mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-17 00:51:58 +00:00
099473cd6d
复核发现质询工件接口相对基线回退:agent 自己的查询快照在工具边界遮好之后,事后取质询工件收不到它另一次查询里的号码, 没设遮罩的备注、联系方式列原样给出。另有几处模型照样看得到遮罩值、或者遮多了的地方。 - 工具边界(tools/datasource、engine/masking.mask_query_result):此前已经认出的遮罩值(这个 agent 前面的查询遮掉的、 调用方交进来的)也用来按值遮后面的结果,数据源没设遮罩的也一样。审批质询、反方审稿在派 agent 之前把报告快照里的 遮罩值交进来(inquiry.snapshot_secrets,run_verify_agent 的 masked_values),模型从第一条查询起就看不到。已经摆在 读者眼前的值(提问人写的、报告指标、被审的结论原文)跟着并进来的值一起免遮(_Secrets.protected);这份结果自己写 的分档标签不按前面收的值换。 - 先查备注、后查出号码的:回答时只靠内存里的值才换掉的地方按快照记成位置(不含值,masking.view_edits),写进质询 事件的 masked_cells,质询工件接口先照着重放(replay_edits)再按原来的办法遮。工具边界按值换掉的整格照样画成遮罩列。 agent 查到的值先收、报告快照的后收:上限收满时不丢 agent 自己查到的。 - _derived_hidden 认 FROM、JOIN、LATERAL 里拆开遮罩列的表值函数(json_each、json_tree、unnest、jsonb_each_text、 regexp_split_to_table、JSON_TABLE、LATERAL VALUES):按出列名认,不认得的函数又没写列名表的整份遮。 - 汇总里的 CASE 只在条件里用到遮罩列的不算取自遮罩列(同写在 WHERE 里)。双盲复算按同一套认法认出指标取自遮罩列 (SUM(amount)、amount 设了遮罩)时,派盲算之前就记为无法复算并说清原因,不白花一次盲算(nodes/dual_check 的 _query_input,改动在文件所有权之外,只加这一处判断)。 - 反方审稿:理由正文和理由文档按值换掉遮罩值(换了写法的也认),这句结论原文里写着的不换;重新组装时引用到的格子按 遮好的快照渲染。 - 换了写法的号码另认不换行空格、窄空格、细空格、制表符、减号,最多四个分隔符;SQL 的字符串、带引号的名字、注释里 也认(不带引号的数照旧是算式),字符串里换掉的号码同一条 SQL 里写成阈值的也不留;列名同样认。 - 查询失败时报错原话里的值换掉(mask_error_text:引号里、不带引号的长数,SQL 里写着的照留)。 - 快照的 mask_columns 只记数据源设的列名,改了名字的结果列只在 masked_at_query:证据面板并各份快照的这一项从表的 字段清单里去列,混进别名会误藏同源别的表里同名的真实字段。 - * 展开数据源里的表时按冻结的表结构认它带不带同名列,没有的照常认分档标签(J2 收紧之后的多遮)。 - 格子引用指到按值换掉的格子时照实说;上万个值的合并、免遮放进线程。 测试:质询 agent 先查号码后查备注、先查备注后查号码、另一张表的号码列两种先后,质询工件、通用工件接口、事件都没有 原值;报告快照里的值在第一条查询就遮;表值函数(SQLite、Postgres、MySQL 的写法)、汇总里 CASE 的条件、* 连着的表按 表结构认标签、新的分隔符、SQL 字符串和注释和列名、报错原话、mask_columns 不混别名、遮罩列汇总的指标不派盲算、反方 理由按值遮且免遮结论原文、上限收满时 agent 的值先收。逐处去掉实现(24 处)对应的用例都会变红。
210 lines
12 KiB
Python
210 lines
12 KiB
Python
"""换了写法的号码(engine/masking.py 的 _Secrets):按值认得出的长数字(SECRET_DIGITS 位以上)被写成带空格、连字符、
|
||
点、括号、全角数字或全角标点、带 +86 / 0086 / 86 前缀的样子,模型写的正文(审批质询的回答)里照样认得出,换成
|
||
「已遮罩」;最后一道关(inquiry._leaked)、质询工件里文字格夹着的号码同样认得出。
|
||
|
||
验收准备时复现过:模型把号码写成 199-1474-2785、199 1474 2785、199.1474.2785、(199)1474-2785、+8619914742785、
|
||
全角数字,回答正文、事件、回答文档原样带着,最后一道关也拦不住,正式运行里永久封存。
|
||
|
||
认法是在正文里找允许夹分隔符的数字串,规范化(全角转半角、去分隔符、去国家码)后在「遮罩值的纯数字写法」集合里
|
||
查——集合查找,不现编正则。日期、小数、更长的数里的一截、连字符连着的编号照旧不动。
|
||
|
||
数据都是虚构的:手机号、金额。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import time
|
||
|
||
import pytest
|
||
|
||
from app.engine import inquiry as inquiry_mod
|
||
