kin/tests/test_kin_arch_014_regression.py

289 lines
12 KiB
Python
Raw Normal View History

2026-03-17 16:02:47 +02:00
"""
Regression tests for KIN-ARCH-014:
blocked_reason из dept_result должен корректно пробрасываться в БД при неуспешном sub-pipeline.
Два уровня цепочки (convention #305 — отдельный тест на каждый уровень):
Уровень 1: _execute_department_head_step должен включать blocked_reason из sub_result
в возвращаемый dict. Без фикса dict возвращался без этого поля.
Уровень 2: run_pipeline должен передавать dept_result.blocked_reason в update_task
(а не generic 'Department X sub-pipeline failed').
"""
import json
import pytest
from unittest.mock import patch, MagicMock
# ---------------------------------------------------------------------------
# Fixtures
# ---------------------------------------------------------------------------
@pytest.fixture
def conn():
from core.db import init_db
from core import models
c = init_db(":memory:")
models.create_project(c, "proj1", "Project1", "/tmp/proj1", tech_stack=["python"])
models.create_task(c, "PROJ1-001", "proj1", "Parent task",
brief={"route_type": "department"})
yield c
c.close()
def _dept_head_result_with_sub_pipeline():
"""Валидный вывод dept head с sub_pipeline — используется в обоих уровнях."""
return {
"raw_output": json.dumps({
"sub_pipeline": [
{"role": "backend_dev", "brief": "implement feature"}
],
"artifacts": {},
"handoff_notes": "proceed",
})
}
def _failing_sub_result(blocked_reason=None, error=None):
return {
"success": False,
"blocked_reason": blocked_reason,
"error": error,
"total_cost_usd": 0,
"total_tokens": 0,
"total_duration_seconds": 0,
"steps_completed": 0,
"results": [],
"pipeline_id": None,
}
def _mock_agent_success(output="done"):
"""Мок успешного subprocess.run для run_agent."""
m = MagicMock()
m.stdout = json.dumps({"result": output})
m.stderr = ""
m.returncode = 0
return m
# ---------------------------------------------------------------------------
# Уровень 1: _execute_department_head_step propagates blocked_reason
# Convention #304: имя теста описывает сломанное поведение
# ---------------------------------------------------------------------------
class TestDeptHeadStepBlockedReasonPropagation:
"""Уровень 1 цепочки: _execute_department_head_step → sub_result → returned dict."""
def test_blocked_reason_dropped_from_returned_dict_when_sub_pipeline_fails(self, conn):
"""Broken behavior: _execute_department_head_step отбрасывал blocked_reason
из sub_result возвращаемый dict не содержал этого поля.
Fixed (KIN-ARCH-014): blocked_reason из sub_result прокидывается в dict.
"""
from agents.runner import _execute_department_head_step
failing = _failing_sub_result(
blocked_reason="Конкретная причина: tester заблокировал задачу",
error="TestError: endpoint not found",
)
with patch("agents.runner.run_pipeline", return_value=failing), \
patch("agents.runner.models.create_pipeline", return_value={"id": 99}), \
patch("agents.runner.models.create_handoff"):
result = _execute_department_head_step(
conn=conn,
task_id="PROJ1-001",
project_id="proj1",
parent_pipeline_id=None,
step={"role": "backend_head", "brief": "plan"},
dept_head_result=_dept_head_result_with_sub_pipeline(),
)
assert result["success"] is False
assert "blocked_reason" in result, (
"blocked_reason должен быть в возвращаемом dict "
"— без фикса KIN-ARCH-014 это поле отсутствовало"
)
assert result["blocked_reason"] == "Конкретная причина: tester заблокировал задачу"
def test_error_field_used_as_fallback_when_blocked_reason_is_none(self, conn):
"""Если sub_result.blocked_reason is None, но есть error —
error должен использоваться как blocked_reason (fallback).
"""
from agents.runner import _execute_department_head_step
failing = _failing_sub_result(
blocked_reason=None,
error="RuntimeError: Claude CLI exited with code 1",
)
with patch("agents.runner.run_pipeline", return_value=failing), \
patch("agents.runner.models.create_pipeline", return_value={"id": 99}), \
patch("agents.runner.models.create_handoff"):
result = _execute_department_head_step(
conn=conn,
task_id="PROJ1-001",
project_id="proj1",
parent_pipeline_id=None,
step={"role": "backend_head", "brief": "plan"},
dept_head_result=_dept_head_result_with_sub_pipeline(),
)
assert result["success"] is False
assert result.get("blocked_reason") == "RuntimeError: Claude CLI exited with code 1"
def test_success_result_does_not_inject_blocked_reason(self, conn):
"""При успешном sub_result blocked_reason не добавляется в dict."""
