mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 02:11:58 +00:00
ThearchyHelios
2f90f0eb3a
- QueryCore 固定执行（K09）：按已发布的不可变口径跑一次，交回类型化标量、发布/manifest 引用与逐查询证据 - 依赖闭包（K08）：按选中的指标算出要拷的取数节点、要的入口字段，接进 derive / compose / metrics - SQL 模板参数完整性（K01）：试跑、组合问数、派生图、运行时取数共用一套语义，缺参数在渲染前就拒 - 受限字段变更服务（K03）与口径领域边界（K01）：改动打在整图深拷贝上，身份与证据归属不可改 - 版本化受控模板绑定（K04）：业务口径按显式绑定落到 SQL，对不上就报 unsupported / needs_update - 验证记录（K05）与最小审阅：样例匹配不等于正式批准，审批人只从服务端身份取 - 发布依据清单（K07）：发布时按不可变图物化内容寻址清单，正式用途按它判 - 增量重导入（K11）与 B/L/R 三方合并、受限自然语言修改提案（K12）：只产候选，落图仍走既有写路径 - 接口与前端：口径草稿页接上字段变更、绑定状态、验证与重导入，新增 caliber-drafts / studio 检查脚本 验证：前端 tsc -b 通过。
1488 lines
84 KiB
Python
1488 lines
84 KiB
Python
"""K10：父 Run 的口径查询工具（`caliber_query`）与父子 Run 的证据接入。
|
||
|
||
这份文件盯住这几条：
|
||
|
||
1. **工具是真的、参数是固定的。** 登记表里有 `caliber_query`（不是 `db_query__` 前缀那类动态工具），参数只有口径身份
|
||
（workflow / card）、要算的指标和这次入口的类型化取值；用途、主体、父 Run 身份、固定的数据版本都由父 Run 的上下文定，
|
||
参数里没有也传不了。
|
||
2. **证据接回父 Run。** 用父 Run 的上下文执行（`ToolContext`），当场 await 收尾；派生的探针运行是**明确的子运行**
|
||
（origin 里记 parent_run_id），父 Run 的节点把每条 query artifact 记进自己的台账（query_id / execution_id / attempt、
|
||
来源、列与行数、截断都在）——一次交回**多条**查询（queries 数组）不会只记第一条。
|
||
3. **伪造、跨运行、别的种类的工件一律拒绝。** 每条 query artifact 必须是这次授权的父运行自己落的，或它明确派生的子运行
|
||
落下的，且 kind=query_snapshot（capabilities.registry.snapshot_owned）。
|
||
4. **同一指标工作台与父 Run 是同一种说法。** 同一份口径、同一组指标与入口，HTTP 工作台（compose）与父 Run 拿到的发布
|
||
清单、结果类型、参数解释一致。
|
||
5. **父 Run 的边界跟着走。** 遮罩、截断、上传/记录源版本固定、只读、预算、复算要的材料；正式父 Run 走 formal judge，
|
||
交互（compose）拿不到正式身份；超时、失败都不是成功。
|
||
|
||
数据全是合成的（一个临时 SQLite 库、几条合成标量指标），不连真实业务库；这些指标不是任何真实业务口径。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import asyncio
|
||
import copy
|
||
import json
|
||
import sqlite3
|
||
import uuid
|
||
from typing import Any
|
||
|
||
import pytest
|
||
from httpx import ASGITransport, AsyncClient
|
||
from langchain_core.messages import AIMessage, ToolMessage
|
||
from sqlalchemy import select
|
||
|
||
from app.apps import publications
|
||
from app.db.base import SessionLocal
|
||
from app.db.models import Artifact, DataSource, Run, RunEvent
|
||
from app.engine.runner import run_manager
|
||
from app.ledger import querycore
|
||
from app.main import app
|
||
from app.records import store
|
||
from app.records.tables import columns_of, path_of
|
||
from tests import ledger_samples as S
|
||
from tests.publication_samples import add_record, add_version, confirm, make_workflow, publish
|
||
|
||
#: 这一版口径卡的节点 id（源头口径卡）
|
||
CARD = "card"
|
||
#: 取数窗口：start（含）/ end（不含）
|
||
WINDOW = "AND paid_at >= '{{ input.start }}' AND paid_at < '{{ input.end }}'"
|
||
ROWS = [
|
||
(1, "east", "paid", 100.0, 2, "2026-09-02"),
|
||
(2, "east", "paid", 50.0, 1, "2026-09-03"),
|
||
(3, "west", "refunded", 30.0, 3, "2026-09-04"),
|
||
(4, "west", "paid", 20.0, 4, "2026-09-07"),
|
||
(5, "east", "open", 12.5, 5, "2026-09-07"),
|
||
(6, "west", "paid", 70.0, 6, "2026-10-02"),
|
||
(7, "east", "paid", 5.0, 7, "2026-08-31"),
|
||
]
|
||
#: 窗口内的期望值（2026-09-01 ≤ paid_at < 2026-10-01，只算 paid）
|
||
WINDOW_SUM = 170.0
|
||
INPUTS = {"start": "2026-09-01", "end": "2026-10-01"}
|
||
|
||
|
||
def build_db(path) -> None:
|
||
db = sqlite3.connect(path)
|
||
db.executescript("CREATE TABLE orders (id INTEGER PRIMARY KEY, region TEXT, status TEXT, amount REAL, "
|
||
"qty INTEGER, paid_at TEXT);")
|
||
db.executemany("INSERT INTO orders VALUES (?, ?, ?, ?, ?, ?)", ROWS)
|
||
db.commit()
|
||
db.close()
|
||
|
||
|
||
async def fresh_source(tmp_path) -> str:
|
||
"""每个用例一个名字不同的数据源（候选口径按数据源预筛，别的用例的口径不会混进来）。"""
|
||
path = tmp_path / "caliber_query.db"
|
||
build_db(path)
|
||
S.CURRENT["source"] = f"cq_{uuid.uuid4().hex[:8]}"
|
||
await S.set_source(path)
|
||
return S.CURRENT["source"]
|
||
|
||
|
||
def graph(*, sales_limit: int | None = None, grouped: bool = False) -> dict[str, Any]:
|
||
"""入口 → 两条取数 → 口径卡 → 报告 → 出口（报告节点只在原图里，探针图不跑它）。
|
||
|
||
grouped=True 时 sales 那条按大区分组、可以给 limit：用来测截断（结果被截短照实记进台账）。"""
|
||
sales_sql = (f"SELECT region, SUM(amount) AS gmv, SUM(qty) AS units FROM orders WHERE status = 'paid' {WINDOW} "
|
||
"GROUP BY region" if grouped else
|
||
f"SELECT SUM(amount) AS gmv, SUM(qty) AS units FROM orders WHERE status = 'paid' {WINDOW}")
|
||
sales_args: dict[str, Any] = {"sql": sales_sql}
|
||
if sales_limit is not None:
|
||
