mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:41:57 +00:00
ThearchyHelios
941723b17c
- 身份与工具：登录页建第一个管理员（认人接口的 401 说明能不能建）；搭建者有只读的人和组名录，选人给候选并核对名字；授权三行带标题；端点请求头名校验（空名、非法名存不进、发不出时说清原因）；写入能力在工具页试运行一律演练、不发送，HTTP 状态照实说成败；写记忆不算只读、查询能力只能选同一端点上的只读能力；审批与遮罩的说法和实际规则一致；开登录时不要署名、导入和发布记登录的人；使用者不显示推演等只给搭建者的操作，403 按是否开登录说话；登录页只认一次人 - 流程与实例：图里用到的岗位（派给谁、超时升级给谁、发送确认派给谁）全部派生、没绑人的发布时报错；写外部系统的参数按能力声明的类型收拢后再算哈希和发送，收不成就不发；发起人的实例页只看这一步的状态，处理人才给待办卡；会签进度带登录核实和时间，升级后照实写规则；实例状态写流程的结论（已退回、未确认、等待核实写入结果），核实生效后不再挂失败原因；流程应用首页不派生最近封存运行的结果块，看板没数据时是安静的空态；样例⑩的金额带类型和下限、待办标题带金额，表单上下限前后端一起生效 - 看板与证据：定时任务模板加 ISO 周、上一 ISO 周、本月、上月；出处链穿过原样解析查询结果的整形节点，追不到原始数据的照实说停在哪里，样例⑨改用 cell()；应用首页和版本历史都能做成应用；指标块单位不重复；版本列表带退役说明；定时任务页头跟着刷新；删视图要确认、新视图能起名；编辑器属性用中文名、术语带解释 - 指标出具（用户 2026-10-15 拍板）：契约只写「指标来自」、上游没有报告撰写节点、成果里没有模型写的话时按指标出具——必需指标齐、指标输入追到原始数据、成果字段里的数回指得上才完整出具；受管级要求每个指标静态地追到原始数据（governed.metric_untraced）；出具横幅和应用印章写明按指标 - 受管工作流降为已发布只给管理员（Q8 顺延，服务端同样拦）；工作流接口带当前发布版的级别，画布上改过之后也认得出受管工作流
273 lines
16 KiB
Python
273 lines
16 KiB
Python
"""能力登记（应用平台 A6）：同步、治理、快照、作证判定、写入能力的位置、正式运行免审批的只读能力、四条发布检查。
|
|
|
|
- 同步：主键是图里的工具名；同步得出的效果只说系统知道的（可写数据源为视参数而定，绑定端点的非 GET 为写入）；
|
|
source_hash 没变不动；来源没了记 missing、不删行；
|
|
- 治理：version 乐观锁、写审计；只读能力才能免审批；查询能力只能是同一端点上、绑定端点的只读能力；
|
|
- 写记忆这类改平台自身状态的内置工具不记成只读（不能免审批、不能当查询能力），照旧能给 Agent 用；
|
|
- is_platform_query：数据源查询，和绑定端点、声明了表格、治理后只读的 http 工具；Python 工具写了表格也不算；
|
|
- build_tools(allow_write=False) 拒绝写入能力，视参数而定的照旧放行；
|
|
- 正式运行：只读 + 免审批的能力直接放行（修 F7），其余照旧问人；总要确认的探索运行里也问；
|
|
- 发布检查：action.write_in_agent、action.unattended_write、action.prod_missing、action.reconcile_unplanned。
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
from langchain_core.tools import StructuredTool
|
|
from pydantic import BaseModel
|
|
from sqlalchemy import delete, select, update
|
|
|
|
from app.capabilities import registry
|
|
from app.db.base import SessionLocal
|
|
from app.db.models import AuditEvent, Capability, CustomTool, DataSource, Endpoint, McpServer
|
|
from app.tools.registry import ToolBuildError, ToolContext, Versions, build_tools
|
|
from app.tools.trust import CONFIRM, EXEMPT, FORMAL_MARK, call_policy, run_snapshot
|
|
from tests.fakes import auth_env
|
|
|
|
|
|
@pytest.fixture
|
|
async def world():
|
|
"""一个登记的端点 erp（只有生产环境）、绑定它的读（GET）和写（POST）两个 http 工具、一个 python 工具、
|
|
只读和可写两个数据源、一个缓存了工具清单的 MCP 服务。"""
|
|
async with SessionLocal() as s:
|
|
erp = Endpoint(name="erp", environments={"prod": {"base_url": "https://erp.example.com", "headers": {}}})
|
|
s.add(erp)
|
|
await s.flush()
|
|
s.add_all([
|
|
CustomTool(name="erp_orders", kind="http", parameters={"type": "object", "properties": {}},
|
|
config={"endpoint": erp.id, "method": "GET", "path": "/orders"}),
|
|
CustomTool(name="erp_pay", kind="http", parameters={"type": "object", "properties": {}},
|
|
config={"endpoint": erp.id, "method": "POST", "path": "/payments"}),
|
|
CustomTool(name="calc_py", kind="python", parameters={"type": "object", "properties": {}},
|
|
config={"code": "print(1)"}),
|
|
DataSource(name="sales_ro", kind="sqlite", readonly=True),
|
|
