279 lines
9.9 KiB
Python
279 lines
9.9 KiB
Python
import threading
|
|
from pathlib import Path
|
|
from unittest.mock import Mock, patch
|
|
|
|
from test_config import settings
|
|
|
|
from lai_automation.git import ChangeSummary
|
|
from lai_automation.models import FailureReport
|
|
from lai_automation.worker import AutomationWorker
|
|
|
|
|
|
def test_dry_run_generates_task_without_agent_or_git() -> None:
|
|
report = FailureReport(
|
|
id=11,
|
|
chat_message_id=1,
|
|
source="automatic",
|
|
category="sql_execution",
|
|
user_feedback=None,
|
|
feedback_type=None,
|
|
expected_behavior=None,
|
|
prompt="List locations",
|
|
response=None,
|
|
sql_query=None,
|
|
error_message="Unknown column",
|
|
context={},
|
|
)
|
|
worker = AutomationWorker(settings())
|
|
worker.reports = Mock()
|
|
worker.reports.claim_next.return_value = report
|
|
|
|
assert worker.run_once()
|
|
worker.reports.release.assert_called_once()
|
|
worker.agent = Mock()
|
|
worker.agent.run.assert_not_called()
|
|
|
|
|
|
def test_stop_signal_interrupts_worker_wait() -> None:
|
|
worker = AutomationWorker(settings())
|
|
|
|
worker._stop(15, None)
|
|
|
|
assert worker.stop_event.is_set()
|
|
|
|
|
|
def test_database_credentials_are_only_passed_to_test_process() -> None:
|
|
worker = AutomationWorker(settings())
|
|
workspace = Mock()
|
|
workspace.run.return_value = Mock(returncode=0, stdout="ok", stderr="")
|
|
|
|
results = worker._run_tests(workspace)
|
|
|
|
assert results[0]["returncode"] == 0
|
|
environment = workspace.run.call_args.kwargs["extra_environment"]
|
|
assert environment["AGENT_TEST_DB_ENABLED"] == "true"
|
|
assert environment["APP_ENV"] == "testing"
|
|
assert environment["AI_MISTRAL_API_KEY"] == "automation-test-key"
|
|
assert environment["TEST_DB_USER"] == "test_runner"
|
|
assert environment["TEST_DB_PASSWORD"] == "test-secret"
|
|
assert environment["DB_DATABASE"] == "lai_agent_test"
|
|
assert environment["ERP_DATABASE"] == "customer_agent_test"
|
|
assert environment["COCKPIT_CONFIG_DATABASE"] == "cockpit_agent_test"
|
|
|
|
|
|
def test_validation_commands_are_logged_without_output(caplog) -> None:
|
|
worker = AutomationWorker(settings())
|
|
workspace = Mock()
|
|
workspace.run.return_value = Mock(returncode=0, stdout="sensitive output", stderr="")
|
|
|
|
with caplog.at_level("INFO"):
|
|
worker._run_tests(workspace, report_id=7, attempt=2)
|
|
|
|
assert "Report 7 attempt 2 validation command: php artisan test starting" in caplog.text
|
|
assert "returncode=0" in caplog.text
|
|
assert "sensitive output" not in caplog.text
|
|
|
|
|
|
def test_failed_validation_preserves_start_and_end_of_long_output() -> None:
|
|
worker = AutomationWorker(settings())
|
|
workspace = Mock()
|
|
workspace.run.return_value = Mock(
|
|
returncode=1,
|
|
stdout="DECISIVE ASSERTION\n" + "x" * 13_000,
|
|
stderr="FINAL ERROR",
|
|
)
|
|
|
|
try:
|
|
worker._run_tests(workspace)
|
|
except RuntimeError as exception:
|
|
failure = str(exception)
|
|
else:
|
|
raise AssertionError("Expected failed validation")
|
|
|
|
assert "DECISIVE ASSERTION" in failure
|
|
assert "...[middle omitted]..." in failure
|
|
assert "FINAL ERROR" in failure
|
|
|
|
|
|
def test_long_stage_emits_heartbeat(caplog) -> None:
|
|
worker = AutomationWorker(settings())
|
|
|
|
