feat: add agent registry manifests and coverage
Introduce a manifest-backed agent registry surface and align graph tests with the new runtime prompt and tool indexing behavior.
This commit is contained in:
@@ -1,470 +1,291 @@
|
||||
from pathlib import Path
|
||||
from types import SimpleNamespace
|
||||
import sys
|
||||
|
||||
from langchain_core.messages import AIMessage, HumanMessage
|
||||
WORKTREE_ROOT = Path(__file__).resolve().parents[4]
|
||||
if str(WORKTREE_ROOT) not in sys.path:
|
||||
sys.path.insert(0, str(WORKTREE_ROOT))
|
||||
for module_name in list(sys.modules):
|
||||
if module_name == "app" or module_name.startswith("app."):
|
||||
del sys.modules[module_name]
|
||||
|
||||
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
|
||||
from langgraph.graph import END
|
||||
|
||||
from app.agents.graph import (
|
||||
_choose_sub_commander,
|
||||
_parse_json_action,
|
||||
_route_agent_from_user_query,
|
||||
_run_sub_commander,
|
||||
JSON_ACTION_FALLBACK_PROMPT,
|
||||
_get_role_tools,
|
||||
call_agent_llm,
|
||||
execute_tools_node,
|
||||
master_node,
|
||||
route_after_agent,
|
||||
route_master,
|
||||
)
|
||||
from app.agents.tools.time_reasoning import resolve_time_expression
|
||||
from app.agents.state import AgentRole
|
||||
from app.agents.tools import SUB_COMMANDER_TOOLSETS
|
||||
from app.agents.prompts import MASTER_SYSTEM_PROMPT
|
||||
|
||||
|
||||
|
||||
|
||||
def _base_state(message: str, user_llm_config: dict | None = None) -> dict:
|
||||
def _base_state(message: str = "帮我安排今天的重点") -> dict:
|
||||
return {
|
||||
'messages': [HumanMessage(content=message)],
|
||||
'user_id': 'u1',
|
||||
'conversation_id': 'c1',
|
||||
'current_agent': AgentRole.MASTER,
|
||||
'active_agents': [AgentRole.MASTER],
|
||||
'current_sub_commander': None,
|
||||
'active_sub_commanders': [],
|
||||
'sub_commander_trace': [],
|
||||
'pending_tasks': [],
|
||||
'completed_tasks': [],
|
||||
'tool_calls': [],
|
||||
'last_tool_result': None,
|
||||
'action_results': [],
|
||||
'created_entities': [],
|
||||
'tool_strategy_used': None,
|
||||
'provider_capabilities': None,
|
||||
'fallback_parse_error': None,
|
||||
'knowledge_context': None,
|
||||
'graph_context': None,
|
||||
'schedule_context_summary': None,
|
||||
'plan': None,
|
||||
'plan_steps': [],
|
||||
'analysis_report': None,
|
||||
'final_response': None,
|
||||
'should_respond': True,
|
||||
'memory_context': None,
|
||||
'current_datetime_context': 'CURRENT_TIME: 2026-03-28T12:00:00+08:00',
|
||||
'current_datetime_reference': {'current_time_iso': '2026-03-28T12:00:00+08:00', 'current_date_iso': '2026-03-28', 'timezone': 'UTC'},
|
||||
'user_llm_config': user_llm_config,
|
||||
"messages": [HumanMessage(content=message)],
|
||||
"user_id": "u1",
|
||||
"conversation_id": "c1",
|
||||
"current_agent": AgentRole.MASTER.value,
|
||||
"next_step": None,
|
||||
"agent_trace": [AgentRole.MASTER.value],
|
||||
"pending_tasks": [],
|
||||
"completed_tasks": [],
|
||||
"created_entities": [],
|
||||
"knowledge_context": None,
|
||||
"schedule_context_summary": None,
|
||||
"analysis_report": None,
|
||||
"final_response": None,
|
||||
"memory_context": None,
|
||||
"current_datetime_context": None,
|
||||
"user_llm_config": None,
|
||||
"provider_capabilities": None,
|
||||
}
|
||||
|
||||
|
||||
class FakeFallbackLLM:
|
||||
def __init__(self, first_content: str, followup_content: str = '已创建提醒：开会，时间为 2026-03-29 09:00（按当前时间理解为“明天早上9点”）。'):
|
||||
self.first_content = first_content
|
||||
self.followup_content = followup_content
|
||||
self.calls = 0
|
||||
|
||||
async def ainvoke(self, messages):
|
||||
self.calls += 1
|
||||
if self.calls == 1:
|
||||
return AIMessage(content=self.first_content)
|
||||
return AIMessage(content=self.followup_content)
|
||||
|
||||
def bind_tools(self, tools):
|
||||
raise AssertionError('bind_tools should not be called in JSON fallback mode')
|
||||
|
||||
|
||||
class FakeNativeBoundLLM:
|
||||
async def ainvoke(self, messages):
|
||||
return AIMessage(
|
||||
content='',
|
||||
tool_calls=[
|
||||
{
|
||||
'id': 'call_1',
|
||||
'name': 'create_reminder',
|
||||
'args': {'title': '开会', 'reminder_at': '明天 09:00'},
|
||||
}
|
||||
],
|
||||
)
|
||||
|
||||
|
||||
class FakeNativeLLM:
|
||||
def __init__(self):
|
||||
self.bound = FakeNativeBoundLLM()
|
||||
self.tool_binding_count = 0
|
||||
self.calls = 0
|
||||
self._jarvis_provider_capabilities = SimpleNamespace(provider='openai', supports_native_tools=True, preferred_tool_strategy='native')
|
||||
|
||||
def bind_tools(self, tools):
|
||||
self.tool_binding_count += 1
|
||||
return self.bound
|
||||
|
||||
async def ainvoke(self, messages):
|
||||
self.calls += 1
|
||||
return AIMessage(content='已创建提醒：开会，时间为 2026-03-29 09:00（按当前时间理解为“明天早上9点”）。')
|
||||
|
||||
|
||||
class FakeTool:
|
||||
def __init__(self, name: str, result: str):
|
||||
self.name = name
|
||||
self.result = result
|
||||
self.invocations: list[dict] = []
|
||||
|
||||
def invoke(self, args: dict):
|
||||
self.invocations.append(args)
|
||||
return self.result
|
||||
|
||||
|
||||
class CapturingLLM:
|
||||
def __init__(self, content: str = '{"mode":"final","final_response":"好的。"}'):
|
||||