DataSource(name="sales_rw", kind="sqlite", readonly=False),
|
|
McpServer(name="kb", transport="stdio", command="kb-server", tools_cache=["search"]),
|
|
])
|
|
await s.commit()
|
|
erp_id = erp.id
|
|
await registry.sync_all(s)
|
|
try:
|
|
yield {"erp": erp_id}
|
|
finally:
|
|
async with SessionLocal() as s:
|
|
await s.execute(delete(CustomTool).where(CustomTool.name.in_(["erp_orders", "erp_pay", "calc_py"])))
|
|
await s.execute(delete(DataSource).where(DataSource.name.in_(["sales_ro", "sales_rw"])))
|
|
await s.execute(delete(McpServer).where(McpServer.name == "kb"))
|
|
await s.execute(delete(Endpoint).where(Endpoint.name == "erp"))
|
|
await s.execute(delete(Capability))
|
|
await s.execute(delete(AuditEvent))
|
|
await s.commit()
|
|
await registry.refresh_cache(s)
|
|
|
|
|
|
async def _row(name: str) -> Capability:
|
|
async with SessionLocal() as s:
|
|
row = await s.get(Capability, name)
|
|
assert row is not None, name
|
|
return row
|
|
|
|
|
|
async def test_sync_derives_effects_and_only_touches_what_changed(world):
|
|
effects = {n: (await _row(n)).effect_system for n in (
|
|
"calculator", "shell_exec", "memory_recall", "memory_write", "db_query__sales_ro", "db_query__sales_rw",
|
|
"db_schema__sales_rw", "erp_orders", "erp_pay", "calc_py", "mcp:kb/search")}
|
|
assert effects == {"calculator": "read", "shell_exec": "dynamic", "memory_recall": "read", "memory_write": "dynamic",
|
|
"db_query__sales_ro": "read", "db_query__sales_rw": "dynamic", "db_schema__sales_rw": "read",
|
|
"erp_orders": "read", "erp_pay": "write", "calc_py": "dynamic", "mcp:kb/search": "dynamic"}, \
|
|
"写记忆改的是平台自己的状态：不是只读"
|
|
pay = await _row("erp_pay")
|
|
assert (pay.system, pay.endpoint_id) == ("erp", world["erp"])
|
|
async with SessionLocal() as s:
|
|
assert await registry.sync_all(s) == 0, "来源没变不改任何行"
|
|
await s.execute(delete(CustomTool).where(CustomTool.name == "calc_py"))
|
|
await s.commit()
|
|
assert await registry.sync_all(s) == 1
|
|
assert (await _row("calc_py")).missing, "来源没了记 missing、不删行"
|
|
|
|
|
|
async def test_governance_lock_audit_and_rules(world):
|
|
c = auth_env.client()
|
|
row = (await c.get("/api/capabilities/mcp:kb/search")).json()
|
|
assert row["version"] == 1 and row["effect"] is None
|
|
r = await c.put("/api/capabilities/mcp:kb/search/governance",
|
|
json={"version": 1, "effect": "read", "approval": "none", "masked": ["/phone"]})
|
|
assert r.status_code == 200, r.text
|
|
assert (r.json()["version"], r.json()["effect"], r.json()["approval"]) == (2, "read", "none")
|
|
stale = await c.put("/api/capabilities/mcp:kb/search/governance", json={"version": 1, "approval": "confirm"})
|
|
assert stale.status_code == 409, "读到的是旧版本：别人已经改过"
|
|
bad = await c.put("/api/capabilities/erp_pay/governance", json={"version": 1, "approval": "none"})
|
|
assert bad.status_code == 422, "写入能力不能免审批"
|
|
async with SessionLocal() as s:
|
|
audit = (await s.execute(select(AuditEvent).where(AuditEvent.action == "capability.govern"))).scalar_one()
|
|
assert audit.subject_id == "mcp:kb/search" and audit.data["after"]["approval"] == "none"