sales_args["limit"] = sales_limit
|
||
return S.chain(
|
||
S.node("start", "input", fields=[
|
||
{"name": "start", "type": "string", "required": True, "description": "开始日期（含）"},
|
||
{"name": "end", "type": "string", "required": True, "description": "结束日期（不含）"},
|
||
]),
|
||
S.node("sales", "tool", tool=S.tool(), args=sales_args),
|
||
S.node("shared", "tool", tool=S.tool(),
|
||
args={"sql": f"SELECT SUM(amount) AS total, MAX(amount) AS peak FROM orders "
|
||
f"WHERE status = 'paid' {WINDOW}"}),
|
||
S.node(CARD, "metrics", caliber="合成周销售", caliber_version="v1", metrics=[
|
||
{"id": "gmv", "name": "销售额", "unit": "元", "decimals": 2, "format": "thousands",
|
||
"expression": "cell(nodes.sales, 0, 'gmv')"},
|
||
{"id": "units", "name": "件数", "unit": "件", "expression": "cell(nodes.sales, 0, 'units')"},
|
||
{"id": "peak", "name": "最大单笔", "unit": "元", "decimals": 2,
|
||
"expression": "cell(nodes.shared, 0, 'peak')"},
|
||
]),
|
||
S.node("write", "report", instructions="写周销售"),
|
||
S.node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}]),
|
||
)
|
||
|
||
|
||
def twin_graph() -> dict[str, Any]:
|
||
"""两个取数节点写**同一条 SQL**、读同一张表、同一批数据（内容寻址下会落到同一个工件 id）：两条不同的逻辑查询。
|
||
|
||
这是「同 artifact 对应不同逻辑 query」的反例：按工件 id 去重的证据索引会把它们并成一条。"""
|
||
sql = f"SELECT SUM(amount) AS gmv FROM orders WHERE status = 'paid' {WINDOW}"
|
||
metrics = [{"id": "left", "name": "甲", "unit": "元", "expression": "cell(nodes.one, 0, 'gmv')"},
|
||
{"id": "right", "name": "乙", "unit": "元", "expression": "cell(nodes.two, 0, 'gmv')"}]
|
||
return S.chain(
|
||
S.node("start", "input", fields=[{"name": "start", "type": "string", "required": True},
|
||
{"name": "end", "type": "string", "required": True}]),
|
||
S.node("one", "tool", tool=S.tool(), args={"sql": sql}),
|
||
S.node("two", "tool", tool=S.tool(), args={"sql": sql}),
|
||
S.node(CARD, "metrics", caliber="合成双查", caliber_version="v1", metrics=metrics),
|
||
S.node("write", "report", instructions="写周销售"),
|
||
S.node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}]),
|
||
)
|
||
|
||
|
||
def parent_graph(workflow_id: str, *, metric_ids: list[str] | None = None, card: str = CARD,
|
||
inputs: dict[str, Any] | None = None, gate: bool = False) -> dict[str, Any]:
|
||
"""父 Run 的图：入口 →（可选的人工确认）→ 调用工具（caliber_query）→ 出口。"""
|
||
parts = [S.node("start", "input")]
|
||
if gate:
|
||
parts.append(S.node("gate", "human", mode="approve", title="看一眼"))
|
||
parts.append(S.node("look", "tool", tool=querycore.CALIBER_QUERY, args={
|
||
"workflow": workflow_id, "card": card,
|
||
"metric_ids": metric_ids or ["gmv", "peak"],
|
||
"inputs": inputs if inputs is not None else dict(INPUTS)}))
|
||
parts.append(S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.look.ok }}"}]))
|
||
return S.chain(*parts)
|
||
|
||
|
||
@pytest.fixture(autouse=True)
|
||
async def engine_up(tmp_path):
|
||
await run_manager.setup()
|
||
await fresh_source(tmp_path)
|
||
yield
|
||
await run_manager.shutdown()
|
||
|
||
|
||
@pytest.fixture
|
||
async def client():
|
||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
|
||
yield c
|
||
|
||
|
||
async def published(client: AsyncClient, *, sales_limit: int | None = None, grouped: bool = False) -> dict:
|
||
"""建一份已发布的口径（走 K07 的真实发布入口），返回资产。"""
|
||
asset = await make_workflow(graph=graph(sales_limit=sales_limit, grouped=grouped))
|
||
await publish(client, asset)
|
||
return asset
|
||
|
||
|
||
async def run_parent(asset: dict, *, inputs: dict[str, Any] | None = None, metric_ids: list[str] | None = None,
|
||
data_versions: dict[str, Any] | None = None, options: dict[str, Any] | None = None,
|
||
run_class: str = "exploratory") -> Run:
|
||
"""起一次父 Run（走调用工具节点那条路），等它收尾。"""
|
||
run = await run_manager.start(graph=parent_graph(asset["workflow_id"], inputs=inputs, metric_ids=metric_ids),
|
||
input_payload={}, run_class=run_class, data_versions=data_versions,
|
||
options=options)
|
||
return await settled(run.id)
|
||
|
||
|
||
async def run_once(asset: dict, *, metric_ids: list[str], inputs: dict[str, Any] | None = None, **kw: Any):
|
||
"""走服务入口跑一次（HTTP 之外的那条路），等 K09 的执行结果。"""
|
||
async with SessionLocal() as session:
|
||
return await querycore.execute(session, workflow_id=asset["workflow_id"], card_id=CARD,
|
||
metric_ids=metric_ids,
|
||
inputs=inputs if inputs is not None else dict(INPUTS), **kw)
