Reapply "feat(observability): integrate NeMo Relay runtime and shared metrics"

Signed-off-by: Alex Fournier <afournier@nvidia.com>

model_tools.py

@@ -1056,7 +1056,7 @@ def _emit_post_tool_call_hook(
     listener will actually consume it).
     """
     try:
-        from hermes_cli.plugins import has_hook, invoke_hook
+        from hermes_cli.lifecycle import has_hook, invoke_hook
         if not has_hook("post_tool_call"):
             return
         if status is None:
@@ -1093,6 +1093,7 @@ def handle_function_call(
     enabled_tools: Optional[List[str]] = None,
     skip_pre_tool_call_hook: bool = False,
     skip_tool_request_middleware: bool = False,
+    skip_tool_execution_middleware: bool = False,
     tool_request_middleware_trace: Optional[List[Dict[str, Any]]] = None,
     enabled_toolsets: Optional[List[str]] = None,
     disabled_toolsets: Optional[List[str]] = None,
@@ -1203,6 +1204,7 @@ def handle_function_call(
                 enabled_tools=enabled_tools,
                 skip_pre_tool_call_hook=skip_pre_tool_call_hook,
                 skip_tool_request_middleware=skip_tool_request_middleware,
+                skip_tool_execution_middleware=skip_tool_execution_middleware,
                 tool_request_middleware_trace=list(_tool_middleware_trace),
                 enabled_toolsets=enabled_toolsets,
                 disabled_toolsets=disabled_toolsets,
@@ -1341,19 +1343,22 @@ def handle_function_call(
                         session_id=session_id,
                         user_task=user_task,
                     )
-            from hermes_cli.middleware import run_tool_execution_middleware
+            if skip_tool_execution_middleware:
+                result = _dispatch(function_args)
+            else:
+                from hermes_cli.middleware import run_tool_execution_middleware
 
-            result = run_tool_execution_middleware(
-                function_name,
-                function_args,
-                _dispatch,
-                original_args=_tool_original_args,
-                task_id=task_id or "",
-                session_id=session_id or "",
-                tool_call_id=tool_call_id or "",
-                turn_id=turn_id or "",
-                api_request_id=api_request_id or "",
-            )
+                result = run_tool_execution_middleware(
+                    function_name,
+                    function_args,
+                    _dispatch,
+                    original_args=_tool_original_args,
+                    task_id=task_id or "",
+                    session_id=session_id or "",
+                    tool_call_id=tool_call_id or "",
+                    turn_id=turn_id or "",
+                    api_request_id=api_request_id or "",
+                )
         finally:
             if _approval_tokens is not None and reset_current_observability_context is not None:
                 try:
@@ -1384,7 +1389,7 @@ def handle_function_call(
         # Gated on has_hook so the no-listener path skips both the result
         # field derivation and the payload dispatch.
         try:
-            from hermes_cli.plugins import has_hook, invoke_hook
+            from hermes_cli.lifecycle import has_hook, invoke_hook
             if has_hook("transform_tool_result"):
                 status, error_type, error_message = _tool_result_observer_fields(result)
                 hook_results = invoke_hook(