fix: resolve codex shim for harness runner

scripts/execute.py

@@ -12,6 +12,7 @@ import fnmatch
 import json
 import os
 import re
+import shutil
 import subprocess
 import sys
 import threading
@@ -373,7 +374,7 @@ class StepExecutor:
 
         prompt = preamble + step_file.read_text(encoding="utf-8")
         result = subprocess.run(
-            ["codex", "exec", "--dangerously-bypass-approvals-and-sandbox", "--json", "-"],
+            [self._codex_command(), "exec", "--dangerously-bypass-approvals-and-sandbox", "--json", "-"],
             cwd=self._root, capture_output=True, text=True, input=prompt, timeout=1800,
         )
 
@@ -395,6 +396,15 @@ class StepExecutor:
 
     # --- 헤더 & 검증 ---
 
+    @staticmethod
+    def _codex_command() -> str:
+        return (
+            shutil.which("codex.cmd")
+            or shutil.which("codex.exe")
+            or shutil.which("codex")
+            or "codex"
+        )
+
     def _print_header(self):
         print(f"\n{'='*60}")
         print(f"  Harness Step Executor")

scripts/test_execute.py

@@ -294,17 +294,31 @@ class ExecuteRunnerSafetyTests(unittest.TestCase):
                 return subprocess.CompletedProcess(cmd, 0, '{"event":"done"}\n', "")
 
             with patch.object(execute.subprocess, "run", side_effect=fake_run) as run_mock:
+                with patch.object(executor, "_codex_command", return_value="codex.cmd"):
                     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", "-"],
+                ["codex.cmd", "exec", "--dangerously-bypass-approvals-and-sandbox", "--json", "-"],
             )
             self.assertEqual(kwargs["input"], long_preamble + "# Step 1\n")
             self.assertEqual(kwargs["cwd"], str(root))
 
+    def test_codex_command_prefers_windows_cmd_or_exe_shim(self):
+        execute = load_execute()
+
+        def fake_which(name):
+            return {
+                "codex.cmd": "C:/tools/codex.cmd",
+                "codex.exe": "C:/tools/codex.exe",
+                "codex": "C:/tools/codex",
+            }.get(name)
+
+        with patch.object(execute.shutil, "which", side_effect=fake_which):
+            self.assertEqual(execute.StepExecutor._codex_command(), "C:/tools/codex.cmd")
+
 
 if __name__ == "__main__":
     unittest.main()