue
+            crisis = state.data.get("crisis")
+            if next_day_check_due(crisis, now):
+                # The day after a crisis the app asks how the person is, the safety plan one tap away.
+                texts = load_dialog_texts()
+                plan_button = Button(label=texts.safety_plan.title.rstrip(":"), data="safety_plan")
+                await integrations.outbox.push(
+                    profile.token, [Reply(text=texts.reach_out.after_crisis, buttons=[[plan_button]], kind="question")]
+                )
+                state.data = {**state.data, "crisis": {**crisis, "followup_sent": now.isoformat()}}
             rows = await service.records(session, profile.token)
             initial = next((r for r in reversed(rows) if r.kind == "intake" and r.status == "complete"), None)
             if initial is None:

app/services/anonymous_care.py

@@ -0,0 +1,24 @@
+"""Pseudonymous care routing; no identity fields or changes to clinical release gates."""
+
+from hashlib import sha256
+
+import yaml
+from pydantic import BaseModel, ConfigDict, Field
+
+from app.config import get_settings
+
+
+class AnonymousCare(BaseModel):
+    model_config = ConfigDict(extra="forbid", str_strip_whitespace=True)
+    doctor_id: str = Field(min_length=1, max_length=64)
+
+
+def doctor_id() -> str:
+    path = get_settings().config_dir / "anonymous_care.yaml"
+    return AnonymousCare.model_validate(yaml.safe_load(path.read_text())).doctor_id
+
+
+def display_name(token: str) -> str:
+    # Do not use a piece of the bearer token itself as the public patient label.
+    label = sha256(("anonymous-patient:" + token).encode()).hexdigest()[:10].upper()
+    return f"Пациент {label}"

app/services/followup.py

@@ -0,0 +1,43 @@
+"""Follow-up after a crisis (clinical review 2026-10-04: «план безопасности и звонок после кризиса — обязательно»).
+
+Every crisis — from the safety lexicon, a rule (PHQ-9 item 9, C-SSRS high) or the questionnaire (red flags, high
+risk) — leaves a task for the doctor on duty: call the patient back, about 24 hours as a guide. The task is an alert
+on the dashboard; «позвонили» is its acknowledgement (`repo.ack_alert`, audited). One open task per patient: a
+repeated crisis signal does not pile up callbacks. On the next day the app itself asks how the person is, with the
+safety plan at hand (`app/scheduler/jobs.py`).
+"""
+
+from datetime import datetime, timedelta
+
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core.types import AlertLevel
+from app.db import repo
+from app.db.clinical import Alert
+
+POST_CRISIS_CALL = "post_crisis_call"
+ON_DUTY = "on-duty"
+CALL_WITHIN = timedelta(hours=24)  # a guide, not a guarantee
+NEXT_DAY_CHECK_AFTER = timedelta(hours=20)
+
+
+async def schedule_post_crisis_call(session: AsyncSession, token: str, now: datetime, *, source: str) -> Alert | None:
+    if POST_CRISIS_CALL in await repo.open_alert_codes(session, token):
+        return None
+    return await repo.add_alert(
+        session,
+        token=token,
+        level=AlertLevel.HIGH,
+        code=POST_CRISIS_CALL,
+        reason="Перезвонить пациенту после кризиса — дежурный специалист, ориентир 24 ч. Отметка «позвонили» — "
+        "подтверждение этого сигнала.",
+        ts=now,
+        params={"due_at": (now + CALL_WITHIN).isoformat(), "assignee": ON_DUTY, "source": source},
+    )
+
+
+def next_day_check_due(crisis: dict | None, now: datetime) -> bool:
+    """The day after a crisis the app asks how the person is — once per crisis episode."""
+    if not crisis or not crisis.get("opened_at") or crisis.get("followup_sent"):
+        return False
+    return now - datetime.fromisoformat(crisis["opened_at"]) >= NEXT_DAY_CHECK_AFTER

app/services/garden.py

@@ -0,0 +1,156 @@
+"""The patient's garden (#118, slide 13): light gamification over all modules.
+
+Deterministic, no LLM (guardrail 3), rules in `config/garden.yaml` (ADR-008). Computed on the fly from existing
+activity rows — no own table, no migration (ADR-003), and changing the rules replays the whole history.
+
+«Without pressure»: points only add up (a skipped day takes nothing away), growth counts the fact of an activity
+and never its content (scale scores, mood), and there is no comparison between patients. Only the patient sees it.
+"""
+
+from collections import defaultdict
+from collections.abc import Iterable
+from datetime import date, datetime
+from functools import lru_cache
+from pathlib import Path
+from typing import Literal
+
+import yaml
+from pydantic import BaseModel
+from sqlalchemy import select
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from app.core import clock
+from app.core.intake_questions import DIARY_KINDS
+from app.db.clinical import DailyCheckin, Dose, ScaleResult
+from app.db.intake import QuestionnaireRecord
+
+Activity = Literal["meds", "diary", "questionnaire"]
+
+
+class ActivityRule(BaseModel):
+    points: int
+    daily_cap: int
+
+
+class TitleRule(BaseModel):
+    from_steps: int
+    title: str
+
+
+class GardenConfig(BaseModel):
+    activities: dict[Activity, ActivityRule]
+    step_points: int
+    stages_per_plant: int
+    max_plants: int
+    titles: list[TitleRule]
+    today: dict[Activity, str]
+
+
+@lru_cache
+def load_config(path: Path | None = None) -> GardenConfig:
+    from app.config import get_settings
+
+    path = path or get_settings().config_dir / "garden.yaml"
+    return GardenConfig.model_validate(yaml.safe_load(path.read_text(encoding="utf-8")))
+
+
+class Plant(BaseModel):
+    stage: int  # 0 = not yet sprouted … stages_per_plant = in bloom
+
+
+class TodayItem(BaseModel):
+    activity: Activity
+    title: str
+    done: bool
+    detail: str | None = None  # e.g. "1 из 2" for doses
+
+
+class Garden(BaseModel):
+    points: int
+    steps: int
+    title: str
+    step_points: int
+    points_in_step: int  # progress towards the next growth step: «8 из 10»
+    plants: list[Plant]
+    today: list[TodayItem]
+
+
+def compute(events: Iterable[tuple[Activity, date]], today: date, cfg: GardenConfig | None = None) -> Garden:
+    """Points from (activity, day) events: per-day caps, events after `today` (demo clock) ignored."""
+    cfg = cfg or load_config()
+    per_day: dict[tuple[Activity, date], int] = defaultdict(int)
+    for kind, day in events:
+        if day <= today:
+            per_day[(kind, day)] += 1
+    points = sum(
+        min(n, cfg.activities[kind].daily_cap) * cfg.activities[kind].points for (kind, _), n in per_day.items()
+    )
+    steps = points // cfg.step_points
+    plants = [
+        Plant(stage=max(0, min(cfg.stages_per_plant, steps - i * cfg.stages_per_plant))) for i in range(cfg.max_plants)
+    ]
+    title = next(t.title for t in sorted(cfg.titles, key=lambda t: -t.from_steps) if steps >= t.from_steps)
+    return Garden(
+        points=points,
+        steps=steps,
+        title=title,
+        step_points=cfg.step_points,
+        points_in_step=points % cfg.step_points,
+        plants=plants,
+        today=[],
+    )
+
+
+def _day(ts: datetime, tz) -> date:
+    return ts.astimezone(tz).date()
+
+
+async def garden_for(s: AsyncSession, token: str) -> Garden:
+    cfg = load_config()
+    now = clock.now()
+    tz, today = now.tzinfo, now.date()
+    events: list[tuple[Activity, date]] = []
+
+    doses = (await s.scalars(select(Dose).where(Dose.token == token))).all()
+    for d in doses:
+        if d.status == "taken":
+            events.append(("meds", _day(d.answered_at or d.planned_at, tz)))
+    checkins = (await s.scalars(select(DailyCheckin).where(DailyCheckin.token == token))).all()
+    for c in checkins:
+        if c.mood is not None:
+            events.append(("diary", c.day))
+    records = (await s.scalars(select(QuestionnaireRecord).where(QuestionnaireRecord.token == token))).all()
+    for r in records:
+        if r.status != "complete":
+            continue
+        if r.kind in DIARY_KINDS:
+            events.append(("diary", r.day))
+        else:
+            events.append(("questionnaire", _day(r.completed_at, tz) if r.completed_at else r.day))
+    scales = (await s.scalars(select(ScaleResult).where(ScaleResult.token == token))).all()
+    for sc in scales:
+        if sc.complete:
+            events.append(("questionnaire", _day(sc.ts, tz)))
+
+    garden = compute(events, today, cfg)
+    done_today = {kind for kind, day in events if day == today}
+
+    items: list[TodayItem] = []
+    due = [d for d in doses if _day(d.planned_at, tz) == today]
+    if due:
+        taken = sum(d.status == "taken" for d in due)
+        due_now = sum(d.planned_at <= now for d in due)
+        items.append(
+            TodayItem(
+                activity="meds",
+                title=cfg.today["meds"],
+                done=taken > 0 and taken >= due_now,
+                detail=f"{taken} из {len(due)}",
+            )
+        )
+    items.append(TodayItem(activity="diary", title=cfg.today["diary"], done="diary" in done_today))
+    # A questionnaire is not due every day: show it only once done, so an empty day is never a «missed» item.
+    if "questionnaire" in done_today:
+        items.append(TodayItem(activity="questionnaire", title=cfg.today["questionnaire"], done=True))
+    garden.today = items
+    return garden

app/services/monitoring.py

@@ -30,6 +30,7 @@ from app.db.audit import audit
 from app.db.session import get_clinical_session
 from app.integrations.base import integrations
 from app.services.contracts import register_monitoring
+from app.services.followup import schedule_post_crisis_call
 from app.services.snapshot import build_snapshot
 
 log = logging.getLogger(__name__)
@@ -50,6 +51,37 @@ async def refresh_status(session: AsyncSession, token: str) -> str:
     return status.color
 
 
+# Which data a crisis rule is triggered by: a crisis for the same code is raised again only when such data is
+# newer than the last alert with that code (acked or not) — otherwise an acked PHQ-9 item 9 re-fired on every
+# data change for 24 h (seen live: AUDIT-C answers 0/0/0 → crisis again).
+_CRISIS_TRIGGERS: dict[str, tuple[str, ...]] = {
+    "phq9_item9": ("phq9",),
+    "cssrs_high": ("cssrs_screen",),
+    "cssrs_moderate": ("cssrs_screen",),
+    "weapon_access_risk": ("phq9", "cssrs_screen"),
+}
+
+
+async def _drop_handled_crises(s: AsyncSession, token: str, actions: list[Action]) -> list[Action]:
+    crisis_codes = {a.code for a in actions if a.type == "crisis"}
+    if not crisis_codes:
+        return actions
+    alerts = await repo.list_alerts(s, token=token)
+    results = await repo.list_scale_results(s, token)
+    handled: set[str] = set()
+    for code in crisis_codes:
+        last_alert = max((a.ts for a in alerts if a.code == code), default=None)
+        if last_alert is None:
+            continue
+        scales = _CRISIS_TRIGGERS.get(code)
+        last_data = max((r.ts for r in results if scales is None or r.scale in scales), default=None)
+        if last_data is None or last_data <= last_alert:
+            handled.add(code)
+    if handled:
+        log.info("monitoring token=%s crisis already handled=%s", token, sorted(handled))
+    return [a for a in actions if a.code not in handled]
+
+
 def _drop_running_crises(actions: list[Action], open_acute: set[str]) -> list[Action]:
     """Silence every action of a crisis rule whose acute alert is still open (see module docstring)."""
     running = {a.code for a in actions if a.type == "crisis" and a.code in open_acute}
@@ -67,6 +99,7 @@ class RuleMonitoring:
                 return []
             open_acute = {a.code for a in await repo.list_alerts(s, token=token, open_only=True) if a.level == "acute"}
             actions = _drop_running_crises(evaluate(snapshot), open_acute)
+            actions = await _drop_handled_crises(s, token, actions)
             for action in actions:
                 if action.type == "enhanced_monitoring":
                     profile.enhanced_monitoring = True
@@ -93,6 +126,7 @@ class RuleMonitoring:
                         ts=now,
                     )
                     notify.append((ON_DUTY_DOCTOR, level, action.reason_ru))
+                    await schedule_post_crisis_call(s, token, now, source=f"rule:{action.code}")
             await refresh_status(s, token)
         log.info("monitoring token=%s actions=%s", token, [a.code for a in actions])
         for doctor_id, level, text in notify:

config/anonymous_care.yaml

@@ -0,0 +1,3 @@
+# Routing for the current prototype: anonymous patients use the same doctor workspace,
+# without an identity record. Set this to the responsible clinic's doctor when deploying.
+doctor_id: doc-smirnova

config/garden.yaml

@@ -0,0 +1,38 @@
+# Сад (#118, слайд 13): лёгкая геймификация поверх модулей. Расчёт детерминированный, без LLM.
+# Принципы «без давления» (вычитка — #120):
+#   1. Сад не вянет: очки только прибавляются, пропуск ничего не отнимает.
+#   2. Рост не зависит от ответов: считается факт активности, а не баллы шкал или настроение.
+#   3. Нет соревнования: только свой сад, никаких сравнений.
+
+# Очки за активность и дневной лимит по каждому типу (чтобы пять одинаковых действий подряд не давали пятикратный рост).
+activities:
+  meds:            # отметка «Принял(а)» по назначенной дозе
+    points: 2
+    daily_cap: 4
+  diary:           # ежедневный чек-ин или запись дневника
+    points: 3
+    daily_cap: 1
+  questionnaire:   # завершённый опрос (шкала, этап опроса)
+    points: 4
+    daily_cap: 1
+
+# Каждые step_points очков — один шаг роста. Растения растут по очереди: у каждого stages_per_plant стадий
+# (росток → стебель → бутон → цветок); распустился один — проклёвывается следующий.
+step_points: 10
+stages_per_plant: 4
+max_plants: 6
+
+# Название сада по числу шагов роста (берётся последний порог ≤ шагов).
+titles:
+  - {from_steps: 0, title: "Сад ждёт первых ростков"}
+  - {from_steps: 1, title: "Появились первые ростки"}
+  - {from_steps: 4, title: "Распустился первый цветок"}
+  - {from_steps: 10, title: "Сад растёт"}
+  - {from_steps: 18, title: "Сад цветёт"}
+  - {from_steps: 24, title: "Цветущий сад"}
+
+# Подписи пунктов «Сегодня».
+today:
+  meds: "Приём лекарств"
+  diary: "Запись в дневнике"
+  questionnaire: "Назначенный опрос"

config/instruments/red_flags.yaml

@@ -0,0 +1,27 @@
+# Red flags — asked first in stage 1, before any scale (clinical review 2026-10-04: «обязательно»).
+# Any «Да» stops the questionnaire: an acute alert to the doctor, a callback task for the doctor on duty and the
+# help screen (crisis script variant by `category`, config/texts/crisis.yaml). Wording — team draft, the doctor
+# approves it (#8). These are not a scale: there is no score, only a stop rule.
+version: red-flags-v1-2026-10-04
+intro: "Сначала несколько коротких вопросов о безопасности. Отвечайте так, как есть сейчас."
+options:
+- {value: 1, label: Да}
+- {value: 0, label: Нет}
+items:
+- id: rf_psychosis
+  text: Слышите ли Вы сейчас голоса или видите то, чего не видят и не слышат другие люди?
+  category: null
+- id: rf_harm_others
+  text: Есть ли у Вас сейчас мысли причинить вред другому человеку?
+  category: harm_others
+- id: rf_weapons
+  text: Есть ли у Вас дома или рядом доступ к оружию?
+  category: null
+  # Asked and stored, but NOT a stop rule (team decision 2026-10-04: weapon ownership is common in Russia).
+  trigger: false
+- id: rf_violence
+  text: Угрожает ли Вам сейчас кто-то дома или применяет к Вам насилие?
+  category: domestic_violence
+- id: rf_intoxication
+  text: Находитесь ли Вы сейчас в состоянии сильного опьянения или тяжёлой отмены алкоголя либо других веществ?
+  category: null

config/questionnaire.yaml

@@ -1,7 +1,8 @@
 # Review candidate from the supplied MVP PDF, not a signed clinical protocol.
 version: mvp-v2-review-2026-10-03
 approval: null
-cssrs_positive_policy: pending
+# C01 decided by the team after the clinical review (2026-10-04, ADR-013): crisis screen at HIGH risk only.
+cssrs_positive_policy: graded
 rus_audit_s_expand_from_zone: null
 phq2_expand_at: 2
 tsq_expand_at: 6
@@ -9,5 +10,4 @@ phq9_psychologist_at: 5
 phq9_specialist_at: 10
 phq9_priority_at: 20
 pcl5_specialist_at: 31
-pcl5_priority_at: 50
 pcl5_behavior_flag_at: 3

config/questionnaire_demo.yaml

@@ -17,5 +17,4 @@ phq9_psychologist_at: 5
 phq9_specialist_at: 10
 phq9_priority_at: 20
 pcl5_specialist_at: 31
-pcl5_priority_at: 50
 pcl5_behavior_flag_at: 3

config/rules.yaml

@@ -47,6 +47,7 @@ rules:
     reason_ru: "C-SSRS: высокий риск (намерение/план или поведение)"
 
   - id: weapon_access_risk
+    enabled: false   # team decision 2026-10-04: weapon ownership is common in Russia — the answer triggers nothing
     description: >
       Access to weapons together with any suicide-risk signal (recent PHQ-9 item 9 >= 1 or C-SSRS >= moderate).
       Means restriction is one of the most evidence-based suicide-prevention measures → crisis safety block.
@@ -62,7 +63,7 @@ rules:
     params: {}
     actions:
       - {type: alert, level: high}
-    reason_ru: "C-SSRS: умеренный риск, врачу в течение 24 ч"
+    reason_ru: "C-SSRS: умеренный риск, психиатр — ориентир 24 ч"
 