|
|
|
|
async with SessionLocal() as s:
|
|
snap = await registry.capability_snapshot(s, ["mcp:kb/search", "erp_pay", "nope"])
|
|
versions = await registry.capability_versions_for(s, ["mcp:kb/search", "erp_pay"])
|
|
assert snap["mcp:kb/search"] == {**snap["mcp:kb/search"], "effect": "read", "approval": "none", "version": 2,
|
|
"masked": ["/phone"]}
|
|
assert "nope" not in snap and versions == {"mcp:kb/search": 2, "erp_pay": 1}
|
|
|
|
|
|
async def test_lookup_is_a_read_on_the_same_endpoint(world):
|
|
c = auth_env.client()
|
|
for wrong in ("memory_write", "db_query__sales_ro", "mcp:kb/search", "erp_pay", "nope"):
|
|
r = await c.put("/api/capabilities/erp_pay/governance", json={"version": 1, "lookup": wrong})
|
|
assert r.status_code == 422 and "同一端点" in r.json()["detail"], wrong
|
|
r = await c.put("/api/capabilities/erp_pay/governance", json={"version": 1, "lookup": "erp_orders"})
|
|
assert r.status_code == 200 and r.json()["lookup"] == "erp_orders", r.text
|
|
r = await c.put("/api/capabilities/erp_pay/governance", json={"version": 2, "lookup": None})
|
|
assert r.status_code == 200 and r.json()["lookup"] is None, "清掉照旧可以"
|
|
async with SessionLocal() as s:
|
|
await s.execute(update(Capability).where(Capability.name == "memory_write").values(effect="read"))
|
|
await s.commit()
|
|
r = await c.put("/api/capabilities/erp_pay/governance", json={"version": 3, "lookup": "memory_write"})
|
|
assert r.status_code == 422, "治理成只读的内置工具也不是端点上的读取，查不了这一笔写入"
|
|
ctx = ToolContext(run_id="t", node_id="n", data_versions=Versions.CURRENT)
|
|
async with SessionLocal() as s:
|
|
await s.execute(update(Capability).where(Capability.name == "memory_write").values(effect=None))
|
|
await s.commit()
|
|
assert [t.name for t in await build_tools(["memory_write"], ctx, session=s)] == ["memory_write"], \
|
|
"写记忆不是对外写入，Agent 照旧能用"
|
|
|
|
|
|
async def test_tool_library_rehearses_anything_governed_as_a_write(world):
|
|
c = auth_env.client()
|
|
r = await c.put("/api/capabilities/mcp:kb/search/governance", json={"version": 1, "effect": "write"})
|
|
assert r.status_code == 200, r.text
|
|
out = await c.post("/api/tools/mcp:kb/search/run", json={"args": {"q": "x"}})
|
|
assert out.status_code == 200, "演练不发送，没有副作用，不先要确认"
|
|
assert out.json()["rehearsal"] is True and out.json()["result"]["args"] == {"q": "x"}
|
|
|
|
|
|
async def test_platform_query_is_by_identity_not_by_shape(world):
|
|
c = auth_env.client()
|
|
tabular = {"rows": "/items", "columns": {"单号": "/id"}}
|
|
assert not registry.is_platform_query("erp_orders")
|
|
r = await c.put("/api/capabilities/erp_orders/governance", json={"version": 1, "tabular": tabular})
|
|
assert r.status_code == 200, r.text
|
|
assert registry.is_platform_query("erp_orders")
|
|
assert registry.is_platform_query("db_query__sales_ro")
|
|
r = await c.put("/api/capabilities/calc_py/governance", json={"version": 1, "tabular": tabular, "effect": "read"})
|
|
assert r.status_code == 200
|
|
assert not registry.is_platform_query("calc_py"), "用户代码在算的工具写了表格也不能作证"
|
|
|
|
|
|
