diff --git a/app/dialog/engine.py b/app/dialog/engine.py
index 495bf48..2f7d43e 100644
--- a/app/dialog/engine.py
+++ b/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
@@ -67,8 +68,12 @@ from app.services.contracts import MonitoringService, get_monitoring
 log = logging.getLogger(__name__)
 
 LAST_PROMPT = "last_prompt"
+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()
@@ -309,12 +314,18 @@ class Engine:
         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"
-        if ctx.text is None or (ctx.state in SIDE_ANSWER_SKIP and not in_crisis):
-            return None
-        if not looks_like_question(ctx.text):
-            return None
-        answer = await knowledge_answer(ctx, explicit=True)
+        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)
diff --git a/app/services/monitoring.py b/app/services/monitoring.py
index 34d82b6..7497116 100644
--- a/app/services/monitoring.py
+++ b/app/services/monitoring.py
@@ -51,6 +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}
@@ -68,6 +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
diff --git a/config/texts/knowledge.yaml b/config/texts/knowledge.yaml
index e9c5cba..ee69678 100644
--- a/config/texts/knowledge.yaml
+++ b/config/texts/knowledge.yaml
@@ -33,6 +33,13 @@ question_starts:
   - "чем"
   - "где"
   - "есть ли"
+  - "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.
diff --git a/tests/test_dialog_engine.py b/tests/test_dialog_engine.py
index 13beb44..6727f89 100644
--- a/tests/test_dialog_engine.py
+++ b/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"]
 
diff --git a/tests/test_dialog_side_answer.py b/tests/test_dialog_side_answer.py
index aa22bd4..eb4441f 100644
--- a/tests/test_dialog_side_answer.py
+++ b/tests/test_dialog_side_answer.py
@@ -3,9 +3,15 @@ buttons; in a crisis the answer carries the emergency contacts and the crisis st
 
 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:
@@ -25,6 +31,7 @@ async def test_question_mid_scale_is_answered_and_question_returns(engine):  # n
 
 
 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()
@@ -43,3 +50,24 @@ async def test_non_question_mid_scale_still_goes_to_the_scale(engine):  # noqa:
     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)
diff --git a/tests/test_services_crisis_refire.py b/tests/test_services_crisis_refire.py
new file mode 100644
index 0000000..e81c0aa
--- /dev/null
+++ b/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]