   # ---------------------------------------------------------------- scale thresholds (yellow)
   - id: phq9_severe
@@ -73,10 +74,12 @@ rules:
     reason_ru: "PHQ-9 = {total} (тяжёлая депрессия, порог {threshold})"
 
   - id: pcl5_severe
-    description: Latest PCL-5 total >= 50 → doctor within 24 h.
+    description: >
+      Latest PCL-5 total >= 50 → planned specialist review. Not a priority appointment by itself: risk and flags
+      come first, scales second (clinical review 2026-10-04).
     params: {threshold: 50}
     actions:
-      - {type: alert, level: high}
+      - {type: alert, level: moderate}
     reason_ru: "PCL-5 = {total} (порог {threshold})"
 
   - id: phq9_rise

config/texts/crisis.yaml

@@ -1,4 +1,6 @@
 # Кризисный скрипт (guardrail №2, issue #4). Отправляется ДОСЛОВНО при safety = acute; LLM не вызывается.
+# `resources` — ЕДИНСТВЕННАЯ таблица телефонов приложения: остальные тексты и промпты ссылаются на неё через
+# `{phone:<ключ>}` (app/core/safety.py::fill_phones). Источники и проверка — docs/clinical/crisis-contacts.md.
 # СТАТУС: черновик, утверждает психиатр (#8).
 #
 # Выбор варианта (app/core/safety.py::crisis_reply): `<audience>.<категория>` по порядку category_priority,
@@ -17,6 +19,11 @@ resources:
     name: "Психологическая помощь МЧС России"
     phone: "+7 (495) 989-50-50"
     note: "круглосуточно; стоимость звонка по тарифу оператора"
+  # For minors only (age gate stop text); not a general adult crisis line (docs/clinical/crisis-contacts.md).
+  child_helpline:
+    name: "Детский телефон доверия"
+    phone: "8-800-2000-122"
+    note: "для детей, подростков и родителей; бесплатно, анонимно"
   police:
     name: "Полиция"
     phone: "102"

config/texts/dialog.yaml

@@ -53,7 +53,7 @@ age_gate:
   unclear: "Не удалось понять ответ. Нажмите кнопку или напишите год рождения четырьмя цифрами, например 1990."
   minor_stop: |-
     Спасибо, что ответили. Этот помощник рассчитан на взрослых, поэтому дальше я не смогу с вами работать и не сохраняю ваши ответы.
-    Если вам тяжело или нужна поддержка — позвоните на Детский телефон доверия 8-800-2000-122 (бесплатно, анонимно, круглосуточно) или расскажите взрослому, которому доверяете: родителям, учителю, школьному психологу.
+    Если вам тяжело или нужна поддержка — позвоните на Детский телефон доверия {phone:child_helpline} (бесплатно, анонимно) или расскажите взрослому, которому доверяете: родителям, учителю, школьному психологу.
     Если есть угроза жизни прямо сейчас — звоните 112.
 
 idle:
@@ -79,9 +79,9 @@ revoked:
 
 crisis:
   # Повторный острый сигнал в уже открытом кризисе: скрипт второй раз не шлём, только это + ресурсы.
-  repeat: "Я рядом. Ваш при угрозе жизни звоните 112. Если вам угрожает опасность прямо сейчас — позвоните:"
+  repeat: "Я рядом. Если вам угрожает опасность прямо сейчас — позвоните:"
   # Любое другое сообщение, пока кризис открыт (до подтверждения врача).
-  hold: "Я на связи и читаю вас. Не ждите ответа врача, если помощь нужна сейчас. Если станет хуже — звоните 112 или на телефон доверия +7 (495) 989-50-50."
+  hold: "Я на связи и читаю вас. Не ждите ответа врача, если помощь нужна сейчас. Если станет хуже — звоните 112 или на телефон доверия {phone:helpline}."
   # Врач подтвердил кризисный алерт (doctor_ack): выходим из кризиса.
   closed: "Ваш врач получил сообщение и в курсе вашей ситуации. Спасибо, что остаётесь на связи. Если снова станет тяжело — напишите мне в любой момент."
 
@@ -101,7 +101,7 @@ safety_plan:
     - "Что помогает мне самому: прогулка, душ, дыхание, любимое дело."
     - "Люди и места, которые отвлекают: к кому можно пойти."
     - "Кому позвонить, если тяжело: близкий человек, сослуживец."
-    - "Специалисты: ваш врач, телефон доверия +7 (495) 989-50-50, экстренные службы 112."
+    - "Специалисты: ваш врач, телефон доверия {phone:helpline}, экстренные службы 112."
     - "Безопасная среда: оружие и лишние лекарства — на хранение близкому человеку."
   means_restriction: "Если дома есть оружие или запас лекарств — попросите близкого человека временно хранить их у себя."
 
@@ -109,10 +109,15 @@ reach_out:
   silence: "Давно не было вестей от вас. Как вы? Можно ответить одним словом — мне важно знать, что с вами всё в порядке."
   missed_doses: "Заметил, что несколько приёмов препарата пропущено. Всё ли в порядке? Если что-то мешает — расскажите, я передам врачу."
   default: "Хотел узнать, как вы. Напишите пару слов, когда будет удобно."
+  # The day after a crisis (app/services/followup.py); the doctor on duty also calls back.
+  after_crisis: "Как вы сегодня? Вчера было очень тяжело — мне важно знать, что вы в порядке. Ваш план безопасности под кнопкой ниже. С вами также свяжется дежурный специалист. Если станет хуже — звоните 112 или на телефон доверия {phone:helpline}."
 
 visit:
   button: "Записаться к врачу раньше"
 
 unavailable: |-
   Сейчас у меня технические неполадки, и я не могу сохранить ваше сообщение. Пожалуйста, напишите чуть позже.
-  Если вам плохо прямо сейчас — звоните 112 или на телефон доверия +7 (495) 989-50-50.
+  Если вам плохо прямо сейчас — звоните 112 или на телефон доверия {phone:helpline}.
+
+# Answer to an off-script question is followed by the pending question again (engine side answer).
+side_answer_return: "Вернёмся к вопросу:"

config/texts/knowledge.yaml

@@ -33,6 +33,21 @@ question_starts:
   - "чем"
   - "где"
   - "есть ли"
+  - "что содержится"
+  - "что содержат"
+  - "что входит"
+  - "что в составе"
+  - "какой состав"
+  - "какие побочки"
+  - "какие побочные"
+  - "состав"
+  - "какие"
+  - "можно"
+  - "из чего"
+  - "в чём"
+  - "в чем"
+  - "расскажи"
+  - "объясни"
 
 # Бытовые фразы — не вопросы о лечении, даже со «?»: «как дела?», «как ты?». База знаний для них не ищется
 # (иначе «как дела» находит «линия делит» — у слов одна основа), бот отвечает обычной фразой поддержки.

config/texts/max.yaml

@@ -0,0 +1,8 @@
+connected: "Уведомления подключены. Здесь будут появляться напоминания с сайта. Отключить их можно командой /stop."
+not_connected: "Чтобы получать уведомления, откройте сайт и подключите MAX в настройках уведомлений."
+invalid: "Не удалось подключить уведомления. Создайте новую ссылку на сайте. Если MAX уже подключён к другому аккаунту, сначала отключите прежнюю привязку."
+stopped: "Уведомления отключены. Подключить их снова можно по новой ссылке с сайта."
+help: "Этот бот присылает уведомления. Пройти тесты, посмотреть результаты и продолжить общение можно на сайте. Команда /stop отключает уведомления."
+notification: "На сайте есть напоминание для вас. Откройте личный кабинет, чтобы продолжить."
+completed: "Прохождение теста завершено. Подробности доступны в личном кабинете на сайте."
+open_site: "Открыть сайт"

config/texts/results.yaml

@@ -32,13 +32,13 @@ findings:
     text: "Рекомендуем консультацию нарколога — параллельно с основным маршрутом."
   suicide_low:
     title: "Тяжёлые мысли"
-    text: "Вы отметили мысли о смерти. Мы добавили план безопасности и будем каждый день коротко спрашивать о самочувствии. Если станет тяжело — звоните 112 или на телефон доверия."
+    text: "Вы отметили мысли о смерти. Мы добавили план безопасности и будем каждый день коротко спрашивать о самочувствии. Если станет тяжело — звоните 112 или на телефон доверия {phone:helpline}."
   tbi:
     title: "Контузия или ЧМТ"
     text: "Рекомендуем консультацию невролога."
 
 routes:
-  urgent_psychiatrist: "Срочно свяжитесь с психиатром — сегодня или завтра."
+  urgent_psychiatrist: "Срочно свяжитесь с психиатром — ориентир: в течение 24 часов."
   priority_psychiatrist: "Рекомендуем приоритетную запись к психиатру в ближайшие дни."
   psychiatrist_or_psychotherapist: "Рекомендуем плановую запись к психиатру или психотерапевту."
   psychologist: "Рекомендуем консультацию психолога."

config/texts/stt.yaml

@@ -0,0 +1,17 @@
+# Распознавание голоса пациента (#116). Модель только расшифровывает речь — не отвечает и не советует.
+# Расшифровка возвращается пациенту на проверку и уходит в диалог обычным текстом (safety → scrub → LLM).
+
+# Системная инструкция модели распознавания.
+system: >-
+  Ты — система распознавания речи. Дословно расшифруй русскую речь из аудио.
+  Выведи только расшифровку: без кавычек, пояснений и ответа на сказанное.
+  Ставь знаки препинания. Если в записи совсем нет речи — выведи ровно [тишина].
+
+# Текст рядом с аудио в сообщении пользователя.
+user: "Расшифруй эту аудиозапись."
+
+# Что модель возвращает, если речи нет (сравнивается без учёта регистра).
+silence_marker: "[тишина]"
+
+# Офлайн-режим (STT_PROVIDER=mock): фиксированная фраза демо-пациента, синтетика.
+mock_transcript: "Сегодня долго не мог уснуть. Хочу обсудить это на следующей встрече."

config/texts/veteran_onboarding.yaml

@@ -14,7 +14,8 @@ status: "черновик, утверждает психиатр (#8)"
 settings:
   # Флаги шкал (config/scales/*.yaml), которые считаются сигналом риска для блока безопасного хранения оружия.
   # Положительный п.9 PHQ-9 и C-SSRS — суицидальный риск; AUDIT-C — алкоголь как фактор риска при доступе к оружию.
-  weapon_risk_flags: [phq9_item9_positive, cssrs_high, cssrs_moderate, auditc_positive]
+  # The weapon question exists but affects nothing (team decision 2026-10-04): no safety block, no trigger.
+  weapon_risk_flags: []
 
 intro:
   text: |-

dashboard/components/patient_card.py

@@ -42,7 +42,10 @@ def dose_strip_html(grid: h.DoseGrid, start: date | None) -> str:
         f'<span><i style="border:1.5px dashed var(--grey)"></i>ожидается · {c["pending"]}</span>'
         "</div>"
     )
-    return f'<div class="mt-doses" style="grid-template-columns:{cols}">{"".join(cells)}</div>{legend}'
+    return (
+        f'<div class="mt-dose-scroll"><div class="mt-doses" style="grid-template-columns:{cols}">'
+        f"{''.join(cells)}</div></div>{legend}"
+    )
 
 
 def _tile(label: str, value: str, hint: str = "", tone: str = "blue") -> str:
@@ -181,7 +184,7 @@ def render(card: PatientCard, today: date | None) -> None:
 
     main, side = st.columns([1.75, 1], gap="large")
     with main:
-        with st.container(border=True):
+        with st.container(border=True, key="panel-scales"):
             c1, c2 = st.columns(2)
             for col, scale_id, title in ((c1, "phq9", "PHQ-9"), (c2, "pcl5", "PCL-5")):
                 fig = charts.scale_figure(card.series, scale_id, title, start)
@@ -191,22 +194,22 @@ def render(card: PatientCard, today: date | None) -> None:
                     else:
                         st.plotly_chart(fig, config=theme.PLOTLY_CONFIG, key=f"chart-{scale_id}")
             st.caption("◇ — внеплановое заполнение (по триггеру правил). Полосы — градации тяжести из config/scales.")
-        with st.container(border=True):
+        with st.container(border=True, key="panel-wellbeing"):
             fig = charts.wellbeing_figure(card.series, start)
             if fig is None:
                 st.caption("Ежедневных чек-инов пока нет")
             else:
                 st.plotly_chart(fig, config=theme.PLOTLY_CONFIG, key="chart-wellbeing")
                 st.caption("Красная зона настроения ≤ 3 (3 дня → внеплановый PHQ-9); пунктир — сон 4 ч.")
-        with st.container(border=True):
+        with st.container(border=True, key="panel-doses"):
             st.html('<div class="mt-panel-title">Календарь приёма</div>')
             st.html(dose_strip_html(grid, start))
     with side:
-        with st.container(border=True):
+        with st.container(border=True, key="panel-alerts"):
             _alerts_feed(card.alerts, today)
-        with st.container(border=True):
+        with st.container(border=True, key="panel-summary"):
             _summary(card.summary)
-        with st.container(border=True):
+        with st.container(border=True, key="panel-safety"):
             _safety_plan(card.safety_plan)
-        with st.container(border=True):
+        with st.container(border=True, key="panel-prescriptions"):
             _prescriptions(card)

dashboard/console/api.py

@@ -103,6 +103,12 @@ class ConsoleApi:
     def reset(self) -> dict[str, Any]:
         return dict(self._request("POST", "/api/v1/demo/reset", timeout=60))
 
+    def demo_settings(self) -> dict[str, Any]:
+        return dict(self._request("GET", "/api/v1/demo/settings"))
+
+    def set_screening(self, enabled: bool) -> dict[str, Any]:
+        return dict(self._request("POST", "/api/v1/demo/settings", json={"screening_enabled": enabled}))
+
     # ------------------------------------------------------------------ patient dialog
     def send_event(
         self,

dashboard/console/playback_ui.py

@@ -51,7 +51,7 @@ def render_playback_panel(persona: dict[str, Any]) -> None:
     token = persona["token"]
     scenario, source = w.scenario_for(persona)
     state = w.playback_state(token)
-    st.subheader("🎬 Сценарий демо")
+    st.subheader("Сценарий демо")
     _scenario_source_controls(persona, source)
     if scenario.errors:
         with st.expander(f"⚠️ Замечания к сценарию: {len(scenario.errors)}", expanded=not scenario.days):

dashboard/console/widgets.py

@@ -368,10 +368,9 @@ def render_doctor_panel(highlight_token: str | None, *, key: str = "doc") -> Non
     ]
     for a in fresh_acute[:1]:
         st.markdown(
-            "<div style='background:#c62828;color:#fff;padding:18px 20px;border-radius:12px;"
-            "font-size:1.35rem;font-weight:700;margin-bottom:8px'>"
+            "<div class='mt-banner t-red mt-console-alert'><div>"
             f"🚨 ОСТРЫЙ АЛЕРТ · {_escape(a.get('patient_name', ''))} · {_fmt_ts(a.get('ts'))}<br>"
-            f"<span style='font-size:1.05rem;font-weight:500'>{_escape(a.get('reason', ''))}</span></div>",
+            f"<div class='mt-sub'>{_escape(a.get('reason', ''))}</div></div></div>",
             unsafe_allow_html=True,
         )
         c1, c2 = st.columns(2)

dashboard/pages/1b_Подключить_пациента.py

@@ -68,7 +68,7 @@ def _reset() -> None:
 
 def _result(res: ConnectPatientOut, name: str) -> None:
     link = h.absolute_link(res.deep_link, os.environ.get("DASHBOARD_PUBLIC_URL", "http://localhost:8501"))
-    with st.container(border=True):
+    with st.container(border=True, key="panel-connect-result"):
         st.html(
             f'<div class="mt-kicker">Пациент подключён · {escape(name)}</div>'
             '<div class="mt-sub" style="margin:4px 0 10px">Покажите QR пациенту или продиктуйте код — бот спросит '
@@ -108,7 +108,7 @@ def main() -> None:
     st.button("Заполнить демо-данными (Андрей)", icon=":material/auto_fix_high:", on_click=_fill_demo, args=(today,))
 
     left, right = st.columns(2, gap="large")
-    with left, st.container(border=True):
+    with left, st.container(border=True, key="panel-connect-person"):
         st.html('<div class="mt-panel-title">Пациент</div>')
         st.text_input("Имя и инициал фамилии", key="cp_name", placeholder="Андрей С.", help="Без полного ФИО")
         a, b = st.columns(2)
@@ -128,7 +128,7 @@ def main() -> None:
             key="cp_modules",
             format_func=lambda m: h.MODULE_LABELS[m],
         )
-    with right, st.container(border=True):
+    with right, st.container(border=True, key="panel-connect-prescription"):
         st.html('<div class="mt-panel-title">Назначение</div>')
         d1, d2 = st.columns([1.4, 1])
         with d1:

dashboard/pages/2_Демо-консоль.py

@@ -14,21 +14,45 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
 from console import widgets as w  # noqa: E402
 from console.playback_ui import render_playback_panel  # noqa: E402
 
+from dashboard import theme  # noqa: E402
+from dashboard.components import ui  # noqa: E402
+
 st.set_page_config(page_title="Демо-консоль", page_icon="🎬", layout="wide")
 
+st.html(theme.CSS)
+st.html(
+    ui.page_header(
+        "Демо-консоль",
+        "Чат пациента и сценарий сопровождения. Прототип на синтетических данных.",
+        kicker="Демонстрация",
+    )
+)
+
 persona = w.select_persona()
 w.render_connect_hero()
 w.show_flash()
 st.sidebar.caption(f"API: {w.get_api().base_url}")
 