async def test_write_capabilities_stay_out_of_agents(world):
|
|
ctx = ToolContext(run_id="t", node_id="n", data_versions=Versions.CURRENT)
|
|
async with SessionLocal() as s:
|
|
with pytest.raises(ToolBuildError, match="erp_pay"):
|
|
await build_tools(["erp_pay"], ctx, session=s)
|
|
assert [t.name for t in await build_tools(["erp_pay"], ctx, session=s, allow_write=True)] == ["erp_pay"]
|
|
# 视参数而定的照旧放行：每次调用按参数问审批
|
|
assert len(await build_tools(["db_query__sales_rw", "shell_exec"], ctx, session=s)) == 2
|
|
|
|
|
|
async def test_an_endpoint_call_that_depends_on_the_arguments_counts_as_a_write(world):
|
|
"""平台执行器不看参数，分不出这一次读还是写：绑定端点的能力不能治理成视参数而定；登记里已经是视参数而定的（先治理、
|
|
后绑定端点）也按写入对待，Agent 里拒绝。MCP 绑定端点只管连到哪儿，不算。"""
|
|
c = auth_env.client()
|
|
r = await c.put("/api/capabilities/erp_pay/governance", json={"version": 1, "effect": "dynamic"})
|
|
assert r.status_code == 422 and "只读或写入" in r.json()["detail"]
|
|
async with SessionLocal() as s:
|
|
await s.execute(update(Capability).where(Capability.name == "erp_pay").values(effect="dynamic"))
|
|
await s.commit()
|
|
ctx = ToolContext(run_id="t", node_id="n", data_versions=Versions.CURRENT)
|
|
with pytest.raises(ToolBuildError, match="erp_pay"):
|
|
await build_tools(["erp_pay"], ctx, session=s)
|
|
assert registry.writes_externally("erp_pay", "dynamic", world["erp"])
|
|
assert not registry.writes_externally("mcp:kb/search", "dynamic", world["erp"])
|
|
assert not registry.writes_externally("calc_py", "dynamic", None)
|
|
|
|
|
|
class _NoArgs(BaseModel):
|
|
pass
|
|
|
|
|
|
def _tool(name: str, key: str | None = None, **meta) -> StructuredTool:
|
|
async def run() -> str:
|
|
return "ok"
|
|
|
|
return StructuredTool(name=name, description="", coroutine=run, func=None, args_schema=_NoArgs,
|
|
metadata={**({"trust_key": key} if key else {}), **meta})
|
|
|
|
|
|
async def test_formal_runs_skip_approval_only_for_governed_read_only(world):
|
|
c = auth_env.client()
|
|
await c.put("/api/capabilities/mcp:kb/search/governance", json={"version": 1, "effect": "read", "approval": "none"})
|
|
await c.put("/api/capabilities/calc_py/governance", json={"version": 1, "approval": "confirm"})
|
|
async with SessionLocal() as s:
|
|
formal = await run_snapshot(s, run_class="formal")
|
|
exploratory = await run_snapshot(s, run_class="exploratory")
|
|
assert formal == {FORMAL_MARK: "1", "mcp:kb/search": EXEMPT, "calc_py": CONFIRM}
|
|
search, other = _tool("search", "mcp:kb/search"), _tool("lookup", "mcp:kb/lookup")
|
|
assert call_policy(search, "search", {}, formal) == "safe", "只读 + 免审批：正式运行直接放行（F7）"
|
|
assert call_policy(other, "lookup", {}, formal) == "ask", "其余照旧问人"
|
|
calc = _tool("calc_py", "calc_py")
|
|
assert call_policy(calc, "calc_py", {}, {**exploratory, "calc_py": CONFIRM}) == "ask"
|
|
assert call_policy(calc, "calc_py", {}, {"calc_py": "always"}) == "safe", "没治理过的照三档"
|
|
# 按参数判定的工具：登记写了只读，DML 照旧问
|
|
dml = _tool("db_query__sales_rw", dangerous_if=lambda args: True)
|
|
assert call_policy(dml, "db_query__sales_rw", {}, {FORMAL_MARK: "1", "db_query__sales_rw": EXEMPT}) == "ask"