self.content = content
|
||||
self.messages = None
|
||||
self._jarvis_provider_capabilities = SimpleNamespace(provider='ollama', supports_native_tools=False, preferred_tool_strategy='json_fallback')
|
||||
|
||||
async def ainvoke(self, messages):
|
||||
self.messages = messages
|
||||
return AIMessage(content=self.content)
|
||||
|
||||
|
||||
class FailIfCalledLLM:
|
||||
async def ainvoke(self, messages):
|
||||
raise AssertionError('LLM should not be called for simple greetings')
|
||||
raise AssertionError("LLM should not be called for greeting fast-path")
|
||||
|
||||
|
||||
async def test_master_node_returns_stable_reply_for_simple_greeting(monkeypatch):
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: FailIfCalledLLM())
|
||||
class StaticResponseLLM:
|
||||
def __init__(self, response: AIMessage):
|
||||
self.response = response
|
||||
self.messages = None
|
||||
|
||||
state = {
|
||||
'messages': [HumanMessage(content='你好')],
|
||||
'user_id': 'u1',
|
||||
'conversation_id': 'c1',
|
||||
'current_agent': AgentRole.MASTER,
|
||||
'active_agents': [AgentRole.MASTER],
|
||||
'pending_tasks': [],
|
||||
'completed_tasks': [],
|
||||
'tool_calls': [],
|
||||
'last_tool_result': None,
|
||||
'knowledge_context': None,
|
||||
'graph_context': None,
|
||||
'plan': None,
|
||||
'plan_steps': [],
|
||||
'analysis_report': None,
|
||||
'final_response': None,
|
||||
'should_respond': True,
|
||||
'memory_context': None,
|
||||
'user_llm_config': None,
|
||||
}
|
||||
|
||||
result = await master_node(state)
|
||||
|
||||
assert result['final_response'] == '您好。我在。\n\n您把问题给我，我先帮您收束重点，再往下推。'
|
||||
assert result['current_agent'] == AgentRole.MASTER
|
||||
assert result['active_agents'] == [AgentRole.MASTER]
|
||||
async def ainvoke(self, messages):
|
||||
self.messages = messages
|
||||
return self.response
|
||||
|
||||
|
||||
async def test_master_node_returns_stable_reply_for_identity_question(monkeypatch):
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: FailIfCalledLLM())
|
||||
class CaptureFallbackLLM:
|
||||
def __init__(self, response: AIMessage):
|
||||
self.response = response
|
||||
self.messages = None
|
||||
self.bind_tools_called = False
|
||||
|
||||
state = {
|
||||
'messages': [HumanMessage(content='你是谁')],
|
||||
'user_id': 'u1',
|
||||
'conversation_id': 'c1',
|
||||
'current_agent': AgentRole.MASTER,
|
||||
'active_agents': [AgentRole.MASTER],
|
||||
'pending_tasks': [],
|
||||
'completed_tasks': [],
|
||||
'tool_calls': [],
|
||||
'last_tool_result': None,
|
||||
'knowledge_context': None,
|
||||
'graph_context': None,
|
||||
'plan': None,
|
||||
'plan_steps': [],
|
||||
'analysis_report': None,
|
||||
'final_response': None,
|
||||
'should_respond': True,
|
||||
'memory_context': None,
|
||||
'user_llm_config': None,
|
||||
}
|
||||
async def ainvoke(self, messages):
|
||||
self.messages = messages
|
||||
return self.response
|
||||
|
||||
result = await master_node(state)
|
||||
|
||||
assert result['final_response'] == '我是 Jarvis。\n\n比起做一个泛泛的助手，我更像您的判断型协作伙伴：帮您看清问题、压缩路径、把事情往前推进。'
|
||||
assert result['current_agent'] == AgentRole.MASTER
|
||||
assert result['active_agents'] == [AgentRole.MASTER]
|
||||
def bind_tools(self, tools):
|
||||
self.bind_tools_called = True
|
||||
raise AssertionError("bind_tools should not be used when native tools are unsupported")
|
||||
|
||||
|
||||
async def test_master_node_returns_stable_reply_for_identity_question_with_punctuation(monkeypatch):
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: FailIfCalledLLM())
|
||||
class AsyncFakeTool:
|
||||
def __init__(self, name: str, result: str):
|
||||
self.name = name
|
||||
self.result = result
|
||||
self.calls: list[dict] = []
|
||||
|
||||
state = {
|
||||
'messages': [HumanMessage(content='你是谁？')],
|
||||
'user_id': 'u1',
|
||||
'conversation_id': 'c1',
|
||||
'current_agent': AgentRole.MASTER,
|
||||
'active_agents': [AgentRole.MASTER],
|
||||
'pending_tasks': [],
|
||||
'completed_tasks': [],
|
||||
'tool_calls': [],
|
||||
'last_tool_result': None,
|
||||
'knowledge_context': None,
|
||||
'graph_context': None,
|
||||
'plan': None,
|
||||
'plan_steps': [],
|
||||
'analysis_report': None,
|
||||
'final_response': None,
|
||||
'should_respond': True,
|
||||
'memory_context': None,
|
||||
'user_llm_config': None,
|
||||
}
|
||||
|
||||
result = await master_node(state)
|
||||
|
||||
assert result['final_response'] == '我是 Jarvis。\n\n比起做一个泛泛的助手，我更像您的判断型协作伙伴：帮您看清问题、压缩路径、把事情往前推进。'
|
||||
assert result['current_agent'] == AgentRole.MASTER
|
||||
assert result['active_agents'] == [AgentRole.MASTER]
|
||||
async def ainvoke(self, args: dict):
|
||||
self.calls.append(args)
|
||||
return self.result
|
||||
|
||||
|
||||
async def test_master_node_returns_stable_reply_for_identity_question_with_particle(monkeypatch):
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: FailIfCalledLLM())
|
||||
class SyncFakeTool:
|
||||
def __init__(self, name: str, result: str):
|
||||
self.name = name
|
||||
self.result = result
|
||||
self.calls: list[dict] = []
|
||||
|
||||
state = {
|
||||
'messages': [HumanMessage(content='你是谁啊')],
|
||||
'user_id': 'u1',
|
||||
'conversation_id': 'c1',
|
||||
'current_agent': AgentRole.MASTER,
