44 lines
2.0 KiB
Python
44 lines
2.0 KiB
Python
# requires: pydantic, pytest
|
|
import tool_loop as tl
|
|
|
|
|
|
def setup_function():
|
|
tl.REFUNDS.clear()
|
|
|
|
|
|
def test_happy_path_with_self_correction():
|
|
out = tl.run_agent(tl.ScriptedModel(tl.demo_script()), "退款", {"user_id": "u1"}, lambda n, a: True)
|
|
assert out["answer"].startswith("已为订单") and tl.REFUNDS == {"A1001": 199.0}
|
|
|
|
|
|
def test_user_rejects_refund():
|
|
out = tl.run_agent(tl.ScriptedModel(tl.demo_script()), "退款", {"user_id": "u1"}, lambda n, a: False)
|
|
assert tl.REFUNDS == {} and out["trace"][2][2] == {"error": "用户拒绝执行该操作"}
|
|
|
|
|
|
def test_other_users_order_hidden():
|
|
script = [{"role": "assistant", "tool_calls": [tl.call("c1", "get_order", {"order_id": "A1001"})]},
|
|
{"role": "assistant", "content": "没有找到订单"}]
|
|
out = tl.run_agent(tl.ScriptedModel(script), "查订单", {"user_id": "u2"}, lambda n, a: True)
|
|
assert out["trace"][0][2] == {"error": "订单不存在"}
|
|
|
|
|
|
def test_amount_over_paid_rejected_and_unknown_tool():
|
|
script = [{"role": "assistant", "tool_calls": [
|
|
tl.call("c1", "refund", {"order_id": "A1001", "amount": 500}),
|
|
tl.call("c2", "drop_table", {})]},
|
|
{"role": "assistant", "content": "无法退款"}]
|
|
model = tl.ScriptedModel(script)
|
|
out = tl.run_agent(model, "退 500", {"user_id": "u1"}, lambda n, a: True)
|
|
assert "不能超过" in out["trace"][0][2]["error"] and "未知工具" in out["trace"][1][2]["error"]
|
|
# 两个工具结果都回传给了模型，并且 tool_call_id 对应
|
|
last = model.seen[-1]
|
|
assert [m["tool_call_id"] for m in last if m["role"] == "tool"] == ["c1", "c2"]
|
|
|
|
|
|
def test_max_steps():
|
|
script = [{"role": "assistant", "tool_calls": [tl.call(f"c{i}", "get_order", {"order_id": "A1001"})]}
|
|
for i in range(10)]
|
|
out = tl.run_agent(tl.ScriptedModel(script), "x", {"user_id": "u1"}, lambda n, a: True, max_steps=3)
|
|
assert out["steps"] == 3 and "转人工" in out["answer"]
|