with caplog.at_level("INFO"), worker._progress(7, "agent attempt 1", interval=0.001):
|
|
threading.Event().wait(0.01)
|
|
|
|
assert "Report 7 agent attempt 1 still running" in caplog.text
|
|
assert "Report 7 agent attempt 1 finished" in caplog.text
|
|
|
|
|
|
def test_php_syntax_failure_discards_attempt_before_full_validation() -> None:
|
|
report = FailureReport(
|
|
id=56,
|
|
chat_message_id=1,
|
|
source="user_feedback",
|
|
category="incorrect_answer",
|
|
user_feedback="Wrong hours",
|
|
feedback_type="wrong_result",
|
|
expected_behavior="Use time registration",
|
|
prompt="Hours in April",
|
|
response="0 hours",
|
|
sql_query="SELECT 0",
|
|
error_message=None,
|
|
context={},
|
|
)
|
|
config = settings(mode="pr", dry_run=False, gitea_token="token")
|
|
worker = AutomationWorker(config)
|
|
worker.reports = Mock()
|
|
worker.reports.claim_next.return_value = report
|
|
worker.agent = Mock()
|
|
worker.agent.run.side_effect = ["broken PHP", "fixed PHP"]
|
|
worker.gitea = Mock()
|
|
worker._run_tests = Mock(return_value=[])
|
|
workspace = Mock()
|
|
workspace.path = Path("/tmp/nonexistent-workspace")
|
|
workspace.branch = "ai/failure-56"
|
|
workspace.summary.return_value = ChangeSummary(
|
|
("config/erp_semantics.php", "tests/Unit/FixTest.php"), 20, 2
|
|
)
|
|
workspace.run_args.side_effect = [
|
|
Mock(returncode=255, stdout="Parse error", stderr="unexpected token"),
|
|
Mock(returncode=0, stdout="No syntax errors", stderr=""),
|
|
Mock(returncode=0, stdout="No syntax errors", stderr=""),
|
|
]
|
|
workspace.commit.return_value = "commit-id"
|
|
|
|
with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
|
|
assert worker.run_once()
|
|
|
|
assert worker.agent.run.call_count == 2
|
|
assert "unexpected token" in worker.agent.run.call_args_list[1].args[1]
|
|
assert "clean repository" in worker.agent.run.call_args_list[1].args[1]
|
|
workspace.discard_changes.assert_called_once()
|
|
worker._run_tests.assert_called_once()
|
|
|
|
|
|
def test_failed_validation_gets_one_correction_before_pr() -> None:
|
|
report = FailureReport(
|
|
id=13,
|
|
chat_message_id=1,
|
|
source="automatic",
|
|
category="sql_execution",
|
|
user_feedback=None,
|
|
feedback_type=None,
|
|
expected_behavior=None,
|
|
prompt="List products",
|
|
response=None,
|
|
sql_query="SELECT missing_column FROM invoice_row",
|
|
error_message="Unknown column",
|
|
context={},
|
|
)
|
|
config = settings(mode="pr", dry_run=False, gitea_token="token")
|
|
worker = AutomationWorker(config)
|
|
worker.reports = Mock()
|
|
worker.reports.claim_next.return_value = report
|
|
worker.agent = Mock()
|
|
worker.agent.run.side_effect = ["first change", "corrected change"]
|
|
worker.gitea = Mock()
|
|
worker._run_tests = Mock(side_effect=[RuntimeError("Missing import"), []])
|
|
workspace = Mock()
|
|
workspace.path = Path("/tmp/nonexistent-workspace")
|
|
workspace.branch = "ai/failure-13"
|
|
workspace.summary.return_value = ChangeSummary(
|
|
("app/Services/Fix.php", "tests/Feature/FixTest.php"), 10, 1
|
|
)
|
|
workspace.run_args.return_value = Mock(returncode=0, stdout="No syntax errors", stderr="")
|
|
workspace.commit.return_value = "commit-id"
|
|
|
|
with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
|
|
assert worker.run_once()
|
|
|
|
assert worker.agent.run.call_count == 2
|
|
assert "Missing import" in worker.agent.run.call_args_list[1].args[1]
