fix: streaming display — show tool feed lines and fix response box framing

Two display bugs when streaming is enabled in the interactive CLI:

1. Tool feed lines (┊ 📖 read, ┊ 💻 $, etc.) were invisible during
   streaming sessions. The sequential tool execution path at line 4840
   had a guard `not self._has_stream_consumers()` that skipped the
   entire spinner + cute message display when a stream_delta_callback
   was registered. But no tokens are streaming during tool execution —
   the _executing_tools flag already handles this for _vprint. Removed
   the unnecessary guard.

2. Response box (╭─ ⚕ Hermes ─╮) could wrap intermediate tool-calling
   turns instead of only the final response. Content tokens arrive
   before tool_call tokens during streaming, so the box opens before
   the agent knows tool calls are coming. Now sends a None sentinel
   through the stream callback when tool_calls are confirmed, which
   tells the CLI to close any open box and reset stream state. Only
   the actual final response gets the Hermes border.

Live tested: reasoning blocks → tool feed → reasoning → response box
now display in the correct sequence.

cli.py

@@ -1511,10 +1511,22 @@ class HermesCLI:
         partial lines and emits complete lines via _cprint to work
         reliably with prompt_toolkit's patch_stdout.
 
+        A ``None`` sentinel signals that an intermediate turn (with tool
+        calls) just completed — close any open response/reasoning box
+        and reset so the final response gets a clean frame.
+
         Reasoning/thinking blocks (<REASONING_SCRATCHPAD>, <think>, etc.)
         are suppressed during streaming since they'd display raw XML tags.
         The agent strips them from the final response anyway.
         """
+        if text is None:
+            # Intermediate turn ended — discard any open box and reset
+            self._close_reasoning_box()
+            if self._stream_box_opened:
+                w = shutil.get_terminal_size().columns
+                _cprint(f"{_GOLD}╰{'─' * (w - 2)}╯{_RST}")
+            self._reset_stream_state()
+            return
         if not text:
             return
 

run_agent.py

@@ -4837,7 +4837,7 @@ class AIAgent:
                         spinner.stop(cute_msg)
                     elif self.quiet_mode:
                         self._vprint(f"  {cute_msg}")
-            elif self.quiet_mode and not self._has_stream_consumers():
+            elif self.quiet_mode:
                 face = random.choice(KawaiiSpinner.KAWAII_WAITING)
                 emoji = _get_tool_emoji(function_name)
                 preview = _build_tool_preview(function_name, function_args) or function_name
@@ -6552,7 +6552,13 @@ class AIAgent:
                     )
 
                     assistant_msg = self._build_assistant_message(assistant_message, finish_reason)
-                    
+
+                    # Signal streaming display that this was an intermediate turn.
+                    # Content may have been streamed (opening the response box) before
+                    # tool_calls were detected.  The None sentinel tells the CLI to
+                    # close any open box and reset so the final response gets a clean frame.
+                    self._fire_stream_delta(None)
+
                     # If this turn has both content AND tool_calls, capture the content
                     # as a fallback final response. Common pattern: model delivers its
                     # answer and calls memory/skill tools as a side-effect in the same