mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-20 15:02:00 +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。
110 lines
7.2 KiB
Python
110 lines
7.2 KiB
Python
"""按 SQL 认取自遮罩列的结果列(engine/masking.py 的 _derived_hidden):FROM、JOIN、LATERAL 里的表值函数,汇总里的 CASE。
|
||
|
||
核查 agent 的查询在工具边界按这套认法整列遮掉(mask_query_result),审批质询、双盲复算也用它:
|
||
|
||
- 表值函数(SQLite 的 json_each / json_tree,Postgres 的 unnest、jsonb_each_text、regexp_split_to_table、LATERAL
|
||
VALUES,MySQL 的 JSON_TABLE)把遮罩列拆开时,它出的列都算取自遮罩列:遮罩列是 JSON 时,用 json_each 拆开是最自然的
|
||
写法,以前只追 SELECT 项、子查询和 *,这些写法的原值原样交给模型、落进快照。出列名认得全的(常见的几个函数、写了
|
||
列名表的)按名字认,认不全的整份遮
|
||
- 汇总里的 CASE / IIF 只在条件里用到遮罩列的(SUM(CASE WHEN buyer_phone IS NOT NULL THEN amount ELSE 0 END)),同
|
||
写在 WHERE 里,不算;结果的某一支取的是遮罩列、不汇总的逐行 CASE 照旧算
|
||
|
||
数据都是虚构的:手机号、区域、金额。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import pytest
|
||
|
||
from app.engine import masking
|
||
|
||
#: 设了遮罩的列:手机号,和装着手机号、姓名的 JSON 列
|
||
MASKED = {"buyer_phone", "profile"}
|
||
SALES = ["id", "region", "buyer_phone", "amount", "profile"]
|
||
#: SQLite 的 json_each / json_tree 出的列
|
||
JSON_EACH = ["key", "value", "type", "atom", "id", "parent", "fullkey", "path"]
|
||
|
||
|
||
def hidden_of(sql: str, columns: list[str], kind: str = "sqlite", masked: set[str] = MASKED) -> set[str]:
|
||
snapshot = {"source": "shop", "sql": sql, "columns": columns, "rows": []}
|
||
return masking._derived_hidden(snapshot, set(masked), kind)[0]
|
||
|
||
|
||
@pytest.mark.parametrize("kind, sql, columns, shown", [
|
||
pytest.param("sqlite", "SELECT j.key, j.value FROM sales, json_each(sales.profile) AS j", ["key", "value"],
|
||
["key", "value"], id="json_each 拆遮罩的 JSON 列"),
|
||
pytest.param("sqlite", "SELECT t.atom AS 值 FROM sales, json_tree(sales.profile) AS t WHERE t.type = 'text'",
|
||
["值"], ["值"], id="json_tree"),
|
||
pytest.param("sqlite", "SELECT value FROM sales, json_each(json_array(buyer_phone))", ["value"], ["value"],
|
||
id="json_each 套 json_array"),
|
||
pytest.param("sqlite", "SELECT r.value AS 号码 FROM sales s JOIN json_each(s.profile) r ON 1 = 1", ["号码"],
|
||
["号码"], id="JOIN 里的 json_each"),
|
||
pytest.param("sqlite", "SELECT value FROM json_each((SELECT json_group_array(buyer_phone) FROM sales))",
|
||
["value"], ["value"], id="参数是子查询"),
|
||
pytest.param("sqlite", "SELECT c.value AS 号码 FROM (SELECT json_group_array(buyer_phone) AS arr FROM sales) t, "
|
||
"json_each(t.arr) c", ["号码"], ["号码"], id="参数取自派生来源里改了名字的遮罩列"),
|
||
pytest.param("sqlite", "SELECT * FROM sales, json_each(sales.profile)", [*SALES, *JSON_EACH],
|
||
["buyer_phone", "profile", "key", "value", "atom", "fullkey", "path"], id="* 带出 json_each 的列"),
|
||
pytest.param("sqlite", "SELECT x FROM (SELECT j.value AS x FROM sales, json_each(sales.profile) j) t", ["x"],
|
||
["x"], id="子查询里拆开再改名"),
|
||
pytest.param("postgres", "SELECT u FROM sales, unnest(ARRAY[buyer_phone]) AS u", ["u"], ["u"], id="unnest"),
|
||
pytest.param("postgres", "SELECT x.p FROM sales s CROSS JOIN LATERAL unnest(ARRAY[s.buyer_phone]) AS x(p)",
|
||
["p"], ["p"], id="LATERAL unnest 带列名表"),
|
||
pytest.param("postgres", "SELECT x FROM sales, LATERAL unnest(ARRAY[buyer_phone]) AS x", ["x"], ["x"],
|
||
id="LATERAL unnest"),
|
||
