Pass the prompt to the Claude Code CLI via stdin

openevolve/llm/claude_code.py

@@ -79,7 +79,8 @@ async def generate_with_context(
         budget = kwargs.get("max_budget_usd", self.max_budget_usd)
         cmd.extend(["--max-budget-usd", str(budget)])
 
-        cmd.append(user_content)
+        # The prompt goes through stdin: as one argv string it hits Linux's 128 KiB
+        # per-argument limit (MAX_ARG_STRLEN) once a few programs are in the prompt.
 
         timeout = kwargs.get("timeout", self.timeout)
         retries = kwargs.get("retries", self.retries)
@@ -89,7 +90,7 @@ async def generate_with_context(
         for attempt in range(retries + 1):
             try:
                 result = await asyncio.wait_for(
-                    loop.run_in_executor(None, lambda: self._run_cli(cmd, timeout)),
+                    loop.run_in_executor(None, lambda: self._run_cli(cmd, timeout, user_content)),
                     timeout=timeout + 30,
                 )
                 return result
@@ -112,10 +113,11 @@ async def generate_with_context(
                     logger.error(f"All {retries + 1} attempts failed with error: {e}")
                     raise
 
-    def _run_cli(self, cmd: list, timeout: int) -> str:
+    def _run_cli(self, cmd: list, timeout: int, prompt: Optional[str] = None) -> str:
         try:
             result = subprocess.run(
                 cmd,
+                input=prompt,
                 capture_output=True,
                 text=True,
                 timeout=timeout,

tests/test_claude_code_llm.py

@@ -50,7 +50,20 @@ def test_generate_calls_cli(self, mock_run):
         self.assertIn("-p", cmd)
         self.assertIn("--model", cmd)
         self.assertIn("sonnet", cmd)
-        self.assertIn("test prompt", cmd)
+        self.assertNotIn("test prompt", cmd)
+        self.assertEqual(mock_run.call_args.kwargs["input"], "test prompt")
+
+    @patch("openevolve.llm.claude_code.subprocess.run")
+    def test_large_prompt_not_passed_as_argument(self, mock_run):
+        # Linux caps a single argv string at 128 KiB (MAX_ARG_STRLEN); a prompt that
+        # carries a few parent programs easily exceeds it, so it must go through stdin.
+        mock_run.return_value = MagicMock(returncode=0, stdout="response", stderr="")
+        llm = ClaudeCodeLLM()
+        prompt = "x" * 200_000
+        asyncio.run(llm.generate(prompt))
+        cmd = mock_run.call_args[0][0]
+        self.assertTrue(all(len(arg) < 128 * 1024 for arg in cmd))
+        self.assertEqual(mock_run.call_args.kwargs["input"], prompt)
 
     @patch("openevolve.llm.claude_code.subprocess.run")
     def test_system_message_passed(self, mock_run):
@@ -102,8 +115,7 @@ def test_generate_with_context(self, mock_run):
             )
         )
         self.assertEqual(result, "ctx response")
-        cmd = mock_run.call_args[0][0]
-        self.assertIn("first\n\nsecond", cmd[-1])
+        self.assertIn("first\n\nsecond", mock_run.call_args.kwargs["input"])
 
 
 class TestMaxBudgetConfig(unittest.TestCase):