|
||
|
||
|
||
async def settled(run_id: str) -> Run:
|
||
for _ in range(600):
|
||
async with SessionLocal() as session:
|
||
row = await session.get(Run, run_id)
|
||
if row.status in querycore.TERMINAL_RUN_STATUSES:
|
||
return row
|
||
await asyncio.sleep(0.05)
|
||
raise AssertionError(f"运行 {run_id} 没有结束：{row.status}")
|
||
|
||
|
||
async def events_of(run_id: str, etype: str) -> list[dict[str, Any]]:
|
||
async with SessionLocal() as session:
|
||
rows = (await session.execute(select(RunEvent).where(RunEvent.run_id == run_id,
|
||
RunEvent.type == etype)
|
||
.order_by(RunEvent.seq))).scalars().all()
|
||
return [dict(r.data or {}) for r in rows]
|
||
|
||
|
||
async def tool_end(run_id: str) -> dict[str, Any]:
|
||
[data] = await events_of(run_id, "tool.end")
|
||
return data
|
||
|
||
|
||
async def tool_result(run_id: str, node_id: str = "look") -> dict[str, Any]:
|
||
"""调用工具节点的**完整**产出：tool.end 的 preview 截到 2000 个字符，读不全 JSON（K10 的结果里带几条查询）。"""
|
||
from app.engine.evidence_reader import read_sealed
|
||
|
||
sealed = await read_sealed(run_id)
|
||
payload = sealed.payload_of([node_id])
|
||
assert isinstance(payload, dict), "调用工具节点没有产出"
|
||
return payload
|
||
|
||
|
||
async def finished_evidence(run_id: str, node_id: str = "look") -> list[dict[str, Any]]:
|
||
async with SessionLocal() as session:
|
||
rows = (await session.execute(select(RunEvent).where(RunEvent.run_id == run_id,
|
||
RunEvent.type == "node.finished")
|
||
.order_by(RunEvent.seq))).scalars().all()
|
||
for row in rows:
|
||
data = dict(row.data or {})
|
||
if row.node_id == node_id and data.get("evidence"):
|
||
return list(data["evidence"])
|
||
return []
|
||
|
||
|
||
# ==========================================================================
|
||
# 1、工具是真的：登记表里有它、参数固定、不是 db_query__ 前缀
|
||
# ==========================================================================
|
||
|
||
|
||
async def test_the_tool_is_registered_with_fixed_parameters():
|
||
from app.capabilities.registry import is_platform_query
|
||
from app.tools.registry import all_specs
|
||
|
||
spec = all_specs()[querycore.CALIBER_QUERY]
|
||
assert spec.category and spec.description and not spec.dangerous and not spec.writes_state
|
||
assert spec.effect == "read", "口径查询是只读的：不写外部系统、不该要确认"
|
||
assert set(spec.args_schema.model_fields) == {"workflow", "card", "metric_ids", "inputs"}
|
||
# 用途、主体、版本、预算都不在参数里：由父 Run 的上下文定，模型改不了
|
||
assert not set(spec.args_schema.model_fields) & {"purpose", "role", "actor", "version", "budget", "trace"}
|
||
assert is_platform_query(querycore.CALIBER_QUERY), "口径查询的结果要能当查询快照作证"
|
||
assert not querycore.CALIBER_QUERY.startswith("db_query__"), "它是真实工具，不是数据源那类动态工具"
|
||
|
||
|
||
async def test_it_needs_a_parent_run():
|
||
from app.tools.registry import ToolContext
|
||
|
||
with pytest.raises(querycore.QueryCoreRefused) as caught:
|
||
await querycore.caliber_query(ToolContext(), workflow="w", card=CARD, metric_ids=["gmv"], inputs={})
|
||
assert caught.value.code == querycore.NO_PARENT
|
||
|
||
|
||
# ==========================================================================
|
||
# 2、证据接回父 Run：子运行、多条查询、执行身份
|
||
# ==========================================================================
|
||
|
||
|
||
async def test_the_parent_run_records_every_query_of_one_call(client):
|
||
asset = await published(client)
|
||
parent = await run_parent(asset, metric_ids=["gmv", "peak"])
|
||
assert parent.status == "succeeded", parent.error
|
||
|
||
end = await tool_end(parent.id)
|
||
assert end["tool"] == querycore.CALIBER_QUERY
|
||
# 一次调用交回两条查询（sales、shared）：tool.end 上两条的工件 id 都在
|
||
assert len(end["query_artifacts"]) == 2 and end["query_artifact"] == end["query_artifacts"][0]
|
||
|
||
entries = [e for e in await finished_evidence(parent.id) if e.get("kind") == "query"]
|
||
assert [e["tool"] for e in entries] == [querycore.CALIBER_QUERY] * 2
|
||
assert {e["artifact"] for e in entries} == set(end["query_artifacts"])
|
||
child_id = (await tool_result(parent.id))["run_id"]
|
||
for entry in entries:
|
||
assert entry["columns"] and entry["rows"] and entry["truncated"] is False
|
||
# 源名写在和数据源查询同一个位置（证据面板、遮罩按它认源）
|
||
assert entry["source"] == S.CURRENT["source"]
|
||
# 出处是这一版这一张图里的那条取数，不是父 Run 的节点
|
||
