Files
Cognitive_Agent/tests/integration/test_scene_template_hybrid.py
T

627 lines
24 KiB
Python

from __future__ import annotations
import hashlib
import json
from collections.abc import Callable, Mapping
from copy import deepcopy
from pathlib import Path
from unittest.mock import MagicMock, patch
import pytest
from cognitive_agent.application.assistant_capabilities import AssistantCapabilityCatalog
from cognitive_agent.application.codex_assistant import CodexAssistantService
from cognitive_agent.application.mode_profile import ModeProfileRegistry
from cognitive_agent.application.scene_composer import normalize_scene_output
from cognitive_agent.application.seat_service import REPOSITORY_ROOT, SeatDemoService
@pytest.fixture
def hybrid_service(
tmp_path: Path,
) -> tuple[SeatDemoService, AssistantCapabilityCatalog, Path]:
service = SeatDemoService(REPOSITORY_ROOT)
service.artifact_root = tmp_path / "runs"
user_root = tmp_path / "user-modes"
service.mode_profiles = ModeProfileRegistry(
REPOSITORY_ROOT / "config" / "modes",
user_root=user_root,
)
try:
yield service, AssistantCapabilityCatalog(service), user_root
finally:
service.close()
def _target(action: Mapping[str, object]) -> str:
target = action.get("target")
assert isinstance(target, Mapping)
return str(target.get("instance") or target.get("zone") or "")
def _actions(result: Mapping[str, object]) -> list[dict[str, object]]:
intent = result.get("intent_ir")
if isinstance(intent, Mapping):
spec = intent.get("spec")
raw_actions = spec.get("actions") if isinstance(spec, Mapping) else None
else:
raw_actions = result.get("actions")
assert isinstance(raw_actions, list)
assert all(isinstance(action, dict) for action in raw_actions)
return raw_actions
def _action_map(
result: Mapping[str, object],
) -> dict[tuple[str, str], dict[str, object]]:
return {
(str(action["capability"]), _target(action)): deepcopy(dict(action.get("parameters") or {}))
for action in _actions(result)
}
def _draft_reference(result: Mapping[str, object]) -> dict[str, str]:
draft = result.get("mode_profile_draft")
assert isinstance(draft, Mapping)
return {
"name": str(draft["name"]),
"source_request_id": str(draft["source_request_id"]),
"vehicle_id": str(draft["vehicle_id"]),
}
def _template_hashes() -> dict[str, str]:
root = REPOSITORY_ROOT / "config" / "scene_templates"
return {
path.name: hashlib.sha256(path.read_bytes()).hexdigest()
for path in sorted(root.glob("*.yaml"))
}
def _assistant_preview(
service: SeatDemoService,
catalog: AssistantCapabilityCatalog,
message: str,
*,
conversation_id: str,
) -> tuple[
dict[str, object],
dict[str, object],
list[dict[str, object]],
MagicMock,
list[dict[str, object] | None],
]:
client = MagicMock()
client.available = False
assistant = CodexAssistantService(client)
snapshot = catalog.snapshot(vehicle_id="denza_n7")
fast_answer = catalog.fast_answer(message, snapshot)
events: list[dict[str, object]] = []
raw_results: list[dict[str, object]] = []
passed_plans: list[dict[str, object] | None] = []
def execute_control(
instruction: str,
progress: Callable[[dict[str, object]], None],
assistant_plan: Mapping[str, object] | None = None,
) -> dict[str, object]:
plan = deepcopy(dict(assistant_plan)) if assistant_plan is not None else None
passed_plans.append(plan)
result = service.preview_native_can(
instruction,
provider="local",
conversation_id=conversation_id,
execution_mode="hardware",
vehicle_id="denza_n7",
on_commands_validated=progress,
assistant_plan=plan,
)
raw_results.append(result)
return result
response = assistant.stream(
message,
conversation_id=conversation_id,
principal="scene-template-hybrid-test",
history=[],
authoritative_answer=fast_answer.answer if fast_answer else None,
emit=lambda event: events.append(event) or True,
