fix(agent): make tool loop guardrails warning-first

hermes_cli/config.py

@@ -607,6 +607,24 @@ DEFAULT_CONFIG = {
         "max_line_length": 2000,
     },
 
+    # Tool loop guardrails nudge models when they repeat failed or
+    # non-progressing tool calls. Soft warnings are always-on by default;
+    # hard stops are opt-in so interactive CLI/TUI sessions keep flowing.
+    "tool_loop_guardrails": {
+        "warnings_enabled": True,
+        "hard_stop_enabled": False,
+        "warn_after": {
+            "exact_failure": 2,
+            "same_tool_failure": 3,
+            "idempotent_no_progress": 2,
+        },
+        "hard_stop_after": {
+            "exact_failure": 5,
+            "same_tool_failure": 8,
+            "idempotent_no_progress": 5,
+        },
+    },
+
     "compression": {
         "enabled": True,
         "threshold": 0.50,            # compress when context usage exceeds this ratio