feat(bot): every free-text message gets the agent's answer; every demo entry is a private copy

- Idle chat: all messages go to the agent (greetings, side-effect and substitute questions included); check-in
  signals are still recorded silently, quit intent keeps its rule reaction. Seen live: «какие будут побочки»
  was logged as a complaint («Спасибо, отметил»), «привет» got a canned phrase.
- Agent prompt: warm short replies to small talk; substitutes → only the doctor changes the drug.
- ЕМИАС demo persona button creates a private copy too (starting from consent + questionnaires, subject to
  the screening switch); the web app drops a stored session of the old shared demo account.

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

app/api/patient_web.py

@@ -148,6 +148,10 @@ 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)
@@ -158,6 +162,10 @@ DEMO_SOURCE = "tok-andrey"
 
 @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."""
@@ -177,12 +185,14 @@ async def auth_demo_fresh() -> AuthOut:
             birth_year=now.year - 34,
             sex="m",
         )
-        await identity_repo.record_consent(s, token=token, version="demo", now=now)
+        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)
-        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)
+        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=persona.display_name)
 

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/llm/agent/prompts.py

@@ -32,6 +32,10 @@ _SYSTEM_PROMPT = """Ты — помощник пациента в сервисе
 - проводить психотерапию; обещать конфиденциальность сверх реальной;
 - выдумывать факты, которых нет в результатах инструментов.
 
+На приветствия и болтовню («привет», «как дела?») отвечай коротко и тепло и предложи, с чем можешь помочь:
+вопросы о лечении, лекарствах, побочных эффектах, самочувствии. Про замену препарата или аналоги: объясни,
+что подбор и замену делает только врач, и при желании назови группу препарата по инструкции.
+
 Стиль: по-русски, на «вы», спокойно и по-человечески, 3–6 коротких предложений, без markdown-заголовков.
 Если в вопросе есть признаки угрозы жизни — коротко скажи, что важно сразу позвонить 112 или на телефон
 доверия {phone:helpline}, и что врач будет уведомлён."""