mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 19:51:59 +00:00
ThearchyHelios
e4f8ee5d08
推断的来源（新 engine/direct_select、data/provenance、data/provenance_db；api/evidence 新增推断来源接口） - 只对直接引用一列的查询下钻：受限文法用和执行时相同的 SQLite 分词读法，别名互换、自连接、表达式、聚合、子查询、集合运算（含藏在注释、字符串、方括号、转义、BOM 之后的）一律只给表级来历；再用 EXPLAIN 字节码和授权回调兜底 - 结果带齐主键、值非空，按主键参数化回查只读快照、类型和值严格一致才给格子；不一致时整份重算库哈希，哈希不符标红 - 从导入清单和溯源找回格子：宽表、交叉表逆透视的长表（日期取自表头格、年份取自统计期、行标签、分段标题）、列表、合计表、按期累积的并集（换算到所属那一期）；原文与规范写法不同时一并列出 - 清单读不到、登记哈希与清单不一致、快照已回收、遮罩列，都只给表级来历或不给，写明原因；原件已清除、接受已作废照实显示；未封存的运行照常推断并标明 - 相关核对：这一行、这一期的核对结论与接受理由 裁判摘录（engine/judge） - 带期 4 标记的报告，上传表格的查询旁加来源行、口径行、接受理由、可见单元格里的区域外文字和人工录入的统计期；限定语写明这些数未经系统核对，长度有上限、截断时写明 - 隐藏行列里的区域外文字不送；设了遮罩的源只写条数 前端证据面板（run/EvidenceProvenance；ManifestBody 从版本页抽出复用） - 「推断的来源」一节与已封存的证据分开标注：格子坐标、取自哪些格、年份来源、规范写法、回查、相关核对、数据版本与各期；标红与拒绝原因按状态显示 不变的部分 - 文档顶层加 provenance 标记，只有新报告走新路径；期 4 之前的运行、非上传源、手工源的面板、复核、裁判摘录逐字不变，由金样测试钉住 - 执行器给区域外文字记录是否在隐藏行列；不建「导入校验」系统表（D6）
285 lines
14 KiB
Python
285 lines
14 KiB
Python
"""裁判摘录的期 4 来历行（P4-SPEC 3.2–3.4）：按文档标记分版本、只对上传表格。
|
||
|
||
- 期 4 之前的文档（没有标记）：摘录逐字等于金样（tests/fixtures/evidence/pre_p4_golden.json），provenance、
|
||
direct_select 两个模块一次都不碰（用桩钉住：桩一被访问就抛错）；
|
||
- 带标记的文档：手工源的摘录不变；上传源的查询摘录只在 SQL 那一行之后、「列：」之前多出 judge_lines 的行，
|
||
去掉这些行就等于金样；同样的输入两次得到同样的字；
|
||
- 按需裁判的 loader：带标记的文档才取得到哈希链上的清单，老文档连坏掉的清单都碰不到（A26）；
|
||
- 真导入的 D28（表下有含数字的说明）：区域外文字带限定语、带统计期；遮罩时不送；单句的点击估价在上限以内。
|
||
|
||
金样的工件逐个 put_json 进临时数据目录，先断言返回的 id 等于金样记下的 id（内容寻址，对不上说明金样被改过）。
|
||
夹具全部合成、假名；不调用任何模型（只比摘录文字和估价）。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import copy
|
||
import json
|
||
import sys
|
||
import types
|
||
from pathlib import Path
|
||
|
||
import pytest
|
||
from httpx import ASGITransport, AsyncClient
|
||
|
||
from app.api import evidence as api
|
||
from app.core import artifact_store
|
||
from app.core.config import settings
|
||
from app.data import provenance as prov
|
||
from app.data.provenance_types import DOC_PROVENANCE, PROV_OUTSIDE_CHARS
|
||
from app.data.tabular import UNSHAPED_NOTE
|
||
from app.engine import judge
|
||
from app.engine.direct_select import tables_in
|
||
from app.main import app
|
||
from tests.test_provenance_api import import_flow, run_sql, sealed_of
|
||
|
||
GOLDEN = Path(__file__).parent / "fixtures" / "evidence" / "pre_p4_golden.json"
|
||
#: 3.2 区域外文字的小标题（限定语每块只写一次）
|
||
OUTSIDE_HEAD = ("区域外文字（原表区域外的文字原文，未经系统核对；其中的数字不是查询结果，不能用来支持或否定报告中的"
|
||
"数字）：")
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 夹具
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
@pytest.fixture
|
||
async def golden(monkeypatch, tmp_path) -> dict:
|
||
"""金样的输入工件写进一个临时数据目录（id 必须等于金样记下的），返回金样本身。"""
|
||
data = tmp_path / "data"
|
||
(data / "uploads").mkdir(parents=True)
|
||
monkeypatch.setattr(settings, "data_dir", data)
|
||
g = json.loads(GOLDEN.read_text("utf-8"))
|
||
for aid, content in g["artifacts"].items():
|
||
assert await artifact_store.put_json(content, kind="golden") == aid
|
||
return g
|
||
|
||
|
||
class _Forbidden(types.ModuleType):
|
||
"""桩模块：任何属性一被访问就记下来并抛错。期 4 之前的文档、手工源的摘录不该碰 provenance、direct_select。"""
|
||
|
||
def __init__(self, name: str, touched: list[str]) -> None:
|
||
super().__init__(name)
|
||
self._touched = touched
|
||
|
||
def __getattr__(self, attr: str):