execute_control=execute_control,
capability_snapshot=snapshot,
capability_context=catalog.compact_context(snapshot),
scene_capability_context=catalog.scene_composition_context(snapshot),
control_terms=catalog.control_terms(snapshot),
)
assert len(raw_results) == 1
return response, raw_results[0], events, client, passed_plans
def test_all_production_scene_templates_complete_the_full_local_dry_run(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, user_root = hybrid_service
snapshot = catalog.snapshot(vehicle_id="denza_n7")
templates = snapshot["scene_templates"]
assert isinstance(templates, list)
assert len(templates) == 13
assert all(template["assistant_callable"] is True for template in templates)
with patch.object(
service.hardware_execution,
"reserve",
side_effect=AssertionError("a scene template preview reached hardware execution"),
):
for index, template in enumerate(templates):
plan = deepcopy(template["plan"])
result = service.preview_native_can(
f"构建一个{template['name']}",
provider="local",
conversation_id=f"conv-template-matrix-{index:02d}",
execution_mode="hardware",
vehicle_id="denza_n7",
assistant_plan=plan,
)
assert result["status"] == "succeeded", template["id"]
assert result["execution_mode"] == "dry_run", template["id"]
assert result["hardware_execution_allowed"] is False, template["id"]
assert result["orchestration_route"] == "mode_profile_create_preview"
draft = result["mode_profile_draft"]
assert draft["generation_source"] == "scene_template_copy"
assert draft["operation_count"] == template["operation_count"]
assert draft["source_template_id"] == template["id"]
assert draft["source_template_revision"] == template["template_revision"]
assert "existing_mode_id" not in draft
expected = {
(
str(request["capability"]),
str(request["target_scope"]),
json.dumps(
{
parameter["name"]: parameter["value"]
for parameter in request["parameters"]
},
ensure_ascii=False,
sort_keys=True,
),
)
for request in plan["requests"]
}
actual = {
(
str(action["capability"]),
_target(action),
json.dumps(
action.get("parameters") or {},
ensure_ascii=False,
sort_keys=True,
),
)
for action in _actions(result)
}
assert actual == expected, template["id"]
assert not list(user_root.rglob("*.yaml"))
def test_viewing_template_creates_an_unbound_audited_copy_without_codex(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, user_root = hybrid_service
with patch.object(
service.hardware_execution,
"reserve",
side_effect=AssertionError("template copy reached hardware execution"),
):
response, raw_result, events, client, passed_plans = _assistant_preview(
service,
catalog,
"构建一个观影模式",
conversation_id="conv-template-viewing-copy",
)
draft = raw_result["mode_profile_draft"]
assert response["route"] == "scene_template"
assert response["success"] is True
assert raw_result["status"] == "succeeded"
assert raw_result["execution_mode"] == "dry_run"
assert raw_result["hardware_execution_allowed"] is False
assert draft["name"] == "观影模式"
assert draft["generation_source"] == "scene_template_copy"
assert draft["source_template_id"] == "cinema"
assert draft["source_template_revision"]
assert "existing_mode_id" not in draft
assert "existing_mode_revision" not in draft
assert raw_result["mode_creation"]["source_template_id"] == "cinema"
assert passed_plans[0]["metadata"]["source"] == "scene_template"
assert events[0]["engine"] == "n7web-scene-template"
assert "独立草稿" in response["answer"]
assert not list(user_root.rglob("*.yaml"))
client.start_structured_turn.assert_not_called()
client.start_turn.assert_not_called()
@pytest.mark.parametrize(
("case_id", "message", "expected_name", "template_id"),
(
(
"reading-purpose",
"帮我构建一个适合看书的功能模式",
"看书模式",
"reading",
),
(
"reading-purpose-preview",
