mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 05:21:57 +00:00
K3b：工具边界的遮罩分析出了意外时整份遮，兜底不再经逐格换值那一步
mask_query_result 的兜底原来又调一遍 _masked_view：出意外的正是它时兜底跟着出意外，结果交不出去，模块说明里「分析出意外就整份遮」也没有测试证明。改成 _masked_whole 直接把每一格换成「已遮罩」、类型记成文本，masked_at_query 记全部的列。补测试：认遮罩列、认标签列、逐格换值三处各自出意外时，交给模型的工具结果、落库的快照、通用工件接口都只有「已遮罩」，格子引用照实说是遮罩列；把兜底改回原样放行，这几项全部变红。
This commit is contained in:
ThearchyHelios
@@ -2020,8 +2020,9 @@ def mask_query_result(payload: dict[str, Any], masked: Iterable[str], kind: str
|
||||
- 列名不换（模型自己写的，换了它引用的列名就对不上）；结果里一列都没取自遮罩列、也没有可按值换的原样交回：数值的
|
||||
汇总、只数个数的查询不受影响
|
||||
|
||||
交回的 secrets 只有这份结果里的遮罩值（known 调用方自己有）。分析本身出了意外，整份结果的列都遮：这一步挡在模型
|
||||
前面，分析的缺陷不能把原值放过去（宁可多遮）。"""
|
||||
交回的 secrets 只有这份结果里的遮罩值（known 调用方自己有）。分析本身出了意外（认列、认标签、逐格换值哪一步都算），
|
||||
整份结果的每一格都遮、全部的列记成查询时遮掉的（_masked_whole）：这一步挡在模型前面，分析的缺陷不能把原值放过去
|
||||
（宁可多遮）。"""
|
||||
base = {str(c).strip().lower() for c in masked if str(c).strip()}
|
||||
columns = [str(c) for c in payload.get("columns") or []]
|
||||
given = known is not None and bool(known.strong)
|
||||
@@ -2046,11 +2047,25 @@ def mask_query_result(payload: dict[str, Any], masked: Iterable[str], kind: str
|
||||
view, _ = _masked_view(payload, hidden, pool, embedded=True, rename=False)
|
||||
except Exception: # noqa: BLE001 - 认不准就整份遮掉，原值不交给模型
|
||||
logger.warning("核查查询的遮罩分析出错，整份结果按遮罩处理", exc_info=True)
|
||||
hidden, own = {c.lower() for c in columns}, _Secrets()
|
||||
view, _ = _masked_view(payload, hidden, own, rename=False)
|
||||
return QueryMask(_masked_whole(payload), columns, _Secrets())
|
||||
return QueryMask(view, [c for c in columns if c.lower() in hidden], own)
|
||||
|
||||
|
||||
def _masked_whole(payload: dict[str, Any]) -> dict[str, Any]:
|
||||
"""整份结果的每一格都换成「已遮罩」、类型都记成文本：遮罩分析出了意外时的兜底。不经 _masked_view——出意外的可能
|
||||
就是它，兜底跟着出意外，原值就挡不住了。列名、SQL 是模型自己写的，照原样。"""
|
||||
word = _masked_word()
|
||||
rows: list[Any] = []
|
||||
for row in payload.get("rows") or []:
|
||||
if isinstance(row, list):
|
||||
rows.append([word] * len(row))
|
||||
elif isinstance(row, dict):
|
||||
rows.append(dict.fromkeys(row, word))
|
||||
else:
|
||||
rows.append(word)
|
||||
return {**payload, "rows": rows, "column_types": {str(c): "text" for c in payload.get("columns") or []}}
|
||||
|
||||
|
||||
#: 查询报错里引号括起来的一段（Postgres、DuckDB、MySQL 把转换失败的值写在引号里：invalid input syntax for type
|
||||
#: bigint: "139-0000-0001"）。引号前后紧挨着字母数字的不算（can't、isn't 里的撇号）
|
||||
_QUOTED_ERROR = re.compile(r"(?<![0-9A-Za-z_])(\"[^\"\n]*\"|'[^'\n]*')(?![0-9A-Za-z_])")
|
||||
|
||||
@@ -0,0 +1,92 @@
|
||||
"""工具边界先遮（engine/masking.py 的 mask_query_result）的收口：认不准就整份遮，多遮也不改写没设遮罩的数值证据。
|
||||
|
||||
- 遮罩分析本身出了意外（认遮罩列、标签列、逐格换值的哪一步都算）：整份结果的每一格都是「已遮罩」，交给模型的、落库的
|
||||
都一样，masked_at_query 记全部的列。兜底不再经出意外的那一步（以前兜底又调了一遍逐格换值的那一步，它出意外时兜底跟着
|
||||
出意外，结果没交出去，也没有测试证明「整份遮」这句保证）
|
||||
- 按值遮只换文字格：没设遮罩的数值列（访问次数正好等于某一笔被遮的金额）原样，存下的证据和数据库里的一样，双盲复算照常
|
||||
取得到；文字格里整格就是遮罩值、夹着遮罩值的照换，换过的格子记进快照（masked_cells），格子引用指到它们照实说是遮罩。
|
||||
占位的短词（未知、暂无……）不当遮罩值，没设遮罩的列、备注里的同一个词照原样
|
||||
|
||||
模型一律是 MockChatModel 剧本，数据源是现造的 SQLite 文件，手机号、姓名、金额、访问次数都是虚构的。
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
|
||||
import pytest
|
||||
from httpx import ASGITransport, AsyncClient
|
||||
|
||||
from app.core import artifact_store
|
||||
from app.data.engine import engines
|
||||
from app.engine import masking
|
||||
from app.main import app
|
||||
from tests.test_verify_agent import Script, q