-st.title("Демо-консоль")
-st.caption("Прототип на синтетических данных. Время — виртуальные демо-часы (ADR-005).")
-w.render_clock_bar(persona, key="console-clock")
-st.divider()
+# One switch for the presentation: initial questionnaires after consent on/off (off → straight to the chat).
+try:
+    _screening = bool(w.get_api().demo_settings().get("screening_enabled", True))
+    _wanted = st.sidebar.toggle(
+        "Начальный тест (скрининг)",
+        value=_screening,
+        help="Выключено: после согласия пациент сразу попадает в чат, без опросников.",
+    )
+    if _wanted != _screening:
+        w.get_api().set_screening(_wanted)
+        st.sidebar.success("Начальный тест " + ("включён" if _wanted else "выключен"))
+except Exception:  # noqa: BLE001 — the switch must never break the console
+    st.sidebar.caption("Переключатель теста недоступен (API)")
+
+with st.container(border=True, key="panel-console-clock"):
+    w.render_clock_bar(persona, key="console-clock")
 
 chat_col, play_col = st.columns([1.05, 1], gap="large")
-with chat_col:
-    st.subheader(f"💬 Чат пациента — {persona.get('display_name') or persona['token']}")
+with chat_col, st.container(border=True, key="panel-console-chat"):
+    st.subheader(f"Чат пациента — {persona.get('display_name') or persona['token']}")
     w.render_chat(persona["token"], height=560, key="console-chat")
-with play_col:
+with play_col, st.container(border=True, key="panel-console-scenario"):
     render_playback_panel(persona)

dashboard/pages/3_Сплит-экран.py

@@ -15,12 +15,12 @@ sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
 
 from console import widgets as w  # noqa: E402
 
+from dashboard import theme  # noqa: E402
+from dashboard.components import ui  # noqa: E402
+
 st.set_page_config(page_title="Сплит-экран", page_icon="🖥️", layout="wide", initial_sidebar_state="collapsed")
-st.markdown(
-    "<style>.block-container{padding-top:1.2rem;max-width:100%} "
-    "[data-testid='stMarkdownContainer'] p, [data-testid='stMarkdownContainer'] li{font-size:1.08rem}</style>",
-    unsafe_allow_html=True,
-)
+st.html(theme.CSS)
+st.html(ui.page_header("Сплит-экран", "Пациент и врач: сопровождение в реальном времени.", kicker="Демонстрация"))
 
 persona = w.select_persona()
 w.show_flash()
@@ -44,15 +44,15 @@ if clicked and nxt is not None:
     st.rerun()
 
 left, right = st.columns([1.1, 1], gap="large")
-with left:
+with left, st.container(border=True, key="panel-split-patient"):
     url = w.patient_web_url()
     if url:
-        st.markdown(f"#### 📱 Пациент — {persona.get('display_name') or token}")
+        st.markdown(f"#### Пациент — {persona.get('display_name') or token}")
         st.iframe(url, height=760)
         st.caption(f"Веб-приложение пациента: {url}")
     else:
-        st.markdown(f"#### 💬 Пациент — {persona.get('display_name') or token}")
+        st.markdown(f"#### Пациент — {persona.get('display_name') or token}")
         w.render_chat(token, height=600, key="split-chat")
-with right:
-    st.markdown("#### 🩺 Кабинет врача — д-р Смирнова Е. А.")
+with right, st.container(border=True, key="panel-split-doctor"):
+    st.markdown("#### Кабинет врача — д-р Смирнова Е. А.")
     w.polled_doctor_panel(token, key="split-doc")

dashboard/theme.py

@@ -138,8 +138,8 @@ div[class*="st-key-prow-"] {{ border-top:1px solid var(--line); padding:18px 0 4
 .mt-empty p {{ font-size:.85rem; margin:8px 0 0; }}
 
 /* ---------- chips, badges, bars ---------- */
-.mt-chip {{ display:inline-block; font-size:.72rem; padding: 1px 8px; border-radius: 999px; margin-right: 4px;
-  background: var(--blue-tint); color: var(--blue-ink); font-weight: 500; }}
+.mt-chip {{ display:inline-block; font-size:.72rem; padding: 1px 8px; border-radius: 5px; margin-right: 4px;
+  background: #edf3ee; color: #38634b; font-weight: 500; }}
 .mt-badge {{ display:inline-flex; align-items:center; gap:6px; font-size:.7rem; font-weight:550; padding: 5px 8px;
   border-radius: 5px; white-space: normal; }}
 .mt-badge.t-red {{ background: var(--red-tint); color: var(--red-ink); }}
@@ -166,7 +166,7 @@ div[class*="st-key-prow-"] {{ border-top:1px solid var(--line); padding:18px 0 4
 .mt-banner b {{ color: inherit; }}
 .mt-pulse {{ width: 12px; height: 12px; border-radius: 50%; background: var(--red); margin-top: 5px; flex: none;
   box-shadow: 0 0 0 4px #faeeee; }}
-.mt-panel-title {{ font-size: .8rem; text-transform: uppercase; letter-spacing: .07em; font-weight: 650;
+.mt-panel-title {{ font-size: .82rem; text-transform: uppercase; letter-spacing: .03em; font-weight: 550;
   color: var(--ink-3); margin: 2px 0 8px; }}
 .mt-alert {{ border-left: 4px solid var(--grey); padding: 4px 0 4px 12px; }}
 .mt-alert.t-red {{ border-left-color: var(--red); }}
@@ -175,9 +175,9 @@ div[class*="st-key-prow-"] {{ border-top:1px solid var(--line); padding:18px 0 4
 .mt-alert .when {{ color: var(--ink-3); font-size: .78rem; }}
 .mt-alert .why {{ color: var(--ink); font-weight: 600; font-size: .92rem; margin: 2px 0; }}
 .mt-alert.acked .why {{ color: var(--ink-3); font-weight: 500; }}
-.mt-summary {{ background: var(--blue-tint); border-radius: 12px; padding: 14px 16px; color: var(--ink);
-  line-height: 1.5; font-size: .93rem; }}
-.mt-summary .disc {{ color: var(--blue-ink); font-size: .76rem; margin-top: 8px; }}
+.mt-summary {{ background: #f4f6f2; border:1px solid var(--line); border-radius: 12px; padding: 14px 16px;
+  color: var(--ink); line-height: 1.5; font-size: .93rem; }}
+.mt-summary .disc {{ color: var(--ink-2); font-size: .76rem; margin-top: 8px; }}
 .mt-plan ol {{ margin: 0; padding-left: 1.2rem; }}
 .mt-plan li {{ margin: 3px 0; color: var(--ink); font-size: .9rem; }}
 
@@ -195,6 +195,21 @@ div[class*="st-key-prow-"] {{ border-top:1px solid var(--line); padding:18px 0 4
 .mt-legend i {{ display:inline-block; width: 12px; height: 12px; border-radius: 3px; margin-right: 5px;
   vertical-align: -1px; }}
 
+/* ---------- internal pages ---------- */
+div[class*="st-key-panel-"] {{ background:var(--surface); border:1px solid var(--line); border-radius:12px;
+  padding:20px; }}
+[data-testid="stExpander"] {{ background:var(--surface); border-color:var(--line); border-radius:12px; }}
+[data-testid="stChatMessage"] {{ background:#f4f6f2; border:1px solid var(--line); border-radius:12px; }}
+[data-testid="stChatInput"] {{ background:var(--surface); border:1px solid var(--line); border-radius:8px; }}
+[data-testid="stNumberInputContainer"], [data-testid="stDateInput"] input {{
+  background:var(--surface); color:var(--ink); border-color:var(--line); border-radius:8px; }}
+[data-testid="stApp"] [data-variant="pills"][data-selected] {{ background:#edf3ee; color:#285e4e;
+  border-color:#b4c6ba; }}
+.mt-console-alert {{ font-size:1.1rem; font-weight:600; }}
+.mt-console-alert .mt-sub {{ color:var(--red-ink); font-weight:400; }}
+.mt-dose-scroll {{ overflow-x:auto; padding-bottom:4px; }}
+.mt-doses {{ min-width:max-content; }}
+
 /* ---------- connect result ---------- */
 .mt-code {{ font-size: 2.4rem; font-weight: 700; letter-spacing: .12em; color: var(--ink);
   font-variant-numeric: tabular-nums; }}
@@ -222,6 +237,8 @@ div[class*="st-key-prow-"] {{ border-top:1px solid var(--line); padding:18px 0 4
   .mt-mobile-label {{ display:block; color:var(--ink-3); font-size:.68rem; margin-bottom:5px; }}
   .mt-row-details {{ grid-template-columns:1fr; gap:12px; }}
   .mt-list-footer {{ flex-wrap:wrap; }}
+  .mt-head .mt-side {{ text-align:left; }}
+  div[class*="st-key-panel-"] {{ padding:16px; }}
   .mt-tip:hover::after, .mt-tip:focus-visible::after {{ display:none; }}
 }}
 @media (prefers-reduced-motion: reduce) {{

data/emias/records.yaml

@@ -8,10 +8,12 @@
 #   allergies        — аллергии и непереносимости; `tags` сверяются с составом препаратов (вспомогательные вещества)
 #   other_medications — назначения других специалистов; `classes` сверяются со взаимодействиями
 #   pregnancy        — для женщин: беременность/лактация (null — нет данных)
+#   tokens           — демо-токены веб-приложения, которым принадлежит та же карта
 # Текущее психотропное назначение приходит из таблицы prescriptions (мок МИС, app/integrations/mis_mock.py).
 
 records:
   "900-000-013 34":   # Андрей С., 34, ветеран СВО — герой демо
+    tokens: [tok-andrey]   # демо-вход веб-приложения пациента
     diagnoses:
       - {code: "F43.1", name: "Посттравматическое стрессовое расстройство"}
       - {code: "F32.1", name: "Депрессивный эпизод средней степени"}

docker-compose.yml

@@ -47,5 +47,14 @@ services:
     ports:
       - "8501:8501"
 
+  web:
+    # Patient web app (React) — built inside Docker, no Node needed on the host.
+    build: ./web
+    depends_on:
+      app:
+        condition: service_healthy
+    ports:
+      - "8080:80"
+
 volumes:
   pgdata:

docs/DECISIONS.md

@@ -3,6 +3,26 @@
 > Живой журнал: **что** решили и **почему**. Новые записи — сверху. Одна запись = одно решение.
 > При конфликте с [`docs/info/SPEC.md`](info/SPEC.md) действует ADR.
 
+## 2026-10-31 — ADR-014: Голосовой ввод — расшифровка отдельно от диалога; OpenRouter только на синтетике
+**Контекст:** для гранта ВШЭ и пилота «Династия» нужен голосовой ввод (слайд 11: «Записать → Проверить → Сводка»). Голос — ПДн, а `scrub()` его не обезличивает. OpenRouter — зарубежный сервис (ADR-004, guardrail 9).
+**Решение:**
+1. `POST /me/{token}/stt` **только расшифровывает**. Текст возвращается пациенту на проверку и правку, в диалог уходит обычным сообщением, поэтому safety, `scrub()` и шлюз LLM работают как с набранным текстом. Автоотправки нет.
+2. Аудио не сохраняется и не логируется. В лог пишутся только длительность, провайдер, время ответа и длина текста.
+3. Формат — WAV 16 кГц mono, его кодирует браузер: ffmpeg на сервере не нужен, одинаково работает в Chrome и iOS Safari. Лимит — `STT_MAX_SECONDS` (60 с).
+4. Провайдеры за одним протоколом `SpeechToText` (`app/llm/stt.py`): `mock` (офлайн), `openrouter` (модель с аудиовходом, `input_audio`). **OpenRouter допустим только на синтетических данных** (демо, жюри, грант). До первого реального пациента нужен РФ-провайдер, SaluteSpeech (#121).
+**Последствия:** голос в демо работает без новых зависимостей. Для пилота нужен договор с SaluteSpeech.
+
+## 2026-10-04 — ADR-013: Правки по клиническому ревью — красные флаги, одно правило кризиса, звонок после кризиса, одна таблица номеров
+**Контекст:** ревью опросников: обязательны блок красных флагов и план безопасности со звонком после кризиса; маршрутизация должна сначала смотреть на риск и флаги, потом на шкалы; в документе два противоречащих правила кризисного экрана («любое Да» против «только высокий риск»); кризисные номера разбросаны по текстам (и чат-агент давал взрослым детский телефон доверия).
+**Решение:**
+1. Красные флаги (`config/instruments/red_flags.yaml`) идут первыми в этапе 1: голоса/видения, мысли причинить вред другому, доступ к оружию, насилие дома, сильное опьянение или отмена. Любое «Да» останавливает опрос: острый алерт врачу и экран помощи нужного варианта (`harm_others`, `domestic_violence`).
+2. Одно правило кризиса во всём приложении: кризисный экран открывается только при **высоком** риске C-SSRS. Умеренный риск — стоп опроса и психиатр, «ориентир 24 ч»; низкий — флаг врачу, план безопасности, опрос продолжается. Пункт 9 PHQ-9 при отрицательном C-SSRS — флаг «расхождение» и план безопасности; без пройденного C-SSRS (плановый PHQ-9, дневник) — сразу C-SSRS, решает его уровень. `config/questionnaire.yaml`: `cssrs_positive_policy: graded` (C01).
+3. После любого кризиса дежурному создаётся задача «перезвонить пациенту» (алерт `post_crisis_call`, ориентир 24 ч; «позвонили» — подтверждение алерта), одна на пациента, пока открыта. На следующий день приложение само спрашивает «как вы» с кнопкой плана безопасности (`app/services/followup.py`).
+4. Маршрут: риск и флаги → шкалы. PCL-5 ≥50 убран из приоритетной записи (и в правилах дашборда `pcl5_severe` — уровень moderate): высокий PCL-5 ведёт к плановой записи к специалисту.
+5. Все телефоны — только в `config/texts/crisis.yaml`; остальные тексты и промпт агента ссылаются через `{phone:<ключ>}`, неизвестный ключ роняет загрузку, тест не пускает «голые» номера в код и тексты.
+**Причина:** врач назвал пункты 1–2 обязательными; одно правило кризиса исключает расхождение демо с документом; неверный номер на кризисном экране недопустим.
+**Последствия:** формулировки красных флагов — черновик команды, утверждает врач (#8). Старое правило дашборда `phq9_item9` (ADR-006, сценарий без опросника) по-прежнему поднимает кризис по п.9 — свести при следующем ревью правил.
+
 ## 2026-10-04 — ADR-012: Два этапа опросов — общий скрининг, предварительный результат пациенту, дневник по результату
 **Контекст:** PDF «Опросники для MVP» (этап 1) и «Этап 2: наблюдение после диагноза». Пациент должен видеть, что у него всё в порядке или есть признаки депрессии/ПТСР, и дальше вести дневник под свой результат. Клиническая политика (`config/questionnaire.yaml`) закрыта до решений врача C01–C09, поэтому в демо опрос всегда заканчивался остановкой «нужна оценка специалиста».
 **Решение:** этап 1 общий для всех и идёт в фиксированном порядке документа: PHQ-2 → фильтр травмы/TSQ → RUS-AUDIT-S → C-SSRS → общие вопросы, затем этап 2 только там, где этап 1 положительный. Пункты внутри шкал не переставляются. Стоп-правила риска срабатывают после каждого ответа. После этапа 1 правила (`app/core/interpretation.py`, тексты в `config/texts/results.yaml`) показывают пациенту предварительный результат: «не выявлено», «предположение депрессии» (PHQ-9 ≥10), «склонность к ПТСР» (PCL-5 ≥31), алкоголь, ЧМТ, маршрут. Это всегда «результат скрининга, не диагноз». Результат включает модули дневника: `depression` и `veteran` (ПТСР). Без врача их определяет скрининг, модули врача сохраняются и дополняются. Дневник этапа 2 состоит из двух частей: утро (часы сна, качество сна, порции алкоголя) и вечер (настроение, лекарства + модули: 3 пункта PHQ-9 в день по кругу, короткая PCL-5 из 4 пунктов). C-SSRS «с последнего контакта» задаётся каждый вечер только при суицидальном риске. Флаги считаются по недельным средним против личной нормы за первые 14 дней и выносят полную шкалу раньше срока. Каждая запись хранит интерпретацию, показанную при завершении; журнал отдаёт `/me/{token}/results`. Демо (`INTAKE_DEMO=1` + демо-часы) читает отдельный `config/questionnaire_demo.yaml`, помеченный «не клиническое согласование», а реальная политика не меняется.

docs/clinical/README.md

@@ -160,7 +160,7 @@ q6 за всю жизнь всем; при положительном q6 уто
 
 | ID | Решение, которое нужно зафиксировать | Статус |
 |---|---|---|
-| C01 | Любой «Да» останавливает обычный опрос или разрешено продолжение после низкого риска? Различить обычный опрос, уточнение риска и кризисную помощь. Новый PDF с.1 против с.12; в старом с.1 против с.11. Скриншот поддерживает остановку. | pending |
+| C01 | Любой «Да» останавливает обычный опрос или разрешено продолжение после низкого риска? Различить обычный опрос, уточнение риска и кризисную помощь. Новый PDF с.1 против с.12; в старом с.1 против с.11. Скриншот поддерживает остановку. | **решено 2026-10-04 (ADR-013):** кризисный экран только при высоком риске; умеренный — стоп и психиатр (ориентир 24 ч); низкий — флаг, план безопасности, продолжение. `cssrs_positive_policy: graded`. Подпись врача — в #8. |
 | C02 | Использовать сумму ≥1 или согласованную границу зоны для расширения RUS-AUDIT-S? Не переносить порог AUDIT-C ≥3/≥4. | pending |
 | C03 | Самозаполнение полного RUS-AUDIT, таблица стандартных порций, порядок дополнения и повторные замеры. | pending |
 | C04 | Новая спецификация: TSQ ≥6 вместо PC-PTSD-5 ≥3; получить приложение Г1, согласовать применение и PCL-5 ≥31 для популяции. TSQ из приложения Г1 редакции 2023 года уже добавлен; актуальность редакции и применение требуют подтверждения. | pending |

docs/clinical/crisis-contacts.md

@@ -1,5 +1,8 @@
 # Проверка контактов: 3 октября 2026
 
+> **Единая таблица номеров приложения — `config/texts/crisis.yaml` (`resources`).** Остальные тексты и промпт
+> чат-агента ссылаются на неё через `{phone:<ключ>}` (ADR-013); номер меняется в одном месте.
+
 Проверены официальные страницы, **не телефонное соединение**. Никаких звонков
 или уведомлений службам/врачу не выполнялось.
 