|
||
if attr.startswith("__"):
|
||
raise AttributeError(attr)
|
||
self._touched.append(f"{self.__name__}.{attr}")
|
||
raise AssertionError(f"不该用到 {self.__name__}.{attr}")
|
||
|
||
|
||
@pytest.fixture
|
||
def forbidden(monkeypatch) -> list[str]:
|
||
touched: list[str] = []
|
||
for name in ("app.data.provenance", "app.engine.direct_select"):
|
||
monkeypatch.setitem(sys.modules, name, _Forbidden(name, touched))
|
||
return touched
|
||
|
||
|
||
def doc_of(g: dict, kind: str, *, marked: bool) -> tuple[dict, dict]:
|
||
report = g["reports"][kind]
|
||
doc = copy.deepcopy(report["outputs"]["doc"])
|
||
assert "provenance" not in doc
|
||
if marked:
|
||
doc["provenance"] = DOC_PROVENANCE
|
||
return doc, copy.deepcopy(report["catalog"])
|
||
|
||
|
||
def excerpts(doc: dict, catalog: dict, **kw) -> dict[str, str]:
|
||
return judge.prepare(doc, catalog, loader=kw.pop("loader", artifact_store.load), **kw).excerpts
|
||
|
||
|
||
def inserted(new: str, old: str) -> list[str]:
|
||
"""新摘录比金样多出的行：只许插在第 2 行（SQL）之后、「列：」之前，其余逐字相同。"""
|
||
a, b = new.split("\n"), old.split("\n")
|
||
extra = len(a) - len(b)
|
||
assert extra >= 0 and a[:2] == b[:2] and a[2 + extra:] == b[2:], (new, old)
|
||
assert b[2].startswith("列：")
|
||
return a[2:2 + extra]
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 按文档标记分版本（3.3）
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
@pytest.mark.parametrize("kind", ["recipe", "simple", "manual"])
|
||
async def test_unmarked_documents_are_unchanged_and_never_touch_the_new_modules(golden, forbidden, kind):
|
||
doc, catalog = doc_of(golden, kind, marked=False)
|
||
assert excerpts(doc, catalog) == golden["reports"][kind]["outputs"]["excerpts"]
|
||
assert forbidden == []
|
||
|
||
|
||
async def test_marked_manual_source_is_unchanged(golden, forbidden):
|
||
"""非上传源（查询快照没有 data_version）：带标记也一字不差，两个模块照样不碰。"""
|
||
doc, catalog = doc_of(golden, "manual", marked=True)
|
||
assert excerpts(doc, catalog) == golden["reports"]["manual"]["outputs"]["excerpts"]
|
||
assert forbidden == []
|
||
|
||
|
||
async def test_marked_recipe_source_adds_lines_only_between_sql_and_columns(golden):
|
||
doc, catalog = doc_of(golden, "recipe", marked=True)
|
||
new, old = excerpts(doc, catalog), golden["reports"]["recipe"]["outputs"]["excerpts"]
|
||
assert set(new) == set(old)
|
||
# 表名实体的摘录不加来历（只改查询摘录）
|
||
assert new["t:日客流"] == old["t:日客流"]
|
||
for alias, table in (("Q1", "日客流"), ("Q2", "时段客流")):
|
||
lines = inserted(new[alias], old[alias])
|
||
snap = artifact_store.load(catalog[alias]["artifact"])
|
||
schema = artifact_store.load(snap["schema_artifact"])
|
||
tables = {t: [c["name"] for c in meta["columns"]] for t, meta in schema["tables"].items()}
|
||
assert tables_in(snap["sql"], tables) == [table]
|
||
assert lines == prov.judge_lines(snap, schema, artifact_store.load, tables=[table], hidden=set())
|
||
assert lines[0].startswith("来源：上传的表格，每期替换，第 1 次导入：统计期 2026-08-01 至 2026-08-31")
|
||
assert f"口径：表 {table} 的说明：" in "\n".join(lines)
|
||
|
||
|
||
async def test_marked_simple_upload_adds_only_caliber_lines(golden):
|
||
"""简单导入（没有导入清单）只多口径行（调整 6）：未规整的表写 UNSHAPED_NOTE；没有来源、接受、区域外文字。"""
|
||
doc, catalog = doc_of(golden, "simple", marked=True)
|
||
