LLM Agent: prevent tool-output leakage; nicer responses (no Tool Result / no JSON dumps)

backend/app/services/llm_agent.py

@@ -39,12 +39,17 @@ def _build_system_prompt(path: str) -> str:
     return (
         "You are the Pounce Hunter Companion (domain trading expert). Always respond in English.\n"
         "You have access to internal tools that return live data. Use tools when needed.\n\n"
+        "OUTPUT STYLE:\n"
+        "- Never show raw tool output to the user.\n"
+        "- 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"
         "TOOL CALLING PROTOCOL:\n"
         "- If you need data, respond with ONLY a JSON object:\n"
         '  {"tool_calls":[{"name":"tool_name","args":{...}}, ...]}\n'
         "- Do not include any other text when requesting tools.\n"
         "- After tools are executed, you will receive TOOL_RESULT messages.\n"
-        "- When you are ready to answer the user, respond normally (not JSON).\n\n"
+        "- When you are ready to answer the user, respond normally (not JSON) and do NOT mention tools.\n\n"
         "AVAILABLE TOOLS (JSON schemas):\n"
         f"{json.dumps(tools, ensure_ascii=False)}\n\n"
         "RULES:\n"
@@ -132,7 +137,10 @@ async def run_agent(
             convo.append(
                 {
                     "role": "system",
-                    "content": f"TOOL_RESULT name={name} json={_truncate_json(result)}",
+                    "content": (
+                        f"TOOL_RESULT_INTERNAL name={name} json={_truncate_json(result)}. "
+                        "This is internal context. Do NOT quote or display this to the user."
+                    ),
                 }
             )
 
@@ -153,7 +161,13 @@ async def stream_final_answer(convo: list[dict[str, Any]], *, model: Optional[st
         + [
             {
                 "role": "system",
-                "content": "Final step: respond to the user. Do NOT output JSON tool_calls. Do NOT request tools.",
+                "content": (
+                    "Final step: respond to the user.\n"
+                    "- Do NOT output JSON tool_calls.\n"
+                    "- Do NOT request tools.\n"
+                    "- Do NOT include raw tool outputs, internal tags, or code-fenced JSON.\n"
+                    "- If you used tools, present only a clean human summary."
+                ),
             }
         ],
         "temperature": temperature,