assert entry["origin"]["workflow_id"] == asset["workflow_id"] and entry["origin"]["version"] == 1
|
||
assert entry["origin"]["node_id"] in ("sales", "shared")
|
||
# 执行身份三件套齐（K10）：执行号记的是真的执行那次 SQL 的运行（派生的子运行），不是父 Run
|
||
assert entry["query_id"] and entry["execution_id"] and entry["attempt"] >= 1
|
||
assert child_id in entry["execution_id"], "执行身份带上真的执行那次，别把两次调用的执行混成一个"
|
||
# 复算要的材料跟着条目走：这次参数输入的哈希与源模板哈希（值本身、SQL 正文都不进台账）
|
||
assert entry["parameters"]["inputs_hash"] and entry["parameters"]["template_hash"]
|
||
assert {e["query_id"] for e in entries} == {f"{asset['workflow_id']}:sales", f"{asset['workflow_id']}:shared"}
|
||
|
||
# 父 Run 的封存证据里认得出这两条查询（报告目录、证据面板从这里出发）
|
||
from app.engine.evidence_reader import read_sealed
|
||
|
||
sealed = await read_sealed(parent.id)
|
||
assert {e["artifact"] for e in entries} <= set(sealed.queries)
|
||
for artifact in end["query_artifacts"]:
|
||
assert sealed.queries[artifact]["node_id"] == "look"
|
||
|
||
|
||
async def test_the_probe_run_is_an_explicit_child_of_the_parent(client):
|
||
asset = await published(client)
|
||
parent = await run_parent(asset)
|
||
end = await tool_end(parent.id)
|
||
child_id = (await tool_result(parent.id))["run_id"]
|
||
assert child_id, "结果里要给出子运行的 id"
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, child_id)
|
||
assert child is not None and child.id != parent.id and child.run_class == "exploratory"
|
||
origin = child.origin or {}
|
||
assert origin["service"] == "querycore" and origin["purpose"] == querycore.DEFAULT_PURPOSE
|
||
assert origin["parent_run_id"] == parent.id, "子运行的来源里必须记着父运行，父子关系不能丢"
|
||
assert origin["manifest"] and origin["selection_hash"] and origin["closure_hash"], "复算要的材料都在"
|
||
assert child.options.get("read_only_data") is True, "派生的子运行照旧一律只读"
|
||
|
||
|
||
async def test_two_identical_calls_get_different_execution_ids(client):
|
||
"""同样的参数、同样的 SQL 调两次：两个子运行，执行身份各不相同（工件按内容寻址，可以是同一件）。"""
|
||
from app.tools.registry import ToolContext
|
||
|
||
asset = await published(client)
|
||
ctx = ToolContext(run_id="parent-fixed", node_id="look")
|
||
first = await querycore.caliber_query(ctx, workflow=asset["workflow_id"], card=CARD,
|
||
metric_ids=["gmv"], inputs=dict(INPUTS))
|
||
second = await querycore.caliber_query(ctx, workflow=asset["workflow_id"], card=CARD,
|
||
metric_ids=["gmv"], inputs=dict(INPUTS))
|
||
assert first["ok"] and second["ok"]
|
||
assert first["run_id"] != second["run_id"], "每次调用派生一次自己的子运行"
|
||
left, right = first["queries"][0], second["queries"][0]
|
||
assert left["query_id"] == right["query_id"], "逻辑查询是同一条"
|
||
assert left["execution_id"] != right["execution_id"], "同一条 SQL 执行两次，执行身份必须不同"
|
||
|
||
|
||
async def test_two_identical_executions_in_one_run_are_not_merged(client):
|
||
"""同一个父 Run 里两个节点跑同一条 SQL：台账里两条各记各的执行，不合并成一条。"""
|
||
asset = await published(client)
|
||
args = {"workflow": asset["workflow_id"], "card": CARD, "metric_ids": ["gmv"], "inputs": dict(INPUTS)}
|
||
graph = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("one", "tool", tool=querycore.CALIBER_QUERY, args=dict(args)),
|
||
S.node("two", "tool", tool=querycore.CALIBER_QUERY, args=dict(args)),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.two.ok }}"}]),
|
||
)
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
row = await settled(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
entries = ([e for e in await finished_evidence(row.id, "one") if e.get("kind") == "query"]
|
||
+ [e for e in await finished_evidence(row.id, "two") if e.get("kind") == "query"])
|
||
assert len(entries) == 2, "两次执行各记一条"
|
||
assert len({e["execution_id"] for e in entries}) == 2, "同一份 SQL 跑两次，各记各的执行身份"
|
||
assert len({e["query_id"] for e in entries}) == 1, "逻辑查询是同一条"
|
||
|
||
|
||
# ==========================================================================
|
||
# 3、伪造 / 跨运行 / 别的种类：一律不当证据
|
||
# ==========================================================================
|
||
|
||
|
||
async def test_a_forged_or_cross_run_artifact_is_refused(client):
|
||
"""名字对了，工件还得是这次父运行（或它明确派生的子运行）落下的 query_snapshot。"""
|
||
from app.capabilities.registry import snapshot_owned
|
||
from app.core.artifact_store import put_json
|
||
|
||
asset = await published(client)
|
||
parent = await run_parent(asset)
|
||
[artifact] = (await tool_end(parent.id))["query_artifacts"][:1]
