Avoid tools for direct answer prompts

prompts/roles/action.md

@@ -18,6 +18,10 @@ Available tools:
   Args: {"task_description": "what to build or analyze", "language": "python", "context": "optional context"}
 
 Return actions=[] when the user can be answered directly without tools.
+Return actions=[] for simple arithmetic, translation, explanation, formatting,
+summarization, brainstorming, general knowledge, or any request where the
+thinker can answer from the conversation alone. Do not run shell commands like
+`echo`, `python`, `bc`, or calculators for simple arithmetic.
 When tool_observations are already present, request only genuinely missing
 follow-up information. Return actions=[] when the observations are sufficient
 for the thinker to answer.

tests/smoke/test_models_config.py

@@ -14,3 +14,10 @@ def test_models_config_maps_roles_to_same_qwen_non_mtp_model():
 
     assert roles["action"]["structured_output"] is True
     assert roles["thinker"]["max_output_tokens"] == 8192
+
+
+def test_action_prompt_forbids_tools_for_simple_arithmetic():
+    prompt = Path("prompts/roles/action.md").read_text()
+
+    assert "Return actions=[] for simple arithmetic" in prompt
+    assert "Do not run shell commands like" in prompt