diff --git a/cli.py b/cli.py
index 3cc4116b49ea1..14cf9b29f3ba4 100644
--- a/cli.py
+++ b/cli.py
@@ -4075,6 +4075,23 @@ def _normalize_moa_model(model: Optional[str]) -> tuple[Optional[str], Optional[
     return None, model
 
 
+class _VoiceInputMessage:
+    """Sentinel wrapper for voice-transcribed messages in ``_pending_input``.
+
+    Distinguishes STT output from manually typed text while voice mode is
+    active, so the concise-voice-response prefix is applied only to messages
+    that actually came from the microphone (#65827).
+    """
+
+    __slots__ = ("text",)
+
+    def __init__(self, text: str):
+        self.text = text
+
+    def __str__(self) -> str:
+        return self.text
+
+
 class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
     """
     Interactive CLI for the Hermes Agent.
@@ -11928,7 +11945,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                 self._attached_images.clear()
                 if hasattr(self, '_app') and self._app:
                     self._app.invalidate()
-                self._pending_input.put(transcript)
+                self._pending_input.put(_VoiceInputMessage(transcript))
                 submitted = True
             elif result.get("success"):
                 _cprint(f"{_DIM}No speech detected.{_RST}")
@@ -12116,7 +12133,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                     _cprint(f"\n{_DIM}Stop phrase detected — ending voice chat.{_RST}")
                     self._disable_voice_mode()
                     return
-                self._pending_input.put(transcript)
+                self._pending_input.put(_VoiceInputMessage(transcript))
                 submitted = True
             elif not result.get("success"):
                 _cprint(f"\n{_DIM}Transcription failed: {result.get('error', 'Unknown error')}{_RST}")
@@ -12820,7 +12837,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
             except Exception:
                 pass
 
-    def chat(self, message, images: list = None) -> Optional[str]:
+    def chat(self, message, images: list = None, voice_input: bool = False) -> Optional[str]:
         """
         Send a message to the agent and get a response.
         
@@ -12835,6 +12852,8 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
         Args:
             message: The user's message (str or multimodal content list)
             images: Optional list of Path objects for attached images
+            voice_input: True when the message came from voice transcription
+                (gates the concise voice-response prefix, #65827)
             
         Returns:
             The agent's response, or None on error
@@ -13062,7 +13081,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
             # model responds concisely. The prefix is API-call-local only —
             # run_conversation persists the original clean user message.
             _voice_prefix = ""
-            if self._voice_mode and isinstance(message, str):
+            if voice_input and isinstance(message, str):
                 _voice_prefix = (
                     "[Voice input — respond concisely and conversationally, "
                     "2-3 sentences max. No code blocks or markdown.] "
@@ -16226,7 +16245,13 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                             except Exception:
                                 pass
                         continue
-                    
+
+                    # Voice-transcribed messages arrive wrapped in a sentinel
+                    # so only genuine STT output gets the voice prefix (#65827).
+                    is_voice_input = isinstance(user_input, _VoiceInputMessage)
+                    if is_voice_input:
+                        user_input = user_input.text
+
                     if not user_input:
                         continue
 
@@ -16322,7 +16347,7 @@ class HermesCLI(CLIAgentSetupMixin, CLICommandsMixin, CLIBillingMixin):
                     app.invalidate()  # Refresh status line
 
                     try:
-                        self.chat(user_input, images=submit_images or None)
+                        self.chat(user_input, images=submit_images or None, voice_input=is_voice_input)
                     finally:
                         self._agent_running = False
                         self._spinner_text = ""