|
||||
'active_agents': [AgentRole.MASTER],
|
||||
'pending_tasks': [],
|
||||
'completed_tasks': [],
|
||||
'tool_calls': [],
|
||||
'last_tool_result': None,
|
||||
'knowledge_context': None,
|
||||
'graph_context': None,
|
||||
'plan': None,
|
||||
'plan_steps': [],
|
||||
'analysis_report': None,
|
||||
'final_response': None,
|
||||
'should_respond': True,
|
||||
'memory_context': None,
|
||||
'user_llm_config': None,
|
||||
}
|
||||
|
||||
result = await master_node(state)
|
||||
|
||||
assert result['final_response'] == '我是 Jarvis。\n\n比起做一个泛泛的助手，我更像您的判断型协作伙伴：帮您看清问题、压缩路径、把事情往前推进。'
|
||||
assert result['current_agent'] == AgentRole.MASTER
|
||||
assert result['active_agents'] == [AgentRole.MASTER]
|
||||
def invoke(self, args: dict):
|
||||
self.calls.append(args)
|
||||
return self.result
|
||||
|
||||
|
||||
async def test_master_node_returns_stable_reply_for_capability_question(monkeypatch):
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: FailIfCalledLLM())
|
||||
async def test_master_node_greeting_fast_path_returns_stable_reply_without_llm(monkeypatch):
|
||||
monkeypatch.setattr("app.agents.graph._get_llm_for_state", lambda state: (FailIfCalledLLM(), SimpleNamespace()))
|
||||
|
||||
state = {
|
||||
'messages': [HumanMessage(content='你能做什么')],
|
||||
'user_id': 'u1',
|
||||
'conversation_id': 'c1',
|
||||
'current_agent': AgentRole.MASTER,
|
||||
'active_agents': [AgentRole.MASTER],
|
||||
'pending_tasks': [],
|
||||
'completed_tasks': [],
|
||||
'tool_calls': [],
|
||||
'last_tool_result': None,
|
||||
'knowledge_context': None,
|
||||
'graph_context': None,
|
||||
'plan': None,
|
||||
'plan_steps': [],
|
||||
'analysis_report': None,
|
||||
'final_response': None,
|
||||
'should_respond': True,
|
||||
'memory_context': None,
|
||||
'user_llm_config': None,
|
||||
}
|
||||
result = await master_node(_base_state("你好"))
|
||||
|
||||
result = await master_node(state)
|
||||
|
||||
assert result['final_response'] == '主要做三件事。\n- 帮您判断：看问题本质、梳理取舍、给出方向\n- 帮您收束：把复杂内容理顺，把重点拎出来\n- 帮您推进：拆任务、定步骤、把下一步变清楚\n\n如果您现在有具体目标，我可以直接进入处理。'
|
||||
assert result['current_agent'] == AgentRole.MASTER
|
||||
assert result['active_agents'] == [AgentRole.MASTER]
|
||||
assert result["final_response"] == "您好。我在。\n\n您把问题给我，我先帮您收束重点，再往下推。"
|
||||
assert result["messages"][0].content == "您好。我在。"
|
||||
|
||||
|
||||
def test_choose_sub_commander_routes_schedule_requests_to_schedule_planning():
|
||||
assert _choose_sub_commander(AgentRole.SCHEDULE_PLANNER, '帮我安排一下这周计划') == 'schedule_planning'
|
||||
|
||||
|
||||
def test_choose_sub_commander_routes_focus_requests_to_schedule_analysis():
|
||||
assert _choose_sub_commander(AgentRole.SCHEDULE_PLANNER, '基于最近对话帮我判断该聚焦什么') == 'schedule_analysis'
|
||||
|
||||
|
||||
def test_route_agent_from_user_query_routes_knowledge_requests_to_librarian():
|
||||
assert _route_agent_from_user_query('帮我搜索知识库里的项目资料') == AgentRole.LIBRARIAN
|
||||
|
||||
|
||||
def test_route_agent_from_user_query_routes_schedule_requests_to_schedule_planner():
|
||||
assert _route_agent_from_user_query('明天提醒我开会') == AgentRole.SCHEDULE_PLANNER
|
||||
|
||||
|
||||
def test_route_agent_from_user_query_routes_explicit_month_day_milestone_to_schedule_planner():
|
||||
assert _route_agent_from_user_query('3月29日，对话系统交付节点') == AgentRole.SCHEDULE_PLANNER
|
||||
|
||||
|
||||
def test_choose_sub_commander_routes_explicit_month_day_milestone_to_schedule_planning():
|
||||
assert _choose_sub_commander(AgentRole.SCHEDULE_PLANNER, '3月29日，对话系统交付节点') == 'schedule_planning'
|
||||
|
||||
|
||||
|
||||
|
||||
def test_parse_json_action_extracts_tool_calls_from_fenced_json():
|
||||
parsed = _parse_json_action(
|
||||
'```json\n{"mode":"tool_call","tool_calls":[{"name":"create_reminder","arguments":{"title":"开会","reminder_at":"明天 09:00"}}]}\n```',
|
||||
['create_reminder'],
|
||||
async def test_master_node_routes_to_agent_when_llm_returns_role_name(monkeypatch):
|
||||
llm = StaticResponseLLM(AIMessage(content="schedule_planner"))
|
||||
monkeypatch.setattr(
|
||||
"app.agents.graph._get_llm_for_state",
|
||||
lambda state: (llm, SimpleNamespace(provider="test", supports_native_tools=True)),
|
||||
)
|
||||
|
||||
assert parsed == {
|
||||
'mode': 'tool_call',
|
||||
'tool_calls': [
|
||||
{
|
||||
'name': 'create_reminder',
|
||||
'args': {'title': '开会', 'reminder_at': '明天 09:00'},
|
||||
'reason': None,
|
||||
}
|
||||
state = _base_state("帮我安排这周重点")
|
||||
result = await master_node(state)
|
||||
|
||||
assert result["current_agent"] == AgentRole.SCHEDULE_PLANNER.value
|
||||
assert result["agent_trace"] == [AgentRole.MASTER.value, AgentRole.SCHEDULE_PLANNER.value]
|
||||
assert result["messages"][0].content == f"已分发至 {AgentRole.SCHEDULE_PLANNER.value} 处理。"
|
||||
assert isinstance(llm.messages[0], SystemMessage)
|
||||
assert MASTER_SYSTEM_PROMPT in llm.messages[0].content
|
||||
|
||||
|
||||
async def test_master_node_returns_final_response_when_llm_answers_directly(monkeypatch):
|
||||
response = AIMessage(content="我建议先收束需求，再拆执行步骤。")
|
||||
llm = StaticResponseLLM(response)
|
||||
monkeypatch.setattr(
|
||||
"app.agents.graph._get_llm_for_state",
|
||||
lambda state: (llm, SimpleNamespace(provider="test", supports_native_tools=True)),
