fix(cli): prepend turn notes to multimodal messages

cli.py

@@ -2355,6 +2355,40 @@ def _trunc(name: str, limit: int) -> str:
     )
 
 
+def _prepend_turn_note_to_message(note: str, message):
+    """Prepend a one-shot CLI note to text or native multimodal content."""
+    note = str(note or "").strip()
+    if not note:
+        return message
+
+    if isinstance(message, str):
+        return f"{note}\n\n{message}" if message else note
+
+    if isinstance(message, list):
+        updated = []
+        inserted = False
+        for part in message:
+            if (
+                not inserted
+                and isinstance(part, dict)
+                and part.get("type") == "text"
+                and isinstance(part.get("text"), str)
+            ):
+                new_part = dict(part)
+                text = new_part.get("text", "")
+                new_part["text"] = f"{note}\n\n{text}" if text else note
+                updated.append(new_part)
+                inserted = True
+            else:
+                updated.append(part)
+
+        if not inserted:
+            return [{"type": "text", "text": note}, *updated]
+        return updated
+
+    return f"{note}\n\n{message}"
+
+
 def _should_auto_attach_clipboard_image_on_paste(pasted_text: str) -> bool:
     """Auto-attach clipboard images only for image-only paste gestures."""
     return not pasted_text.strip()
@@ -12138,14 +12172,14 @@ def run_agent():
                 # Prepend pending model switch note so the model knows about the switch
                 _msn = getattr(self, '_pending_model_switch_note', None)
                 if _msn:
-                    agent_message = _msn + "\n\n" + agent_message
+                    agent_message = _prepend_turn_note_to_message(_msn, agent_message)
                     self._pending_model_switch_note = None
                 # Prepend pending /reload-skills note so the model sees which
                 # skills were added/removed before handling this turn. Same
                 # one-shot queue pattern as the model-switch note above.
                 _srn = getattr(self, '_pending_skills_reload_note', None)
                 if _srn:
-                    agent_message = _srn + "\n\n" + agent_message
+                    agent_message = _prepend_turn_note_to_message(_srn, agent_message)
                     self._pending_skills_reload_note = None
                 try:
                     result = self.agent.run_conversation(

tests/cli/test_cli_image_command.py

@@ -5,6 +5,7 @@
     HermesCLI,
     _collect_query_images,
     _format_image_attachment_badges,
+    _prepend_turn_note_to_message,
     _termux_example_image_path,
 )
 
@@ -107,3 +108,33 @@ def test_compact_badges_summarize_multiple_images(self, tmp_path):
         badges = _format_image_attachment_badges([img1, img2], image_counter=2, width=45)
 
         assert badges == "[📎 2 images attached]"
+
+
+class TestPrependTurnNoteToMessage:
+    def test_prepends_note_to_text_message(self):
+        message = _prepend_turn_note_to_message("[MODEL SWITCHED]", "describe this")
+
+        assert message == "[MODEL SWITCHED]\n\ndescribe this"
+
+    def test_prepends_note_to_native_image_text_part_without_mutating_original(self):
+        original = [
+            {"type": "text", "text": "What do you see?\n\n[Image attached at: /tmp/a.png]"},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}},
+        ]
+
+        message = _prepend_turn_note_to_message("[MODEL SWITCHED]", original)
+
+        assert isinstance(message, list)
+        assert message[0]["text"].startswith("[MODEL SWITCHED]\n\nWhat do you see?")
+        assert message[1] == original[1]
+        assert original[0]["text"].startswith("What do you see?")
+
+    def test_inserts_text_part_when_native_message_has_only_images(self):
+        original = [
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}},
+        ]
+
+        message = _prepend_turn_note_to_message("[MODEL SWITCHED]", original)
+
+        assert message[0] == {"type": "text", "text": "[MODEL SWITCHED]"}
+        assert message[1] == original[0]