infra/deploy/deploy.sh

@@ -1,24 +1,36 @@
 #!/usr/bin/env bash
-# Deploy origin/dev to the shared server (ssh alias `vobot`, dir /srv/ptsr). Run from a clean checkout:
+# Deploy a git ref to the shared server (dir /srv/ptsr). Used by hand and by CD (.gitea/workflows/deploy.yml).
 #   infra/deploy/deploy.sh [git-ref]        default: origin/dev
+# Env: DEPLOY_HOST (ssh host/alias, default `vobot`), DEPLOY_SSH_OPTS (extra ssh/scp options).
 # Keeps /srv/ptsr/.env (secrets live only on the server) and the Postgres volume.
+# The host drops some ssh connections (rate limiting) — every remote step is retried.
 set -euo pipefail
 REF=${1:-origin/dev}
 HOST=${DEPLOY_HOST:-vobot}
+OPTS=${DEPLOY_SSH_OPTS:-}
 DIR=/srv/ptsr
-git fetch -q origin || true
+git fetch -q origin 2>/dev/null || true
 SHA=$(git rev-parse --short "$REF")
+TGZ=$(mktemp --suffix=.tgz)
+trap 'rm -f "$TGZ"' EXIT
+git archive --format=tar.gz -o "$TGZ" "$REF"
+
+retry() { local n=0; until "$@"; do n=$((n + 1)); [ $n -ge 5 ] && return 1; echo "retry $n: $1" >&2; sleep $((n * 10)); done; }
+
 echo "deploy $REF ($SHA) -> $HOST:$DIR"
-git archive --format=tar "$REF" | ssh "$HOST" "rm -rf $DIR/app.new && mkdir -p $DIR/app.new && tar -x -C $DIR/app.new"
-ssh "$HOST" bash -s <<SH
+# shellcheck disable=SC2086
+retry scp -q $OPTS "$TGZ" "$HOST:$DIR/release.tgz"
+# shellcheck disable=SC2086
+retry ssh $OPTS "$HOST" "bash -s" <<SH
 set -euo pipefail
 cd $DIR
+rm -rf app.new && mkdir app.new && tar -xzf release.tgz -C app.new && rm -f release.tgz
 cp .env app.new/.env
 rm -rf app.old; [ -d app ] && mv app app.old; mv app.new app
+chmod 755 $DIR
 cd app
-echo "$SHA" > REVISION
-docker run --rm -v "\$PWD/web:/w" -w /w node:20-alpine sh -c "npm ci --no-audit --no-fund >/dev/null && npm run build >/dev/null"
+docker run --rm --user "\$(id -u):\$(id -g)" -e HOME=/tmp -e npm_config_cache=/tmp/.npm -v "\$PWD/web:/w" -w /w node:20-alpine sh -c "npm ci --no-audit --no-fund >/dev/null && npm run build >/dev/null"
 docker compose -p ptsr -f docker-compose.yml -f infra/deploy/docker-compose.server.yml up -d --build </dev/null
 for i in \$(seq 1 40); do curl -sf 127.0.0.1:8010/health >/dev/null && break; sleep 3; done
-curl -sf 127.0.0.1:8010/health && echo " api ok"
+curl -sf 127.0.0.1:8010/health >/dev/null && echo "$SHA" > REVISION && echo "api ok @ $SHA"
 SH

infra/deploy/docker-compose.server.yml

@@ -23,8 +23,12 @@ services:
       - --browser.gatherUsageStats=false
     environment:
       API_BASE_URL: http://app:8000
-      DASHBOARD_PUBLIC_URL: ${PUBLIC_URL:-http://localhost}/doctor
+      DASHBOARD_PUBLIC_URL: ${PUBLIC_URL:-https://dev.meditron.tenvex.ru}/doctor
     ports: !override
       - "127.0.0.1:8511:8501"
     restart: unless-stopped
     mem_limit: 768m
+
+  web:
+    # On the server the host nginx serves the web bundle itself (deploy.sh builds web/dist).
+    profiles: ["local-only"]

infra/deploy/nginx-meditron.conf

@@ -1,8 +1,9 @@
-# meditron.tenvex.ru — PTSRaiAgent prototype (synthetic data only).
+# PTSRaiAgent prototype (synthetic data only) — shared host with vobot.
 #   /          patient web app (static, web/dist)
 #   /api/      FastAPI (127.0.0.1:8010)
 #   /doctor/   Streamlit doctor dashboard + demo console (127.0.0.1:8511, baseUrlPath=doctor)
-# TLS: after the DNS A-record exists — certbot --nginx -d meditron.tenvex.ru
+# dev.meditron.tenvex.ru — HTTPS (Let's Encrypt, certbot certonly --nginx, auto-renew).
+# meditron.tenvex.ru     — HTTPS (same stack for now; will serve the main branch after promotion).
 
 map $http_upgrade $ptsr_connection_upgrade {
     default upgrade;
@@ -12,7 +13,73 @@ map $http_upgrade $ptsr_connection_upgrade {
 server {
     listen 80;
     listen [::]:80;
+    server_name dev.meditron.tenvex.ru;
+    return 301 https://$host$request_uri;
+}
+
+server {
+    listen 443 ssl;
+    listen [::]:443 ssl;
+    server_name dev.meditron.tenvex.ru;
+
+    ssl_certificate     /etc/letsencrypt/live/dev.meditron.tenvex.ru/fullchain.pem;
+    ssl_certificate_key /etc/letsencrypt/live/dev.meditron.tenvex.ru/privkey.pem;
+    include             /etc/letsencrypt/options-ssl-nginx.conf;
+    ssl_dhparam         /etc/letsencrypt/ssl-dhparams.pem;
+
+    access_log /var/log/nginx/ptsr-access.log;
+    client_max_body_size 5m;
+
+    root /srv/ptsr/app/web/dist;
+    index index.html;
+
+    location / {
+        try_files $uri $uri/ /index.html;
+    }
+
+    location /api/ {
+        proxy_pass http://127.0.0.1:8010;
+        proxy_set_header Host $host;
+        proxy_set_header X-Real-IP $remote_addr;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_read_timeout 120s;   # chat agent + scheduler replay on clock advance
+    }
+
+    location ~ ^/(docs|openapi.json|health)$ {
+        proxy_pass http://127.0.0.1:8010;
+        proxy_set_header Host $host;
+    }
+
+    location /doctor/ {
+        proxy_pass http://127.0.0.1:8511;
+        proxy_http_version 1.1;
+        proxy_set_header Upgrade $http_upgrade;
+        proxy_set_header Connection $ptsr_connection_upgrade;
+        proxy_set_header Host $host;
+        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
+        proxy_set_header X-Forwarded-Proto $scheme;
+        proxy_read_timeout 86400s;
+        proxy_buffering off;
+    }
+}
+
+server {
+    listen 80;
+    listen [::]:80;
     server_name meditron.tenvex.ru;
+    return 301 https://$host$request_uri;
+}
+
+server {
+    listen 443 ssl;
+    listen [::]:443 ssl;
+    server_name meditron.tenvex.ru;
+
+    ssl_certificate     /etc/letsencrypt/live/meditron.tenvex.ru/fullchain.pem;
+    ssl_certificate_key /etc/letsencrypt/live/meditron.tenvex.ru/privkey.pem;
+    include             /etc/letsencrypt/options-ssl-nginx.conf;
+    ssl_dhparam         /etc/letsencrypt/ssl-dhparams.pem;
 
     access_log /var/log/nginx/ptsr-access.log;
     client_max_body_size 5m;

pyproject.toml

@@ -20,6 +20,7 @@ dependencies = [
     "pandas>=2.2",
     "qrcode[pil]>=7.4",
     "tzdata>=2026.5",
+    "aiohttp-socks>=0.8",
 ]
 
 [dependency-groups]

tests/conftest.py

@@ -12,6 +12,8 @@ def _no_live_llm(monkeypatch: pytest.MonkeyPatch):
     monkeypatch.setenv("CHAT_AGENT", "off")
     monkeypatch.setenv("OPENROUTER_API_KEY", "")
     monkeypatch.setenv("LLM_PROVIDER", "mock")
+    # Tests cover the clinical protocol by default (blocking crisis); demo-mode tests flip it explicitly.
+    monkeypatch.setenv("CRISIS_BLOCKING", "true")
     get_settings.cache_clear()
     yield
     get_settings.cache_clear()

tests/test_agent_stream.py

@@ -0,0 +1,83 @@
+"""Live agent output: OpenRouter stream parsing and the SSE endpoint (status → delta → done)."""
+
+import asyncio
+import json
+
+import httpx
+import pytest
+from httpx import ASGITransport, AsyncClient
+
+from app.config import Settings
+from app.db import session
+from app.db.bootstrap import init_db
+from app.llm import agent
+from app.llm.agent import stream
+from app.llm.agent.providers import OpenRouterProvider
+from app.main import app
+
+MEM = "sqlite+aiosqlite:///:memory:"
+
+
+def _sse(*chunks: dict) -> bytes:
+    lines = [": OPENROUTER PROCESSING", ""]
+    lines += [f"data: {json.dumps(c, ensure_ascii=False)}" for c in chunks] + ["data: [DONE]", ""]
+    return "\n\n".join(lines).encode()
+
+
+async def test_openrouter_stream_deltas_and_tool_calls():
+    body = _sse(
+        {"choices": [{"delta": {"content": "Сейчас "}}]},
+        {
+            "choices": [
+                {
+                    "delta": {
+                        "tool_calls": [{"index": 0, "id": "c1", "function": {"name": "get_patient_", "arguments": ""}}]
+                    }
+                }
+            ]
+        },
+        {"choices": [{"delta": {"tool_calls": [{"index": 0, "function": {"name": "record", "arguments": "{}"}}]}}]},
+    )
+    client = httpx.AsyncClient(transport=httpx.MockTransport(lambda r: httpx.Response(200, content=body)))
+    provider = OpenRouterProvider(Settings(_env_file=None, openrouter_api_key="k"), client=client)
+    q: asyncio.Queue = asyncio.Queue()
+    token = stream.STREAM.set(q)
+    try:
+        turn = await provider.chat([{"role": "user", "content": "x"}], [])
+    finally:
+        stream.STREAM.reset(token)
+    assert [c.name for c in turn.tool_calls] == ["get_patient_record"]
+    events = [q.get_nowait() for _ in range(q.qsize())]
+    assert ("delta", "Сейчас ") in events and ("reset", None) in events
+
+
+@pytest.fixture
+async def client():
+    await session.configure(identity_url=MEM, clinical_url=MEM)
+    await init_db()
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://t") as c:
+        yield c
+    await session.configure()
+
+
+async def test_sse_endpoint_streams_steps_then_final(client, monkeypatch):
+    async def fake(token, text, **kw):
+        stream.emit("status", stream.TOOL_STATUS["get_patient_record"])
+        stream.emit("delta", "Ответ ")
+        stream.emit("delta", "агента")
+        return agent.AgentAnswer(text="Ответ агента", sources=[], tools=["get_patient_record"], provider="fake")
+
+    monkeypatch.setattr(agent, "answer_chat", fake)
+    token = (await client.post("/api/v1/auth/demo/fresh", json={})).json()["token"]
+    kinds = []
+    async with client.stream(
+        "POST", f"/api/v1/dialog/{token}/events/stream", json={"type": "text", "text": "Можно ибупрофен?"}
+    ) as resp:
+        assert resp.headers["content-type"].startswith("text/event-stream")
+        async for line in resp.aiter_lines():
+            if line.startswith("event:"):
+                kinds.append(line[6:].strip())
+            if line.startswith("data:") and kinds and kinds[-1] == "done":
+                final = json.loads(line[5:])
+    assert kinds[0] == "status" and "delta" in kinds and kinds[-1] == "done"
+    assert any("Ответ агента" in r["text"] for r in final)

tests/test_api_patient_web.py

@@ -12,6 +12,7 @@ from app.db import identity_repo, repo, session
 from app.db.bootstrap import init_db
 from app.db.clinical import Dose, Prescription
 from app.main import app
+from app.services.anonymous_care import doctor_id as anonymous_doctor
 
 MEM = "sqlite+aiosqlite:///:memory:"
 TOKEN = "b" * 32
@@ -48,17 +49,17 @@ async def test_anonymous_tokens_are_fresh_and_store_no_personal_data(client: Asy
     async with session.get_identity_session() as s:
         assert await identity_repo.get_patient(s, first["token"]) is None
     async with session.get_clinical_session() as s:
-        # Only a doctor-less pseudonymous profile: the dialog API serves known profiles only.
+        # The care team sees a pseudonymous profile without an identity record.
         profile = await repo.get_profile(s, first["token"])
-        assert profile.doctor_id == "" and profile.modules == []
+        assert profile.doctor_id == anonymous_doctor() and profile.modules == []
 
 
-async def test_anonymous_focus_picks_the_module_without_a_doctor(client: AsyncClient) -> None:
+async def test_anonymous_focus_does_not_remove_access_to_doctor(client: AsyncClient) -> None:
     token = (await client.post("/api/v1/auth/anonymous", json={"focus": "ptsd"})).json()["token"]
     async with session.get_clinical_session() as s:
         profile = await repo.get_profile(s, token)
-        assert profile.modules == ["veteran"] and not profile.doctor_id
-    assert (await client.get(f"/api/v1/me/{token}/today")).json()["has_doctor"] is False
+        assert profile.modules == ["veteran"] and profile.doctor_id == anonymous_doctor()
+    assert (await client.get(f"/api/v1/me/{token}/today")).json()["has_doctor"] is True
     assert (await client.post("/api/v1/auth/anonymous", json={"focus": "flu"})).status_code == 422
 
 
@@ -86,10 +87,14 @@ async def test_emias_rejects_unknown_code(client: AsyncClient) -> None:
 
 async def test_emias_demo_persona(client: AsyncClient) -> None:
     persona = fixtures.DEMO_PERSONAS[0]
-    res = await client.post("/api/v1/auth/emias/demo", json={"persona_token": persona.token})
-    assert res.json() == {"token": persona.token, "mode": "emias", "display_name": persona.display_name}
+    res = (await client.post("/api/v1/auth/emias/demo", json={"persona_token": persona.token})).json()
+    # Every visitor gets a private copy of the persona, never the shared account.
+    assert res["token"].startswith(persona.token + "-") and res["mode"] == "emias"
+    assert res["display_name"] == persona.display_name
+    again = (await client.post("/api/v1/auth/emias/demo", json={"persona_token": persona.token})).json()
+    assert again["token"] != res["token"]
     async with session.get_clinical_session() as s:
-        assert (await repo.get_profile(s, persona.token)).doctor_id == fixtures.DOCTOR_ID
+        assert (await repo.get_profile(s, res["token"])).doctor_id == fixtures.DOCTOR_ID
     missing = await client.post("/api/v1/auth/emias/demo", json={"persona_token": "tok-nobody"})
     assert missing.status_code == 404
 
@@ -122,7 +127,7 @@ async def test_today_plan_and_week_marks(client: AsyncClient) -> None:
 async def test_today_for_anonymous_is_empty(client: AsyncClient) -> None:
     token = (await client.post("/api/v1/auth/anonymous")).json()["token"]
     body = (await client.get(f"/api/v1/me/{token}/today")).json()
-    assert body["has_doctor"] is False and body["doses"] == [] and body["prescriptions"] == []
+    assert body["has_doctor"] is True and body["doses"] == [] and body["prescriptions"] == []
 
 
 async def test_crisis_resources_come_from_config(client: AsyncClient) -> None:
@@ -140,8 +145,8 @@ async def test_today_links_prescription_to_reference_and_therapy_day(client: Asy
 
 
 async def test_demo_persona_week_mirrors_dashboard_story(client: AsyncClient) -> None:
-    persona = fixtures.DEMO_PERSONAS[0].token
-    await client.post("/api/v1/auth/emias/demo", json={"persona_token": persona})
+    auth = await client.post("/api/v1/auth/emias/demo", json={"persona_token": fixtures.DEMO_PERSONAS[0].token})
+    persona = auth.json()["token"]
     rx = (await client.get(f"/api/v1/me/{persona}/today")).json()["prescriptions"][0]
     assert rx["ref_id"] == "sertraline" and rx["therapy_day"] > 1
 