pytest.param("postgres", "SELECT u FROM sales CROSS JOIN LATERAL (VALUES (buyer_phone)) AS v(u)", ["u"], ["u"],
|
||
id="LATERAL VALUES"),
|
||
pytest.param("postgres", "SELECT d FROM sales, regexp_split_to_table(buyer_phone, '-') AS d", ["d"], ["d"],
|
||
id="regexp_split_to_table"),
|
||
pytest.param("postgres", "SELECT e.key, e.value FROM sales, jsonb_each_text(profile) AS e", ["key", "value"],
|
||
["key", "value"], id="jsonb_each_text"),
|
||
pytest.param("postgres", "SELECT x FROM sales, unnest(string_to_array(buyer_phone, ',')) AS x", ["x"], ["x"],
|
||
id="unnest 套 string_to_array"),
|
||
pytest.param("postgres", "SELECT * FROM sales, unnest(ARRAY[buyer_phone])", [*SALES, "unnest"],
|
||
["buyer_phone", "profile", "unnest"], id="没起别名的 unnest 按函数名出列"),
|
||
pytest.param("mysql", "SELECT j.p FROM sales s, JSON_TABLE(JSON_ARRAY(s.buyer_phone), '$[*]' "
|
||
"COLUMNS (p VARCHAR(20) PATH '$')) j", ["p"], ["p"], id="JSON_TABLE"),
|
||
])
|
||
def test_a_table_function_over_a_masked_column_hides_what_it_returns(kind, sql, columns, shown):
|
||
assert {c.lower() for c in shown} <= hidden_of(sql, columns, kind)
|
||
|
||
|
||
def test_columns_a_table_function_does_not_return_stay_visible():
|
||
"""拆的是遮罩列,别的列(区域、金额)照常;拆的不是遮罩列的,一列都不遮。"""
|
||
hidden = hidden_of("SELECT s.region, s.amount, j.value AS 值 FROM sales s, json_each(s.profile) j",
|
||
["region", "amount", "值"])
|
||
assert "值" in hidden and not {"region", "amount"} & hidden
|
||
assert hidden_of("SELECT s.region, j.value AS 标签 FROM sales s, json_each(s.tags) j", ["region", "标签"]) == MASKED
|
||
|
||
|
||
def test_a_table_function_that_cannot_be_read_hides_the_whole_snapshot():
|
||
"""不认得的函数拆遮罩列、又没写列名表:说不清出哪几列,整份快照的列都遮(同 SQL 解析不了)。"""
|
||
columns = ["region", "piece"]
|
||
assert hidden_of("SELECT s.region, t.piece FROM sales s, my_split(s.buyer_phone) t", columns, "postgres") == {
|
||
*MASKED, "region", "piece"}
|
||
|
||
|
||
@pytest.mark.parametrize("sql", [
|
||
"SELECT SUM(CASE WHEN buyer_phone IS NOT NULL THEN amount ELSE 0 END) AS total FROM sales",
|
||
"SELECT SUM(CASE buyer_phone WHEN '13900000001' THEN amount END) AS total FROM sales",
|
||
"SELECT AVG(IIF(buyer_phone LIKE '139%', amount, NULL)) AS total FROM sales",
|
||
"SELECT MAX(CASE WHEN buyer_phone IS NULL THEN 0 ELSE amount END) AS total FROM sales",
|
||
"SELECT SUM(CASE WHEN buyer_phone IS NULL THEN 0 ELSE CASE WHEN region = '华东' THEN amount END END) AS total "
|
||
"FROM sales",
|
||
])
|
||
def test_a_masked_column_only_in_the_condition_of_an_aggregated_case_is_like_a_filter(sql):
|
||
"""值全部来自金额,手机号只决定这一行算不算进来:同写在 WHERE 里,不算取自遮罩列。"""
|
||
assert hidden_of(sql, ["total"]) == MASKED
|
||
assert hidden_of(sql.replace("CASE WHEN buyer_phone IS NOT NULL THEN amount ELSE 0 END", "amount")
|
||
+ " WHERE buyer_phone IS NOT NULL", ["total"]) == MASKED
|
||
|
||
|
||
@pytest.mark.parametrize("sql", [
|
||
# 结果的某一支取的就是遮罩列
|
||
"SELECT SUM(CASE WHEN region = '华东' THEN buyer_phone END) AS total FROM sales",
|
||
"SELECT MAX(IIF(amount > 1000, buyer_phone, NULL)) AS total FROM sales",
|
||
# 不汇总的逐行 CASE:每一行各自落在哪一支
|
||
"SELECT CASE WHEN buyer_phone IS NULL THEN 0 ELSE amount END AS total FROM sales",
|
||
])
|
||
def test_a_case_returning_the_masked_value_or_not_aggregated_is_still_derived(sql):
|
||
assert "total" in hidden_of(sql, ["total"])
|