fix: normalize Windows subprocess exit statuses (#35)

Map unsigned and NTSTATUS-style Windows subprocess failures to exit code 1 before sys.exit(), and preserve POSIX signal semantics on Unix.

Co-authored-by: Codex <noreply@openai.com>

Author: jbeckwith-oai

src/promptfoo/cli.py

@@ -15,6 +15,7 @@
 
 _WRAPPER_ENV = "PROMPTFOO_PY_WRAPPER"
 _WINDOWS_SHELL_EXTENSIONS = (".bat", ".cmd")
+_WINDOWS_STATUS_SIGN_BIT = 1 << 31
 _VERSION_ENV = "PROMPTFOO_VERSION"
 
 
@@ -173,6 +174,18 @@ def _run_command(cmd: list[str], env: Optional[dict[str, str]] = None) -> subpro
     return subprocess.run(cmd, env=env)
 
 
+def _normalize_exit_code(returncode: int) -> int:
+    """Normalize subprocess return codes into portable process exit codes."""
+    if os.name == "nt":
+        if returncode < 0 or returncode & _WINDOWS_STATUS_SIGN_BIT:
+            return 1
+        return returncode
+
+    if returncode < 0:
+        return 128 + min(abs(returncode), 127)
+    return returncode
+
+
 def main() -> NoReturn:
     """
     Main entry point for the promptfoo CLI wrapper.
@@ -207,7 +220,7 @@ def main() -> NoReturn:
                 print("Or ensure Node.js is properly installed.", file=sys.stderr)
                 sys.exit(1)
 
-        sys.exit(result.returncode)
+        sys.exit(_normalize_exit_code(result.returncode))
     except KeyboardInterrupt:
         sys.exit(130)
 

tests/test_cli.py

@@ -22,6 +22,7 @@
     _WRAPPER_ENV,
     _find_external_promptfoo,
     _find_windows_promptfoo,
+    _normalize_exit_code,
     _normalize_path,
     _requires_shell,
     _resolve_argv0,
@@ -358,6 +359,34 @@ def test_run_command_passes_environment(self, monkeypatch: pytest.MonkeyPatch) -
         assert call_args.kwargs.get("env") == env
 
 
+class TestExitCodeNormalization:
+    """Test subprocess exit code normalization."""
+
+    @pytest.mark.parametrize("returncode", [0, 1, 100, 255])
+    def test_normalize_exit_code_preserves_standard_codes(
+        self, returncode: int, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        """Standard shell exit codes pass through unchanged."""
+        monkeypatch.setattr(os, "name", "nt")
+
+        assert _normalize_exit_code(returncode) == returncode
+
+    @pytest.mark.parametrize("returncode", [4294967295, 3221226505, -1])
+    def test_normalize_exit_code_maps_windows_error_statuses(
+        self, returncode: int, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        """Windows unsigned and NTSTATUS failure values map to exit code 1."""
+        monkeypatch.setattr(os, "name", "nt")
+
+        assert _normalize_exit_code(returncode) == 1
+
+    def test_normalize_exit_code_maps_unix_signal_status(self, monkeypatch: pytest.MonkeyPatch) -> None:
+        """Unix signal-style negative return codes map to 128 + signal number."""
+        monkeypatch.setattr(os, "name", "posix")
+
+        assert _normalize_exit_code(-15) == 143
+
+
 # =============================================================================
 # Integration Tests for main()
 # =============================================================================
@@ -533,6 +562,30 @@ def test_main_returns_subprocess_exit_code(self, monkeypatch: pytest.MonkeyPatch
 
         assert exc_info.value.code == 42
 
+    @pytest.mark.parametrize("raw_returncode", [4294967295, 3221226505])
+    def test_main_normalizes_windows_error_statuses(self, raw_returncode: int, monkeypatch: pytest.MonkeyPatch) -> None:
+        """Converts Windows-specific subprocess statuses into a stable exit code."""
+        monkeypatch.setattr(os, "name", "nt")
+        monkeypatch.setattr(sys, "argv", ["promptfoo", "eval", "-c", "missing.yaml"])
+        monkeypatch.setattr(
+            "shutil.which",
+            lambda cmd, path=None: {
+                "node": "C:\\Program Files\\nodejs\\node.exe",
+                "npx": "C:\\Program Files\\nodejs\\npx.cmd",
+            }.get(cmd),
+        )
+        monkeypatch.setattr("promptfoo.cli.record_wrapper_used", lambda mode: None)
+        monkeypatch.setattr(
+            subprocess,
+            "run",
+            MagicMock(return_value=subprocess.CompletedProcess([], raw_returncode)),
+        )
+
+        with pytest.raises(SystemExit) as exc_info:
+            main()
+
+        assert exc_info.value.code == 1
+
 
 # =============================================================================
 # Platform-Specific Tests