fix: pass harness prompts through stdin

scripts/execute.py

@@ -373,8 +373,8 @@ class StepExecutor:
 
         prompt = preamble + step_file.read_text(encoding="utf-8")
         result = subprocess.run(
-            ["codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "--json", prompt],
+            ["codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "--json", "-"],
-            cwd=self._root, capture_output=True, text=True, timeout=1800,
+            cwd=self._root, capture_output=True, text=True, input=prompt, timeout=1800,
         )
 
         if result.returncode != 0:

scripts/test_execute.py

@@ -281,6 +281,30 @@ class ExecuteRunnerSafetyTests(unittest.TestCase):
 
             self.assertEqual(cm.exception.code, 1)
 
+    def test_invoke_codex_passes_prompt_through_stdin(self):
+        execute = load_execute()
+        with tempfile.TemporaryDirectory() as tmp:
+            root = Path(tmp)
+            write_phase(root)
+            executor = make_executor(execute, root)
+            step = {"step": 1, "name": "Docs"}
+            long_preamble = "x" * 40000
+
+            def fake_run(cmd, **kwargs):
+                return subprocess.CompletedProcess(cmd, 0, '{"event":"done"}\n', "")
+
+            with patch.object(execute.subprocess, "run", side_effect=fake_run) as run_mock:
+                executor._invoke_codex(step, long_preamble)
+
+            cmd = run_mock.call_args.args[0]
+            kwargs = run_mock.call_args.kwargs
+            self.assertEqual(
+                cmd,
+                ["codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "--json", "-"],
+            )
+            self.assertEqual(kwargs["input"], long_preamble + "# Step 1\n")
+            self.assertEqual(kwargs["cwd"], str(root))
+
 
 if __name__ == "__main__":
     unittest.main()