|
||||
)
|
||||
|
||||
result = await master_node(_base_state("现在应该怎么推进这个项目？"))
|
||||
|
||||
assert result["final_response"] == response.content
|
||||
assert result["messages"] == [response]
|
||||
|
||||
|
||||
def test_route_after_agent_sends_tool_calls_to_tools_node():
|
||||
state = _base_state()
|
||||
state["messages"] = [AIMessage(content="", tool_calls=[{"id": "1", "name": "create_task", "args": {}}])]
|
||||
|
||||
assert route_after_agent(state) == "tools"
|
||||
|
||||
|
||||
def test_route_after_agent_ends_when_no_tool_calls_exist():
|
||||
state = _base_state()
|
||||
state["messages"] = [AIMessage(content="done")]
|
||||
|
||||
assert route_after_agent(state) == END
|
||||
|
||||
|
||||
def test_route_master_ends_when_final_response_exists():
|
||||
state = _base_state()
|
||||
state["final_response"] = "done"
|
||||
state["current_agent"] = AgentRole.EXECUTOR.value
|
||||
|
||||
assert route_master(state) == END
|
||||
|
||||
|
||||
def test_route_master_returns_current_agent_when_more_work_remains():
|
||||
state = _base_state()
|
||||
state["current_agent"] = AgentRole.LIBRARIAN.value
|
||||
|
||||
assert route_master(state) == AgentRole.LIBRARIAN.value
|
||||
|
||||
|
||||
def test_get_role_tools_returns_expected_semantic_tool_sets():
|
||||
expected_by_role = {
|
||||
AgentRole.SCHEDULE_PLANNER: [
|
||||
"get_schedule_day",
|
||||
"get_tasks",
|
||||
"resolve_time_expression",
|
||||
"create_todo",
|
||||
"create_schedule_task",
|
||||
"create_reminder",
|
||||
"create_goal",
|
||||
],
|
||||
AgentRole.EXECUTOR: [
|
||||
"get_tasks",
|
||||
"create_task",
|
||||
"update_task_status",
|
||||
"resolve_time_expression",
|
||||
"create_todo",
|
||||
"create_schedule_task",
|
||||
"create_reminder",
|
||||
"create_goal",
|
||||
"get_forum_posts",
|
||||
"create_forum_post",
|
||||
"scan_forum_for_instructions",
|
||||
],
|
||||
AgentRole.LIBRARIAN: [
|
||||
"search_knowledge",
|
||||
"hybrid_search",
|
||||
"web_search",
|
||||
"get_knowledge_graph_context",
|
||||
"build_knowledge_graph",
|
||||
],
|
||||
AgentRole.ANALYST: [
|
||||
"get_tasks",
|
||||
"get_forum_posts",
|
||||
"scan_forum_for_instructions",
|
||||
"search_knowledge",
|
||||
"hybrid_search",
|
||||
"web_search",
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def test_parse_json_action_returns_none_for_invalid_or_unknown_payload():
|
||||
assert _parse_json_action('not json', ['create_reminder']) is None
|
||||
assert _parse_json_action('{"mode":"tool_call","tool_calls":[{"name":"unknown","arguments":{}}]}', ['create_reminder']) is None
|
||||
for role, expected_tool_names in expected_by_role.items():
|
||||
actual_tools = _get_role_tools(role)
|
||||
actual_tool_names = [tool.name for tool in actual_tools]
|
||||
assert actual_tool_names == expected_tool_names
|
||||
assert len(actual_tool_names) == len(set(actual_tool_names))
|
||||
|
||||
|
||||
def test_parse_json_action_tolerates_prefix_and_suffix_text():
|
||||
parsed = _parse_json_action(
|
||||
'好的，下面是 JSON：\n```json\n{"mode":"tool_call","tool_calls":[{"name":"create_reminder","arguments":{"title":"开会","reminder_at":"明天 09:00"}}]}\n```\n谢谢',
|
||||
['create_reminder'],
|
||||
)
|
||||
assert parsed is not None
|
||||
assert parsed['mode'] == 'tool_call'
|
||||
assert parsed['tool_calls'][0]['name'] == 'create_reminder'
|
||||
async def test_execute_tools_node_executes_tool_calls_and_tracks_created_entities(monkeypatch):
|
||||
create_tool = AsyncFakeTool("create_task", "created task 123")
|
||||
read_tool = SyncFakeTool("get_tasks", "[]")
|
||||
|
||||
|
||||
def test_parse_json_action_accepts_parameters_alias_for_tool_calls():
|
||||
parsed = _parse_json_action(
|
||||
'{"mode":"tool_call","tool_calls":[{"name":"create_reminder","parameters":{"title":"收被子","reminder_at":"2026-03-29T09:00:00+08:00"}}]}',
|
||||
['create_reminder'],
|
||||
monkeypatch.setattr("app.agents.graph.ALL_TOOLS", [create_tool, read_tool])
|
||||
monkeypatch.setattr(
|
||||
"app.agents.graph.normalize_tool_time_arguments",
|
||||
lambda tool_name, tool_args, current_datetime_context: {**tool_args, "normalized": True},
|
||||
)
|
||||
|
||||
assert parsed == {
|
||||
'mode': 'tool_call',
|
||||
'tool_calls': [
|
||||
{
|
||||
'name': 'create_reminder',
|
||||
'args': {'title': '收被子', 'reminder_at': '2026-03-29T09:00:00+08:00'},
|
||||
'reason': None,
|
||||
}
|
||||
],
|
||||
}
|
||||
state = _base_state()
|
||||
state["created_entities"] = [{"tool": "existing", "result": "already there"}]
|
||||
state["current_datetime_context"] = "2026-04-02T09:00:00+08:00"
|
||||
state["messages"] = [
|
||||
AIMessage(
|
||||
content="",
|
||||
tool_calls=[
|
||||
{"id": "tool-1", "name": "create_task", "args": {"title": "Write tests"}},
|
||||
{"id": "tool-2", "name": "get_tasks", "args": {"status": "open"}},
|
||||
],
|
||||
)
|
||||
]
|
||||
|
||||
result = await execute_tools_node(state)
|
||||
|
||||
assert create_tool.calls == [{"title": "Write tests", "normalized": True}]
|
||||
assert read_tool.calls == [{"status": "open", "normalized": True}]
|
||||
assert [type(message) for message in result["messages"]] == [ToolMessage, ToolMessage]
|
||||
assert result["messages"][0].tool_call_id == "tool-1"