new, old = excerpts(doc, catalog), golden["reports"]["simple"]["outputs"]["excerpts"]
|
||
lines = inserted(new["Q1"], old["Q1"])
|
||
assert lines and all(line.startswith("口径：") for line in lines)
|
||
assert f"口径：表 月报 的说明：{UNSHAPED_NOTE}" in lines
|
||
assert new["t:月报"] == old["t:月报"]
|
||
|
||
|
||
async def test_excerpts_are_deterministic(golden):
|
||
"""摘录进断点指纹和逐句复用键：同样的输入两次得到同样的字（不取当前时间、不取数据库列）。"""
|
||
doc, catalog = doc_of(golden, "recipe", marked=True)
|
||
assert excerpts(doc, catalog) == excerpts(copy.deepcopy(doc), copy.deepcopy(catalog))
|
||
|
||
|
||
async def test_masks_and_tables_are_handed_to_judge_lines(golden, monkeypatch):
|
||
"""hidden 就是查询摘录现成的那份遮罩（数据源现在设的 ∪ 快照记下的，小写）；tables 由 tables_in 按冻结表结构算。"""
|
||
calls: list[dict] = []
|
||
real = prov.judge_lines
|
||
|
||
def spy(query, schema, load, *, tables, hidden):
|
||
calls.append({"sql": query["sql"], "tables": tables, "hidden": set(hidden), "schema": schema is not None})
|
||
return real(query, schema, load, tables=tables, hidden=hidden)
|
||
|
||
monkeypatch.setattr(prov, "judge_lines", spy)
|
||
doc, catalog = doc_of(golden, "recipe", marked=True)
|
||
out = excerpts(doc, catalog, masked={"flow_demo": ["分区甲"], "别的源": ["日期"]})
|
||
assert [(c["tables"], c["hidden"], c["schema"]) for c in calls] == [
|
||
(["日客流"], {"分区甲"}, True), (["时段客流"], {"分区甲"}, True)]
|
||
# 遮罩的列：结果里不给，列的说明也不写
|
||
assert "口径：列 分区甲 的说明" not in out["Q1"] and "分区甲" not in out["Q1"].split("\n")[-1]
|
||
|
||
|
||
async def test_a_broken_artifact_degrades_to_one_line(golden, monkeypatch):
|
||
"""生成来历行时出了意外（工件形状不对）：这条查询只写「导入清单无法读取…」一行，整份报告照样能判。"""
|
||
def boom(*_a, **_kw):
|
||
raise TypeError("坏形状")
|
||
|
||
monkeypatch.setattr(prov, "judge_lines", boom)
|
||
doc, catalog = doc_of(golden, "recipe", marked=True)
|
||
new, old = excerpts(doc, catalog), golden["reports"]["recipe"]["outputs"]["excerpts"]
|
||
assert inserted(new["Q1"], old["Q1"]) == [prov.SOURCE_BROKEN]
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 按需裁判的取证范围（3.4、2.8.3）
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def golden_sealed(g: dict, kind: str) -> api._Sealed:
|
||
info = g["reports"][kind]["sealed"]
|
||
sealed = sealed_of(queries=list(info["queries"]), schemas=list(info["schemas"]))
|
||
sealed.queries.update(info["queries"])
|
||
return sealed
|
||
|
||
|
||
async def test_on_demand_loader_reaches_the_chain_only_for_marked_docs(golden):
|
||
doc, catalog = doc_of(golden, "recipe", marked=True)
|
||
sealed = golden_sealed(golden, "recipe")
|
||
old = golden["reports"]["recipe"]["outputs"]["excerpts"]
|
||
# 带标记：loader 扩展到表结构快照经哈希链列出的导入清单，来源行照常写
|
||
chained = excerpts(doc, catalog, loader=api._sealed_loader(sealed, chain=True))
|
||
assert inserted(chained["Q1"], old["Q1"])[0].startswith("来源：上传的表格，每期替换")
|
||
assert chained == excerpts(doc, catalog)
|
||
# 不扩展的 loader 取不到清单：只写「导入清单无法读取…」一行，口径照写（来自封存范围内的表结构快照）
|
||
plain = excerpts(doc, catalog, loader=api._sealed_loader(sealed))
|
||
lines = inserted(plain["Q1"], old["Q1"])
|
||
assert lines[0] == prov.SOURCE_BROKEN and any(line.startswith("口径：表 日客流") for line in lines)
|
||
|
||
|
||
async def test_pre_p4_on_demand_judging_never_reads_a_broken_manifest(golden, forbidden):
|
||
"""A26：老文档（chain 为假）+ 表结构快照列出的导入清单文件被改坏：摘录与金样逐字相同，根本不去取清单。"""
