fix(bot): agent polish — instant small talk, capped tool calls, analogues, plain text, definitions allowed

- Greetings / chit-chat answered without the LLM (instant, free, no medical findings dumped into «привет»).
- At most 4 tool calls per turn (one «как дела?» turn fired 23 parallel searches).
- «My drug» is never asked back; analogues → the drug's group, replacement only by the doctor.
- Markdown stripped to plain text with «•»; no sources footer under short replies.
- Diagnosis filter no longer blocks definitions («ПТСР — это расстройство…»); diagnosis-like statements about
  the patient are still blocked.

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

app/llm/agent/__init__.py

@@ -14,6 +14,7 @@ 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
@@ -35,6 +36,19 @@ 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 +80,15 @@ 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 +125,8 @@ 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")
     ctx = ToolContext(token=token)
     messages: list[dict[str, Any]] = [
         {"role": "system", "content": SYSTEM_PROMPT + await _patient_brief(ctx)},
@@ -123,17 +148,17 @@ 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]:
                     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 +184,14 @@ 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)