|
||
# 真的：父 Run 明确派生的子运行落下的查询快照
|
||
assert await snapshot_owned(artifact, parent.id, querycore.CALIBER_QUERY)
|
||
# 别的运行落下的同一件：不认
|
||
assert not await snapshot_owned(artifact, "another-run", querycore.CALIBER_QUERY)
|
||
# 伪造的工件 id：不认
|
||
assert not await snapshot_owned("f" * 64, parent.id, querycore.CALIBER_QUERY)
|
||
|
||
# 种类不对（同一件内容、父运行自己落下的 tool_snapshot）：不认
|
||
async with SessionLocal() as session:
|
||
snap = (await session.execute(select(Artifact).where(Artifact.id == artifact))).first()
|
||
wrong_kind = await put_json({"x": 1}, kind="tool_snapshot", run_id=parent.id)
|
||
assert wrong_kind != artifact
|
||
assert not await snapshot_owned(wrong_kind, parent.id, querycore.CALIBER_QUERY), "不是 query_snapshot 的不当作证"
|
||
assert snap is not None, "子运行那件工件在工件表里记着"
|
||
|
||
|
||
async def test_a_forged_tool_result_does_not_become_evidence(client, monkeypatch):
|
||
"""调用工具节点交回一份拼出来的 caliber 结果（工件是别人的）：不进台账、tool.end 上也不带。"""
|
||
from app.engine.nodes import tools as tool_node
|
||
|
||
asset = await published(client)
|
||
|
||
async def forged_call_tool(*_a, **_k):
|
||
return {"tool": querycore.CALIBER_QUERY, "kind": querycore.CALIBER_QUERY, "ok": True, "run_id": "made-up",
|
||
"queries": [{"artifact": "a" * 64, "source": {"workflow_id": asset["workflow_id"], "version": 1,
|
||
"node_id": "sales"},
|
||
"columns": ["gmv"], "row_count": 1, "truncated": False,
|
||
"query_id": "made-up:sales", "execution_id": "made-up:sales#e1", "attempt": 1}]}
|
||
|
||
monkeypatch.setattr(tool_node, "call_tool", forged_call_tool)
|
||
parent = await run_parent(asset)
|
||
assert parent.status == "succeeded", parent.error
|
||
end = await tool_end(parent.id)
|
||
assert "query_artifact" not in end and "query_artifacts" not in end
|
||
assert [e for e in await finished_evidence(parent.id) if e.get("kind") == "query"] == []
|
||
|
||
|
||
# ==========================================================================
|
||
# 4、同一指标工作台与父 Run 是同一种说法
|
||
# ==========================================================================
|
||
|
||
|
||
async def test_the_workbench_and_the_parent_run_agree(client):
|
||
asset = await published(client)
|
||
res = await client.post("/api/ledger/querycore", json={
|
||
"workflow_id": asset["workflow_id"], "card_id": CARD, "metric_ids": ["gmv", "peak"], "inputs": INPUTS})
|
||
assert res.status_code == 200, res.text
|
||
bench = res.json()
|
||
assert bench["ok"] and bench["purpose"] == publications.PURPOSE_COMPOSE
|
||
|
||
parent = await run_parent(asset, metric_ids=["gmv", "peak"])
|
||
end = await tool_end(parent.id)
|
||
tool = await tool_result(parent.id)
|
||
|
||
# 同一份发布清单、同一种结果类型、同一套参数解释
|
||
assert tool["publication"] == bench["publication"]
|
||
assert tool["version"] == bench["version"] and tool["explicit_version"] == bench["explicit_version"]
|
||
assert [m["id"] for m in tool["metrics"]] == [m["id"] for m in bench["metrics"]]
|
||
assert [m["value"] for m in tool["metrics"]] == [m["value"] for m in bench["metrics"]]
|
||
assert tool["metrics"] == bench["metrics"]
|
||
assert tool["parameters"]["values"] == bench["parameters"]["values"] == INPUTS
|
||
assert tool["parameters"]["contract"] == bench["parameters"]["contract"]
|
||
assert tool["parameters"]["adapter"] == bench["parameters"]["adapter"]
|
||
assert tool["selection"]["closure_hash"] == bench["selection"]["closure_hash"]
|
||
assert tool["data"]["tables"] == bench["data"]["tables"] == ["orders"]
|
||
# 每条查询：逻辑身份、列、参数输入哈希、行数与截断一致（工件是两次运行各自的，不要求相同）
|
||
assert [(q["query_id"], q["columns"], q["truncated"]) for q in tool["queries"]] \
|
||
== [(q["query_id"], q["columns"], q["truncated"]) for q in bench["queries"]]
|
||
assert [q["parameters"]["inputs_hash"] for q in tool["queries"]] \
|
||
== [q["parameters"]["inputs_hash"] for q in bench["queries"]]
|
||
|
||
|
||
# ==========================================================================
|
||
# 5、父 Run 的边界：遮罩、截断、版本固定、只读、预算、formal
|
||
# ==========================================================================
|
||
|
||
|
||
async def test_masked_values_are_not_handed_out(client, tmp_path):
|
||
asset = await published(client)
|
||
await S.set_source(tmp_path / "caliber_query.db", options={"mask_columns": ["gmv"]})
|
||
parent = await run_parent(asset, metric_ids=["gmv"])