@@ -158,8 +163,8 @@ async def test_questionnaires_are_off_without_release_or_demo_flag(
 
 async def test_demo_flag_opens_intake_in_the_engine(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("INTAKE_DEMO", "1")
-    persona = fixtures.DEMO_PERSONAS[0].token
-    await client.post("/api/v1/auth/emias/demo", json={"persona_token": persona})
+    auth = await client.post("/api/v1/auth/emias/demo", json={"persona_token": fixtures.DEMO_PERSONAS[0].token})
+    persona = auth.json()["token"]
 
     async def press(button: str | None = None) -> list[str]:
         event = {"type": "button", "button": button} if button else {"type": "start"}
@@ -169,8 +174,8 @@ async def test_demo_flag_opens_intake_in_the_engine(client: AsyncClient, monkeyp
     await press()
     await press("consent:yes")
     buttons = await press("age:adult")
-    # Stage 1 starts with PHQ-2, in the order of the MVP questionnaire.
-    assert any(b.startswith("iq:") and ":phq9_1:" in b for b in buttons)
+    # Stage 1 starts with the red-flag block, then the order of the MVP questionnaire.
+    assert any(b.startswith("iq:") and ":rf_psychosis:" in b for b in buttons)
     assert "intake:help" in buttons and "intake:pause" in buttons
     q = (await client.get(f"/api/v1/me/{persona}/today")).json()["questionnaires"]
     assert q["enabled"] and q["intake"] == "active"
@@ -214,9 +219,73 @@ async def test_stage_one_starts_without_modules(client: AsyncClient, monkeypatch
     assert (await client.get(f"/api/v1/me/{token}/today")).json()["questionnaires"]["enabled"]
     await press()
     await press("consent:yes")
-    assert any(":phq9_1:" in b for b in await press("age:adult"))
+    assert any(":rf_psychosis:" in b for b in await press("age:adult"))
 
 
 async def test_patient_app_has_its_own_demo_personas(client: AsyncClient) -> None:
     body = (await client.get("/api/v1/patient/demo-personas")).json()
     assert [p["token"] for p in body] == [p.token for p in fixtures.DEMO_PERSONAS]
+
+
+async def test_anonymous_is_visible_only_to_assigned_doctor(client: AsyncClient) -> None:
+    first = (await client.post("/api/v1/auth/anonymous")).json()["token"]
+    second = (await client.post("/api/v1/auth/anonymous")).json()["token"]
+    items = (await client.get(f"/api/v1/doctors/{anonymous_doctor()}/patients")).json()
+    names = {p["token"]: p["display_name"] for p in items}
+    assert names[first].startswith("Пациент ") and names[first] != names[second]
+    assert first not in names[first]
+    own = await client.get(f"/api/v1/patients/{first}", params={"doctor_id": anonymous_doctor()})
+    assert own.status_code == 200 and own.json()["patient"]["display_name"] == names[first]
+    denied = await client.get(f"/api/v1/patients/{first}", params={"doctor_id": "another-doctor"})
+    assert denied.status_code == 403
+    assert (await client.get("/api/v1/doctors/another-doctor/patients")).json() == []
+
+
+async def test_old_anonymous_profile_requires_explicit_care_opt_in(client: AsyncClient) -> None:
+    async with session.get_clinical_session() as s:
+        await repo.upsert_profile(s, token=TOKEN, doctor_id="", modules=["veteran"])
+    # Merely opening Today must not reveal an old private profile to a doctor.
+    assert not (await client.get(f"/api/v1/me/{TOKEN}/today")).json()["has_doctor"]
+    assert (await client.get(f"/api/v1/doctors/{anonymous_doctor()}/patients")).json() == []
+    for _ in range(2):
+        assert (await client.post(f"/api/v1/me/{TOKEN}/doctor/connect")).json() == {"has_doctor": True}
+    async with session.get_clinical_session() as s:
+        p = await repo.get_profile(s, TOKEN)
+        assert p.doctor_id == anonymous_doctor() and p.modules == ["veteran"]
+    async with session.get_identity_session() as s:
+        assert await identity_repo.get_patient(s, TOKEN) is None
+
+
+async def test_connect_care_does_not_reassign_or_create_patients(client: AsyncClient) -> None:
+    assert (await client.post(f"/api/v1/me/{TOKEN}/doctor/connect")).status_code == 404
+    async with session.get_clinical_session() as s:
+        await repo.upsert_profile(s, token=TOKEN, doctor_id="own-doctor")
+    assert (await client.post(f"/api/v1/me/{TOKEN}/doctor/connect")).status_code == 200
+    async with session.get_clinical_session() as s:
+        assert (await repo.get_profile(s, TOKEN)).doctor_id == "own-doctor"
+
+
+async def test_screening_capability_respects_release_gate(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("INTAKE_DEMO", "0")
+    assert (await client.get("/api/v1/patient/capabilities")).json()["primary_screening"] is False
+    monkeypatch.setenv("INTAKE_DEMO", "1")
+    assert (await client.get("/api/v1/patient/capabilities")).json()["primary_screening"] is True
+
+
+async def test_anonymous_contact_request_reaches_doctor(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("INTAKE_DEMO", "0")
+    token = (await client.post("/api/v1/auth/anonymous")).json()["token"]
+    for event in [
+        {"type": "start"},
+        {"type": "button", "button": "consent:yes"},
+        {"type": "button", "button": "age:adult"},
+        {"type": "button", "button": "menu:contact_doctor"},
+    ]:
+        response = await client.post(f"/api/v1/dialog/{token}/events", json=event)
+        assert response.status_code == 200
+    items = (await client.get(f"/api/v1/doctors/{anonymous_doctor()}/patients")).json()
+    assert items[0]["last_event"] == "Пациент просит врача связаться"
+    card = (await client.get(f"/api/v1/patients/{token}", params={"doctor_id": anonymous_doctor()})).json()
+    assert card["patient"]["last_event"] == items[0]["last_event"]
+    async with session.get_identity_session() as s:
+        assert await identity_repo.get_patient(s, token) is None

tests/test_api_stt.py

@@ -0,0 +1,71 @@
+"""Voice input endpoint (#116): transcribes only, validates the audio, never touches the dialog."""
+
+from collections.abc import AsyncIterator
+
+import pytest
+from httpx import ASGITransport, AsyncClient
+from sqlalchemy import select
+
+from app.api import patient_web
+from app.db import session
+from app.db.bootstrap import init_db
+from app.db.clinical import Message
+from app.llm.stt import MockSTT, SttError, SttNoSpeech, load_texts
+from app.main import app
+from tests.test_llm_stt import make_wav
+
+MEM = "sqlite+aiosqlite:///:memory:"
+WAV = {"Content-Type": "audio/wav"}
+
+
+@pytest.fixture
+async def client(monkeypatch: pytest.MonkeyPatch) -> AsyncIterator[AsyncClient]:
+    monkeypatch.setattr(patient_web, "get_stt", lambda _settings: MockSTT())  # never the network in tests
+    await session.configure(identity_url=MEM, clinical_url=MEM)
+    await init_db()
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
+        yield c
+    await session.configure()
+
+
+async def _token(client: AsyncClient) -> str:
+    return (await client.post("/api/v1/auth/anonymous")).json()["token"]
+
+
+async def test_transcribes_without_writing_to_dialog(client: AsyncClient) -> None:
+    token = await _token(client)
+    resp = await client.post(f"/api/v1/me/{token}/stt", content=make_wav(2), headers=WAV)
+    assert resp.status_code == 200
+    assert resp.json() == {"text": load_texts().mock_transcript, "provider": "mock"}
+    async with session.get_clinical_session() as s:
+        assert (await s.execute(select(Message))).scalars().all() == []
+
+
+async def test_capability_and_off(client: AsyncClient, monkeypatch: pytest.MonkeyPatch) -> None:
+    assert (await client.get("/api/v1/patient/capabilities")).json()["voice_input"] is True
+    monkeypatch.setattr(patient_web, "get_stt", lambda _settings: None)
+    assert (await client.get("/api/v1/patient/capabilities")).json()["voice_input"] is False
+    token = await _token(client)
+    assert (await client.post(f"/api/v1/me/{token}/stt", content=make_wav(1), headers=WAV)).status_code == 404
+
+
+async def test_rejects_unknown_patient_bad_format_and_long_audio(client: AsyncClient) -> None:
+    assert (await client.post("/api/v1/me/nobody/stt", content=make_wav(1), headers=WAV)).status_code == 404
+    token = await _token(client)
+    url = f"/api/v1/me/{token}/stt"
+    assert (await client.post(url, content=b"OggS" + b"\0" * 200)).status_code == 415
+    assert (await client.post(url, content=make_wav(70), headers=WAV)).status_code == 413
+    assert (await client.post(url, content=make_wav(10, rate=48_000), headers=WAV)).status_code == 200
+
+
+@pytest.mark.parametrize(("error", "status"), [(SttNoSpeech("x"), 422), (SttError("x"), 503)])
+async def test_provider_errors(client: AsyncClient, monkeypatch: pytest.MonkeyPatch, error, status) -> None:
+    class Failing:
+        name = "failing"
+
+        async def transcribe(self, audio: bytes, fmt: str = "wav") -> str:
+            raise error
+
+    monkeypatch.setattr(patient_web, "get_stt", lambda _settings: Failing())
+    token = await _token(client)
+    assert (await client.post(f"/api/v1/me/{token}/stt", content=make_wav(1), headers=WAV)).status_code == status

tests/test_contracts_settings_privacy.py

@@ -0,0 +1,21 @@
+"""Startup validation errors must not print configuration credentials."""
+
+import pytest
+from pydantic import ValidationError
+
+from app.config import Settings
+
+
+def test_incomplete_max_configuration_does_not_expose_token():
+    with pytest.raises(ValidationError) as error:
+        Settings(_env_file=None, max_enabled=True, max_bot_token="SYNTHETIC_SECRET")
+    message = str(error.value)
+    assert "MAX_BOT_USERNAME" in message
+    assert "SYNTHETIC_SECRET" not in message and "input_value" not in message
+
+
+def test_invalid_field_does_not_expose_accidentally_pasted_secret():
+    with pytest.raises(ValidationError) as error:
+        Settings(_env_file=None, max_enabled="SYNTHETIC_SECRET")
+    assert "max_enabled" in str(error.value)
+    assert "SYNTHETIC_SECRET" not in str(error.value)

tests/test_crisis_numbers.py

@@ -0,0 +1,31 @@
+"""Crisis numbers live in one table (config/texts/crisis.yaml); every other text refers to it by `{phone:<key>}`."""
+
+import re
+from pathlib import Path
+
+import pytest
+
+from app.core.safety import CRISIS_PATH, fill_phones
+from app.llm.agent.prompts import SYSTEM_PROMPT
+
+ROOT = Path(__file__).resolve().parents[1]
+# Any phone number except the universal short emergency numbers 112 / 101–103.
+NUMBER = re.compile(r"(\+7|\b8)[\s(-]*\d{3}[\s)-]*\d{3}[\s-]*\d{2}[\s-]*\d{2}")
+
+
+def test_no_number_outside_the_crisis_table():
+    texts = [p for p in (ROOT / "config" / "texts").glob("*.yaml") if p != CRISIS_PATH]
+    code = [*(ROOT / "app").rglob("*.py"), *(ROOT / "web" / "src").rglob("*.ts*")]
+    offenders = [
+        str(p.relative_to(ROOT))
+        for p in [*texts, *code]
+        if p.name != "privacy.py"  # the PII scrubber documents the phone format with an example
+        and NUMBER.search("\n".join(line for line in p.read_text(encoding="utf-8").splitlines() if "#" not in line))
+    ]
+    assert offenders == []
+
+
+def test_references_resolve_and_unknown_keys_fail():
+    assert "989-50-50" in fill_phones("{phone:helpline}") and "989-50-50" in SYSTEM_PROMPT
+    with pytest.raises(ValueError):
+        fill_phones("{phone:no_such_line}")

tests/test_db.py

@@ -79,6 +79,7 @@ async def test_create_all_puts_tables_in_the_right_db(dbs: None) -> None:
         "questionnaire_notices",
         "questionnaire_tasks",
         "scheduler_events",
+        "demo_settings",
     } == clinical_tables
     await init_db()  # idempotent
 

tests/test_demo_crisis_deferred.py

@@ -0,0 +1,36 @@
+"""Demo mode (crisis_blocking off): a crisis signal inside a questionnaire does not stop it; the crisis message
+comes at the end and blocks nothing; the weapon question triggers nothing."""
+
+import pytest
+
+from app.db.clinical import Alert
+from tests.test_dialog_intake import answer_button, complete_existing, press, released, rows  # noqa: F401
+
+
+@pytest.fixture
+def non_blocking(monkeypatch):
+    from app.config import get_settings
+
+    monkeypatch.setattr(get_settings(), "crisis_blocking", False)
+
+
+async def test_red_flag_does_not_stop_the_questionnaire(released, non_blocking):  # noqa: F811
+    replies = await press(released, "intake:start")
+    replies = await press(released, answer_button(replies, {"rf_psychosis": 0}))  # «Да» on a red flag
+    assert not any(r.kind == "crisis" for r in replies), "no crisis window in the middle of the test"
+    assert any(r.buttons for r in replies), "the next question is asked"
+    (row,) = await rows()
+    assert row.status != "interrupted" and row.details["pending_urgent"]["code"] == "red_flag_psychosis"
+    replies = await complete_existing(released, replies, {})
+    assert any(r.kind == "crisis" for r in replies), "the crisis message comes at the end"
+    assert any(a.code == "red_flag_psychosis" and a.level == "acute" for a in await rows(Alert))
+
+
+async def test_weapon_answer_triggers_nothing(released, non_blocking):  # noqa: F811
+    replies = await press(released, "intake:start")
+    for _ in range(2):  # psychosis, harm to others: «Нет»
+        replies = await press(released, answer_button(replies))
+    replies = await press(released, answer_button(replies, {"rf_weapons": 0}))  # «Да» on weapons
+    assert not any(r.kind == "crisis" for r in replies)
+    (row,) = await rows()
+    assert not row.details.get("pending_urgent")

tests/test_demo_mode.py

@@ -0,0 +1,45 @@
+"""Presentation mode: fresh private demo patient straight into the chat; screening switch skips onboarding."""
+
+from collections.abc import AsyncIterator
+
+import pytest
+from httpx import ASGITransport, AsyncClient
+
+from app.core import demo_settings
+from app.db import repo, session
+from app.db.bootstrap import init_db
+from app.integrations.emias_record import emias_records
+from app.main import app
+
+MEM = "sqlite+aiosqlite:///:memory:"
+
+
+@pytest.fixture
+async def client() -> AsyncIterator[AsyncClient]:
+    await session.configure(identity_url=MEM, clinical_url=MEM)
+    await init_db()
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://t") as c:
+        yield c
+    await session.configure()
+
+
+async def test_fresh_demo_patients_are_private_and_ready(client):
+    a = (await client.post("/api/v1/auth/demo/fresh", json={})).json()["token"]
+    b = (await client.post("/api/v1/auth/demo/fresh", json={})).json()["token"]
+    assert a != b and a.startswith("tok-andrey-")
+    async with session.get_clinical_session() as s:
+        st = await repo.get_dialog_state(s, a)
+        assert st.state == "idle" and st.data["consent"] is True
+    rec = await emias_records.get(a)
+    assert rec is not None and any("язвенная" in c.name.lower() for c in rec.conditions)
+    r = await client.post(f"/api/v1/dialog/{a}/events", json={"type": "text", "text": "привет"})
+    assert r.status_code == 200 and r.json()["replies"]
+
+
+async def test_screening_switch(client):
+    assert (await client.get("/api/v1/demo/settings")).json() == {"screening_enabled": True}
+    assert (await client.post("/api/v1/demo/settings", json={"screening_enabled": False})).json() == {
+        "screening_enabled": False
+    }
+    async with session.get_clinical_session() as s:
+        assert await demo_settings.screening_enabled(s) is False

tests/test_dialog_engine.py

@@ -270,7 +270,8 @@ async def test_repeat_crisis_does_not_duplicate_script(dbs: None, engine: Engine
     assert r[0].text.startswith(TEXTS.crisis.repeat) and "112" in r[0].text
     assert spy.calls == []
     r = await text(engine, "ладно")
-    assert r[0].text == TEXTS.crisis.hold
+    # In a crisis every message gets a human reply, followed by the emergency contacts (chat never goes silent).
+    assert r[-1].kind == "crisis" and "112" in r[-1].text and r[0].kind != "crisis"
     _, data = await get_state()
     assert [e["handled"] for e in data["safety_events"]] == ["crisis_script", "crisis_repeat"]
 
@@ -482,4 +483,6 @@ def test_dialog_texts_avoid_forbidden_words() -> None:
     body = "\n".join(line for line in raw.splitlines() if not line.lstrip().startswith("#"))
     for word in ("учёт", "учет", "контрол", "надзор", "диагноз:"):
         assert word not in body
-    assert "152-фз" in body and "8-800-2000-122" in body
+    assert "152-фз" in body
+    # Numbers live in one table (config/texts/crisis.yaml); the minors' stop text gets the children's line from it.
+    assert "8-800-2000-122" not in body and "8-800-2000-122" in load_dialog_texts().age_gate.minor_stop

tests/test_dialog_idle_agent.py

@@ -0,0 +1,24 @@
+"""In idle every message goes to the chat agent (when it is on); quit intent keeps the rule reaction."""
+
+from app.llm import agent
+from app.llm.agent import AgentAnswer
+from tests.test_dialog_knowledge import TOKEN, engine, joined, say  # noqa: F401
+
+
+async def test_idle_text_goes_to_agent(engine, monkeypatch):  # noqa: F811
+    seen = []
+
+    async def fake(token, text, **kw):
+        seen.append(str(text))
+        return AgentAnswer(text="ОТВЕТ АГЕНТА", sources=[], tools=["search_guidelines"], provider="fake")
+
+    monkeypatch.setattr(agent, "answer_chat", fake)
+    for msg in ("прив", "какие у меня будут ощущения и побочки от лекарств", "а какие аналоги моего лекарства есть"):
+        replies = await say(engine, msg)
+        assert "ОТВЕТ АГЕНТА" in joined(replies), msg
+    assert len(seen) == 3 and all(t == TOKEN for t in [TOKEN])
+
+
+async def test_idle_falls_back_when_agent_off(engine):  # noqa: F811
+    replies = await say(engine, "прив")
+    assert replies and "ОТВЕТ АГЕНТА" not in joined(replies)

tests/test_dialog_intake.py

@@ -13,6 +13,7 @@ from app.core import clock
 from app.core.scales import clear_cache
 from app.db import repo, session
 from app.db.bootstrap import init_db
+from app.db.clinical import Alert
 from app.db.intake import QuestionnaireNotice, QuestionnaireRecord, QuestionnaireTask
 from app.dialog.contracts import Event
 from app.dialog.engine import Engine
@@ -76,7 +77,8 @@ def answer_button(replies, choices=None):
             for b in line:
                 if b.data.startswith("iq:"):
                     item = b.data.split(":")[2]
-                    index = choices.get(item, 1 if item.startswith("cssrs_") else 0)
+                    # «Нет» by default for the safety items (C-SSRS, red flags), the first option otherwise.
+                    index = choices.get(item, 1 if item.startswith(("cssrs_", "rf_")) else 0)
                     for line2 in reply.buttons:
                         for other in line2:
                             if other.data.endswith(":" + str(index)) and other.data.startswith("iq:"):
@@ -122,17 +124,17 @@ async def test_pause_restart_and_duplicate_callbacks(released):
     old = answer_button(replies)
     await asyncio.gather(press(released, old), press(released, old))
     (row,) = await rows()
-    assert row.answers == {"phq9_1": 0}
+    assert row.answers == {"rf_psychosis": 0}
     await press(released, "intake:pause")
     restarted = Engine(gateway=LLMGateway(MockTasks()))
     replies = await press(restarted, "intake:resume")
-    assert "phq9_2" in answer_button(replies)
+    assert "rf_harm_others" in answer_button(replies)
     await press(restarted, old)
-    assert (await rows())[0].answers == {"phq9_1": 0}
+    assert (await rows())[0].answers == {"rf_psychosis": 0}
     assert all(len(b.data.encode()) <= 64 for r in replies for line in r.buttons for b in line)
 
