docs: Update _split_prompt docstring to explain guardrail scanning behavior

Author: sudsali

scripts/issue_bot/bedrock_client.py

@@ -93,11 +93,17 @@ def invoke(self, prompt, max_tokens=4096, temperature=0.3, json_schema=None):
             return None
 
     def _split_prompt(self, prompt):
-        """Split rendered prompt into system prompt and user message.
+        """Split rendered prompt into system prompt and user message blocks.
 
-        Everything before <user_input> goes in system prompt (exempt from guardrail).
-        Only <user_input> content (issue title/body/comments) is guarded.
-        This keeps KB, diffs, and codebase maps out of guardrail scanning.
+        When guardrail is configured and both text and guardContent blocks are
+        present in the user message, the guardrail only scans guardContent blocks.
+        Text blocks are passed through unscanned.
+
+        Split at <user_input>:
+        - Before <user_input> → system prompt (cached, never scanned)
+        - <user_input> content → guardContent (scanned by guardrail)
+
+        If no <user_input> marker, everything goes as guardContent.
         """
         marker = "<user_input>"
         idx = prompt.find(marker)
@@ -110,5 +116,6 @@ def _split_prompt(self, prompt):
         user_data = prompt[idx:].strip()
 
         if self._guardrail_id:
+            # guardContent is scanned; the system prompt (with KB/diff) is exempt
             return system, [{"guardContent": {"text": {"text": user_data}}}]
         return system, [{"text": user_data}]