Merge pull request 'release: dev → main (финальная сдача)' (#114) from dev into main

.env.example

@@ -23,6 +23,8 @@ 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)

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**.
+> Хакатон MedITron · кейс SberMed AI «ИИ-помощник пациента». Прототип работает **только на синтетических данных**.
-> Telegram используется только для прототипа на синтетических данных; прод-канал — MAX / Госуслуги.
+> 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
+## 1. Цели и задачи
-2. Что реализовано
-3. Развёртывание, запуск, использование
-4. Входные и выходные данные, API
-5. Ограничения и особенности медицинской тематики
 
-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/fixtures.py

@@ -25,7 +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"),
 ]
 
 

app/api/patient_web.py

@@ -158,6 +158,7 @@ async def auth_emias_demo(body: EmiasDemoIn) -> AuthOut:
 
 
 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)
@@ -172,12 +173,13 @@ async def _fresh_demo_copy(*, start_in_chat: bool) -> AuthOut:
     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"{persona.display_name} (демо)",
+            display_name=f"{name} (демо)",
             snils=snils,
             clinic_id="demo",
             doctor_id=fixtures.DOCTOR_ID,
@@ -194,7 +196,7 @@ async def _fresh_demo_copy(*, start_in_chat: bool) -> AuthOut:
             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=persona.display_name)
+    return AuthOut(token=token, mode="emias", display_name=name)
 
 
 @router.post("/auth/anonymous", response_model=AuthOut)

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
@@ -42,10 +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)
@@ -219,6 +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."""

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

@@ -44,6 +44,8 @@ class Settings(BaseSettings):
 
     # 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.
@@ -80,6 +82,9 @@ class Settings(BaseSettings):
     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/intake_questions.py

@@ -61,6 +61,11 @@ 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: 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

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",

app/dialog/engine.py

@@ -37,6 +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
@@ -68,6 +69,9 @@ 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
 )
@@ -145,6 +149,8 @@ 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.
@@ -177,6 +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):

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,
@@ -15,6 +16,7 @@ from app.core.intake_questions import (
     group_answers,
     next_intake,
     red_flag_category,
+    red_flag_triggers,
     scale_question,
     texts,
 )
@@ -116,6 +118,39 @@ 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:
@@ -170,18 +205,27 @@ 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()
     # 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:
+        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)

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/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

@@ -21,6 +21,7 @@ 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,
@@ -127,6 +128,7 @@ async def answer_chat(
         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)},
@@ -159,6 +161,7 @@ async def answer_chat(
                     }
                 )
                 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))

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__)
 
@@ -53,6 +54,8 @@ class OpenRouterProvider:
         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):
@@ -174,6 +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))

config/instruments/red_flags.yaml

@@ -17,6 +17,8 @@ items:
 - 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

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.

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: |-

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:

infra/deploy/docker-compose.server.yml

@@ -28,3 +28,7 @@ services:
       - "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"]

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_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_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_rules.py

@@ -125,8 +125,9 @@ def test_cssrs_levels() -> None:
 @pytest.mark.parametrize(
     ("weapon", "cssrs", "item9", "fires"),
     [
-        (True, "moderate", None, True),
+        # Team decision 2026-10-04: weapon access triggers nothing (rule disabled in config/rules.yaml).
-        (True, None, 1, True),
+        (True, "moderate", None, False),
+        (True, None, 1, False),
         (True, "low", 0, False),
         (False, "high", 2, False),
         (None, "moderate", None, False),

tests/test_services_monitoring.py

@@ -223,10 +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"}
+    # Team decision 2026-10-04: weapon access is stored but triggers nothing (no crisis, no safety block).
-    plan = next(a for a in actions if a.type == "show_safety_plan")
+    assert "weapon_access_risk" not in {a.code for a in actions}
-    assert plan.params["block"] == "means_restriction"
+    assert not any(a.type == "show_safety_plan" and a.params.get("block") == "means_restriction" for a in actions)
-    assert await color() == "red"
 
 
 async def test_snapshot_mapping_and_unknown_patient(db: MemoryNotifier) -> None:

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

web/Dockerfile

@@ -0,0 +1,12 @@
+# Patient web app: build with Node in a throwaway stage, serve the static bundle with nginx (+ /api → app).
+# The jury only needs Docker: `make up` builds and starts it (http://localhost:8080).
+FROM node:20-alpine AS build
+WORKDIR /w
+COPY package.json package-lock.json ./
+RUN npm ci --no-audit --no-fund
+COPY . .
+RUN npm run build
+
+FROM nginx:1.27-alpine
+COPY nginx.conf /etc/nginx/conf.d/default.conf
+COPY --from=build /w/dist /usr/share/nginx/html

web/nginx.conf

@@ -0,0 +1,18 @@
+# Same-origin API: the app calls /api/v1/... — proxied to the FastAPI container (SSE streaming unbuffered).
+server {
+    listen 80;
+    root /usr/share/nginx/html;
+    index index.html;
+
+    location / {
+        try_files $uri $uri/ /index.html;
+    }
+
+    location /api/ {
+        proxy_pass http://app:8000;
+        proxy_set_header Host $host;
+        proxy_http_version 1.1;
+        proxy_buffering off;
+        proxy_read_timeout 120s;
+    }
+}

web/src/api/client.ts

@@ -59,3 +59,53 @@ export const api = {
   outbox: async (token: string) =>
     (await request<{ replies: Reply[] }>(`/dialog/${encodeURIComponent(token)}/outbox`)).replies,
 };
+
+export type StreamEvent =
+  | { kind: "status"; text: string }
+  | { kind: "delta"; text: string }
+  | { kind: "reset" }
+  | { kind: "done"; replies: Reply[] };
+
+/** POST /dialog/{token}/events/stream (Server-Sent Events): agent steps and the answer as it is written,
+ *  then the final replies. Resolves with the final replies; throws if the stream breaks (caller falls back). */
+export async function sendEventStream(
+  token: string,
+  event: DialogEvent,
+  onEvent: (e: StreamEvent) => void,
+): Promise<Reply[]> {
+  const res = await fetch(`${BASE}/dialog/${token}/events/stream`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json", Accept: "text/event-stream" },
+    body: JSON.stringify(event),
+  });
+  if (!res.ok || !res.body) throw new ApiError(res.status, res.statusText);
+  const reader = res.body.getReader();
+  const decoder = new TextDecoder();
+  let buf = "";
+  for (;;) {
+    const { value, done } = await reader.read();
+    if (done) break;
+    buf += decoder.decode(value, { stream: true });
+    let cut: number;
+    while ((cut = buf.indexOf("\n\n")) >= 0) {
+      const block = buf.slice(0, cut);
+      buf = buf.slice(cut + 2);
+      let kind = "message";
+      let data = "";
+      for (const line of block.split("\n")) {
+        if (line.startsWith("event:")) kind = line.slice(6).trim();
+        else if (line.startsWith("data:")) data += line.slice(5).trim();
+      }
+      if (!data) continue;
+      const payload = JSON.parse(data);
+      if (kind === "status") onEvent({ kind, text: String(payload) });
+      else if (kind === "delta") onEvent({ kind, text: String(payload) });
+      else if (kind === "reset") onEvent({ kind: "reset" });
+      else if (kind === "done") {
+        onEvent({ kind: "done", replies: payload as Reply[] });
+        return payload as Reply[];
+      } else if (kind === "error") throw new ApiError(500, "stream error");
+    }
+  }
+  throw new ApiError(0, "stream closed");
+}

web/src/components/AppShell.tsx

@@ -16,6 +16,7 @@ import { Icon, LogoMark, type IconName } from "./Icon";
 
 const TABS: { to: string; label: string; icon: IconName }[] = [
   { to: "/today", label: ru.nav.today, icon: "home" },
+  { to: "/messages", label: ru.nav.chat, icon: "message" },
   { to: "/meds", label: ru.nav.meds, icon: "pill" },
   { to: "/doctor", label: ru.nav.doctor, icon: "doctor" },
   { to: "/profile", label: ru.nav.profile, icon: "user" },

