feat(test): recursive test discovery + subfolder selection

cli/nao_core/commands/test/case.py

@@ -39,7 +39,11 @@ def discover_tests(project_path: Path) -> list[TestCase]:
         UI.warn(f"Tests folder not found: {tests_dir}")
         return []
 
-    test_files = list(tests_dir.glob("*.yml")) + list(tests_dir.glob("*.yaml"))
+    test_files = [
+        p
+        for p in (*tests_dir.rglob("*.yml"), *tests_dir.rglob("*.yaml"))
+        if "outputs" not in p.relative_to(tests_dir).parts
+    ]
 
     if not test_files:
         UI.warn(f"No test files found in {tests_dir}")

cli/nao_core/commands/test/runner.py

@@ -289,8 +289,17 @@ def save_results(results: list[TestRunResult], output_dir: Path) -> Path:
     return output_file
 
 
-def filter_test_cases(test_cases: list[TestCase], selected_tests: str | None) -> list[TestCase]:
-    """Filter test cases to the selected tests, if provided."""
+def filter_test_cases(
+    test_cases: list[TestCase],
+    selected_tests: str | None,
+    tests_dir: Path | None = None,
+) -> list[TestCase]:
+    """Filter test cases to the selected tests, if provided.
+
+    Each comma-separated selection matches either a single test (by ``name`` or
+    file stem) or, when ``tests_dir`` is given, every test under a subfolder of
+    tests/ (e.g. ``contracts`` selects all tests in tests/contracts/).
+    """
     if not selected_tests:
         return test_cases
 
@@ -302,18 +311,28 @@ def filter_test_cases(test_cases: list[TestCase], selected_tests: str | None) ->
     selected: list[TestCase] = []
     seen: set[Path] = set()
     for selection in selections:
-        matches = [tc for tc in test_cases if tc.name == selection or tc.file_path.stem == selection]
+        folder_matches: list[TestCase] = []
+        if tests_dir is not None:
+            for tc in test_cases:
+                try:
+                    parts = tc.file_path.relative_to(tests_dir).parent.parts
+                except ValueError:
+                    continue
+                if selection in parts:
+                    folder_matches.append(tc)
+        name_matches = [tc for tc in test_cases if tc.name == selection or tc.file_path.stem == selection]
+        matches = folder_matches or name_matches
         if not matches:
             available = ", ".join(tc.name for tc in test_cases)
             raise ValueError(f"Test not found: {selection}. Available tests: {available}")
-        if len(matches) > 1:
+        if not folder_matches and len(matches) > 1:
             names = ", ".join(f"{tc.name} ({tc.file_path.name})" for tc in matches)
             raise ValueError(f"Multiple tests match '{selection}': {names}")
-        match = matches[0]
-        if match.file_path in seen:
-            continue
-        seen.add(match.file_path)
-        selected.append(match)
+        for match in matches:
+            if match.file_path in seen:
+                continue
+            seen.add(match.file_path)
+            selected.append(match)
 
     return selected
 
@@ -337,7 +356,7 @@ def test(
         str | None,
         Parameter(
             name=["-s", "--select"],
-            help="Run only the selected tests by name or yaml filename stem. Comma-separated (e.g. '12,13,14').",
+            help="Run only selected tests by name, yaml stem, or subfolder. Comma-separated (e.g. 'contracts' or '12,13,14').",
         ),
     ] = None,
     username: Annotated[
@@ -384,8 +403,9 @@ def test(
 
     project_path = Path.cwd()
     model_costs = config.llm.meta.costs if config.llm and config.llm.meta else None
+    tests_dir = project_path / TESTS_FOLDER
     UI.print(f"[dim]Project: {config.project_name}[/dim]")
-    UI.print(f"[dim]Tests folder: {project_path / TESTS_FOLDER}[/dim]")
+    UI.print(f"[dim]Tests folder: {tests_dir}[/dim]")
     UI.print(f"[dim]Models: {', '.join(str(m) for m in model_configs)}[/dim]\n")
 
     test_cases = discover_tests(project_path)
@@ -395,7 +415,7 @@ def test(
         return
 
     try:
-        test_cases = filter_test_cases(test_cases, select)
+        test_cases = filter_test_cases(test_cases, select, tests_dir)
     except ValueError as e:
         UI.error(str(e))
         return

cli/tests/nao_core/commands/test_case.py

@@ -0,0 +1,24 @@
+from nao_core.commands.test.case import discover_tests
+
+
+def test_discover_tests_is_recursive(tmp_path):
+    (tmp_path / "tests" / "revenue").mkdir(parents=True)
+    (tmp_path / "tests" / "revenue" / "mrr.yml").write_text("prompt: how much mrr\n")
+    (tmp_path / "tests" / "ops" / "sla").mkdir(parents=True)
+    (tmp_path / "tests" / "ops" / "sla" / "uptime.yaml").write_text("prompt: uptime\n")
+
+    cases = discover_tests(tmp_path)
+
+    names = {c.name for c in cases}
+    assert names == {"mrr", "uptime"}
+
+
+def test_discover_tests_ignores_outputs_dir(tmp_path):
+    (tmp_path / "tests").mkdir()
+    (tmp_path / "tests" / "real.yml").write_text("prompt: real\n")
+    (tmp_path / "tests" / "outputs").mkdir()
+    (tmp_path / "tests" / "outputs" / "results.yml").write_text("prompt: not a test\n")
+
+    cases = discover_tests(tmp_path)
+
+    assert {c.name for c in cases} == {"real"}

cli/tests/nao_core/commands/test_runner.py

@@ -227,3 +227,37 @@ def test_run_test_records_reference_sql_on_client_error(monkeypatch):
     assert result.error == "backend unreachable"
     assert result.details is not None
     assert result.details.reference_sql == "select 1"
+
+
+def test_filter_test_cases_by_folder(tmp_path):
+    tests_dir = tmp_path / "tests"
+    tc_orders = NaoTestCase(name="orders", prompt="p1", file_path=tests_dir / "revenue" / "orders.yml", sql="select 1")
+    tc_mrr = NaoTestCase(name="mrr", prompt="p2", file_path=tests_dir / "revenue" / "mrr.yml", sql="select 1")
+    tc_users = NaoTestCase(name="users", prompt="p3", file_path=tests_dir / "ops" / "users.yml", sql="select 1")
+
+    filtered = filter_test_cases([tc_orders, tc_mrr, tc_users], "revenue", tests_dir)
+
+    assert {tc.name for tc in filtered} == {"orders", "mrr"}
+
+
+def test_filter_test_cases_folder_and_name_combined(tmp_path):
+    tests_dir = tmp_path / "tests"
+    tc_orders = NaoTestCase(name="orders", prompt="p1", file_path=tests_dir / "revenue" / "orders.yml", sql="select 1")
+    tc_users = NaoTestCase(name="users", prompt="p2", file_path=tests_dir / "ops" / "users.yml", sql="select 1")
+
+    filtered = filter_test_cases([tc_orders, tc_users], "revenue,users", tests_dir)
+
+    assert {tc.name for tc in filtered} == {"orders", "users"}
+
+
+def test_filter_test_cases_by_name_without_tests_dir_unchanged():
+    # Backward-compat: two-arg call still filters by name/stem.
+    test_cases = [
+        NaoTestCase(name="orders", prompt="p1", file_path=Path("tests/orders.yml"), sql="select 1"),
+        NaoTestCase(name="users", prompt="p2", file_path=Path("tests/users.yml"), sql="select 1"),
+    ]
+
+    filtered = filter_test_cases(test_cases, "users")
+
+    assert len(filtered) == 1
+    assert filtered[0].name == "users"