smartgridbench-review-artifact / code /tests /test_aat_tools_direct.py
garn-garn's picture
Refresh SmartGridBench reviewer artifact
d606d10 verified
"""Tests for the Cell A tool builder.
The builder wraps every entry in mcp_servers.direct_adapter as an
agents SDK function_tool, preserving the MCP-visible name and description.
"""
from __future__ import annotations
import asyncio
import pathlib
from types import SimpleNamespace
import pytest
def test_build_direct_tools_returns_one_per_registry_entry() -> None:
from mcp_servers import direct_adapter
from scripts.aat_tools_direct import build_direct_tools
tools = build_direct_tools()
registry = direct_adapter.get_tools()
assert len(tools) == len(
registry
), f"Expected {len(registry)} tools wrapped, got {len(tools)}"
def test_build_direct_tools_uses_mcp_visible_names() -> None:
from mcp_servers import direct_adapter
from scripts.aat_tools_direct import build_direct_tools
tools = build_direct_tools()
registry_names = {
spec.name.rsplit(".", 1)[-1] for spec in direct_adapter.get_tools()
}
wrapped_names = {getattr(tool, "name", None) for tool in tools}
assert wrapped_names == registry_names, (
f"Missing or renamed tools.\n"
f"Only in registry: {registry_names - wrapped_names}\n"
f"Only in wrapped: {wrapped_names - registry_names}"
)
def test_build_direct_tools_rejects_duplicate_visible_names(monkeypatch) -> None:
from scripts import aat_tools_direct
def first_tool() -> None:
pass
def second_tool() -> None:
pass
monkeypatch.setattr(
aat_tools_direct.direct_adapter,
"get_tools",
lambda: [
SimpleNamespace(name="iot.duplicate", fn=first_tool, doc="first"),
SimpleNamespace(name="fmsr.duplicate", fn=second_tool, doc="second"),
],
)
with pytest.raises(ValueError, match="Duplicate Agent-visible AaT tool name"):
aat_tools_direct.build_direct_tools()
@pytest.mark.slow
def test_direct_and_mcp_tool_schemas_match(monkeypatch):
"""Fairness-contract enforcer for Experiment 1.
The Cell A tool surface (in-process callables) and the Cell B tool
surface (MCP stdio) must expose the same tool names and parameter
requiredness. If this test
fails, (Cell B - Cell A) is no longer a clean measurement of MCP
transport overhead — there's an additional tool-surface delta.
Marked @pytest.mark.slow because it launches 4 MCP stdio subprocesses.
"""
from mcp_servers import direct_adapter
from scripts.aat_tools_mcp import build_mcp_servers
from scripts.aat_tools_direct import build_direct_tools
monkeypatch.setenv("AAT_MCP_SERVER_LAUNCH_MODE", "uv")
monkeypatch.delenv("AAT_MCP_SERVER_PYTHON", raising=False)
def schema_contract(schema: dict) -> dict[str, bool]:
properties = schema.get("properties", {}) or {}
required = set(schema.get("required", []) or [])
return {name: name in required for name in properties}
direct_tools = build_direct_tools()
direct_names = {getattr(tool, "name", None) for tool in direct_tools}
direct_schemas = {
getattr(tool, "name", None): schema_contract(tool.params_json_schema)
for tool in direct_tools
}
async def collect_mcp_schemas() -> tuple[set[str], dict[str, dict[str, bool]]]:
servers = await build_mcp_servers(pathlib.Path.cwd())
try:
names: set[str] = set()
schemas: dict[str, dict[str, bool]] = {}
for srv in servers:
tools_result = await srv.list_tools()
for t in tools_result:
names.add(t.name)
schemas[t.name] = schema_contract(t.inputSchema)
return names, schemas
finally:
for srv in servers:
try:
await srv.cleanup()
except (asyncio.CancelledError, Exception):
# asyncio.CancelledError is a BaseException in Py3.8+, not
# an Exception. The openai-agents MCP stdio teardown can
# leak a benign CancelledError from the anyio subprocess
# wait() — absorb it here so the parity assertion still runs.
pass
mcp_names, mcp_schemas = asyncio.run(collect_mcp_schemas())
assert direct_names == mcp_names, (
f"Cell A and Cell B tool surfaces diverge.\n"
f"Only in direct_adapter: {direct_names - mcp_names}\n"
f"Only in MCP stdio: {mcp_names - direct_names}\n"
f"If MCP server hardening renamed or added tools, "
f"resync mcp_servers/direct_adapter.py."
)
assert direct_schemas == mcp_schemas