|
||
tool = await tool_result(parent.id)
|
||
[metric] = [m for m in tool["metrics"] if m["id"] == "gmv"]
|
||
assert metric.get("masked") is True and metric["value"] is None
|
||
assert tool["ok"] is False and querycore.METRIC_MASKED in [p["code"] for p in tool["problems"]]
|
||
assert "masked" not in json.dumps([{"id": m["id"], "v": m["value"]} for m in tool["metrics"]])
|
||
|
||
|
||
async def test_truncated_results_are_reported_as_truncated(client):
|
||
asset = await published(client, grouped=True, sales_limit=1)
|
||
parent = await run_parent(asset, metric_ids=["gmv"])
|
||
[entry] = [e for e in await finished_evidence(parent.id) if e.get("kind") == "query"]
|
||
assert entry["truncated"] is True, "结果被截短照实记进台账，不悄悄当完整结果"
|
||
assert entry["rows"] == 1, "台账里记的是快照里真的有多少行"
|
||
|
||
|
||
async def test_the_parent_run_pins_the_source_versions_into_the_child(client):
|
||
"""父 Run 固定的上传/记录源版本照原样带进子运行：父子两段查的是同一版数据。"""
|
||
asset = await published(client)
|
||
pinned = {"src-1": {"snapshot": "snap-1", "name": "表格"}}
|
||
parent = await run_parent(asset, data_versions=pinned)
|
||
assert parent.data_versions == pinned
|
||
child_id = (await tool_result(parent.id))["run_id"]
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, child_id)
|
||
assert child.data_versions == pinned, "子运行必须照父 Run 那一份固定，不能改用当前版本"
|
||
assert child.options.get("read_only_data") is True
|
||
|
||
|
||
async def test_a_timeout_is_not_a_success(client):
|
||
"""父 Run 给的预算到点没跑完：交回 running + 子运行 id，ok=False，台账里一条查询都没有。"""
|
||
asset = await published(client)
|
||
parent = await run_parent(asset, options={querycore.BUDGET_OPTION: {"timeout_s": 0.0}})
|
||
assert parent.status == "succeeded", parent.error
|
||
end = await tool_end(parent.id)
|
||
tool = await tool_result(parent.id)
|
||
assert tool["ok"] is False and tool["status"] == "running" and tool["run_id"]
|
||
assert [p["code"] for p in tool["problems"]] == [querycore.TIMEOUT]
|
||
assert "query_artifact" not in end
|
||
assert [e for e in await finished_evidence(parent.id) if e.get("kind") == "query"] == []
|
||
|
||
|
||
async def test_a_refusal_is_a_tool_failure(client):
|
||
"""口径不在（卡的名字写错）：工具当场拒，节点失败并说清原因——不交回一份空结果冒充成功。"""
|
||
asset = await published(client)
|
||
parent = await run_parent(asset, metric_ids=["gmv"], inputs=dict(INPUTS))
|
||
assert parent.status == "succeeded"
|
||
run = await run_manager.start(graph=parent_graph(asset["workflow_id"], card="nope", inputs=dict(INPUTS)),
|
||
input_payload={})
|
||
row = await settled(run.id)
|
||
assert row.status == "failed" and querycore.CARD_MISSING in (row.error or "")
|
||
|
||
|
||
async def test_a_formal_parent_uses_the_formal_judge(client):
|
||
"""正式父 Run：走 execute_formal 那一套（发布依据不齐就拒）；交互（compose）拿不到正式身份。"""
|
||
from app.tools.registry import ToolContext
|
||
|
||
asset = await published(client) # 有清单、没有验证记录与审阅
|
||
compose = ToolContext(run_id="p1", node_id="look")
|
||
outcome = await querycore.caliber_query(compose, workflow=asset["workflow_id"], card=CARD,
|
||
metric_ids=["gmv"], inputs=dict(INPUTS))
|
||
assert outcome["ok"] and outcome["purpose"] == publications.PURPOSE_COMPOSE
|
||
|
||
formal = ToolContext(run_id="p2", node_id="look", run_class="formal")
|
||
with pytest.raises(ValueError) as caught:
|
||
await querycore.caliber_query(formal, workflow=asset["workflow_id"], card=CARD,
|
||
metric_ids=["gmv"], inputs=dict(INPUTS))
|
||
assert publications.RECORD_MISSING in str(caught.value), "正式父 Run 用的是 publications.judge 的原码"
|
||
|
||
record_id = await add_record(asset)
|
||
await confirm(asset, record_id)
|
||
await publish(client, asset)
|
||
proven = await querycore.caliber_query(ToolContext(run_id="p3", node_id="look", run_class="formal"),
|
||
workflow=asset["workflow_id"], card=CARD, metric_ids=["gmv"],
|
||
inputs=dict(INPUTS))
|
||
assert proven["ok"] and proven["purpose"] == publications.PURPOSE_FORMAL
|
||
assert proven["publication"]["identity"] == publications.PROVEN
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, proven["run_id"])
|
||
assert child.origin["purpose"] == publications.PURPOSE_FORMAL and child.origin["parent_run_id"] == "p3"
|
||
|
||
|
||
async def test_the_actor_comes_from_the_parent_run_not_the_parameters(client):
|
||
"""认证主体来自父 Run，工具参数里没有这一项，也传不了。"""
|
||
