refactor: move legacy judge message stripping to client._judge_config()

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: jsonbailey

packages/sdk/server-ai/src/ldai/client.py

@@ -9,7 +9,7 @@
 from ldai import log
 from ldai.agent_graph import AgentGraphDefinition
 from ldai.evaluator import Evaluator
-from ldai.judge import Judge
+from ldai.judge import Judge, _strip_legacy_judge_messages
 from ldai.managed_agent import ManagedAgent
 from ldai.managed_agent_graph import ManagedAgentGraph
 from ldai.managed_model import ManagedModel
@@ -237,6 +237,10 @@ def _extract_evaluation_metric_key(variation: Dict[str, Any]) -> Optional[str]:
 
         evaluation_metric_key = _extract_evaluation_metric_key(variation)
 
+        # strip legacy judge template messages before creating config
+        if messages:
+            messages = _strip_legacy_judge_messages(messages)
+
         config = AIJudgeConfig(
             key=key,
             enabled=bool(enabled),

packages/sdk/server-ai/src/ldai/judge/__init__.py

@@ -160,19 +160,6 @@ def get_model_runner(self) -> Runner:
         return self._model_runner
 
     def _build_evaluation_input(self, input_text: str, output_text: str) -> str:
-        """
-        Build the string input for the judge runner.
-
-        Legacy messages (assistant/user messages containing ``{{message_history}}``
-        or ``{{response_to_evaluate}}``) are stripped from the config; the runner
-        was already created from the judge AI config (which carries the system
-        message), so only the plain-text evaluation payload is needed here.
-
-        :param input_text: The input text (message history)
-        :param output_text: The output text to evaluate
-        :return: Formatted evaluation input string
-        """
-        _strip_legacy_judge_messages(self._ai_config.messages or [])
         return f"MESSAGE HISTORY:\n{input_text}\n\nRESPONSE TO EVALUATE:\n{output_text}"
 
     def _parse_evaluation_response(self, data: Dict[str, Any]) -> Optional[Tuple[float, str]]:

packages/sdk/server-ai/tests/test_judge.py

@@ -292,10 +292,14 @@ async def test_evaluate_string_input_format(
         assert input_arg == expected
 
     @pytest.mark.asyncio
-    async def test_evaluate_legacy_config_strips_template_messages(
+    async def test_evaluate_legacy_config_passes_string_input_to_runner(
         self, tracker: LDAIConfigTracker, mock_runner
     ):
-        """Legacy config with assistant/user template messages: runner still gets string input."""
+        """
+        Judge built directly with legacy messages (bypassing the client) still passes
+        a formatted string to the runner.  Legacy message stripping is the client's
+        responsibility; the Judge itself does not strip.
+        """
         legacy_messages = [
             LDMessage(role='system', content='You are a strict judge.'),
             LDMessage(role='assistant', content='{{message_history}}'),