|
||||
assert result["messages"][0].name == "create_task"
|
||||
assert result["messages"][0].content == "created task 123"
|
||||
assert result["messages"][1].tool_call_id == "tool-2"
|
||||
assert result["messages"][1].name == "get_tasks"
|
||||
assert result["messages"][1].content == "[]"
|
||||
assert result["created_entities"] == [
|
||||
{"tool": "existing", "result": "already there"},
|
||||
{"tool": "create_task", "result": "created task 123"},
|
||||
]
|
||||
|
||||
|
||||
async def test_run_sub_commander_uses_json_fallback_for_non_native_provider(monkeypatch):
|
||||
fake_llm = FakeFallbackLLM(
|
||||
'{"mode":"tool_call","tool_calls":[{"name":"create_reminder","arguments":{"title":"开会","reminder_at":"明天 09:00"}}]}'
|
||||
async def test_call_agent_llm_includes_context_messages_and_uses_json_fallback(monkeypatch):
|
||||
llm = CaptureFallbackLLM(AIMessage(content='{"mode":"final","final_response":"好的。"}'))
|
||||
capabilities = SimpleNamespace(
|
||||
provider="ollama",
|
||||
supports_native_tools=False,
|
||||
preferred_tool_strategy="json_fallback",
|
||||
)
|
||||
fake_tool = FakeTool('create_reminder', '成功创建 reminder: 开会 @ 明天 09:00')
|
||||
fake_tools = [SimpleNamespace(name="create_reminder"), SimpleNamespace(name="get_tasks")]
|
||||
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: fake_llm)
|
||||
monkeypatch.setitem(
|
||||
__import__('app.agents.graph', fromlist=['SUB_COMMANDER_TOOLSETS']).SUB_COMMANDER_TOOLSETS,
|
||||
'schedule_planning',
|
||||
[fake_tool],
|
||||
)
|
||||
monkeypatch.setattr("app.agents.graph._get_llm_for_state", lambda state: (llm, capabilities))
|
||||
monkeypatch.setattr("app.agents.graph._get_role_tools", lambda role: fake_tools)
|
||||
monkeypatch.setattr("app.agents.graph.build_skill_context", lambda role_key: "技能上下文: 先判断，再执行")
|
||||
|
||||
state = _base_state('明天 9 点提醒我开会', {'provider': 'ollama', 'model': 'qwen2.5'})
|
||||
state['current_agent'] = AgentRole.SCHEDULE_PLANNER
|
||||
state = _base_state("明天提醒我开会")
|
||||
state["messages"] = [HumanMessage(content="明天提醒我开会")]
|
||||
state["current_datetime_context"] = "CURRENT_TIME: 2026-04-02T09:00:00+08:00"
|
||||
state["memory_context"] = "用户偏好早上处理深度工作。"
|
||||
|
||||
result = await _run_sub_commander(
|
||||
state,
|
||||
AgentRole.SCHEDULE_PLANNER,
|
||||
'manager prompt',
|
||||
'明天 9 点提醒我开会',
|
||||
use_tools=True,
|
||||
)
|
||||
result = await call_agent_llm(state, AgentRole.EXECUTOR, "executor system prompt")
|
||||
|
||||
assert result['tool_strategy_used'] == 'json_fallback'
|
||||
assert fake_tool.invocations == [{'title': '开会', 'reminder_at': '2026-03-29T09:00:00'}]
|
||||
assert result['tool_calls'][0]['name'] == 'create_reminder'
|
||||
assert result['created_entities'][0]['type'] == 'reminder'
|
||||
assert result['fallback_parse_error'] is None
|
||||
assert result['final_response'] == '已创建提醒：开会，时间为 2026-03-29 09:00（按当前时间理解为“明天早上9点”）。'
|
||||
assert result["messages"][0].content == '{"mode":"final","final_response":"好的。"}'
|
||||
assert llm.bind_tools_called is False
|
||||
assert llm.messages is not None
|
||||
|
||||
|
||||
async def test_run_sub_commander_includes_current_datetime_context_in_system_messages(monkeypatch):
|
||||
fake_llm = CapturingLLM('{"mode":"final","final_response":"好的。"}')
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: fake_llm)
|
||||
|
||||
state = _base_state('明天 9 点提醒我开会', {'provider': 'ollama', 'model': 'qwen2.5'})
|
||||
state['current_agent'] = AgentRole.SCHEDULE_PLANNER
|
||||
state['current_datetime_context'] = 'CURRENT_TIME: 2026-03-28T12:00:00+08:00'
|
||||
|
||||
await _run_sub_commander(
|
||||
state,
|
||||
AgentRole.SCHEDULE_PLANNER,
|
||||
'manager prompt',
|
||||
'明天 9 点提醒我开会',
|
||||
use_tools=True,
|
||||
)
|
||||
|
||||
assert fake_llm.messages is not None
|
||||
assert any(
|
||||
getattr(m, 'type', None) == 'system' and 'CURRENT_TIME:' in str(getattr(m, 'content', ''))
|
||||
for m in fake_llm.messages
|
||||
)
|
||||
|
||||
|
||||
async def test_run_sub_commander_uses_web_search_in_json_fallback(monkeypatch):
|
||||
fake_llm = FakeFallbackLLM(
|
||||
'{"mode":"tool_call","tool_calls":[{"name":"web_search","arguments":{"query":"Jarvis 最新模型更新","top_k":2}}]}',
|
||||
'我查了外部网页，下面是最新结果摘要。',
|
||||
)
|
||||
fake_tool = FakeTool('web_search', '成功搜索到 2 条网页结果')
|
||||
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: fake_llm)
|
||||
monkeypatch.setitem(
|
||||
__import__('app.agents.graph', fromlist=['SUB_COMMANDER_TOOLSETS']).SUB_COMMANDER_TOOLSETS,
|
||||
'librarian_retrieval',
|
||||
[fake_tool],
|
||||
)
|
||||
|
||||
state = _base_state('帮我上网查一下 Jarvis 最新模型更新', {'provider': 'ollama', 'model': 'qwen2.5'})
|
||||
state['current_agent'] = AgentRole.LIBRARIAN
|
||||
|
||||
result = await _run_sub_commander(
|
||||
state,
|
||||
AgentRole.LIBRARIAN,
|
||||
'manager prompt',
|
||||
'帮我上网查一下 Jarvis 最新模型更新',
|
||||
use_tools=True,
|
||||
summary_target='knowledge_context',
|
||||
)
|
||||
|
||||
assert result['tool_strategy_used'] == 'json_fallback'
|
||||
assert fake_tool.invocations == [{'query': 'Jarvis 最新模型更新', 'top_k': 2}]
|
||||
assert result['tool_calls'][0]['name'] == 'web_search'
|
||||
assert result['last_tool_result'] == '[web_search] 成功搜索到 2 条网页结果'