from app.engine.masking import _masked_view, _Secrets
|
||
|
||
MASKED = "已遮罩"
|
||
PHONE = "19914742785"
|
||
|
||
#: 同一个号码的各种写法(换了写法的号码)
|
||
REFORMATTED = [
|
||
pytest.param("199-1474-2785", id="连字符"),
|
||
pytest.param("199 1474 2785", id="空格"),
|
||
pytest.param("199.1474.2785", id="点"),
|
||
pytest.param("(199)1474-2785", id="括号"),
|
||
pytest.param("(199) 1474 2785", id="括号加空格"),
|
||
pytest.param("+8619914742785", id="+86"),
|
||
pytest.param("+86 199 1474 2785", id="+86 加空格"),
|
||
pytest.param("(+86)19914742785", id="(+86)"),
|
||
pytest.param("0086 199 1474 2785", id="0086"),
|
||
pytest.param("0086-19914742785", id="0086 连字符"),
|
||
pytest.param("86-199-1474-2785", id="86"),
|
||
pytest.param("8619914742785", id="86 连写"),
|
||
pytest.param("19914742785", id="全角数字"),
|
||
pytest.param("199-1474-2785", id="全角数字和连字符"),
|
||
pytest.param("199 1474 2785", id="全角空格"),
|
||
pytest.param("199.1474.2785", id="全角点"),
|
||
pytest.param("(199)1474-2785", id="全角括号"),
|
||
pytest.param("+86 199 1474 2785", id="全角加号"),
|
||
]
|
||
|
||
|
||
def secrets_of(values: list, *, hidden: str = "phone") -> _Secrets:
|
||
secrets = _Secrets()
|
||
secrets.add({"columns": [hidden], "rows": [[v] for v in values]}, {hidden})
|
||
return secrets
|
||
|
||
|
||
@pytest.mark.parametrize("written", REFORMATTED)
|
||
@pytest.mark.parametrize("stored", [PHONE, int(PHONE), "199-1474-2785"], ids=["文本", "整数", "存的就带连字符"])
|
||
def test_a_long_number_in_another_shape_is_masked_in_prose(stored, written):
|
||
secrets = secrets_of([stored])
|
||
assert secrets.scrub(f"金额最高的一笔来自 {written},共 3 笔", prose=True) == f"金额最高的一笔来自 {MASKED},共 3 笔"
|
||
assert secrets.scrub(f"客户{written}来电", prose=True) == f"客户{MASKED}来电"
|
||
|
||
|
||
@pytest.mark.parametrize("text, shown", [
|
||
("来自(19914742785),共 3 笔", f"来自({MASKED}),共 3 笔"),
|
||
("来自(+86 199 1474 2785),共 3 笔", f"来自({MASKED}),共 3 笔"),
|
||
("电话(+86)199 1474 2785。", f"电话{MASKED}。"),
|
||
("(电话 199-1474-2785)", f"(电话 {MASKED})"),
|
||
])
|
||
def test_brackets_around_a_number_are_kept_and_brackets_inside_it_are_replaced(text, shown):
|
||
"""整个号码外面套着的一对括号是标点,留着;号码里的括号(区号、国家码)和号码一起换掉。"""
|
||
assert secrets_of([PHONE]).scrub(text, prose=True) == shown
|
||
|
||
|
||
@pytest.mark.parametrize("text", [
|
||
# 日期、时间、小数、千分位、更长的数里的一截、连字符连着的编号、别的号码
|
||
"2026-10-06 19:14 下单", "金额 19914.74 元", "共 1,991,474 笔", "号码 1991474278512", "批次 19914742785-2",
|
||