 
-@pytest.mark.parametrize("signal", ["help", "phq", "cssrs"])
+@pytest.mark.parametrize("signal", ["help", "red_flag", "cssrs"])
 async def test_immediate_stop_preserves_partial_answers(released, signal):
     replies = await press(released, "intake:start")
     if signal == "help":
@@ -142,13 +144,45 @@ async def test_immediate_stop_preserves_partial_answers(released, signal):
         for item in ("cssrs_q1", "cssrs_q2", "cssrs_q3", "cssrs_q4"):
             replies = await press(released, answer_button(replies, {item: 0}))
     else:
-        replies = await complete_existing(released, replies, {"trauma": 1, "phq9_1": 2, "phq9_9": 1})
+        replies = await press(released, answer_button(replies, {"rf_psychosis": 0}))
     assert any(r.kind == "crisis" for r in replies)
     (row,) = await rows()
     assert row.status == "interrupted" and row.completed_at is None
     notices = await rows(QuestionnaireNotice)
     assert len(notices) == 1 and notices[0].status == "test_sink"
     assert "уведомлён" not in replies[0].text
+    # Every crisis leaves a callback task for the doctor on duty (one per patient while it is open).
+    calls = [a for a in await rows(Alert) if a.code == "post_crisis_call"]
+    assert len(calls) == 1 and calls[0].params["assignee"] == "on-duty"
+
+
+async def test_red_flag_violence_uses_its_crisis_variant(released):
+    replies = await press(released, "intake:start")
+    for _ in range(3):  # psychosis, harm to others, weapons: «Нет»
+        replies = await press(released, answer_button(replies))
+    replies = await press(released, answer_button(replies, {"rf_violence": 0}))
+    crisis = next(r for r in replies if r.kind == "crisis")
+    assert "насили" in crisis.text.lower()
+    (row,) = await rows()
+    assert row.details["stop_reason"] == "red_flag_violence"
+
+
+async def test_item9_with_a_negative_screener_is_a_flag_not_a_crisis(released):
+    # One crisis rule (ADR-013): C-SSRS was negative, so PHQ-9 item 9 is a discrepancy for the doctor.
+    choices = {f"phq9_{i}": 2 for i in range(1, 9)}
+    choices.update(trauma=1, tbi=1, psychotropics=1, phq9_9=1)
+    replies = await press(released, "intake:start")
+    for _ in range(60):
+        button = answer_button(replies, choices)
+        if button is None or any(r.kind == "safety_plan" for r in replies):
+            break
+        replies = await press(released, button)
+    assert not any(r.kind == "crisis" for r in replies)
+    assert any(r.kind == "safety_plan" for r in replies)
+    codes = {a.code for a in await rows(Alert)}
+    assert "phq9_cssrs_discrepancy" in codes and "post_crisis_call" not in codes
+    (row,) = await rows()
+    assert row.status == "complete" and "phq9_cssrs_discrepancy" in row.details["decision"]["flags"]
 
 
 async def complete_existing(e, replies, choices):
@@ -350,16 +384,24 @@ async def test_diary_morning_then_evening_with_the_ptsd_module(released):
     assert any("уже заполнен" in r.text for r in replies)
 
 
-async def test_depression_module_item9_stops_with_crisis(released, monkeypatch):
+async def test_depression_module_item9_asks_the_screener_at_once(released, monkeypatch):
     await complete_intake(released, {"trauma": 1, "tbi": 1, "psychotropics": 1})
     async with session.get_clinical_session() as s:
         await repo.upsert_profile(s, token=TOKEN, modules=["depression"])
     monkeypatch.setattr(clock, "now", lambda: NOW + timedelta(days=2))
     await run_diary(released, await press(released, "menu:checkin"))
-    replies = await run_diary(released, await press(released, "menu:checkin"), {"medication": 2, "dphq_9": 1})
-    assert any(r.kind == "crisis" for r in replies)
+    replies = await press(released, "menu:checkin")
+    for _ in range(10):
+        button = answer_button(replies, {"medication": 2, "dphq_9": 1})
+        replies = await press(released, button)
+        if ":cssrs_" in (answer_button(replies) or ""):
+            break
+    # Item 9 → the C-SSRS right away («Этап 2»): no crisis screen before the screener says the risk is high.
+    assert not any(r.kind == "crisis" for r in replies) and any("последнего ответа" in r.text for r in replies)
     evening = next(r for r in await rows() if r.kind == "diary_evening")
-    assert evening.answers["dphq_9"] == 1 and evening.status == "interrupted"
+    assert evening.answers["dphq_9"] == 1 and evening.details["stop_reason"] == "phq9_item9_cssrs"
+    replies = await complete_existing(released, replies, {"cssrs_q1": 0, "cssrs_q2": 0, "cssrs_q3": 0, "cssrs_q4": 0})
+    assert any(r.kind == "crisis" for r in replies)
 
 
 async def test_unknown_sex_goes_to_review_without_a_zone(released):
@@ -380,7 +422,7 @@ async def test_version_change_blocks_resume_and_preserves_answers(released):
     path.write_text(yaml.safe_dump(data, allow_unicode=True))
     await press(released, "intake:resume")
     (row,) = await rows()
-    assert row.status == "review_required" and row.answers == {"phq9_1": 0}
+    assert row.status == "review_required" and row.answers == {"rf_psychosis": 0}
 
 
 async def test_followup_delta_uses_comparable_complete_baseline(released):
@@ -434,6 +476,8 @@ async def test_diary_schedule_uses_moscow_date(released, monkeypatch):
 
 async def test_resumed_answer_keeps_original_period(released, monkeypatch):
     replies = await press(released, "intake:start")
+    for _ in range(5):  # the red-flag block
+        replies = await press(released, answer_button(replies))
     await press(released, answer_button(replies))
     await press(released, "intake:pause")
     monkeypatch.setattr(clock, "now", lambda: NOW + timedelta(days=1))
@@ -472,3 +516,16 @@ async def test_without_a_doctor_the_screening_defines_the_modules(released):
     assert any("не выявлено" in r.text for r in replies)
     async with session.get_clinical_session() as s:
         assert (await repo.get_profile(s, TOKEN)).modules == []
+
+
+async def test_next_day_after_a_crisis_the_app_asks_once(released, monkeypatch):
+    replies = await press(released, "intake:start")
+    await press(released, answer_button(replies, {"rf_psychosis": 0}))  # red flag → crisis
+    outbox = MemoryOutbox()
+    monkeypatch.setattr(integrations, "outbox", outbox)
+    monkeypatch.setattr(clock, "now", lambda: NOW + timedelta(hours=21))
+    await jobs.plan()
+    await jobs.plan()
+    sent = outbox.drain(TOKEN)
+    assert len(sent) == 1 and "Как вы сегодня" in sent[0].text
+    assert any(b.data == "safety_plan" for line in sent[0].buttons for b in line)

tests/test_dialog_side_answer.py

@@ -0,0 +1,73 @@
+"""Questions are answered in any state (side answer): mid-scale the pending question comes back with its
+buttons; in a crisis the answer carries the emergency contacts and the crisis stays open."""
+
+from app.core import clock
+from app.db import repo, session
+from app.services.contracts import NoopMonitoring
+from tests.test_dialog_knowledge import TOKEN, buttons, engine, joined, press, say, state  # noqa: F401
+
+
+class _InCrisis(NoopMonitoring):
+    async def is_in_crisis(self, token: str) -> bool:  # an open acute alert, as RuleMonitoring would report
+        return True
+
+
+async def test_question_mid_scale_is_answered_and_question_returns(engine):  # noqa: F811
+    await press(engine, "menu:checkin")  # some flow with buttons; then force a scale for determinism
+    async with session.get_clinical_session() as s:
+        await repo.set_dialog_state(s, token=TOKEN, state="idle", data={"consent": True}, now=clock.now())
+    from app.dialog.contracts import Event
+
+    first = await engine.handle(TOKEN, Event(type="system", system="scale_due", params={"scale": "phq9"}))
+    assert buttons(first), "scale question with buttons expected"
+    st_before, _ = await state()
+    replies = await say(engine, "Через сколько начнёт действовать лекарство?")
+    st_after, data = await state()
+    assert st_after == st_before  # still inside the scale
+    assert "2–4 недели" in joined(replies) or "врач" in joined(replies).lower()
+    assert buttons(replies), "the pending question is shown again with its buttons"
+    assert "Вернёмся к вопросу" in joined(replies)
+    assert data.get("last_prompt")
+
+
+async def test_question_in_crisis_gets_answer_and_contacts(engine):  # noqa: F811
+    engine._monitoring = _InCrisis()
+    async with session.get_clinical_session() as s:
+        await repo.set_dialog_state(
+            s, token=TOKEN, state="crisis", data={"consent": True, "crisis": {"open": True}}, now=clock.now()
+        )
+    replies = await say(engine, "Через сколько начнёт действовать лекарство?")
+    st, _ = await state()
+    assert st == "crisis"
+    kinds = [r.kind for r in replies]
+    assert "crisis" in kinds and len(replies) >= 2
+    assert "112" in joined(replies)
+
+
+async def test_non_question_mid_scale_still_goes_to_the_scale(engine):  # noqa: F811
+    from app.dialog.contracts import Event
+
+    await engine.handle(TOKEN, Event(type="system", system="scale_due", params={"scale": "phq9"}))
+    replies = await say(engine, "ну так себе")
+    assert "Вернёмся к вопросу" not in joined(replies)
+
+
+async def test_crisis_any_text_gets_a_reply_with_contacts(engine):  # noqa: F811
+    engine._monitoring = _InCrisis()
+    async with session.get_clinical_session() as s:
+        await repo.set_dialog_state(
+            s, token=TOKEN, state="crisis", data={"consent": True, "crisis": {"open": True}}, now=clock.now()
+        )
+    for text in ("привет", "мне страшно", "что содержится в моих лекарствах"):
+        replies = await say(engine, text)
+        assert len(replies) >= 2, text
+        assert replies[-1].kind == "crisis" and "112" in joined(replies)
+        assert replies[0].kind != "crisis"  # a real reply first, not only the stub
+
+
+async def test_medicine_mention_without_question_mark_mid_scale(engine):  # noqa: F811
+    from app.dialog.contracts import Event
+
+    await engine.handle(TOKEN, Event(type="system", system="scale_due", params={"scale": "phq9"}))
+    replies = await say(engine, "что содержится в моих лекарствах")
+    assert "Вернёмся к вопросу" in joined(replies) and buttons(replies)

tests/test_dialog_veteran_onboarding.py

@@ -268,7 +268,7 @@ async def test_auditc_sex_declined_uses_default_threshold(dbs_nosex: None, engin
     assert (await results())["auditc"].severity == "positive"  # unknown sex → sensitive threshold ≥ 3
     # Weapon at home + AUDIT-C positive = risk signal → safe-storage block.
     r = await press(engine, "vonb:weapon:self")
-    assert VT.weapon.safety_block_self in [x.text for x in r]
+    assert VT.weapon.safety_block_self not in [x.text for x in r]  # the weapon answer affects nothing
 
 
 # ----------------------------------------------------------------------------------------------- weapon / safety

tests/test_llm_agent.py

@@ -269,3 +269,32 @@ def test_build_provider_prefers_gigachat():
 
     s = Settings(_env_file=None, chat_agent="auto", gigachat_credentials="x", openrouter_api_key="y")
     assert isinstance(agent.build_provider(s), GigaChatToolProvider)
+
+
+async def test_web_demo_token_shares_andreys_record():
+    rec = await emias_records.get("tok-andrey")
+    assert rec is not None and any("язвенная" in c.name.lower() for c in rec.conditions)
+
+
+@pytest.mark.parametrize("text", ["прив", "Привет!", "как дела у тебя ? что делаешь ? спишь ?", "добрый вечер"])
+async def test_small_talk_without_llm(text):
+    provider, sent = _fake_openrouter("x", _CALLS)
+    ans = await agent.answer_chat(ANDREY, scrub(text), provider=provider, settings=_settings())
+    assert ans is not None and ans.provider == "small_talk" and sent == []
+
+
+async def test_medical_question_is_not_small_talk():
+    assert not agent.is_small_talk("привет, можно мне ибупрофен?")
+
+
+async def test_tool_calls_are_capped(sertraline):
+    many = [
+        {"id": f"s{i}", "type": "function", "function": {"name": "search_guidelines", "arguments": '{"query": "x"}'}}
+        for i in range(20)
+    ]
+    provider, sent = _fake_openrouter("Ответ по рекомендациям, обсудите с врачом.", many)
+    ans = await agent.answer_chat(ANDREY, scrub("что такое ПТСР?"), provider=provider, settings=_settings())
+    assert ans is not None and len(ans.tools) == agent.MAX_TOOL_CALLS_PER_TURN
+    tool_msgs = [m for m in sent[1]["messages"] if m["role"] == "tool"]
+    assistant = [m for m in sent[1]["messages"] if m["role"] == "assistant"][0]
+    assert len(tool_msgs) == len(assistant["tool_calls"]) == agent.MAX_TOOL_CALLS_PER_TURN

tests/test_llm_stt.py

@@ -0,0 +1,107 @@
+"""Speech-to-text providers (#116): OpenRouter request shape, retries, silence, WAV parsing, provider choice."""
+
+import base64
+import io
+import json
+import wave
+
+import httpx
+import pytest
+
+from app.config import Settings
+from app.llm.stt import (
+    InvalidAudio,
+    MockSTT,
+    OpenRouterSTT,
+    SttError,
+    SttNoSpeech,
+    get_stt,
+    load_texts,
+    wav_seconds,
+)
+
+
+def make_wav(seconds: float, rate: int = 16_000) -> bytes:
+    buf = io.BytesIO()
+    with wave.open(buf, "wb") as w:
+        w.setnchannels(1)
+        w.setsampwidth(2)
+        w.setframerate(rate)
+        w.writeframes(b"\x00\x00" * int(rate * seconds))
+    return buf.getvalue()
+
+
+def _settings(**kw) -> Settings:
+    base = {"openrouter_api_key": "test", "stt_provider": "openrouter", "stt_timeout_s": 5.0}
+    return Settings(_env_file=None, **{**base, **kw})
+
+
+def _fake(responses: list[httpx.Response]) -> tuple[OpenRouterSTT, list[dict]]:
+    sent: list[dict] = []
+    replies = iter(responses)
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        sent.append(json.loads(request.content))
+        return next(replies)
+
+    stt = OpenRouterSTT(_settings(), client=httpx.AsyncClient(transport=httpx.MockTransport(handler)))
+    return stt, sent
+
+
+def _ok(text: str) -> httpx.Response:
+    return httpx.Response(200, json={"choices": [{"message": {"content": text}}]})
+
+
+def test_wav_seconds() -> None:
+    assert wav_seconds(make_wav(2.5)) == pytest.approx(2.5)
+    with pytest.raises(InvalidAudio):
+        wav_seconds(b"OggS" + b"\x00" * 100)
+    with pytest.raises(InvalidAudio):
+        wav_seconds(b"")
+
+
+async def test_openrouter_sends_audio_and_returns_transcript() -> None:
+    audio = make_wav(1)
+    stt, sent = _fake([_ok("«Сегодня плохо спал.»")])
+    assert await stt.transcribe(audio) == "Сегодня плохо спал."
+    body = sent[0]
+    assert body["model"] == "google/gemini-2.5-flash" and body["temperature"] == 0
+    assert body["reasoning"] == {"enabled": False}
+    part = body["messages"][1]["content"][1]
+    assert part["type"] == "input_audio" and part["input_audio"]["format"] == "wav"
+    assert base64.b64decode(part["input_audio"]["data"]) == audio
+
+
+async def test_silence_marker_is_no_speech() -> None:
+    stt, _ = _fake([_ok(load_texts().silence_marker.upper())])
+    with pytest.raises(SttNoSpeech):
+        await stt.transcribe(make_wav(1))
+
+
+async def test_retries_then_succeeds(monkeypatch) -> None:
+    monkeypatch.setattr("app.llm.stt.asyncio.sleep", _no_sleep)
+    stt, sent = _fake([httpx.Response(503), _ok("текст")])
+    assert await stt.transcribe(make_wav(1)) == "текст"
+    assert len(sent) == 2
+
+
+async def test_http_error_raises(monkeypatch) -> None:
+    monkeypatch.setattr("app.llm.stt.asyncio.sleep", _no_sleep)
+    stt, _ = _fake([httpx.Response(401, json={"error": {"message": "bad key"}})])
+    with pytest.raises(SttError):
+        await stt.transcribe(make_wav(1))
+
+
+async def _no_sleep(_: float) -> None:
+    return None
+
+
+async def test_mock_is_offline() -> None:
+    assert await MockSTT().transcribe(b"") == load_texts().mock_transcript
+
+
+def test_provider_choice() -> None:
+    assert isinstance(get_stt(_settings(stt_provider="auto")), OpenRouterSTT)
+    assert isinstance(get_stt(_settings(stt_provider="auto", openrouter_api_key="")), MockSTT)
+    assert get_stt(_settings(stt_provider="openrouter", openrouter_api_key="")) is None
+    assert get_stt(_settings(stt_provider="off")) is None