web/src/content/ru.ts

@@ -16,6 +16,7 @@ export const ru = {
     doctor: "Врач",
     profile: "Профиль",
     messages: "Сообщения",
+    chat: "Чат",
     signIn: "Вход",
     help: "Помощь",
   },
@@ -251,6 +252,7 @@ export const ru = {
     noticeTitle: "Опрос приостановлен",
   },
   messages: {
+    typing: "Помощник печатает",
     title: "Сообщения",
     inputLabel: "Сообщение",
     inputPlaceholder: "Напишите сообщение",

web/src/dialog/DialogContext.tsx

@@ -1,6 +1,6 @@
 import { createContext, useCallback, useContext, useEffect, useMemo, useRef, useState, type ReactNode } from "react";
 
-import { api } from "../api/client";
+import { api, sendEventStream } from "../api/client";
 import type { DialogEvent, Reply } from "../api/types";
 import { useSession } from "../session/SessionContext";
 import { sessionStore, type ChatEntry, type DialogSnapshot } from "../session/store";
@@ -17,8 +17,15 @@ let seq = 0;
 const newId = () => `${Date.now().toString(36)}-${(seq++).toString(36)}`;
 const asEntries = (replies: Reply[]): ChatEntry[] => replies.map((reply) => ({ id: newId(), from: "assistant", reply }));
 
+/** The agent at work: its steps and the answer as it is being written (cleared when the final reply lands). */
+export interface LiveReply {
+  status: string[];
+  text: string;
+}
+
 interface DialogApi extends DialogSnapshot {
   busy: boolean;
+  live: LiveReply | null;
   error: boolean;
   started: boolean;
   /** Sends an event; `echo` is what the patient «said» (button label / text) for the history. */
@@ -36,6 +43,7 @@ export function DialogProvider({ children }: { children: ReactNode }) {
   const [owner, setOwner] = useState(token);
   const [snap, setSnap] = useState<DialogSnapshot>(() => (token ? sessionStore.loadDialog(token) : EMPTY));
   const [busy, setBusy] = useState(false);
+  const [live, setLive] = useState<LiveReply | null>(null);
   const [error, setError] = useState(false);
   const tokenRef = useRef(token);
   tokenRef.current = token;
@@ -58,7 +66,35 @@ export function DialogProvider({ children }: { children: ReactNode }) {
       setBusy(true);
       setError(false);
       try {
-        const replies = await api.sendEvent(t, event);
+        let replies: Reply[];
+        if (event.type === "text") {
+          // Free text may go to the agent: stream its steps and the answer so the viewer sees it working.
+          const said = echo;
+          if (said) {
+            // Show the patient's message right away, while the agent is still working.
+            setSnap((prev) => ({
+              ...prev,
+              entries: [...prev.entries, { id: newId(), from: "patient" as const, text: said }],
+            }));
+            echo = undefined;
+          }
+          setLive({ status: [], text: "" });
+          try {
+            replies = await sendEventStream(t, event, (e) =>
+              setLive((prev) => {
+                const cur = prev ?? { status: [], text: "" };
+                if (e.kind === "status") return { ...cur, status: [...cur.status, e.text] };
+                if (e.kind === "delta") return { ...cur, text: cur.text + e.text };
+                if (e.kind === "reset") return { ...cur, text: "" };
+                return cur;
+              }),
+            );
+          } catch {
+            replies = await api.sendEvent(t, event); // proxies without SSE support: plain request
+          }
+        } else {
+          replies = await api.sendEvent(t, event);
+        }
         setSnap((prev) => ({
           entries: [
             ...prev.entries,
@@ -73,6 +109,7 @@ export function DialogProvider({ children }: { children: ReactNode }) {
         setError(true);
         return false;
       } finally {
+        setLive(null);
         setBusy(false);
       }
     },
@@ -102,8 +139,8 @@ export function DialogProvider({ children }: { children: ReactNode }) {
   }, [token]);
 