"199-1474-2786", "单号 2026199-1474-2785", "共 199 笔 1474 元 2785 次",
|
||
])
|
||
def test_other_numbers_and_number_like_text_are_left_alone(text):
|
||
secrets = secrets_of([PHONE])
|
||
assert secrets.scrub(text, prose=True) == text
|
||
|
||
|
||
def test_reformatted_numbers_are_only_read_in_prose_and_text_cells():
|
||
"""SQL 里 199-1474-2785 是一个算式,不当成号码换(SQL 里的号码按字面量的用法换,见 inquiry._scrub_sql)。"""
|
||
secrets = secrets_of([PHONE])
|
||
assert secrets.scrub("SELECT 199-1474-2785 AS x") == "SELECT 199-1474-2785 AS x"
|
||
assert secrets.scrub("备注 199-1474-2785", loose=True) == f"备注 {MASKED}"
|
||
|
||
|
||
def test_a_text_cell_holding_a_reformatted_number_is_masked_whole_or_in_part():
|
||
"""快照里别的列(没设遮罩)写着换了写法的号码:整格就是号码的整格遮;质询工件里夹在一段文字里的只换那一截。"""
|
||
secrets = secrets_of([PHONE])
|
||
snapshot = {"columns": ["联系方式", "备注"], "rows": [["199-1474-2785", "客户 (199)1474-2785 来电改址"],
|
||
["138-0000-0000", "例行回访"]]}
|
||
view, masked = _masked_view(snapshot, set(), secrets)
|
||
assert view["rows"] == [[MASKED, "客户 (199)1474-2785 来电改址"], ["138-0000-0000", "例行回访"]]
|
||
assert masked == ["联系方式"]
|
||
view, masked = _masked_view(snapshot, set(), secrets, embedded=True)
|
||
assert view["rows"] == [[MASKED, f"客户 {MASKED} 来电改址"], ["138-0000-0000", "例行回访"]]
|
||
assert masked == ["联系方式"]
|
||
|
||
|
||
@pytest.mark.parametrize("written", REFORMATTED)
|
||
def test_the_last_gate_stops_an_answer_still_showing_a_reformatted_number(written):
|
||
"""组装之后还认得出(渲染进单元格引用的一格文字里夹着换了写法的号码):不交出去。正文里的在组装之前就换掉了。"""
|
||
secrets = secrets_of([PHONE])
|
||
entry = {"alias": "Q1", "kind": "query", "artifact": "a1", "source": "shop", "columns": ["备注"], "rows": 1,
|
||
"truncated": False}
|
||
views = {"a1": _masked_view({"columns": ["备注"], "rows": [[f"客户{written}来电改址"]]}, set(), secrets)[0]}
|
||
with pytest.raises(inquiry_mod._Leak):
|
||
inquiry_mod._compose("回访备注是 [[v:Q1.r0.备注]]。", {"Q1": entry}, run_id="r1", views=views, secrets=secrets,
|
||
allow=[])
|
||
doc = inquiry_mod._compose(f"来电的是 {written}。", {}, run_id="r1", views={}, secrets=secrets, allow=[])
|
||
assert doc["markdown"] == f"来电的是 {MASKED}。"
|
||
assert inquiry_mod._leaked({"markdown": f"来电的是 {written}。", "blocks": []}, secrets)
|
||
|
||
|
||
def test_a_number_the_asker_wrote_in_another_shape_is_left_alone():
|
||
"""提问人自己写的号码(换了写法也一样)已经摆在读者眼前:回答里照原样,别的号码照遮。"""
|
||
secrets = secrets_of([PHONE, "13800000000"])
|
||
secrets.protect(["199-1474-2785 这位客户还买过别的吗?"])