tests/test_max_bot.py

@@ -0,0 +1,281 @@
+"""End-to-end server-only MAX flow, real SQLite transactions and mocked network."""
+
+from datetime import timedelta
+from unittest.mock import AsyncMock
+from urllib.parse import parse_qs, urlsplit
+
+import pytest
+from httpx import ASGITransport, AsyncClient
+from sqlalchemy import select
+
+from app.channels import max_bot
+from app.channels.max_client import MaxApiError
+from app.config import get_settings
+from app.core import clock
+from app.db import repo, session
+from app.db.bootstrap import init_db
+from app.db.clinical import ScaleResult
+from app.db.intake import QuestionnaireRecord
+from app.db.max_notifications import MaxBinding, MaxDelivery, MaxReceipt
+from app.dialog.contracts import Reply
+from app.integrations import max_store as store
+from app.integrations.base import MemoryOutbox, integrations
+from app.main import app
+
+TOKEN = "synthetic-patient"
+SECRET = "synthetic-webhook-secret-" * 2
+
+
+@pytest.fixture(autouse=True)
+async def database(monkeypatch):
+    for key, value in {
+        "MAX_ENABLED": "true",
+        "MAX_BOT_TOKEN": "fake-token",
+        "MAX_BOT_USERNAME": "test_bot",
+        "MAX_WEBHOOK_SECRET": SECRET,
+        "MAX_WEBHOOK_URL": "https://api.example.org/api/v1/webhooks/max",
+        "MAX_SITE_URL": "https://example.org",
+    }.items():
+        monkeypatch.setenv(key, value)
+    get_settings.cache_clear()
+    max_bot.texts.cache_clear()
+    await session.configure(identity_url="sqlite+aiosqlite:///:memory:", clinical_url="sqlite+aiosqlite:///:memory:")
+    await init_db()
+    async with session.get_clinical_session() as s:
+        await repo.upsert_profile(s, token=TOKEN, doctor_id="")
+        await repo.set_dialog_state(
+            s, token=TOKEN, state="idle", data={"consent": True, "adult": True}, now=clock.now()
+        )
+    yield
+    await max_bot.stop()
+    await session.configure()
+    get_settings.cache_clear()
+
+
+def event(kind="bot_started", text="", uid=100, *, stamp=1000):
+    user = {"user_id": uid, "is_bot": False, "first_name": "Synthetic"}
+    if kind == "message_created":
+        return {
+            "update_type": kind,
+            "timestamp": stamp,
+            "message": {
+                "sender": user,
+                "recipient": {"chat_type": "dialog", "chat_id": 123},
+                "body": {"mid": f"mid-{stamp}", "text": text},
+            },
+        }
+    return {"update_type": kind, "timestamp": stamp, "user": user, "chat_id": 123, "payload": text}
+
+
+async def connect():
+    async with session.get_identity_session() as s:
+        code, _ = await store.issue_code(s, TOKEN, clock.now(), 600)
+    await max_bot.handle_update(event(text=code))
+    return code
+
+
+async def test_http_connect_webhook_duplicate_notification_and_stop():
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as client:
+        auth = {"Authorization": f"Bearer {TOKEN}"}
+        linked = await client.post("/api/v1/me/notifications/max/link", headers=auth, json={"consent": True})
+        code = parse_qs(urlsplit(linked.json()["url"]).query)["start"][0]
+        headers = {"X-Max-Bot-Api-Secret": SECRET}
+        for _ in range(2):
+            assert (
+                await client.post("/api/v1/webhooks/max", headers=headers, json=event(text=code))
+            ).status_code == 200
+        assert (await client.get("/api/v1/me/notifications/max", headers=auth)).json()["connected"]
+        async with session.get_identity_session() as s:
+            assert len((await s.scalars(select(MaxReceipt))).all()) == 1
+            assert len((await s.scalars(select(MaxDelivery))).all()) == 1
+        sender = AsyncMock()
+        assert await max_bot.deliver_one(sender)
+        assert sender.send.call_args.args == (100, max_bot.texts().connected)
+        await client.post("/api/v1/webhooks/max", headers=headers, json=event("message_created", "/stop", stamp=2000))
+        assert not (await client.get("/api/v1/me/notifications/max", headers=auth)).json()["connected"]
+
+
+async def test_unbound_start_help_and_invalid_code_are_generic():
+    for e in (event(), event("message_created", "private patient text", stamp=2000), event(text="expired", stamp=3000)):
+        await max_bot.handle_update(e)
+    async with session.get_identity_session() as s:
+        rows = (await s.scalars(select(MaxDelivery))).all()
+        assert {r.kind for r in rows} == {"not_connected", "help", "invalid"}
+        assert all(r.token is None for r in rows)
+        assert not (await s.scalars(select(MaxBinding))).all()
+
+
+@pytest.mark.parametrize("kind", ["bot_stopped", "dialog_removed", "message_created"])
+async def test_stopping_cancels_pending_confirmation_and_reminders(kind):
+    await connect()
+    await max_bot.MaxOutbox(MemoryOutbox()).push(TOKEN, [Reply(text="private scheduled question")])
+    await max_bot.handle_update(event(kind, "/stop", stamp=2000))
+    async with session.get_identity_session() as s:
+        assert not (await s.get(MaxBinding, TOKEN)).enabled
+        rows = (await s.scalars(select(MaxDelivery).where(MaxDelivery.token == TOKEN))).all()
+        assert rows and all(r.state == "cancelled" for r in rows)
+    sender = AsyncMock()
+    await max_bot.deliver_one(sender)
+    assert sender.send.call_count == (1 if kind == "message_created" else 0)
+
+
+async def test_private_patient_content_stays_in_fallback_only():
+    await connect()
+    fallback = MemoryOutbox()
+    wrapper = max_bot.MaxOutbox(fallback)
+    replies = [Reply(text="SYNTHETIC confidential result 42", kind="question")]
+    await wrapper.push(TOKEN, replies)
+    await wrapper.push(TOKEN, replies)
+    assert fallback.drain(TOKEN) == replies * 2
+    async with session.get_identity_session() as s:
+        rows = (await s.scalars(select(MaxDelivery).where(MaxDelivery.kind == "notification"))).all()
+        assert len(rows) == 1
+        assert "confidential" not in repr(rows[0].__dict__)
+    sender = AsyncMock()
+    while await max_bot.deliver_one(sender):
+        pass
+    assert any(c.args[1] == max_bot.texts().notification for c in sender.send.call_args_list)
+    assert all("confidential" not in c.args[1] and TOKEN not in str(c.kwargs) for c in sender.send.call_args_list)
+
+
+async def test_revoked_consent_prevents_binding_and_delivery():
+    await connect()
+    async with session.get_clinical_session() as s:
+        await repo.set_dialog_state(
+            s, token=TOKEN, state="idle", data={"consent": False, "adult": True}, now=clock.now()
+        )
+    sender = AsyncMock()
+    await max_bot.deliver_one(sender)
+    sender.send.assert_not_called()
+    async with session.get_identity_session() as s:
+        await store.disconnect(s, TOKEN)
+        code, _ = await store.issue_code(s, TOKEN, clock.now(), 600)
+    await max_bot.handle_update(event(text=code, stamp=2000))
+    async with session.get_identity_session() as s:
+        assert not (await s.get(MaxBinding, TOKEN)).enabled
+
+
+@pytest.mark.parametrize(
+    "bad",
+    [
+        {},
+        {"update_type": "other", "timestamp": 1},
+        event() | {"timestamp": True},
+        event() | {"user": {"user_id": True, "is_bot": False}},
+        event() | {"user": {"user_id": 100, "is_bot": True}},
+        event() | {"payload": {"wrong": "type"}},
+        {"update_type": "message_created", "timestamp": 1, "message": {}},
+    ],
+)
+async def test_malformed_or_bot_events_are_ignored(bad):
+    await max_bot.handle_update(bad)
+    async with session.get_identity_session() as s:
+        assert not (await s.scalars(select(MaxReceipt))).all()
+
+
+async def test_group_cannot_claim_a_patient_link():
+    async with session.get_identity_session() as s:
+        code, _ = await store.issue_code(s, TOKEN, clock.now(), 600)
+    update = event("message_created", "/start " + code)
+    update["message"]["recipient"]["chat_type"] = "chat"
+    await max_bot.handle_update(update)
+    async with session.get_identity_session() as s:
+        assert await s.get(MaxBinding, TOKEN) is None
+        assert await store.code_token(s, code, clock.now()) == TOKEN
+
+
+async def test_429_respects_retry_deadline_and_then_succeeds(monkeypatch):
+    await max_bot.handle_update(event())
+    now = clock.now()
+    monkeypatch.setattr(clock, "now", lambda: now)
+    sender = AsyncMock()
+    sender.send.side_effect = MaxApiError("http_429", retryable=True, retry_after=120)
+    assert await max_bot.deliver_one(sender)
+    async with session.get_identity_session() as s:
+        row = (await s.scalars(select(MaxDelivery))).one()
+        assert row.state == "queued" and row.next_attempt_at == now + timedelta(seconds=120)
+    assert not await max_bot.deliver_one(sender)
+    now += timedelta(seconds=121)
+    sender.send.side_effect = None
+    assert await max_bot.deliver_one(sender)
+    async with session.get_identity_session() as s:
+        row = (await s.scalars(select(MaxDelivery))).one()
+        assert row.state == "sent" and row.attempts == 2
+
+
+async def test_permanent_error_does_not_retry():
+    await max_bot.handle_update(event())
+    sender = AsyncMock()
+    sender.send.side_effect = MaxApiError("http_403")
+    await max_bot.deliver_one(sender)
+    assert not await max_bot.deliver_one(sender)
+    async with session.get_identity_session() as s:
+        assert (await s.scalars(select(MaxDelivery))).one().state == "failed"
+
+
+@pytest.mark.parametrize("source", ["scale", "questionnaire"])
+async def test_completed_test_produces_one_neutral_notification(source):
+    await connect()
+    sender = AsyncMock()
+    await max_bot.deliver_one(sender)
+    sender.reset_mock()
+    async with session.get_clinical_session() as s:
+        now = clock.now()
+        if source == "scale":
+            s.add(ScaleResult(token=TOKEN, ts=now, scale="phq9", total=1, severity="synthetic", complete=True))
+        else:
+            s.add(
+                QuestionnaireRecord(
+                    id="synthetic-form",
+                    token=TOKEN,
+                    kind="intake",
+                    version="test",
+                    policy_version="test",
+                    started_at=now,
+                    completed_at=now,
+                    day=now.date(),
+                    status="complete",
+                    answers={},
+                    details={},
+                )
+            )
+    for _ in range(2):
+        async with session.get_identity_session() as s:
+            await store.collect_completed(s, clock.now())
+    assert await max_bot.deliver_one(sender)
+    assert not await max_bot.deliver_one(sender)
+    sender.send.assert_awaited_once_with(
+        100, max_bot.texts().completed, site_url="https://example.org", button=max_bot.texts().open_site
+    )
+
+
+async def test_private_start_command_accepts_issued_code():
+    async with session.get_identity_session() as s:
+        code, _ = await store.issue_code(s, TOKEN, clock.now(), 600)
+    await max_bot.handle_update(event("message_created", "/start " + code))
+    async with session.get_identity_session() as s:
+        assert (await s.get(MaxBinding, TOKEN)).user_id == 100
+
+
+async def test_disabled_lifecycle_never_constructs_client(monkeypatch):
+    monkeypatch.setenv("MAX_ENABLED", "false")
+    get_settings.cache_clear()
+    monkeypatch.setattr(max_bot, "MaxClient", lambda _: pytest.fail("disabled MAX must not create a network client"))
+    before = integrations.outbox
+    await max_bot.start()
+    await max_bot.handle_update(event())
+    await max_bot.stop()
+    assert integrations.outbox is before
+
+
+async def test_lifecycle_wraps_once_restores_and_closes(monkeypatch):
+    sender = AsyncMock()
+    monkeypatch.setattr(max_bot, "MaxClient", lambda _: sender)
+    before = integrations.outbox
+    await max_bot.start()
+    first = integrations.outbox
+    await max_bot.start()
+    assert integrations.outbox is first and first.fallback is before
+    await max_bot.stop()
+    assert integrations.outbox is before
+    sender.close.assert_awaited_once()

tests/test_max_client.py

@@ -0,0 +1,95 @@
+"""Wire contract with MAX, using an in-process HTTP fake (no real bot/token)."""
+
+import json
+import logging
+
+import httpx
+import pytest
+
+from app.channels.max_client import API_URL, UPDATE_TYPES, MaxApiError, MaxClient
+from app.config import Settings
+
+
+def settings():
+    return Settings(
+        _env_file=None,
+        max_enabled=True,
+        max_bot_token="synthetic-secret-token",
+        max_bot_username="synthetic_bot",
+        max_webhook_secret="synthetic-webhook-secret-" * 2,
+        max_webhook_url="https://api.example.org/api/v1/webhooks/max",
+        max_site_url="https://example.org",
+    )
+
+
+async def test_message_wire_contract_and_no_identifiers_in_logs(caplog):
+    def handler(request):
+        assert str(request.url).startswith(API_URL + "/messages?")
+        assert request.url.params["user_id"] == "1234567890123"
+        assert "access_token" not in request.url.params
+        assert request.headers["Authorization"] == "synthetic-secret-token"
+        body = json.loads(request.content)
+        assert body["text"] == "neutral reminder" and "format" not in body
+        button = body["attachments"][0]["payload"]["buttons"][0][0]
+        assert button == {"type": "link", "text": "Open", "url": "https://example.org"}
+        return httpx.Response(200, json={"message": {"body": {"mid": "synthetic-id"}}})
+
+    client = MaxClient(settings(), transport=httpx.MockTransport(handler))
+    with caplog.at_level(logging.INFO):
+        await client.send(1234567890123, "neutral reminder", site_url="https://example.org", button="Open")
+    await client.close()
+    assert "1234567890123" not in caplog.text and "synthetic-secret-token" not in caplog.text
+
+
+@pytest.mark.parametrize(
+    "status,retry", [(429, True), (500, True), (503, True), (408, True), (401, False), (403, False)]
+)
+async def test_error_codes_are_redacted_and_classified(status, retry):
+    client = MaxClient(
+        settings(),
+        transport=httpx.MockTransport(
+            lambda r: httpx.Response(
+                status, json={"message": "sensitive provider text"}, headers={"Retry-After": "120"}
+            )
+        ),
+    )
+    with pytest.raises(MaxApiError) as error:
+        await client.send(100, "neutral", site_url="https://example.org", button="Open")
+    assert error.value.retryable is retry and error.value.retry_after == 120
+    assert "sensitive" not in str(error.value)
+    await client.close()
+
+
+@pytest.mark.parametrize("body", [{"success": False, "message": "private"}, {}, {"success": True}, []])
+async def test_http_200_is_not_enough_to_mark_delivered(body):
+    client = MaxClient(settings(), transport=httpx.MockTransport(lambda r: httpx.Response(200, json=body)))
+    with pytest.raises(MaxApiError):
+        await client.send(100, "neutral", site_url="https://example.org", button="Open")
+    await client.close()
+
+
+async def test_timeout_is_retryable_without_request_details():
+    def handler(request):
+        raise httpx.ReadTimeout("private URL", request=request)
+
+    client = MaxClient(settings(), transport=httpx.MockTransport(handler))
+    with pytest.raises(MaxApiError, match="^transport$") as error:
+        await client.request("GET", "/me")
+    assert error.value.retryable
+    await client.close()
+
+
+async def test_subscription_includes_secret_and_exact_event_types():
+    def handler(request):
+        assert request.url.path == "/subscriptions"
+        body = json.loads(request.content)
+        assert body == {
+            "url": settings().max_webhook_url,
+            "update_types": UPDATE_TYPES,
+            "secret": settings().max_webhook_secret.get_secret_value(),
+        }
+        return httpx.Response(200, json={"success": True})
+
+    client = MaxClient(settings(), transport=httpx.MockTransport(handler))
+    await client.subscribe(settings())
+    await client.close()

tests/test_max_setup.py

@@ -0,0 +1,51 @@
+"""Operator tools do not send messages or expose credentials."""
+
+from unittest.mock import AsyncMock
+
+import pytest
+
+from app.channels import max_setup
+from app.channels.max_client import MaxApiError
+from app.config import Settings
+
+
+@pytest.fixture
+def client(monkeypatch):
+    settings = Settings(_env_file=None, max_bot_username="test_bot", max_webhook_url="https://api.example.org/hook")
+    sender = AsyncMock()
+    sender.request.return_value = {"is_bot": True, "username": "test_bot"}
+    monkeypatch.setattr(max_setup, "Settings", lambda **kwargs: settings)
+    monkeypatch.setattr(max_setup, "MaxClient", lambda cfg: sender)
+    return sender
+
+
+async def test_check_only_reads_bot_identity(client):
+    await max_setup.run("check")
+    client.request.assert_awaited_once_with("GET", "/me")
+    client.subscribe.assert_not_called()
+    client.send.assert_not_called()
+    client.close.assert_awaited_once()
+
+
+async def test_subscribe_is_explicit_and_checks_bot_first(client):
+    await max_setup.run("subscribe")
+    client.subscribe.assert_awaited_once()
+    client.send.assert_not_called()
+
+
+async def test_mismatched_bot_does_not_change_webhooks(client):
+    client.request.return_value = {"is_bot": True, "username": "another_bot"}
+    with pytest.raises(MaxApiError, match="bot_username_mismatch"):
+        await max_setup.run("subscribe")
+    client.subscribe.assert_not_called()
+    client.close.assert_awaited_once()
+
+
+async def test_status_checks_exact_configured_webhook(client):
+    client.request.side_effect = [
+        {"is_bot": True, "username": "test_bot"},
+        {"subscriptions": [{"url": "https://elsewhere.example.org"}]},
+    ]
+    with pytest.raises(MaxApiError, match="webhook_missing"):
+        await max_setup.run("status")
+    client.subscribe.assert_not_called()