|
||
report = golden["reports"]["recipe"]
|
||
snap = artifact_store.load(report["catalog"]["Q1"]["artifact"])
|
||
[mid] = artifact_store.load(snap["schema_artifact"])["import_manifests"]
|
||
path = settings.data_dir / "artifacts" / mid[:2] / f"{mid}.json"
|
||
path.write_text(path.read_text("utf-8").replace('"seq":1', '"seq":9', 1), encoding="utf-8")
|
||
with pytest.raises(ValueError):
|
||
artifact_store.load(mid)
|
||
doc, catalog = doc_of(golden, "recipe", marked=False)
|
||
sealed = golden_sealed(golden, "recipe")
|
||
loader = api._sealed_loader(sealed, chain=doc.get("provenance") == DOC_PROVENANCE)
|
||
assert excerpts(doc, catalog, loader=loader) == report["outputs"]["excerpts"]
|
||
assert forbidden == []
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 真导入：区域外文字、遮罩、估价（3.2）
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
@pytest.fixture
|
||
async def client():
|
||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
|
||
yield c
|
||
|
||
|
||
@pytest.fixture
|
||
def no_model(monkeypatch, tmp_path):
|
||
from app.data import recipe_ai
|
||
from app.data.recipe_types import AiAvailability
|
||
|
||
data = tmp_path / "data"
|
||
(data / "uploads").mkdir(parents=True)
|
||
monkeypatch.setattr(settings, "data_dir", data)
|
||
|
||
async def none(session):
|
||
return None, "", AiAvailability(False, reason="测试不接模型")
|
||
|
||
monkeypatch.setattr(recipe_ai, "resolve_draft_model", none)
|
||
|
||
|
||
async def test_outside_text_masks_and_cost_on_a_real_import(no_model, client):
|
||
"""D28（9 月，表下 B32 有一句含数字的说明，新导入的清单记了 hidden=False）：摘录送全文、带限定语和统计期；
|
||
数据源设遮罩时一格都不送、只写条数；只判一句、引用这一条查询的点击估价不超过 0.01 美元。"""
|
||
from app.data.engine import engines
|
||
from tests.fixtures.xlsx.drift import SEP
|
||
|
||
up = await import_flow(client, seed=4201, start=SEP, days=30, variant="D28")
|
||
try:
|
||
qid, _sid = await run_sql(client, up, 'SELECT "日期", "全日客流" FROM "日客流" WHERE "日期" = \'2026-09-15\'')
|
||
g = json.loads(GOLDEN.read_text("utf-8"))
|
||
doc = copy.deepcopy(g["reports"]["recipe"]["outputs"]["doc"])
|
||
doc["provenance"] = DOC_PROVENANCE
|
||
catalog = {"Q1": {**g["reports"]["recipe"]["catalog"]["Q1"], "artifact": qid, "source": up.name,
|
||
"tool": f"db_query__{up.name}"}}
|
||
request = judge.prepare(doc, catalog)
|
||
text = request.excerpts["Q1"]
|
||
lines = text.split("\n")
|
||
assert lines[2].startswith("来源：上传的表格，每期替换，第 1 次导入：统计期 2026-09-01 至 2026-09-30")
|
||
assert OUTSIDE_HEAD in lines
|
||
note = next(line for line in lines if "B32：" in line)
|
||
assert note.startswith(" 第 1 次导入（统计期 2026-09-01 至 2026-09-30）客流汇总!B32：「") and "9月15日" in note
|
||
assert len(note) < PROV_OUTSIDE_CHARS + 60
|
||
assert judge.prepare(doc, catalog).excerpts == request.excerpts # 确定
|
||
# 遮罩：区域外文字一格都不送，只写条数
|
||
masked = judge.prepare(doc, catalog, masked={up.name: ["分区甲"]}).excerpts["Q1"]
|
||
assert "9月15日" not in masked and OUTSIDE_HEAD not in masked
|
||
assert "（数据源设置了遮罩，未列出）" in masked
|
||
# 估价：只判一句、引用这一条查询（claude-sonnet-5 的目录价）
|
||
one = [c for c in request.cands if "Q1" in c.aliases][:1]
|
||
assert one
|
||
cost = request.cost("claude-sonnet-5", one)
|
||
assert cost is not None and cost <= 0.01, cost
|
||
finally:
|
||
await engines.invalidate(up.source_id)
|