|
||||
from tests.test_verify_agent_masking import ( # noqa: F401 - clean_spend 是自动用上的夹具
|
||||
MASKED,
|
||||
SALES,
|
||||
VALUE_SCHEMA,
|
||||
_cells,
|
||||
_clean,
|
||||
_source,
|
||||
ask,
|
||||
clean_spend,
|
||||
masked_shop,
|
||||
)
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
# 分析出意外：整份遮
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
|
||||
def _broken(*args, **kwargs):
|
||||
raise RuntimeError("遮罩分析里的意外")
|
||||
|
||||
|
||||
PAYLOAD = {"sql": "SELECT id, buyer_phone, note FROM sales", "columns": ["id", "buyer_phone", "note"],
|
||||
"rows": [[1, "13900000001", "客户来电"], [2, "13900000002", "例行回访"]],
|
||||
"column_types": {"id": "number", "buyer_phone": "text", "note": "text"}}
|
||||
|
||||
|
||||
@pytest.mark.parametrize("step", ["_derived_hidden", "_snapshot_labels", "_masked_view"])
|
||||
def test_a_failing_analysis_masks_every_cell(monkeypatch, step):
|
||||
"""认遮罩列、认标签列、逐格换值，哪一步出了意外都整份遮：每一格都是「已遮罩」、类型记成文本，整列遮掉的是全部的列。"""
|
||||
monkeypatch.setattr(masking, step, _broken)
|
||||
done = masking.mask_query_result(json.loads(json.dumps(PAYLOAD)), ["buyer_phone"], "sqlite")
|
||||
assert done.payload["rows"] == [[MASKED] * 3, [MASKED] * 3]
|
||||
assert done.payload["column_types"] == {"id": "text", "buyer_phone": "text", "note": "text"}
|
||||
assert done.columns == ["id", "buyer_phone", "note"]
|
||||
assert done.payload["sql"] == PAYLOAD["sql"], "SQL 是模型自己写的，照原样"
|
||||
assert not done.secrets.strong, "认不准，不收值"
|
||||
|
||||
|
||||
def test_a_failing_analysis_masks_rows_of_any_shape(monkeypatch):
|
||||
"""行是字典的结果同样整份遮。"""
|
||||
monkeypatch.setattr(masking, "_masked_view", _broken)
|
||||
payload = {"sql": "SELECT id, buyer_phone FROM sales", "columns": ["id", "buyer_phone"],
|
||||
"rows": [{"id": 1, "buyer_phone": "13900000001"}]}
|
||||
done = masking.mask_query_result(payload, ["buyer_phone"], "sqlite")
|
||||
assert done.payload["rows"] == [{"id": MASKED, "buyer_phone": MASKED}]
|
||||
|
||||
|
||||
@pytest.mark.parametrize("step", ["_derived_hidden", "_masked_view"])
|
||||
async def test_the_model_and_the_snapshot_see_only_masks_when_the_analysis_fails(monkeypatch, masked_shop, step):
|
||||
"""端到端：分析出了意外，交给模型的工具结果、落库的快照、通用接口给的都只有「已遮罩」，masked_at_query 是全部的列；
|
||||
格子引用照实说是遮罩列，不采信。"""
|
||||
monkeypatch.setattr(masking, step, _broken)
|
||||
name = masked_shop
|
||||
sql = "SELECT id, region, buyer_phone AS 手机号, amount FROM sales ORDER BY id"
|
||||
script = Script(monkeypatch, [q(name, sql, "c1")], {"value": {"alias": "Q1", "row": 0, "column": "amount"},
|
||||
"rationale": "x"})
|
||||
out, _ = await ask(name)
|
||||
seen = json.loads(script.tool_results()["c1"].split("\n", 1)[1])
|
||||
assert seen["rows"] == [[MASKED] * 4] * 4, seen
|
||||
snapshot = artifact_store.load(out.queries[0]["artifact"])
|
||||
assert snapshot["rows"] == [[MASKED] * 4] * 4
|
||||
assert snapshot["masked_at_query"] == ["id", "region", "手机号", "amount"]
|
||||
assert _clean(json.dumps(snapshot, ensure_ascii=False))
|
||||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as client:
|
||||
got = await client.get(f"/api/artifacts/{out.queries[0]['artifact']}")
|
||||
assert got.status_code == 200 and _clean(got.text) and "1200" not in got.text
|
||||
assert out.result["value"] is None
|
||||
assert out.cells[0]["reason"] == "Q1 的「amount」取自设置了遮罩的列，查询时已遮罩，不采信"
|
||||
Reference in New Issue
Block a user