diff --git a/openevolve/llm/claude_code.py b/openevolve/llm/claude_code.py
index 5bd3c9fb3..c42ae8dcf 100644
--- a/openevolve/llm/claude_code.py
+++ b/openevolve/llm/claude_code.py
@@ -125,8 +125,9 @@ def _run_cli(self, cmd: list, timeout: int, prompt: Optional[str] = None) -> str
             )
             if result.returncode != 0:
                 stderr = result.stderr.strip()
-                if stderr:
-                    logger.warning(f"Claude CLI stderr: {stderr[:500]}")
+                raise RuntimeError(
+                    f"Claude CLI exited with code {result.returncode}. stderr: {stderr[:500]}"
+                )
             output = result.stdout.strip()
             if not output:
                 raise RuntimeError(f"Empty response from Claude CLI. stderr: {result.stderr[:500]}")
diff --git a/tests/test_claude_code_llm.py b/tests/test_claude_code_llm.py
index d803aa4a6..7c5829726 100644
--- a/tests/test_claude_code_llm.py
+++ b/tests/test_claude_code_llm.py
@@ -1,6 +1,7 @@
 """Tests for the Claude Code CLI LLM backend."""
 
 import asyncio
+import sys
 import unittest
 from unittest.mock import MagicMock, patch
 
@@ -81,6 +82,38 @@ def test_empty_response_raises(self, mock_run):
         with self.assertRaises(RuntimeError):
             asyncio.run(llm.generate("test prompt"))
 
+    def test_nonzero_exit_with_stdout_raises(self):
+        llm = ClaudeCodeLLM(self._make_cfg())
+        command = [
+            sys.executable,
+            "-c",
+            "import sys; print('partial output'); "
+            "print('provider failed', file=sys.stderr); sys.exit(7)",
+        ]
+        with self.assertRaisesRegex(RuntimeError, "exited with code 7.*provider failed"):
+            llm._run_cli(command, timeout=5)
+
+    @patch("openevolve.llm.claude_code.subprocess.run")
+    def test_retry_after_failure_with_stdout(self, mock_run):
+        mock_run.side_effect = [
+            MagicMock(returncode=7, stdout="partial output", stderr="transient error"),
+            MagicMock(returncode=0, stdout="complete response", stderr=""),
+        ]
+        llm = ClaudeCodeLLM(self._make_cfg(timeout=10, retries=1, retry_delay=0))
+        result = asyncio.run(llm.generate("test prompt"))
+        self.assertEqual(result, "complete response")
+        self.assertEqual(mock_run.call_count, 2)
+
+    @patch("openevolve.llm.claude_code.subprocess.run")
+    def test_nonzero_exit_with_stdout_exhausts_retries(self, mock_run):
+        mock_run.return_value = MagicMock(
+            returncode=7, stdout="partial output", stderr="persistent error"
+        )
+        llm = ClaudeCodeLLM(self._make_cfg(timeout=10, retries=2, retry_delay=0))
+        with self.assertRaisesRegex(RuntimeError, "exited with code 7"):
+            asyncio.run(llm.generate("test prompt"))
+        self.assertEqual(mock_run.call_count, 3)
+
     @patch("openevolve.llm.claude_code.subprocess.run")
     def test_retry_on_failure(self, mock_run):
         mock_run.side_effect = [