|
||
assert secrets.scrub("199 1474 2785 另有 138-0000-0000 的单", prose=True) == f"199 1474 2785 另有 {MASKED} 的单"
|
||
|
||
|
||
def test_a_threshold_left_in_the_sql_is_shown_in_the_prose_whatever_the_shape():
|
||
"""SQL 里照原样留着的阈值,回答正文里复述不遮,换了写法也一样(只免正文)。"""
|
||
secrets = secrets_of([20000, 35000], hidden="amount")
|
||
secrets.show({"20000"})
|
||
assert secrets.scrub("金额不低于 20 000 元,最高一笔 35 000 元", prose=True) == f"金额不低于 20 000 元,最高一笔 {MASKED} 元"
|
||
|
||
|
||
def test_many_masked_numbers_are_matched_without_stalling():
|
||
"""遮罩值上万个:集合查找,不现编上万个分支的正则。"""
|
||
secrets = secrets_of([str(19900000000 + i) for i in range(20_000)])
|
||
text = ";".join(f"客户 199-0000-{i:04d} 来电" for i in range(0, 4000, 7)) * 3
|
||
began = time.perf_counter()
|
||
out = secrets.scrub(text, prose=True)
|
||
assert time.perf_counter() - began < 2.0
|
||
assert "199-0000" not in out and out.count(MASKED) == text.count("199-0000")
|
||
|
||
|
||
#: 复核第 1 轮:别的空白、减号作分隔符的写法(模型输出里不换行空格、窄空格并不少见)
|
||
OTHER_SEPARATORS = [
|
||
pytest.param("199 1474 2785", id="不换行空格"),
|
||
pytest.param("199 1474 2785", id="窄不换行空格"),
|
||
pytest.param("199 1474 2785", id="细空格"),
|
||
pytest.param("199 1474 2785", id="发丝空格"),
|
||
pytest.param("199\t1474\t2785", id="制表符"),
|
||
pytest.param("199−1474−2785", id="减号"),
|
||
pytest.param("199 1474 2785", id="四个空格"),
|
||
pytest.param("199 - 1474 - 2785", id="连字符两边带空格"),
|
||
pytest.param("+86 199 1474 2785", id="+86 加不换行空格"),
|
||
]
|
||
|
||
|
||
@pytest.mark.parametrize("written", OTHER_SEPARATORS)
|
||
def test_numbers_split_by_other_spaces_and_minus_signs_are_masked_too(written):
|
||
secrets = secrets_of([PHONE])
|
||
assert secrets.scrub(f"客户 {written} 来电", prose=True) == f"客户 {MASKED} 来电"
|
||
assert inquiry_mod._leaked({"markdown": f"来电的是 {written}。", "blocks": []}, secrets)
|
||
assert secrets.hit(written), "整格就是这个号码的,格子按值认"
|
||
|
||
|
||
def test_a_line_break_still_ends_a_number():
|
||
"""不跨行:两行各写一截的不拼成一个号码。"""
|
||
secrets = secrets_of([PHONE])
|
||
assert secrets.scrub("编号 199\n1474 2785", prose=True) == "编号 199\n1474 2785"
|
||
|
||
|
||
@pytest.mark.parametrize("sql, sealed", [
|
||
pytest.param("SELECT COUNT(*) AS 笔数 FROM callbacks WHERE note LIKE '%199-1474-2785%'",
|
||
f"SELECT COUNT(*) AS 笔数 FROM callbacks WHERE note LIKE '%{MASKED}%'", id="LIKE 里的一截"),