"我想生成一个用来读书的模式并预览",
"读书模式",
"reading",
),
(
"relax-purpose",
"构建一个适合放松的功能模式",
"放松模式",
"relaxation",
),
(
"relax-purpose-polite",
"请生成用于放松的模式,先看看",
"放松模式",
"relaxation",
),
(
"cinema-purpose",
"帮我构建一个适合看电影的功能模式",
"看电影模式",
"cinema",
),
),
)
def test_purpose_descriptions_use_a_concise_name_and_zero_model_template_route(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
case_id: str,
message: str,
expected_name: str,
template_id: str,
) -> None:
service, catalog, user_root = hybrid_service
with patch.object(
service.hardware_execution,
"reserve",
side_effect=AssertionError("purpose template reached hardware execution"),
):
response, raw_result, events, client, passed_plans = _assistant_preview(
service,
catalog,
message,
conversation_id=f"conv-template-{case_id}",
)
template = service.scene_templates.get_template(template_id)
draft = raw_result["mode_profile_draft"]
assert response["route"] == "scene_template"
assert response["success"] is True
assert expected_name in response["answer"]
assert raw_result["status"] == "succeeded"
assert raw_result["execution_mode"] == "dry_run"
assert raw_result["hardware_execution_allowed"] is False
assert draft["name"] == expected_name
assert draft["generation_source"] == "scene_template_copy"
assert draft["source_template_id"] == template_id
assert draft["operation_count"] == len(template.operations)
assert passed_plans[0]["metadata"]["template_id"] == template_id
assert events[0]["engine"] == "n7web-scene-template"
assert not list(user_root.rglob("*.yaml"))
client.start_structured_turn.assert_not_called()
client.start_turn.assert_not_called()
def test_explicit_settings_on_a_purpose_scene_bypass_template_defaults(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, _user_root = hybrid_service
response, raw_result, events, client, passed_plans = _assistant_preview(
service,
catalog,
"构建一个适合看书的功能模式,空调温度调整到25度",
conversation_id="conv-template-purpose-explicit-settings",
)
draft = raw_result["mode_profile_draft"]
assert response["route"] == "vehicle_control"
assert response["success"] is True
assert draft["name"] == "看书模式"
assert draft["generation_source"] == "explicit_scene_requirements"
assert draft["operation_count"] == 1
assert passed_plans == [None]
assert events[0]["engine"] == "n7web-mode-builder"
client.start_structured_turn.assert_not_called()
client.start_turn.assert_not_called()
@pytest.mark.parametrize(
("message", "expected_name", "template_id"),
(
("构建一个座椅舒适的功能模式", "座椅舒适模式", "seat-comfort"),
("构建一个让人坐着放松的功能模式", "坐着放松模式", "relaxation"),
("请帮我创建一个适合乘坐舒适的模式吧", "乘坐舒适模式", "seat-comfort"),
("帮我构建一个让我们坐着舒服的功能模式", "坐着舒服模式", "seat-comfort"),
("创建一个座椅放松模式并预览", "座椅放松模式", "relaxation"),
("构建一个长途舒适模式", "长途舒适模式", "long-trip-comfort"),
),
)
def test_seating_comfort_requests_preview_both_seats_with_no_model_or_hardware_call(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
message: str,
expected_name: str,
template_id: str,
) -> None:
service, catalog, user_root = hybrid_service
with patch.object(
service.hardware_execution,
"reserve",
side_effect=AssertionError("seat comfort preview reached hardware execution"),
):
response, result, _events, client, _plans = _assistant_preview(
service, catalog, message, conversation_id="conv-seat-comfort-preview",
)
assert response["route"] == "scene_template"
assert response["success"] is True
assert result["status"] == "succeeded"
assert result["execution_mode"] == "dry_run"
assert result["hardware_execution_allowed"] is False
assert result["mode_profile_draft"]["name"] == expected_name
assert result["mode_profile_draft"]["source_template_id"] == template_id
seat_actions = {
(capability, target): parameters
for (capability, target), parameters in _action_map(result).items()
if capability.startswith("seat.")