|
|
assert worker._run_tests.call_count == 2
|
|
workspace.push_branch.assert_called_once()
|
|
worker.reports.complete.assert_called_once()
|
|
|
|
|
|
def test_no_change_response_gets_one_correction() -> None:
|
|
report = FailureReport(
|
|
id=14,
|
|
chat_message_id=1,
|
|
source="automatic",
|
|
category="sql_execution",
|
|
user_feedback=None,
|
|
feedback_type=None,
|
|
expected_behavior=None,
|
|
prompt="List products",
|
|
response=None,
|
|
sql_query=None,
|
|
error_message="Unknown column",
|
|
context={},
|
|
)
|
|
config = settings(mode="pr", dry_run=False, gitea_token="token")
|
|
worker = AutomationWorker(config)
|
|
worker.reports = Mock()
|
|
worker.reports.claim_next.return_value = report
|
|
worker.agent = Mock()
|
|
worker.agent.run.side_effect = ["no edit", "applied edit"]
|
|
worker.gitea = Mock()
|
|
worker._run_tests = Mock(return_value=[])
|
|
workspace = Mock()
|
|
workspace.path = Path("/tmp/nonexistent-workspace")
|
|
workspace.branch = "ai/failure-14"
|
|
workspace.summary.side_effect = [
|
|
ChangeSummary((), 0, 0),
|
|
ChangeSummary(("app/Services/Fix.php", "tests/Feature/FixTest.php"), 10, 1),
|
|
ChangeSummary(("app/Services/Fix.php", "tests/Feature/FixTest.php"), 10, 1),
|
|
]
|
|
workspace.run_args.return_value = Mock(returncode=0, stdout="No syntax errors", stderr="")
|
|
workspace.commit.return_value = "commit-id"
|
|
|
|
with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
|
|
assert worker.run_once()
|
|
|
|
assert worker.agent.run.call_count == 2
|
|
assert "Agent produced no file changes" in worker.agent.run.call_args_list[1].args[1]
|
|
worker._run_tests.assert_called_once()
|
|
workspace.push_branch.assert_called_once()
|
|
|
|
|
|
def test_test_only_change_gets_correction_before_validation() -> None:
|
|
report = FailureReport(
|
|
id=49,
|
|
chat_message_id=1,
|
|
source="user_feedback",
|
|
category="incorrect_answer",
|
|
user_feedback="Token was not restored",
|
|
feedback_type="wrong_result",
|
|
expected_behavior="Return the real value",
|
|
prompt="Average price",
|
|
response="__LAI_UNKNOWN__",
|
|
sql_query="SELECT AVG(price)",
|
|
error_message=None,
|
|
context={},
|
|
)
|
|
config = settings(mode="pr", dry_run=False, gitea_token="token")
|
|
worker = AutomationWorker(config)
|
|
worker.reports = Mock()
|
|
worker.reports.claim_next.return_value = report
|
|
worker.agent = Mock()
|
|
worker.agent.run.side_effect = ["tests only", "production fix"]
|
|
worker.gitea = Mock()
|
|
worker._run_tests = Mock(return_value=[])
|
|
workspace = Mock()
|
|
workspace.path = Path("/tmp/nonexistent-workspace")
|
|
workspace.branch = "ai/failure-49"
|
|
workspace.summary.side_effect = [
|
|
ChangeSummary(("tests/Unit/FixTest.php",), 8, 0),
|
|
ChangeSummary(("app/Services/Fix.php", "tests/Unit/FixTest.php"), 15, 1),
|
|
ChangeSummary(("app/Services/Fix.php", "tests/Unit/FixTest.php"), 15, 1),
|
|
]
|
|
workspace.run_args.return_value = Mock(returncode=0, stdout="No syntax errors", stderr="")
|
|
workspace.commit.return_value = "commit-id"
|
|
|
|
with patch("lai_automation.worker.GitWorkspace", return_value=workspace):
|
|
assert worker.run_once()
|
|
|
|
assert worker.agent.run.call_count == 2
|
|
assert "changed only tests" in worker.agent.run.call_args_list[1].args[1]
|
|
worker._run_tests.assert_called_once()
|