Hunter Companion: natural greetings + no invented tasks; lower temperature

backend/app/services/llm_agent.py

@@ -15,6 +15,32 @@ from app.services.llm_tools import execute_tool, tool_catalog_for_prompt
 settings = get_settings()
 
 
+def _is_greeting(text: str) -> bool:
+    t = (text or "").strip().lower()
+    if not t:
+        return False
+    # common minimal greetings
+    greetings = {
+        "hi",
+        "hello",
+        "hey",
+        "yo",
+        "sup",
+        "hola",
+        "hallo",
+        "guten tag",
+        "good morning",
+        "good evening",
+        "good afternoon",
+    }
+    if t in greetings:
+        return True
+    # very short greeting-like messages
+    if len(t) <= 6 and t.replace("!", "").replace(".", "") in greetings:
+        return True
+    return False
+
+
 def _tier_level(tier: str) -> int:
     t = (tier or "").lower()
     if t == "tycoon":
@@ -44,6 +70,13 @@ def _build_system_prompt(path: str) -> str:
         "- Never print phrases like 'Tool Result', 'TOOL_RESULT', or code-fenced JSON.\n"
         "- If you used tools, silently incorporate the data and present ONLY a clean summary.\n"
         "- Keep formatting simple: short paragraphs and bullet points. Avoid dumping structured data.\n\n"
+        "BEHAVIOR:\n"
+        "- Do NOT invent user preferences, keywords, TLDs, budgets, or tasks.\n"
+        "- If the user greets you or sends a minimal message (e.g., 'hi', 'hello'), respond naturally and ask what they want help with.\n"
+        "- Ask 1–2 clarifying questions when the user request is ambiguous.\n\n"
+        "WHEN TO USE TOOLS:\n"
+        "- Use tools only when the user explicitly asks about their account data, their current page, their lists (watchlist/portfolio/listings/inbox/yield), or when they provide a specific domain to analyze.\n"
+        "- Never proactively 'fetch' domains or run scans based on guessed keywords.\n\n"
         "TOOL CALLING PROTOCOL:\n"
         "- If you need data, respond with ONLY a JSON object:\n"
         '  {"tool_calls":[{"name":"tool_name","args":{...}}, ...]}\n'
@@ -54,7 +87,7 @@ def _build_system_prompt(path: str) -> str:
         f"{json.dumps(tools, ensure_ascii=False)}\n\n"
         "RULES:\n"
         "- Never claim you checked external sources unless the user provided the data.\n"
-        "- Keep answers practical and decisive. If domain-related: include BUY/CONSIDER/SKIP + bullets.\n"
+        "- Keep answers practical and decisive. If (and only if) the user is asking about a specific domain: include BUY/CONSIDER/SKIP + bullets.\n"
     )
 
 
@@ -111,6 +144,23 @@ async def run_agent(
     ]
     convo = base + (messages or [])
 
+    # If the user just greets, answer naturally without tool-looping.
+    last_user = next((m for m in reversed(messages or []) if m.get("role") == "user"), None)
+    if last_user and _is_greeting(str(last_user.get("content") or "")):
+        convo.append(
+            {
+                "role": "assistant",
+                "content": (
+                    "Hey — how can I help?\n\n"
+                    "If you want, tell me:\n"
+                    "- which Terminal page you’re on (or what you’re trying to do)\n"
+                    "- or a specific domain you’re considering\n"
+                    "- or what outcome you want (find deals, assess a name, manage leads, etc.)"
+                ),
+            }
+        )
+        return convo
+
     for _ in range(max_steps):
         payload = {
             "model": model or settings.llm_default_model,

frontend/src/components/chat/HunterCompanion.tsx

@@ -192,7 +192,7 @@ export function HunterCompanion() {
     try {
       await streamChatCompletion({
         messages: [...system, ...history, { role: 'user', content: text }],
-        temperature: 0.7,
+        temperature: 0.5,
         path: pathname || '/terminal/hunt',
         onDelta: (delta) => {
           setMessages((prev) =>