|
||||
assert result['final_response'] == '我查了外部网页，下面是最新结果摘要。'
|
||||
|
||||
|
||||
fake_llm = FakeNativeLLM()
|
||||
fake_tool = FakeTool('create_reminder', '成功创建 reminder: 开会 @ 明天 09:00')
|
||||
|
||||
monkeypatch.setattr('app.agents.graph._get_llm_for_state', lambda state: fake_llm)
|
||||
monkeypatch.setitem(
|
||||
__import__('app.agents.graph', fromlist=['SUB_COMMANDER_TOOLSETS']).SUB_COMMANDER_TOOLSETS,
|
||||
'schedule_planning',
|
||||
[fake_tool],
|
||||
)
|
||||
|
||||
state = _base_state('明天 9 点提醒我开会', {'provider': 'openai', 'model': 'gpt-4o'})
|
||||
state['current_agent'] = AgentRole.SCHEDULE_PLANNER
|
||||
|
||||
result = await _run_sub_commander(
|
||||
state,
|
||||
AgentRole.SCHEDULE_PLANNER,
|
||||
'manager prompt',
|
||||
'明天 9 点提醒我开会',
|
||||
use_tools=True,
|
||||
)
|
||||
|
||||
assert result['tool_strategy_used'] == 'native'
|
||||
assert fake_llm.tool_binding_count == 1
|
||||
assert fake_tool.invocations == [{'title': '开会', 'reminder_at': '2026-03-29T09:00:00'}]
|
||||
assert result['created_entities'][0]['type'] == 'reminder'
|
||||
assert result['final_response'] == '已创建提醒：开会，时间为 2026-03-29 09:00（按当前时间理解为“明天早上9点”）。'
|
||||
system_contents = [message.content for message in llm.messages if isinstance(message, SystemMessage)]
|
||||
assert "executor system prompt" in system_contents[0]
|
||||
assert any("当前时间上下文: CURRENT_TIME: 2026-04-02T09:00:00+08:00" == content for content in system_contents)
|
||||
assert any("长期记忆上下文: 用户偏好早上处理深度工作。" == content for content in system_contents)
|
||||
assert any("技能上下文: 先判断，再执行" == content for content in system_contents)
|
||||
assert any(content == JSON_ACTION_FALLBACK_PROMPT for content in system_contents)
|
||||
assert any(content == "本次可用工具列表: create_reminder, get_tasks" for content in system_contents)
|
||||
assert any(isinstance(message, HumanMessage) and message.content == "明天提醒我开会" for message in llm.messages)
|
||||
|
||||
360
backend/tests/backend/app/agents/test_registry.py
Normal file
360
backend/tests/backend/app/agents/test_registry.py
Normal file
@@ -0,0 +1,360 @@
|
||||
import pytest
|
||||
from collections.abc import Mapping
|
||||
|
||||
from app.agents.prompts import (
|
||||
SUB_COMMANDER_PROMPTS_BY_KEY,
|
||||
TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY,
|
||||
)
|
||||
from app.agents.registry import build_registry_indexes, load_builtin_registry_bundle
|
||||
from app.agents.registry.indexes import summarize_registry_indexes
|
||||
from app.agents.registry.models import (
|
||||
AgentManifest,
|
||||
CapabilityManifest,
|
||||
SpecialistTemplateManifest,
|
||||
SubCommanderManifest,
|
||||
)
|
||||
from app.agents.registry.validator import validate_registry_bundle
|
||||
from app.agents.registry.builtins import (
|
||||
BUILTIN_AGENT_MANIFESTS,
|
||||
BUILTIN_CAPABILITY_MANIFESTS,
|
||||
BUILTIN_SPECIALIST_TEMPLATE_MANIFESTS,
|
||||
BUILTIN_SUB_COMMANDER_MANIFESTS,
|
||||
)
|
||||
from app.agents.state import AgentRole
|
||||
from app.agents.tools import SUB_COMMANDER_TOOLSETS
|
||||
|
||||
|
||||
def make_agent(
|
||||
agent_id: str = "master",
|
||||
*,
|
||||
display_name: str = "Master",
|
||||
role_value: str = "master",
|
||||
system_prompt_key: str = "master",
|
||||
default_sub_commanders: list[str] | None = None,
|
||||
) -> AgentManifest:
|
||||
return AgentManifest(
|
||||
agent_id=agent_id,
|
||||
display_name=display_name,
|
||||
role_value=role_value,
|
||||
system_prompt_key=system_prompt_key,
|
||||
routing_hints=["route"],
|
||||
default_sub_commanders=default_sub_commanders or [],
|
||||
)
|
||||
|
||||
|
||||
def make_sub_commander(
|
||||
sub_commander_id: str = "planner",
|
||||
*,
|
||||
parent_agent_id: str = "master",
|
||||
capability_ids: list[str] | None = None,
|
||||
) -> SubCommanderManifest:
|
||||
return SubCommanderManifest(
|
||||
sub_commander_id=sub_commander_id,
|
||||
parent_agent_id=parent_agent_id,
|
||||
prompt_text="Plan the work.",
|
||||
capability_ids=capability_ids or [],
|
||||
)
|
||||
|
||||
|
||||
def make_capability(capability_id: str = "calendar") -> CapabilityManifest:
|
||||
return CapabilityManifest(capability_id=capability_id, tool_name=f"{capability_id}_tool")
|
||||
|
||||
|
||||
def make_specialist_template(
|
||||
template_id: str = "researcher",
|
||||
*,
|
||||
allowed_capability_ids: list[str] | None = None,
|
||||
) -> SpecialistTemplateManifest:
|
||||
return SpecialistTemplateManifest(
|
||||
template_id=template_id,
|
||||
display_name="Researcher",
|
||||
description="Research specialist",
|
||||
allowed_capability_ids=allowed_capability_ids,
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_accepts_valid_bundle() -> None:
|
||||
validate_registry_bundle(
|
||||
agents=[make_agent(default_sub_commanders=["planner"])],
|
||||
sub_commanders=[make_sub_commander(capability_ids=["calendar"])],
|
||||
capabilities=[make_capability()],
|
||||
specialist_templates=[make_specialist_template(allowed_capability_ids=["calendar"])],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_duplicate_agent_ids() -> None:
|
||||
agents = [
|
||||
make_agent(default_sub_commanders=["planner"]),
|
||||
make_agent(
|
||||
display_name="Duplicate Master",
|
||||