}
assert {target for _capability, target in seat_actions} == {"driver", "passenger"}
assert all(
"driver_axis" not in capability and "driver_position" not in capability
for capability, _target in seat_actions
)
assert ("seat.set_driver_massage_native", "passenger") not in seat_actions
if template_id in {"seat-comfort", "long-trip-comfort"}:
assert seat_actions["seat.set_driver_ventilation_native", "driver"] == {
"state": "on", "level": 2,
}
assert seat_actions["seat.set_ventilation", "passenger"] == {
"state": "on", "level": 1,
}
client.start_structured_turn.assert_not_called()
client.start_turn.assert_not_called()
assert not list(user_root.rglob("*.yaml"))
@pytest.mark.parametrize("name", ("云端座椅舒展模式", "晚间乘坐释压模式"))
def test_unregistered_seat_scene_keeps_both_explicit_targets_through_native_merge(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
name: str,
) -> None:
service, catalog, user_root = hybrid_service
plan = normalize_scene_output({
"requests": [
{"capability": "seat.set_driver_massage_native", "target_scope": "driver",
"parameters": {"action": "mode", "mode": "relax"}},
{"capability": "seat.set_driver_heating_native", "target_scope": "driver",
"parameters": {"state": "on", "level": 2}},
{"capability": "seat.set_backrest_recline", "target_scope": "passenger",
"parameters": {"position_pct": 45}},
{"capability": "seat.set_heating", "target_scope": "passenger",
"parameters": {"state": "on", "level": 1}},
],
"ambiguities": [], "confidence": 0.95,
}, catalog.scene_composition_context())
with patch.object(
service.hardware_execution, "reserve",
side_effect=AssertionError("scene composition reached hardware execution"),
):
result = service.preview_native_can(
f"构建一个{name}", provider="local", conversation_id="conv-seat-generated",
execution_mode="hardware", vehicle_id="denza_n7", assistant_plan=plan,
)
assert result["status"] == "succeeded"
assert result["execution_mode"] == "dry_run"
assert result["mode_profile_draft"]["generation_source"] == "codex_scene_composition"
assert set(_action_map(result)) == {
("seat.set_driver_massage_native", "driver"),
("seat.set_driver_heating_native", "driver"),
("seat.set_backrest_recline", "passenger"),
("seat.set_heating", "passenger"),
}
assert not list(user_root.rglob("*.yaml"))
def test_explicit_unscoped_seat_instruction_keeps_target_ambiguity_guard(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, _catalog, _user_root = hybrid_service
result = service.preview_native_can(
"构建一个舒展模式,座椅靠背调整到45%", provider="local",
conversation_id="conv-seat-explicit-target", execution_mode="hardware",
vehicle_id="denza_n7",
)
assert result["status"] != "succeeded"
assert result.get("mode_profile_draft") is None
assert result["hardware_execution_allowed"] is False
def test_template_metadata_cannot_authorize_modified_or_stale_template_actions(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, user_root = hybrid_service
snapshot = catalog.snapshot(vehicle_id="denza_n7")
template = next(
item for item in snapshot["scene_templates"] if item["id"] == "cinema"
)
modified = deepcopy(template["plan"])
modified["requests"][0]["parameters"][0]["value"] = "off"
with pytest.raises(ValueError, match="模板已更新"):
service.preview_native_can(
"构建一个观影模式",
provider="local",
conversation_id="conv-template-plan-tamper",
execution_mode="hardware",
vehicle_id="denza_n7",
assistant_plan=modified,
)
stale = deepcopy(template["plan"])
stale["metadata"]["template_revision"] = "0123456789abcdef"
with pytest.raises(ValueError, match="模板已更新"):
service.preview_native_can(
"构建一个观影模式",
provider="local",
conversation_id="conv-template-plan-stale",
execution_mode="hardware",
vehicle_id="denza_n7",
assistant_plan=stale,
)
assert not list(user_root.rglob("*.yaml"))
def test_template_alias_keeps_the_user_requested_draft_name(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, _user_root = hybrid_service
response, raw_result, _events, client, passed_plans = _assistant_preview(
service,
catalog,
"创建一个影院模式",
conversation_id="conv-template-alias-name",
)
draft = raw_result["mode_profile_draft"]
assert response["route"] == "scene_template"
assert draft["name"] == "影院模式"
assert draft["source_template_id"] == "cinema"
assert passed_plans[0]["metadata"]["template_id"] == "cinema"
client.start_structured_turn.assert_not_called()
client.start_turn.assert_not_called()