asset = await published(client)
|
||
run = await run_manager.start(graph=parent_graph(asset["workflow_id"]), input_payload={}, started_by="老王")
|
||
row = await settled(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
child_id = (await tool_result(row.id))["run_id"]
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, child_id)
|
||
assert child.started_by == "老王"
|
||
|
||
|
||
# ==========================================================================
|
||
# 6、协作成员路径：成员绑口径查询时，多查询逐条编号、逐条认工件
|
||
# ==========================================================================
|
||
|
||
|
||
def team_graph(workflow_id: str) -> dict[str, Any]:
|
||
"""协作团队：成员「查数」绑 caliber_query 与分析工具（绑了分析工具才编 Qn），另一个只是说话。"""
|
||
agents = [
|
||
{"name": "查数", "description": "查口径", "system": "你会查口径", "max_steps": 1,
|
||
"tools": [querycore.CALIBER_QUERY, "analysis_comparable"]},
|
||
{"name": "复核", "description": "复核别人算的数", "system": "你复核", "max_steps": 1, "tools": []},
|
||
]
|
||
return S.chain(
|
||
S.node("start", "input"),
|
||
S.node("team", "supervisor", goal="算一下销售额与最大单笔", agents=agents, max_rounds=2, max_parallel=1,
|
||
approval="never"),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.team.text }}"}]),
|
||
)
|
||
|
||
|
||
async def test_a_cooperative_member_numbers_every_query_of_one_call(client):
|
||
"""成员绑的口径查询交回两条查询：逐条编号、逐条进成员自己的分析容器，编号与执行身份记在调用记录里。
|
||
|
||
协作节点不记证据台账时编号只给成员自己下钻用，所以编号要和这次调用的调用记录一起留下来；否则事后看不出
|
||
某个 Qn 是哪一次执行的哪条查询。给了 ledger 时（K10 收尾）同时逐条记一条运行台账的 query 条目。"""
|
||
from app.capabilities.registry import snapshot_owned
|
||
from app.engine.nodes import multi
|
||
from app.tools.analysis import AnalysisScope
|
||
|
||
asset = await published(client)
|
||
parent = await run_parent(asset, metric_ids=["gmv", "peak"])
|
||
payload = await tool_result(parent.id)
|
||
content = json.dumps(payload, ensure_ascii=False)
|
||
scope = AnalysisScope(node_id="team")
|
||
record: dict[str, Any] = {}
|
||
ledger: list[dict[str, Any]] = []
|
||
|
||
out = await multi._numbered(scope, querycore.CALIBER_QUERY, content, None, parent.id, "team",
|
||
call_id="call-1", member="查数", exec_no=1, record=record, ledger=ledger)
|
||
assert out.startswith("【证据 Q1、Q2】"), "一次交回多条查询：两条都要编号（以前只认第一条）"
|
||
assert list(scope.queries) == ["Q1", "Q2"]
|
||
traced = record["queries"]
|
||
assert [q["alias"] for q in traced] == ["Q1", "Q2"]
|
||
assert record["alias"] == "Q1" and record["artifact"] == payload["queries"][0]["artifact"]
|
||
assert [q["query_id"] for q in traced] == [q["query_id"] for q in payload["queries"]]
|
||
assert [q["artifact"] for q in traced] == [q["artifact"] for q in payload["queries"]]
|
||
# 执行身份照抄口径查询自己带着的那一份（K09 算好的）：真的执行那次 SQL 的子运行、那条查询的第几次
|
||
child = payload["run_id"]
|
||
assert all(q["execution_id"] == p["execution_id"] and child in q["execution_id"]
|
||
for q, p in zip(traced, payload["queries"]))
|
||
assert all(q["attempt"] == p["attempt"] >= 1 for q, p in zip(traced, payload["queries"]))
|
||
# 两条工件都是这次父运行（明确派生的子运行）落下的查询快照
|
||
assert all([await snapshot_owned(q["artifact"], parent.id, querycore.CALIBER_QUERY) for q in traced])
|
||
# 运行台账的条目：与 tools/llm 记的那一种同形，另带成员归属（node_id 前缀 + member）
|
||
assert [e["node_id"] for e in ledger] == ["team:查数", "team:查数"]
|
||
assert [e["member"] for e in ledger] == ["查数", "查数"]
|
||
assert [e["artifact"] for e in ledger] == [q["artifact"] for q in payload["queries"]]
|
||
assert all(e["kind"] == "query" and e["tool"] == querycore.CALIBER_QUERY and e["exec"] == 1
|
||
and e["call_id"] == "call-1" for e in ledger)
|
||
assert all(e["source"] and e["columns"] and e["rows"] >= 1 and e["truncated"] is False for e in ledger)
|
||
assert [e["query_id"] for e in ledger] == [q["query_id"] for q in payload["queries"]]
|
||
assert [e["execution_id"] for e in ledger] == [q["execution_id"] for q in payload["queries"]]
|
||
assert all(e["origin"]["workflow_id"] == asset["workflow_id"] for e in ledger)
|
||
assert all(e["parameters"]["inputs_hash"] for e in ledger), "复算要的材料跟着条目走"
|
||
|
||
|
||
async def test_a_cooperative_member_refuses_forged_and_other_runs_artifacts(client):
|
||
"""名字对了，工件也得对：伪造的、别的运行落的、不是查询快照的都不进成员的分析容器、也不进运行台账。"""
|
||
from app.engine.nodes import multi
|
||
from app.tools.analysis import AnalysisScope
|
||
|
||
asset = await published(client)
|
||