|
||
pytest.param("SELECT COUNT(*) AS 笔数 FROM callbacks WHERE note = '客户 199 1474 2785 来电'",
|
||
f"SELECT COUNT(*) AS 笔数 FROM callbacks WHERE note = '客户 {MASKED} 来电'", id="字符串里夹着"),
|
||
pytest.param('SELECT note AS "199-1474-2785 的备注" FROM callbacks',
|
||
f'SELECT note AS "{MASKED} 的备注" FROM callbacks', id="带引号的列名"),
|
||
pytest.param("SELECT COUNT(*) AS 笔数 FROM callbacks /* 客户 199-1474-2785 */",
|
||
f"SELECT COUNT(*) AS 笔数 FROM callbacks /* 客户 {MASKED} */", id="块注释"),
|
||
pytest.param("SELECT COUNT(*) AS 笔数 FROM callbacks -- 客户 199 1474 2785\n WHERE id = 1",
|
||
f"SELECT COUNT(*) AS 笔数 FROM callbacks -- 客户 {MASKED}\n WHERE id = 1", id="行注释"),
|
||
])
|
||
def test_a_reformatted_number_in_sql_strings_names_and_comments_is_masked(sql, sealed):
|
||
"""SQL 里字符串、带引号的名字、注释里换了写法的号码也认(以前只有整个字面量就是号码的才换)。"""
|
||
assert inquiry_mod._scrub_sql(sql, secrets_of([PHONE]), set(), "sqlite") == sealed
|
||
|
||
|
||
def test_a_bare_number_expression_in_sql_is_still_left_alone():
|
||
"""不带引号的 199-1474-2785 在 SQL 里是一个算式(三个数相减),照旧不按号码认。"""
|
||
sql = "SELECT 199-1474-2785 AS x FROM callbacks"
|
||
assert inquiry_mod._scrub_sql(sql, secrets_of([PHONE]), set(), "sqlite") == sql
|
||
|
||
|
||
def test_a_threshold_matching_a_number_swapped_in_a_string_is_not_kept():
|
||
"""字符串里按换了写法认出、换掉的号码,同一条 SQL 里写成阈值的同一个数也不留(不然阈值把它露了)。"""
|
||
sql = f"SELECT COUNT(*) AS 笔数 FROM callbacks WHERE note LIKE '%199-1474-2785%' AND phone_no >= {PHONE}"
|
||
sealed = inquiry_mod._scrub_sql(sql, secrets_of([PHONE]), set(), "sqlite")
|
||
assert PHONE not in sealed and "1474" not in sealed
|
||
|
||
|
||
def test_a_reformatted_number_in_a_column_name_is_masked_like_the_sql():
|
||
"""agent 把换了写法的号码写进列名(AS "139-0000-0001 的单"):质询工件的列名、回答里引用这一列的标记照 SQL 里带引号
|
||
的名字一样换,三处对得上。"""
|
||
secrets = secrets_of(["13900000001"])
|
||
snapshot = {"columns": ["139-0000-0001 的单"], "rows": [[3]]}
|
||
view, _ = _masked_view(snapshot, set(), secrets)
|
||
assert view["columns"] == [f"{MASKED} 的单"]
|
||
assert inquiry_mod._scrub_sql('SELECT COUNT(*) AS "139-0000-0001 的单" FROM sales', secrets, set(), "sqlite") == (
|
||
f'SELECT COUNT(*) AS "{MASKED} 的单" FROM sales')
|
||
assert inquiry_mod._scrubbed_source("共 [[v:Q1.r0.139-0000-0001 的单]] 单。", secrets) == f"共 [[v:Q1.r0.{MASKED} 的单]] 单。"
|