def test_template_copy_can_be_revised_twice_without_losing_sibling_actions(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, user_root = hybrid_service
conversation_id = "conv-template-viewing-revision-chain"
with patch.object(
service.hardware_execution,
"reserve",
side_effect=AssertionError("template draft revision reached hardware execution"),
):
_response, built, _events, _client, _plans = _assistant_preview(
service,
catalog,
"构建一个观影模式",
conversation_id=conversation_id,
)
shade_revised = service.preview_native_can(
"把观影模式的遮阳帘变成100%",
provider="local",
conversation_id=conversation_id,
execution_mode="hardware",
vehicle_id="denza_n7",
mode_profile_draft_context=_draft_reference(built),
)
final = service.preview_native_can(
"把观影模式的空调温度调整到25度",
provider="local",
conversation_id=conversation_id,
execution_mode="hardware",
vehicle_id="denza_n7",
mode_profile_draft_context=_draft_reference(shade_revised),
)
built_actions = _action_map(built)
shade_actions = _action_map(shade_revised)
final_actions = _action_map(final)
shade_key = ("roof.set_shade_position", "vehicle")
temperature_key = ("climate.set_temperature", "vehicle")
assert shade_revised["status"] == final["status"] == "succeeded"
assert shade_revised["execution_mode"] == final["execution_mode"] == "dry_run"
assert shade_revised["hardware_execution_allowed"] is False
assert final["hardware_execution_allowed"] is False
assert len(built_actions) == len(shade_actions) == len(final_actions) == 7
assert built_actions[shade_key] == {"position_pct": 0}
assert shade_actions[shade_key] == final_actions[shade_key] == {"position_pct": 100}
assert built_actions[temperature_key] == shade_actions[temperature_key] == {"temperature_c": 24}
assert final_actions[temperature_key] == {"temperature_c": 25}
untouched_keys = set(built_actions) - {shade_key, temperature_key}
assert {key: final_actions[key] for key in untouched_keys} == {
key: built_actions[key] for key in untouched_keys
}
first_draft = built["mode_profile_draft"]
second_draft = shade_revised["mode_profile_draft"]
final_draft = final["mode_profile_draft"]
assert first_draft["draft_revision"] == 1
assert second_draft["draft_revision"] == 2
assert final_draft["draft_revision"] == 3
assert second_draft["generation_source"] == "draft_revision"
assert final_draft["generation_source"] == "draft_revision"
assert second_draft["source_template_id"] == final_draft["source_template_id"] == ("cinema")
assert "existing_mode_id" not in final_draft
assert not list(user_root.rglob("*.yaml"))
def test_saving_a_template_copy_never_mutates_template_and_registered_mode_wins_next(
hybrid_service: tuple[SeatDemoService, AssistantCapabilityCatalog, Path],
) -> None:
service, catalog, user_root = hybrid_service
hashes_before = _template_hashes()
template_revision = service.scene_templates.get_template("cinema").revision
_response, built, _events, first_client, first_plans = _assistant_preview(
service,
catalog,
"构建一个观影模式",
conversation_id="conv-template-save-source",
)
draft = built["mode_profile_draft"]
saved = service.create_mode_profile_card(
str(draft["name"]),
str(draft["instruction"]),
str(draft["source_request_id"]),
str(draft["vehicle_id"]),
)
saved_profile = saved["mode_profile"]
assert saved["status"] == "succeeded"
assert saved_profile["name"] == "观影模式"
assert saved_profile["origin"] == "user"
assert saved_profile["id"] != "cinema"
assert list(user_root.rglob(f"{saved_profile['id']}.yaml"))
assert _template_hashes() == hashes_before
assert service.scene_templates.get_template("cinema").revision == template_revision
assert first_plans[0]["metadata"]["template_id"] == "cinema"
first_client.start_structured_turn.assert_not_called()
response, registered, events, second_client, passed_plans = _assistant_preview(
service,
catalog,
"构建一个观影模式",
conversation_id="conv-template-saved-mode-priority",
)
registered_draft = registered["mode_profile_draft"]
assert response["route"] == "vehicle_control"
assert response["success"] is True
assert events[0]["engine"] == "n7web-mode-builder"
assert passed_plans == [None]
assert registered_draft["generation_source"] == "registered_mode_profile"
assert registered_draft["existing_mode_id"] == saved_profile["id"]
assert "source_template_id" not in registered_draft
second_client.start_structured_turn.assert_not_called()
second_client.start_turn.assert_not_called()