fix: use clean user message for auto-title instead of skill-bloated input

cli.py

@@ -7798,7 +7798,7 @@ def run_agent():
                     maybe_auto_title(
                         self._session_db,
                         self.session_id,
-                        message,
+                        result.get("original_user_message") or message,
                         response,
                         self.conversation_history,
                     )

gateway/run.py

@@ -8801,10 +8801,11 @@ def _approval_notify_sync(approval_data: dict) -> None:
                 try:
                     from agent.title_generator import maybe_auto_title
                     all_msgs = result_holder[0].get("messages", []) if result_holder[0] else []
+                    _clean_msg = (result_holder[0] or {}).get("original_user_message") or message
                     maybe_auto_title(
                         self._session_db,
                         effective_session_id,
-                        message,
+                        _clean_msg,
                         final_response,
                         all_msgs,
                     )

run_agent.py

@@ -10708,6 +10708,7 @@ def _stop_spinner():
             "partial": False,  # True only when stopped due to invalid tool calls
             "interrupted": interrupted,
             "response_previewed": getattr(self, "_response_was_previewed", False),
+            "original_user_message": original_user_message,
             "model": self.model,
             "provider": self.provider,
             "base_url": self.base_url,

tests/test_auto_title_clean_message.py

@@ -0,0 +1,49 @@
+"""Tests for auto-title using clean user message instead of skill-bloated input.
+
+When a skill is active, the message passed to run_conversation contains
+the full skill content (often 1K+ chars). The title generator truncates
+to 500 chars, so it only sees skill boilerplate and generates a wrong
+title. The fix passes original_user_message (the clean user input) to
+maybe_auto_title instead.
+"""
+
+from agent.title_generator import _TITLE_PROMPT
+
+
+class TestResultIncludesOriginalMessage:
+
+    def test_result_dict_has_original_user_message(self):
+        """run_conversation result must include original_user_message."""
+        # Simulate a result dict
+        result = {
+            "final_response": "Here are some cat gifs!",
+            "original_user_message": "find me funny cat gifs",
+            "messages": [],
+        }
+        assert "original_user_message" in result
+        assert result["original_user_message"] == "find me funny cat gifs"
+
+    def test_clean_message_preferred_over_bloated(self):
+        """When original_user_message exists, use it over raw message."""
+        bloated_message = "[SYSTEM: skill content 1000 chars...] find cats"
+        result = {"original_user_message": "find cats"}
+
+        title_input = result.get("original_user_message") or bloated_message
+        assert title_input == "find cats"
+        assert len(title_input) < 50  # not the 1000+ char bloated version
+
+    def test_fallback_to_message_when_no_original(self):
+        """When original_user_message is missing, fall back to raw message."""
+        message = "hello world"
+        result = {}
+
+        title_input = result.get("original_user_message") or message
+        assert title_input == "hello world"
+
+    def test_title_prompt_fits_clean_input(self):
+        """Title prompt truncates to 500 chars — clean input fits easily."""
+        clean_input = "find me funny cat gifs"
+        assert len(clean_input) < 500  # will be fully visible to title LLM
+
+        bloated = "[SYSTEM: ...skill...] " * 50 + clean_input
+        assert len(bloated) > 500  # would be truncated, hiding the actual intent