   const value = useMemo<DialogApi>(
-    () => ({ ...snap, busy, error, started: snap.entries.length > 0, send, start, markRead }),
+    () => ({ ...snap, busy, live, error, started: snap.entries.length > 0, send, start, markRead }),
-    [snap, busy, error, send, start, markRead],
+    [snap, busy, live, error, send, start, markRead],
   );
   return <Ctx.Provider value={value}>{children}</Ctx.Provider>;
 }

web/src/screens/Messages.tsx

@@ -18,7 +18,7 @@ export function Messages() {
   useEffect(() => {
     dialog.markRead();
     endRef.current?.scrollIntoView({ block: "end" });
-  }, [dialog.entries.length]);
+  }, [dialog.entries.length, dialog.live?.text, dialog.live?.status.length]);
 
   const send = async (e?: FormEvent) => {
     e?.preventDefault();
@@ -56,6 +56,32 @@ export function Messages() {
               </li>
             ),
           )}
+          {dialog.live && (
+            <li className="bubble bubble--bot bubble--live" aria-busy="true">
+              <span className="bubble__avatar">
+                <Icon name="smile" size={20} />
+              </span>
+              <div className="live">
+                {dialog.live.status.map((st, i) => (
+                  <p key={i} className={`live__step${i === dialog.live!.status.length - 1 && !dialog.live!.text ? " is-active" : ""}`}>
+                    {st}
+                  </p>
+                ))}
+                {dialog.live.text ? (
+                  <p className="live__text">
+                    {dialog.live.text}
+                    <span className="live__cursor" aria-hidden="true" />
+                  </p>
+                ) : (
+                  <p className="live__dots" aria-label={t.typing}>
+                    <span />
+                    <span />
+                    <span />
+                  </p>
+                )}
+              </div>
+            </li>
+          )}
         </ol>
         {quick.length > 0 && (
           <div className="quickreplies">

web/src/theme/global.css

@@ -103,7 +103,7 @@ button:disabled {
   inset: auto 0 0 0;
   z-index: 20;
   display: grid;
-  grid-template-columns: repeat(4, 1fr);
+  grid-template-columns: repeat(5, 1fr);
   height: calc(var(--nav-h) + var(--safe-bottom));
   padding-bottom: var(--safe-bottom);
   background: var(--c-surface);
@@ -2937,3 +2937,19 @@ button.row-card:hover:not(:disabled) {
     animation: none;
   }
 }
+
+
+/* Live agent reply: steps + the answer as it is written (SSE) */
+.bubble--live .live { display: grid; gap: 4px; }
+.live__step { margin: 0; font-size: 13px; color: var(--muted, #6b7280); }
+.live__step.is-active::after { content: ""; display: inline-block; width: 1em; animation: live-ellipsis 1.2s steps(4) infinite; overflow: hidden; vertical-align: bottom; }
+.live__text { margin: 2px 0 0; white-space: pre-wrap; }
+.live__cursor { display: inline-block; width: 7px; height: 1em; margin-left: 2px; background: currentColor; vertical-align: text-bottom; animation: live-blink 1s steps(2) infinite; }
+.live__dots { display: inline-flex; gap: 4px; margin: 4px 0 0; }
+.live__dots span { width: 6px; height: 6px; border-radius: 50%; background: var(--muted, #9ca3af); animation: live-dot 1.2s infinite ease-in-out; }
+.live__dots span:nth-child(2) { animation-delay: .15s; }
+.live__dots span:nth-child(3) { animation-delay: .3s; }
+@keyframes live-blink { 50% { opacity: 0; } }
+@keyframes live-dot { 0%, 80%, 100% { opacity: .3; transform: translateY(0); } 40% { opacity: 1; transform: translateY(-3px); } }
+@keyframes live-ellipsis { from { width: 0; } to { width: 1em; } }
+.tabbar__item span { font-size: 11px; }