|
|
|
|
|
|
def _spec(*nodes, defaults=None):
|
|
from app.engine.schema import GraphSpec
|
|
|
|
graph = {"nodes": [{"id": "in", "type": "input", "position": {"x": 0, "y": 0}, "data": {"label": "入口", "config": {}}},
|
|
*nodes], "edges": [], "defaults": defaults or {}}
|
|
return GraphSpec.model_validate(graph)
|
|
|
|
|
|
def _node(nid, ntype, **config):
|
|
return {"id": nid, "type": ntype, "position": {"x": 0, "y": 0}, "data": {"label": nid, "config": config}}
|
|
|
|
|
|
def _codes(spec, level="published"):
|
|
from app.engine.governance import lint_for_publish
|
|
|
|
return {(i.code, i.level) for i in lint_for_publish(spec, level=level).issues if (i.code or "").startswith("action.")}
|
|
|
|
|
|
async def test_publish_lints(world):
|
|
agent = _spec(_node("a", "agent", tools=["erp_pay", "db_query__sales_rw"]))
|
|
assert ("action.write_in_agent", "error") in _codes(agent, "governed")
|
|
assert any(code == "action.write_in_agent" for code, _ in _codes(agent, "published"))
|
|
team = _spec(_node("t", "supervisor", agents=[{"name": "x", "tools": ["erp_pay"]}]))
|
|
assert any(code == "action.write_in_agent" for code, _ in _codes(team, "governed"))
|
|
|
|
tool = _spec(_node("w", "tool", tool="erp_pay", approval="never"))
|
|
governed = _codes(tool, "governed")
|
|
assert ("action.unattended_write", "error") in governed
|
|
assert ("action.reconcile_unplanned", "warning") in governed
|
|
assert ("action.unattended_write", "warning") in _codes(tool, "published")
|
|
assert not any(code == "action.prod_missing" for code, _ in governed)
|
|
|
|
async with SessionLocal() as s:
|
|
row = await s.get(Endpoint, world["erp"])
|
|
row.environments = {"test": {"base_url": "https://erp-test.example.com", "headers": {}}}
|
|
pay = await s.get(Capability, "erp_pay")
|
|
pay.unattended_write, pay.lookup = True, "erp_orders"
|
|
await s.commit()
|
|
await registry.refresh_cache(s)
|
|
governed = _codes(tool, "governed")
|
|
assert ("action.prod_missing", "error") in governed
|
|
assert not {"action.unattended_write", "action.reconcile_unplanned"} & {c for c, _ in governed}
|
|
assert not _codes(_spec(_node("r", "tool", tool="erp_orders", approval="never")), "governed"), "读取不报"
|
|
|
|
|
|
async def test_snapshots_for_launch_and_publish_never_write(world):
|
|
"""发起运行、发布时调用方的会话开着写事务：快照只读，登记表里还没有的名字按系统此刻知道的样子补，不插行。"""
|
|
async with SessionLocal() as s:
|
|
s.add(DataSource(name="sales_new", kind="sqlite", readonly=False))
|
|
await s.flush() # 调用方自己的写事务还开着
|
|
snap = await registry.capability_snapshot(s, ["db_query__sales_new", "erp_pay"])
|
|
versions = await registry.capability_versions_for(s, ["db_query__sales_new"])
|
|
await s.commit()
|
|
assert snap["db_query__sales_new"] == {**snap["db_query__sales_new"], "effect": "dynamic", "version": 1}
|
|
assert versions == {"db_query__sales_new": 1}
|
|
async with SessionLocal() as s:
|
|
assert await s.get(Capability, "db_query__sales_new") is None, "快照不往登记表里写"
|
|
await s.execute(delete(DataSource).where(DataSource.name == "sales_new"))
|
|
await s.commit()
|