tests/test_questionnaire.py

@@ -172,12 +172,16 @@ def test_unapproved_policy_never_continues_even_a_negative_screen():
 def test_immediate_safety_preempts_unfinished_or_invalid_unrelated_scales():
     for overrides in [
         dict(help_now=True, cssrs={"broken": "data"}),
-        dict(phq9={"phq9_9": 1}),
+        dict(phq9={"phq9_9": 1}, cssrs={}),  # item 9 before the screener is complete
         dict(cssrs={"q4": 1}, alcohol={"bad": 50}),
     ]:
         assert intake(policy=load_policy(), **overrides).status == "crisis"
+
+
+def test_item9_with_a_negative_screener_is_a_discrepancy_flag_not_a_crisis():
+    # One crisis rule (ADR-013): the crisis screen opens at a high C-SSRS risk only.
     r = intake(phq9={"phq9_9": 1})
-    assert "phq9_cssrs_discrepancy" in r.flags
+    assert r.status != "crisis" and {"phq9_item9_positive", "phq9_cssrs_discrepancy"} <= set(r.flags)
 
 
 def test_physician_must_choose_between_stop_and_graded():
@@ -208,7 +212,8 @@ def test_phq2_and_trauma_expansion_do_not_turn_missing_into_negative():
         (30, "self_help"),
         (31, "psychiatrist_or_psychotherapist"),
         (49, "psychiatrist_or_psychotherapist"),
-        (50, "priority_psychiatrist"),
+        (50, "psychiatrist_or_psychotherapist"),
+        (80, "psychiatrist_or_psychotherapist"),
     ],
 )
 def test_pdf_pcl_boundaries_are_distinct_from_legacy_33(total, route):

tests/test_rules.py

@@ -125,8 +125,9 @@ def test_cssrs_levels() -> None:
 @pytest.mark.parametrize(
     ("weapon", "cssrs", "item9", "fires"),
     [
-        (True, "moderate", None, True),
-        (True, None, 1, True),
+        # Team decision 2026-10-04: weapon access triggers nothing (rule disabled in config/rules.yaml).
+        (True, "moderate", None, False),
+        (True, None, 1, False),
         (True, "low", 0, False),
         (False, "high", 2, False),
         (None, "moderate", None, False),
@@ -155,7 +156,9 @@ def test_severe_totals(scale: str, total: int, code: str, fires: bool) -> None:
     acts = evaluate(s)
     assert (code in codes(acts)) is fires
     if fires:
-        assert next(a for a in acts if a.code == code).level == AlertLevel.HIGH
+        # PCL-5 alone is a planned specialist review, not a priority (clinical review 2026-10-04).
+        expected = AlertLevel.MODERATE if scale == "pcl5" else AlertLevel.HIGH
+        assert next(a for a in acts if a.code == code).level == expected
 
 
 @pytest.mark.parametrize(

tests/test_services_crisis_refire.py

@@ -0,0 +1,32 @@
+"""An acked crisis is not raised again for the same signal; a newer PHQ-9 with item 9 raises a new one."""
+
+from datetime import datetime, timedelta
+
+from app.core.rules import Action
+from app.core.types import AlertLevel
+from app.services import monitoring
+
+
+class _A:
+    def __init__(self, code, ts, scale=None):
+        self.code, self.ts, self.scale = code, ts, scale
+
+
+async def test_drop_handled_crises(monkeypatch):
+    t0 = datetime(2026, 10, 4, 16, 32)
+    alerts = [_A("phq9_item9", t0 + timedelta(seconds=1))]
+    results = [_A(None, t0, "phq9"), _A(None, t0 + timedelta(hours=2), "auditc")]
+
+    async def list_alerts(s, token):
+        return alerts
+
+    async def list_scale_results(s, token):
+        return results
+
+    monkeypatch.setattr(monitoring.repo, "list_alerts", list_alerts)
+    monkeypatch.setattr(monitoring.repo, "list_scale_results", list_scale_results)
+    crisis = Action(type="crisis", level=AlertLevel.ACUTE, code="phq9_item9")
+    other = Action(type="alert", level=AlertLevel.HIGH, code="phq9_severe")
+    assert await monitoring._drop_handled_crises(None, "t", [crisis, other]) == [other]  # AUDIT-C is not a trigger
+    results.append(_A(None, t0 + timedelta(hours=3), "phq9"))  # a new PHQ-9 after the alert
+    assert await monitoring._drop_handled_crises(None, "t", [crisis]) == [crisis]

tests/test_services_garden.py

@@ -0,0 +1,88 @@
+"""Garden (#118): deterministic growth from activities, daily caps, never shrinks, ignores answers."""
+
+from collections.abc import AsyncIterator
+from datetime import date, datetime, timedelta
+
+import pytest
+from httpx import ASGITransport, AsyncClient
+from sqlalchemy import select
+
+from app.config import Settings
+from app.core import clock
+from app.db import session
+from app.db.clinical import DailyCheckin, Dose
+from app.main import app
+from app.services.garden import compute, load_config
+from scripts.seed_demo import seed
+
+MEM = "sqlite+aiosqlite:///:memory:"
+NOW = datetime(2026, 10, 5, 14, 0)
+D = date(2026, 10, 5)
+
+
+def test_points_steps_and_daily_caps() -> None:
+    cfg = load_config()
+    one_diary = compute([("diary", D)], D)
+    assert one_diary.points == cfg.activities["diary"].points
+    # Five diary entries on one day count as one; on five days — five times.
+    assert compute([("diary", D)] * 5, D).points == one_diary.points
+    five_days = compute([("diary", D - timedelta(days=i)) for i in range(5)], D)
+    assert five_days.points == 5 * one_diary.points
+    assert five_days.steps == five_days.points // cfg.step_points
+    assert five_days.points_in_step == five_days.points % cfg.step_points
+
+
+def test_never_shrinks_and_future_ignored() -> None:
+    events = [("meds", D), ("diary", D), ("questionnaire", D)]
+    now = compute(events, D)
+    # A week without any activity takes nothing away.
+    assert compute(events, D + timedelta(days=7)).points == now.points
+    # Rows dated after the (demo) today are not counted yet.
+    assert compute([*events, ("diary", D + timedelta(days=1))], D).points == now.points
+
+
+def test_plants_grow_one_after_another() -> None:
+    cfg = load_config()
+    per_plant = cfg.stages_per_plant
+    g = compute([("questionnaire", D - timedelta(days=i)) for i in range(100)], D)
+    stages = [p.stage for p in g.plants]
+    assert stages == sorted(stages, reverse=True) and stages[0] == per_plant
+    assert sum(stages) == min(g.steps, per_plant * cfg.max_plants)
+    assert compute([], D).title == cfg.titles[0].title
+
+
+@pytest.fixture
+async def client(monkeypatch: pytest.MonkeyPatch) -> AsyncIterator[AsyncClient]:
+    monkeypatch.setattr(clock, "get_settings", lambda: Settings(_env_file=None, demo_clock=True))
+    clock.set_now(NOW)
+    await session.configure(identity_url=MEM, clinical_url=MEM)
+    async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
+        yield c
+    await session.configure()
+
+
+async def test_andrey_has_a_living_garden_that_grows_and_ignores_mood(client: AsyncClient) -> None:
+    token = (await seed(12)).andrey_token
+    url = f"/api/v1/me/{token}/garden"
+    before = (await client.get(url)).json()
+    assert before["steps"] >= 4 and before["plants"][0]["stage"] > 0
+
+    # Mood is an answer: a bad day and a good day grow the garden the same way.
+    async with session.get_clinical_session() as s:
+        rows = (await s.scalars(select(DailyCheckin).where(DailyCheckin.token == token))).all()
+        for r in rows:
+            r.mood = 1
+    assert (await client.get(url)).json()["points"] == before["points"]
+
+    # Taking a dose today adds points.
+    async with session.get_clinical_session() as s:
+        dose = Dose(token=token, prescription_id=1, planned_at=clock.now(), status="taken", answered_at=clock.now())
+        s.add(dose)
+    after = (await client.get(url)).json()
+    assert after["points"] >= before["points"]
+    assert any(i["activity"] == "diary" for i in after["today"])
+
+
+async def test_unknown_patient_404(client: AsyncClient) -> None:
+    await seed(0)
+    assert (await client.get("/api/v1/me/nobody/garden")).status_code == 404

tests/test_services_monitoring.py

@@ -178,6 +178,7 @@ async def test_andrey_13_days(db: MemoryNotifier) -> None:
         "sleep_lt4_3n",
         "alcohol_3d",
         "phq9_item9",  # same evaluation: YAML order
+        "post_crisis_call",  # the crisis leaves a callback task for the doctor on duty
         "phq9_rise",
     ]
     assert len(crisis_audit) == 1 and crisis_audit[0].token == TOKEN and crisis_audit[0].actor == "system"
@@ -222,10 +223,9 @@ async def test_weapon_access_from_dialog_facts(db: MemoryNotifier) -> None:
     assert snap is not None and snap.weapon_access is True
     await scale("cssrs_screen", 3, severity="moderate", scheduled=False)
     actions = await monitoring.RuleMonitoring().after_data_change(TOKEN)
-    assert {a.code for a in actions if a.type == "crisis"} == {"weapon_access_risk"}
-    plan = next(a for a in actions if a.type == "show_safety_plan")
-    assert plan.params["block"] == "means_restriction"
-    assert await color() == "red"
+    # Team decision 2026-10-04: weapon access is stored but triggers nothing (no crisis, no safety block).
+    assert "weapon_access_risk" not in {a.code for a in actions}
+    assert not any(a.type == "show_safety_plan" and a.params.get("block") == "means_restriction" for a in actions)
 
 
 async def test_snapshot_mapping_and_unknown_patient(db: MemoryNotifier) -> None:
@@ -260,7 +260,7 @@ async def test_notifier_failure_keeps_alert(db: MemoryNotifier, monkeypatch: pyt
     await scale("pcl5", 55)
     await monitoring.RuleMonitoring().after_data_change(TOKEN)
     async with session.get_clinical_session() as s:
-        assert [(a.code, a.level) for a in await repo.list_alerts(s, token=TOKEN)] == [("pcl5_severe", "high")]
+        assert [(a.code, a.level) for a in await repo.list_alerts(s, token=TOKEN)] == [("pcl5_severe", "moderate")]
     assert await color() == "yellow"
 
 

tests/test_telegram.py

@@ -22,6 +22,7 @@ from app.channels.telegram import (
     render_markup,
     split_text,
 )
+from app.config import get_settings
 from app.db import identity_repo, session
 from app.db.bootstrap import init_db
 from app.dialog.contracts import Button, Event, Reply
@@ -163,7 +164,10 @@ async def test_verified_token_without_identity_record(dbs: None, engine: Recordi
     assert sender.texts == [load_texts().code_invalid]
 
 
-async def test_unbound_chat_gets_not_connected(dbs: None, engine: RecordingEngine) -> None:
+async def test_unbound_chat_gets_not_connected(
+    dbs: None, engine: RecordingEngine, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.setattr(get_settings(), "demo_clock", False)
     sender = FakeSender()
     channel = TelegramChannel(sender)
     await channel.on_start(CHAT, None)
@@ -251,6 +255,18 @@ async def test_outbox_pushes_to_bound_chat_and_falls_back(dbs: None, engine: Rec
     assert [r.text for r in fallback.drain(TOKEN2)] == ["Как вы сегодня?"]
 
 
+async def test_start_without_code_on_demo_stand_binds_a_demo_copy(
+    dbs: None, engine: RecordingEngine, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.setattr(get_settings(), "demo_clock", True)
+    sender = FakeSender()
+    channel = TelegramChannel(sender)
+    await channel.on_start(CHAT, None)
+    token = await channel.directory.token_for(CHAT)
+    assert token is not None and token.startswith("tok-andrey-")
+    assert [c[1].type for c in engine.calls] == ["start"]
+
+
 async def test_start_is_noop_without_token(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setattr(telegram, "get_settings", lambda: type("S", (), {"tg_bot_token": ""})())
     before = integrations.outbox
@@ -261,7 +277,9 @@ async def test_start_is_noop_without_token(monkeypatch: pytest.MonkeyPatch) -> N
 
 
 async def test_start_stop_swaps_outbox(monkeypatch: pytest.MonkeyPatch) -> None:
-    monkeypatch.setattr(telegram, "get_settings", lambda: type("S", (), {"tg_bot_token": "123:fake"})())
+    monkeypatch.setattr(
+        telegram, "get_settings", lambda: type("S", (), {"tg_bot_token": "123:fake", "tg_proxy_url": ""})()
+    )
 
     async def fake_polling(self: object, *bots: object, **kwargs: object) -> None:
         import asyncio

tests/test_telegram_proxy.py

@@ -0,0 +1,16 @@
+"""TG_PROXY_URL: the bot session goes through the proxy when it is set."""
+
+from app.channels import telegram
+from app.config import get_settings
+
+TOKEN = "123456:" + "A" * 35
+
+
+def test_make_bot_without_proxy(monkeypatch):
+    monkeypatch.setattr(get_settings(), "tg_proxy_url", "")
+    assert telegram.make_bot(TOKEN).session.proxy is None
+
+
+def test_make_bot_with_proxy(monkeypatch):
+    monkeypatch.setattr(get_settings(), "tg_proxy_url", "http://user:pass@127.0.0.1:8888")
+    assert telegram.make_bot(TOKEN).session.proxy == "http://user:pass@127.0.0.1:8888"

uv.lock

@@ -122,6 +122,19 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/b2/0e/50d6e6471cd31edce8b282bdec59375a3a69124d8a989a0b1313355cae52/aiohttp-3.14.3-cp313-cp313-win_arm64.whl", hash = "sha256:ddcac3c6b382e81f1dd0499199d4136b877beb4cb5ef770bbbfba56c4b8f55d2", size = 447676, upload-time = "2026-07-23T01:55:40.451Z" },
 ]
 
+[[package]]
+name = "aiohttp-socks"
+version = "0.12.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "aiohttp" },
+    { name = "python-socks" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/18/1d/a306e0111222180e60f17131a3f5d9bc694dd999a8115959a7dd76c2238e/aiohttp_socks-0.12.0.tar.gz", hash = "sha256:3caf9f5a4164611122d412bc11b2f9114fd29c85e1ba27bb38060d3c236bdc8d", size = 12061, upload-time = "2026-08-12T04:43:15.791Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/86/64/ca6289632020523ea1841f01363f83ada10eeb0f21dd931fc1118dd85668/aiohttp_socks-0.12.0-py3-none-any.whl", hash = "sha256:ba6f95ec775c761d87f8578ab48f137d0457c676da104984202bf75e747d5ee6", size = 10693, upload-time = "2026-08-12T04:43:14.524Z" },
+]
+
 [[package]]
 name = "aiosignal"
 version = "1.4.0"
@@ -1075,6 +1088,7 @@ version = "0.1.0"
 source = { editable = "." }
 dependencies = [
     { name = "aiogram" },
+    { name = "aiohttp-socks" },
     { name = "apscheduler" },
     { name = "asyncpg" },
     { name = "fastapi" },
@@ -1103,6 +1117,7 @@ dev = [
 [package.metadata]
 requires-dist = [
     { name = "aiogram", specifier = ">=3.10" },
+    { name = "aiohttp-socks", specifier = ">=0.8" },
     { name = "apscheduler", specifier = ">=3.10,<4" },
     { name = "asyncpg", specifier = ">=0.29" },
     { name = "fastapi", specifier = ">=0.115" },
@@ -1340,6 +1355,15 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/e1/04/e8135ebd1ad02c56ec633277529b2602ff99ff634be76cdba5744cf554fd/python_multipart-0.0.32-py3-none-any.whl", hash = "sha256:ff6d3f776f16878c894e52e107296ffc890e913c611b1a4ec6c44e2821fe2e23", size = 30042, upload-time = "2026-06-04T16:18:57.319Z" },
 ]
 
+[[package]]
+name = "python-socks"
+version = "3.1.1"
+source = { registry = "https://pypi.org/simple" }
+sdist = { url = "https://files.pythonhosted.org/packages/04/ad/484ffb79532517b11a90af38647c38652224650b31a7ae1cedd5a418d8ab/python_socks-3.1.1.tar.gz", hash = "sha256:8d3e817cdbe858dc0bb8c8fdc8e79b6ce37acce110d33374c6f57a675cc9029e", size = 232781, upload-time = "2026-09-08T13:03:31.058Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/3b/23/2c2cef1b4313c55d1713201acd4ee2043fb2cf22e546ca9d36bf4317faea/python_socks-3.1.1-py3-none-any.whl", hash = "sha256:327e0d6378702c73a7790bf732e9f01392f17b48c7348a50b5bd1f710c2df1be", size = 49566, upload-time = "2026-09-08T13:03:29.604Z" },
+]
+
 [[package]]
 name = "pyyaml"
 version = "6.0.3"

web/.dockerignore

@@ -0,0 +1,3 @@
+node_modules
+dist
+.vite