role_value="master_duplicate",
|
||||
system_prompt_key="master_duplicate",
|
||||
),
|
||||
]
|
||||
|
||||
with pytest.raises(ValueError, match="duplicate agent id: master"):
|
||||
validate_registry_bundle(
|
||||
agents=agents,
|
||||
sub_commanders=[],
|
||||
capabilities=[],
|
||||
specialist_templates=[],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_duplicate_sub_commander_ids() -> None:
|
||||
with pytest.raises(ValueError, match="duplicate sub commander id: planner"):
|
||||
validate_registry_bundle(
|
||||
agents=[make_agent()],
|
||||
sub_commanders=[make_sub_commander(), make_sub_commander()],
|
||||
capabilities=[],
|
||||
specialist_templates=[],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_duplicate_capability_ids() -> None:
|
||||
with pytest.raises(ValueError, match="duplicate capability id: calendar"):
|
||||
validate_registry_bundle(
|
||||
agents=[],
|
||||
sub_commanders=[],
|
||||
capabilities=[make_capability(), make_capability()],
|
||||
specialist_templates=[],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_duplicate_template_ids() -> None:
|
||||
with pytest.raises(ValueError, match="duplicate template id: researcher"):
|
||||
validate_registry_bundle(
|
||||
agents=[],
|
||||
sub_commanders=[],
|
||||
capabilities=[],
|
||||
specialist_templates=[make_specialist_template(), make_specialist_template()],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_unknown_sub_commander_parent_agent_ids() -> None:
|
||||
sub_commanders = [make_sub_commander(parent_agent_id="missing-agent")]
|
||||
|
||||
with pytest.raises(ValueError, match="unknown parent agent id: missing-agent"):
|
||||
validate_registry_bundle(
|
||||
agents=[],
|
||||
sub_commanders=sub_commanders,
|
||||
capabilities=[],
|
||||
specialist_templates=[],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_unknown_sub_commander_capability_references() -> None:
|
||||
with pytest.raises(ValueError, match="unknown capability id: search"):
|
||||
validate_registry_bundle(
|
||||
agents=[make_agent(default_sub_commanders=["planner"])],
|
||||
sub_commanders=[make_sub_commander(capability_ids=["search"])],
|
||||
capabilities=[make_capability()],
|
||||
specialist_templates=[],
|
||||
)
|
||||
|
||||
|
||||
def test_validate_registry_bundle_rejects_unknown_specialist_template_capability_references() -> None:
|
||||
with pytest.raises(ValueError, match="unknown capability id: missing-capability"):
|
||||
validate_registry_bundle(
|
||||
agents=[],
|
||||
sub_commanders=[],
|
||||
capabilities=[make_capability()],
|
||||
specialist_templates=[
|
||||
make_specialist_template(allowed_capability_ids=["missing-capability"])
|
||||
],
|
||||
)
|
||||
|
||||
|
||||
def test_registry_bundle_agent_roles_match_runtime_agent_role_enum_values() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
assert set(indexes.agent_by_id) == {role.value for role in AgentRole}
|
||||
assert {agent.role_value for agent in bundle.agents} == {role.value for role in AgentRole}
|
||||
|
||||
|
||||
def test_registry_bundle_agent_system_prompt_keys_match_runtime_top_level_prompt_surface() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
expected_prompt_keys_by_agent_id = {
|
||||
role.value: role.value for role in AgentRole if role.value in TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY
|
||||
}
|
||||
|
||||
assert set(TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY) == {role.value for role in AgentRole}
|
||||
assert indexes.agent_prompt_key_by_id == expected_prompt_keys_by_agent_id
|
||||
assert {
|
||||
agent.agent_id: TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY[agent.system_prompt_key]
|
||||
for agent in bundle.agents
|
||||
} == {
|
||||
role.value: TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY[role.value]
|
||||
for role in AgentRole
|
||||
}
|
||||
|
||||
|
||||
def test_registry_bundle_skill_context_keys_match_graph_role_derivation_rule() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
expected_skill_context_keys = {
|
||||
role.value: role.value.replace("agent_", "")
|
||||
for role in AgentRole
|
||||
}
|
||||
|
||||
assert indexes.skill_context_key_by_agent_id == expected_skill_context_keys
|
||||
assert {
|
||||
agent.agent_id: agent.skill_context_key for agent in bundle.agents
|
||||
} == expected_skill_context_keys
|
||||
|
||||
|
||||
def test_registry_bundle_sub_commander_prompt_texts_match_runtime_prompt_map() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
assert set(indexes.sub_commander_by_id) == set(SUB_COMMANDER_PROMPTS_BY_KEY)
|
||||
assert indexes.sub_commander_prompt_key_by_id == {
|
||||
sub_commander_id: sub_commander_id
|
||||
for sub_commander_id in SUB_COMMANDER_PROMPTS_BY_KEY
|
||||
}
|
||||
assert {
|
||||
sub_commander.sub_commander_id: sub_commander.prompt_text
|
||||
for sub_commander in bundle.sub_commanders
|
||||
} == SUB_COMMANDER_PROMPTS_BY_KEY
|
||||
|
||||
|
||||
def test_registry_bundle_sub_commander_tool_membership_and_order_match_runtime_toolsets() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
assert set(indexes.sub_commander_by_id) == set(SUB_COMMANDER_TOOLSETS)
|
||||
assert indexes.capability_ids_by_sub_commander_id == {
|
||||
