From 65f163c9d8b02c4e3024c91de02a59e796f90366 Mon Sep 17 00:00:00 2001
From: Tobias Reiher <reiher@adacore.com>
Date: Mon, 12 Feb 2024 09:11:40 +0100
Subject: [PATCH] Fix error reporting in CLI

Ref. eng/recordflux/RecordFlux#1530
---
 Makefile                                     |  2 +-
 rflx/cli.py                                  |  4 ++--
 tests/README.md                              |  2 +-
 tests/{end-to-end => end_to_end}/__init__.py |  0
 tests/{end-to-end => end_to_end}/cli_test.py | 19 +++++++++++++++++++
 5 files changed, 23 insertions(+), 4 deletions(-)
 rename tests/{end-to-end => end_to_end}/__init__.py (100%)
 rename tests/{end-to-end => end_to_end}/cli_test.py (91%)

diff --git a/Makefile b/Makefile
index bf5c428d8..5edd3b451 100644
--- a/Makefile
+++ b/Makefile
@@ -155,7 +155,7 @@ test_language_coverage:
 	timeout -k 60 7200 $(PYTEST) --cov=rflx/lang --cov-branch --cov-fail-under=73.8 --cov-report=term-missing:skip-covered tests/language
 
 test_end_to_end:
-	$(PYTEST) tests/end-to-end
+	$(PYTEST) tests/end_to_end
 
 test_property:
 	$(PYTEST) tests/property
diff --git a/rflx/cli.py b/rflx/cli.py
index 594c4a609..5d1aca9d6 100644
--- a/rflx/cli.py
+++ b/rflx/cli.py
@@ -129,8 +129,8 @@ def __call__(  # type: ignore [no-untyped-def]
             setattr(namespace, self.dest, values)
 
 
-def run() -> None:
-    main(sys.argv)
+def run() -> Union[int, str]:
+    return main(sys.argv)
 
 
 def main(  # noqa: PLR0915
diff --git a/tests/README.md b/tests/README.md
index 9ed4de6ed..853f6a2c9 100644
--- a/tests/README.md
+++ b/tests/README.md
@@ -29,7 +29,7 @@ The executability and provability tests require the definition of a session call
 
 Session functions can be defined by putting a custom implementation of the `Session` package inside the `src` directory.
 
-### End-to-End Tests (`tests/end-to-end`)
+### End-to-End Tests (`tests/end_to_end`)
 
 End-to-end tests verify the functionality of the entire application. These tests ensure that the application behaves as intended, particularly focusing on its primary use cases.
 
diff --git a/tests/end-to-end/__init__.py b/tests/end_to_end/__init__.py
similarity index 100%
rename from tests/end-to-end/__init__.py
rename to tests/end_to_end/__init__.py
diff --git a/tests/end-to-end/cli_test.py b/tests/end_to_end/cli_test.py
similarity index 91%
rename from tests/end-to-end/cli_test.py
rename to tests/end_to_end/cli_test.py
index ca575d8a9..1c1ac2d35 100644
--- a/tests/end-to-end/cli_test.py
+++ b/tests/end_to_end/cli_test.py
@@ -27,6 +27,25 @@ def test_check() -> None:
     )
 
 
+def test_check_error() -> None:
+    p = subprocess.run(
+        ["rflx", "--no-caching", "check", SPEC_DIR / "invalid" / "incorrect_name.rflx"],
+        capture_output=True,
+    )
+    assert p.returncode == 1
+    assert p.stdout.decode("utf-8") == ""
+    assert p.stderr.decode("utf-8") == textwrap.dedent(
+        """\
+        Parsing tests/data/specs/invalid/incorrect_name.rflx
+        Processing Test
+        Verifying __BUILTINS__::Boolean
+        Verifying __INTERNAL__::Opaque
+        tests/data/specs/invalid/incorrect_name.rflx:1:9: parser: error: file name does not match\
+ unit name "Test", should be "test.rflx"
+        """,
+    )
+
+
 def test_check_no_verification() -> None:
     p = subprocess.run(
         [