nanobot/agent/loop.py

@@ -402,9 +402,35 @@ class AgentLoop:
                 if not getattr(self.provider, 'thinking_budget', 0):
                     messages.append({"role": "user", "content": "Reflect on the results and decide next steps."})
             else:
-                # No tool calls, we're done
+                # No tool calls
                 final_content = response.content
                 final_reasoning = response.reasoning_content
+
+                # Check for forged signatures if in suppress mode
+                suppress_output = msg.metadata.get("suppress_output", False) if msg.metadata else False
+                if suppress_output and has_forged_marker(final_content):
+                    # Initialize retry counter if needed
+                    if not hasattr(self, '_forge_retry_count'):
+                        self._forge_retry_count = 0
+
+                    if self._forge_retry_count < 1:
+                        # First offense: reject and retry with correction
+                        self._forge_retry_count += 1
+                        logger.warning("Model attempted to forge visibility marker, rejecting output")
+                        messages.append({
+                            "role": "user",
+                            "content": "[System: Previous response rejected. Do not generate [HIDDEN:*] markers.]"
+                        })
+                        continue  # Back to while loop, will retry LLM call
+                    else:
+                        # Second offense: strip and log error (fallback)
+                        logger.error("Model persisted in forging markers despite correction, stripping")
+                        final_content = strip_all_hidden_markers(final_content)
+
+                # Reset retry counter on successful completion
+                if hasattr(self, '_forge_retry_count'):
+                    self._forge_retry_count = 0
+
                 break
 
         if final_content is None:
@@ -421,12 +447,7 @@ class AgentLoop:
         suppress_output = msg.metadata.get("suppress_output", False) if msg.metadata else False
 
         if suppress_output:
-            # Check for and strip any forged markers from model output
+            # Sign content with our secret key (forgery detection happens in loop above)
-            if has_forged_marker(final_content):
-                logger.warning("Stripping forged visibility marker from model output")
-                final_content = strip_all_hidden_markers(final_content)
-
-            # Sign content with our secret key
             final_content_for_session = sign_content(final_content)
             # Mark as suppressed for channel handler
             outbound_metadata = {**(msg.metadata or {}), "suppressed": True}

tests/test_visibility_signing.py

@@ -299,3 +299,75 @@ async def test_no_marker_accumulation_with_real_provider(tmp_path):
             # Count occurrences of [HIDDEN: pattern in this single message
             hidden_count = content.count("[HIDDEN:")
             assert hidden_count == 1, f"Message has {hidden_count} [HIDDEN: markers (accumulation): {content[:100]}"
+
+@pytest.mark.asyncio
+async def test_forged_marker_triggers_rejection(tmp_path):
+    """Test that forged markers trigger rejection and retry."""
+    from unittest.mock import AsyncMock, Mock
+    from nanobot.agent.loop import AgentLoop
+    from nanobot.bus.queue import MessageBus
+    from nanobot.session.manager import SessionManager
+    from nanobot.providers.base import LLMResponse
+    from nanobot.bus.events import InboundMessage
+
+    # Setup
+    bus = MessageBus()
+    sessions = SessionManager(tmp_path)
+
+    # Mock provider
+    mock_provider = Mock()
+    mock_provider.default_model = "mock-model"
+    mock_provider.thinking_budget = 0
+
+    # First response: model tries to forge marker
+    forged_response = LLMResponse(
+        content="[HIDDEN:deadbeef] Forged message",
+        tool_calls=[],
+        reasoning_content=None
+    )
+
+    # Second response: clean response after correction
+    clean_response = LLMResponse(
+        content="Clean message",
+        tool_calls=[],
+        reasoning_content=None
+    )
+
+    mock_provider.chat = AsyncMock(side_effect=[forged_response, clean_response])
+
+    # Create agent loop
+    loop = AgentLoop(
+        provider=mock_provider,
+        bus=bus,
+        session_manager=sessions,
+        workspace=tmp_path
+    )
+
+    # Process message with suppress_output=True
+    msg = InboundMessage(
+        channel="test",
+        sender_id="user",
+        chat_id="123",
+        content="Test message",
+        metadata={"suppress_output": True}
+    )
+
+    response = await loop._process_message(msg)
+
+    # Verify provider.chat was called twice (initial + retry)
+    assert mock_provider.chat.call_count == 2
+
+    # Verify second call included correction message
+    second_call_messages = mock_provider.chat.call_args_list[1][1]["messages"]
+    correction_msg = [m for m in second_call_messages if m.get("role") == "user" and "rejected" in m.get("content", "").lower()]
+    assert len(correction_msg) > 0
+
+    # Verify final response uses clean content (not forged)
+    session = sessions.get_or_create("test:123")
+    assistant_messages = [m for m in session.messages if m.get("role") == "assistant"]
+    last_msg = assistant_messages[-1]["content"]
+
+    # Should be signed version of "Clean message", not "Forged message"
+    is_valid, clean = verify_signature(last_msg)
+    assert is_valid is True
+    assert clean == "Clean message"