sub_commander_id: tuple(tool.name for tool in tools)
|
||||
for sub_commander_id, tools in SUB_COMMANDER_TOOLSETS.items()
|
||||
}
|
||||
assert {
|
||||
sub_commander.sub_commander_id: tuple(sub_commander.capability_ids)
|
||||
for sub_commander in bundle.sub_commanders
|
||||
} == {
|
||||
sub_commander_id: tuple(tool.name for tool in tools)
|
||||
for sub_commander_id, tools in SUB_COMMANDER_TOOLSETS.items()
|
||||
}
|
||||
|
||||
|
||||
def test_builtin_capabilities_reference_actual_runtime_tool_names() -> None:
|
||||
expected_tool_names = {
|
||||
tool.name
|
||||
for tools in SUB_COMMANDER_TOOLSETS.values()
|
||||
for tool in tools
|
||||
}
|
||||
manifest_tool_names = {manifest.tool_name for manifest in BUILTIN_CAPABILITY_MANIFESTS}
|
||||
|
||||
assert manifest_tool_names == expected_tool_names
|
||||
|
||||
|
||||
def test_builtin_sub_commander_capabilities_match_runtime_toolsets() -> None:
|
||||
capabilities_by_tool_name = {
|
||||
manifest.tool_name: manifest.capability_id for manifest in BUILTIN_CAPABILITY_MANIFESTS
|
||||
}
|
||||
|
||||
for sub_commander in BUILTIN_SUB_COMMANDER_MANIFESTS:
|
||||
expected_capability_ids = {
|
||||
capabilities_by_tool_name[tool.name]
|
||||
for tool in SUB_COMMANDER_TOOLSETS[sub_commander.sub_commander_id]
|
||||
}
|
||||
assert set(sub_commander.capability_ids) == expected_capability_ids
|
||||
|
||||
|
||||
def test_builtin_manifests_form_a_valid_registry_bundle() -> None:
|
||||
validate_registry_bundle(
|
||||
agents=list(BUILTIN_AGENT_MANIFESTS),
|
||||
sub_commanders=list(BUILTIN_SUB_COMMANDER_MANIFESTS),
|
||||
capabilities=list(BUILTIN_CAPABILITY_MANIFESTS),
|
||||
specialist_templates=list(BUILTIN_SPECIALIST_TEMPLATE_MANIFESTS),
|
||||
)
|
||||
|
||||
|
||||
def test_load_builtin_registry_bundle_returns_non_empty_manifest_sets() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
|
||||
assert bundle.agents
|
||||
assert bundle.sub_commanders
|
||||
assert bundle.capabilities
|
||||
assert isinstance(bundle.specialist_templates, tuple)
|
||||
|
||||
|
||||
def test_build_registry_indexes_exposes_manifest_lookups_by_id() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
assert indexes.agent_by_id
|
||||
assert indexes.sub_commander_by_id
|
||||
assert indexes.capability_by_id
|
||||
assert isinstance(indexes.specialist_template_by_id, Mapping)
|
||||
assert set(indexes.agent_by_id) == {agent.agent_id for agent in bundle.agents}
|
||||
assert set(indexes.sub_commander_by_id) == {
|
||||
sub_commander.sub_commander_id for sub_commander in bundle.sub_commanders
|
||||
}
|
||||
assert set(indexes.capability_by_id) == {
|
||||
capability.capability_id for capability in bundle.capabilities
|
||||
}
|
||||
assert set(indexes.specialist_template_by_id) == {
|
||||
template.template_id for template in bundle.specialist_templates
|
||||
}
|
||||
|
||||
|
||||
def test_summarize_registry_indexes_returns_read_only_debug_counts() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
assert summarize_registry_indexes(indexes) == {
|
||||
"agent_count": len(bundle.agents),
|
||||
"sub_commander_count": len(bundle.sub_commanders),
|
||||
"capability_count": len(bundle.capabilities),
|
||||
"specialist_template_count": len(bundle.specialist_templates),
|
||||
}
|
||||
|
||||
|
||||
def test_build_registry_indexes_exposes_prompt_keys_skill_context_keys_and_capability_mappings() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
|
||||
indexes = build_registry_indexes(bundle)
|
||||
|
||||
assert indexes.agent_prompt_key_by_id == {
|
||||
agent.agent_id: agent.system_prompt_key for agent in bundle.agents
|
||||
}
|
||||
assert indexes.agent_prompt_key_by_id == {
|
||||
agent.agent_id: agent.system_prompt_key for agent in BUILTIN_AGENT_MANIFESTS
|
||||
}
|
||||
assert set(indexes.agent_prompt_key_by_id.values()) == set(TOP_LEVEL_SYSTEM_PROMPTS_BY_KEY)
|
||||
assert indexes.sub_commander_prompt_key_by_id == {
|
||||
sub_commander.sub_commander_id: sub_commander.sub_commander_id
|
||||
for sub_commander in bundle.sub_commanders
|
||||
}
|
||||
assert set(indexes.sub_commander_prompt_key_by_id.values()) == {
|
||||
sub_commander.sub_commander_id for sub_commander in bundle.sub_commanders
|
||||
}
|
||||
assert indexes.skill_context_key_by_agent_id == {
|
||||
agent.agent_id: agent.skill_context_key
|
||||
for agent in bundle.agents
|
||||
if agent.skill_context_key is not None
|
||||
}
|
||||
assert indexes.capability_ids_by_sub_commander_id == {
|
||||
sub_commander.sub_commander_id: tuple(sub_commander.capability_ids)
|
||||
for sub_commander in bundle.sub_commanders
|
||||
}
|
||||
|
||||
|
||||
def test_validate_registry_bundle_accepts_loaded_builtin_registry_bundle() -> None:
|
||||
bundle = load_builtin_registry_bundle()
|
||||
|
||||
validate_registry_bundle(
|
||||
agents=list(bundle.agents),
|
||||
sub_commanders=list(bundle.sub_commanders),
|
||||
capabilities=list(bundle.capabilities),
|
||||
specialist_templates=list(bundle.specialist_templates),
|
||||
)
|
||||
|
||||
|
||||
def test_phase_one_still_declares_specialist_template_surface_even_if_runtime_is_deferred() -> None:
|
||||
assert isinstance(BUILTIN_SPECIALIST_TEMPLATE_MANIFESTS, tuple)
|
||||
Reference in New Issue
Block a user