from agents.runner import _execute_department_head_step
success_sub_result = {
"success": True,
"total_cost_usd": 0.01,
"total_tokens": 100,
"total_duration_seconds": 5.0,
"steps_completed": 1,
"results": [{"role": "backend_dev", "output": "done"}],
"pipeline_id": 99,
}
with patch("agents.runner.run_pipeline", return_value=success_sub_result), \
patch("agents.runner.models.create_pipeline", return_value={"id": 99}), \
patch("agents.runner.models.create_handoff"):
result = _execute_department_head_step(
conn=conn,
task_id="PROJ1-001",
project_id="proj1",
parent_pipeline_id=None,
step={"role": "backend_head", "brief": "plan"},
dept_head_result=_dept_head_result_with_sub_pipeline(),
)
assert result["success"] is True
assert not result.get("blocked_reason"), (
"blocked_reason не должен появляться при успешном sub_result"
)
# ---------------------------------------------------------------------------
# Уровень 2: run_pipeline passes dept_result.blocked_reason to update_task
# Convention #305: отдельный класс для каждого уровня
# ---------------------------------------------------------------------------
class TestRunPipelineDeptBlockedReasonSavedToDb:
"""Уровень 2 цепочки: run_pipeline → dept_result → models.update_task → БД."""
@patch("agents.runner._run_autocommit")
@patch("agents.runner._execute_department_head_step")
@patch("agents.runner._is_department_head")
@patch("agents.runner.subprocess.run")
@patch("agents.runner.check_claude_auth")
def test_generic_error_msg_saved_instead_of_dept_blocked_reason(
self, mock_auth, mock_run, mock_is_dept, mock_dept_step, mock_autocommit, conn
):
"""Broken behavior: run_pipeline сохранял в БД generic строку
'Department X sub-pipeline failed' вместо реального blocked_reason из dept_result.
Fixed (KIN-ARCH-014): в БД сохраняется dept_result["blocked_reason"].
"""
from agents.runner import run_pipeline
from core import models
mock_run.return_value = _mock_agent_success()
mock_is_dept.return_value = True
mock_dept_step.return_value = {
"success": False,
"blocked_reason": "Специфическая причина: backend_dev упал с OOM",
"error": "OOMError",
"output": "",
"cost_usd": 0,
"tokens_used": 0,
"duration_seconds": 0,
}
steps = [{"role": "backend_head", "brief": "plan the work"}]
result = run_pipeline(conn, "PROJ1-001", steps)
assert result["success"] is False
task = models.get_task(conn, "PROJ1-001")
assert task["status"] == "blocked"
assert task["blocked_reason"] == "Специфическая причина: backend_dev упал с OOM", (
f"Ожидался реальный blocked_reason из dept_result, "
f"получено: {task['blocked_reason']!r}"
)
@patch("agents.runner._run_autocommit")
@patch("agents.runner._execute_department_head_step")
@patch("agents.runner._is_department_head")
@patch("agents.runner.subprocess.run")
@patch("agents.runner.check_claude_auth")
def test_generic_string_not_saved_when_specific_reason_available(
self, mock_auth, mock_run, mock_is_dept, mock_dept_step, mock_autocommit, conn
):
"""run_pipeline НЕ должен сохранять generic 'Department X sub-pipeline failed'
в blocked_reason когда dept_result содержит конкретный blocked_reason.
"""
from agents.runner import run_pipeline
from core import models
specific_reason = "Tester blocked: API endpoint /api/users не реализован"
mock_run.return_value = _mock_agent_success()
mock_is_dept.return_value = True
mock_dept_step.return_value = {
"success": False,
"blocked_reason": specific_reason,
"error": specific_reason,
"output": "",
"cost_usd": 0,
"tokens_used": 0,
"duration_seconds": 0,
}
steps = [{"role": "backend_head", "brief": "plan"}]
run_pipeline(conn, "PROJ1-001", steps)
task = models.get_task(conn, "PROJ1-001")
# Generic строка из old code не должна быть в БД
assert "sub-pipeline failed" not in (task["blocked_reason"] or ""), (
f"Найдена generic строка в blocked_reason: {task['blocked_reason']!r}. "
f"Ожидался специфичный blocked_reason из dept_result."
)
assert task["blocked_reason"] == specific_reason
@patch("agents.runner._run_autocommit")
@patch("agents.runner._execute_department_head_step")
@patch("agents.runner._is_department_head")
@patch("agents.runner.subprocess.run")
@patch("agents.runner.check_claude_auth")
def test_fallback_to_generic_when_dept_result_has_no_blocked_reason(
self, mock_auth, mock_run, mock_is_dept, mock_dept_step, mock_autocommit, conn
):
"""Если dept_result не содержит ни blocked_reason, ни error —
используется fallback generic строка (не KeyError, задача блокируется).
"""
from agents.runner import run_pipeline
from core import models
mock_run.return_value = _mock_agent_success()
mock_is_dept.return_value = True
mock_dept_step.return_value = {
"success": False,
"blocked_reason": None,
"error": None,
"output": "",
"cost_usd": 0,
"tokens_used": 0,
"duration_seconds": 0,
}
steps = [{"role": "backend_head", "brief": "plan"}]
result = run_pipeline(conn, "PROJ1-001", steps)
assert result["success"] is False
task = models.get_task(conn, "PROJ1-001")
assert task["status"] == "blocked"
# Должен быть хоть какой-то blocked_reason (fallback, не None)
assert task["blocked_reason"] is not None
assert len(task["blocked_reason"]) > 0