parent = await run_parent(asset, metric_ids=["gmv", "peak"])
|
||
payload = await tool_result(parent.id)
|
||
# 一份真的、一份工件 id 换掉的
|
||
forged = json.loads(json.dumps(payload))
|
||
forged["queries"][0]["artifact"] = "a" * 64
|
||
content = json.dumps(forged, ensure_ascii=False)
|
||
scope = AnalysisScope(node_id="team")
|
||
record: dict[str, Any] = {}
|
||
ledger: list[dict[str, Any]] = []
|
||
out = await multi._numbered(scope, querycore.CALIBER_QUERY, content, None, parent.id, "team",
|
||
call_id="call-1", member="查数", exec_no=1, record=record, ledger=ledger)
|
||
assert out == f"【证据 Q1】\n{content}", "只给认下来的那条编号"
|
||
assert list(scope.queries) == ["Q1"], "伪造的那条不进容器，真的那条照记"
|
||
assert [q["artifact"] for q in record["queries"]] == [payload["queries"][1]["artifact"]]
|
||
assert [e["artifact"] for e in ledger] == [payload["queries"][1]["artifact"]], "台账里也只记真的那条"
|
||
|
||
# 同一份真工件，换一个运行 id：不是这次运行（也不是它派生的子运行）落的，整条都不认
|
||
real = json.dumps(payload, ensure_ascii=False)
|
||
other = AnalysisScope(node_id="team")
|
||
other_record: dict[str, Any] = {}
|
||
other_ledger: list[dict[str, Any]] = []
|
||
out = await multi._numbered(other, querycore.CALIBER_QUERY, real, None, "another-run", "team",
|
||
call_id="call-1", member="查数", exec_no=1, record=other_record,
|
||
ledger=other_ledger)
|
||
assert out == real and not other.queries and not other_record and not other_ledger
|
||
|
||
|
||
async def test_a_cooperative_member_keeps_the_single_query_behavior():
|
||
"""单条数据源查询照旧：一个 Qn、内容前面一行编号，执行身份按这次调用的 id 算（同 tools/registry）。"""
|
||
from app.engine.nodes import multi
|
||
from app.tools.analysis import AnalysisScope
|
||
|
||
scope = AnalysisScope(node_id="team")
|
||
record: dict[str, Any] = {}
|
||
ledger: list[dict[str, Any]] = []
|
||
payload = {"artifact": "b" * 64, "columns": ["gmv"], "rows": [[1.0]], "truncated": False, "source": "shop",
|
||
"sql": "SELECT 1"}
|
||
content = json.dumps(payload)
|
||
out = await multi._numbered(scope, "db_query__shop", content, None, "run-1", "team",
|
||
call_id="call-1", member="查数", exec_no=2, record=record, ledger=ledger)
|
||
assert out == f"【证据 Q1】\n{content}" and list(scope.queries) == ["Q1"]
|
||
assert [q["artifact"] for q in record["queries"]] == ["b" * 64]
|
||
assert record["queries"][0]["execution_id"] == "run-1:team#call-1"
|
||
assert record["queries"][0]["query_id"] == "team"
|
||
[entry] = ledger
|
||
assert entry["node_id"] == "team:查数" and entry["member"] == "查数" and entry["exec"] == 2
|
||
assert entry["query_id"] == "team" and entry["attempt"] == 2
|
||
assert entry["source"] == "shop" and entry["columns"] == ["gmv"] and entry["rows"] == 1
|
||
|
||
|
||
async def test_a_plain_tool_result_leaves_no_ledger_behind():
|
||
"""不是平台执行的查询（普通工具、编造的同样的 JSON）：容器、调用记录、台账都不动。"""
|
||
from app.engine.nodes import multi
|
||
from app.tools.analysis import AnalysisScope
|
||
|
||
scope = AnalysisScope(node_id="team")
|
||
record: dict[str, Any] = {}
|
||
ledger: list[dict[str, Any]] = []
|
||
content = json.dumps({"artifact": "c" * 64, "columns": ["x"], "rows": [[1]]})
|
||
out = await multi._numbered(scope, "calculator", content, None, "run-1", "team",
|
||
call_id="call-1", member="查数", exec_no=1, record=record, ledger=ledger)
|
||
assert out == content and not scope.queries and not record and not ledger
|
||
|
||
|
||
async def test_the_team_member_path_records_the_queries_it_numbered(client, monkeypatch):
|
||
"""走真的一次协作运行：成员绑口径查询，编号与执行身份随团队产出留下来，成员看到的是「【证据 Q1、Q2】」。"""
|
||
from app.providers import mock_model
|
||
|
||
asset = await published(client)
|
||
args = {"workflow": asset["workflow_id"], "card": CARD, "metric_ids": ["gmv", "peak"], "inputs": dict(INPUTS)}
|
||
decisions = [
|
||
{"assignments": [{"agent": "查数", "instruction": "算 gmv 与 peak"}], "done": False, "reason": "先查"},
|
||
{"assignments": [], "done": True, "reason": "够了"},
|
||
]
|
||
turns = {"n": 0}
|
||
seen: list[str] = []
|
||
|
||
def _decide(self, messages):
|
||
seen.extend(str(m.content) for m in messages if isinstance(m, ToolMessage))
|
||
if self.response_format:
|
||
i = turns["n"]
|
||
turns["n"] += 1
|
||
return AIMessage(content=json.dumps(decisions[i] if i < len(decisions) else decisions[-1],
|
||
ensure_ascii=False))
|
||
if self.tools and not any(isinstance(m, ToolMessage) for m in messages):
|
||
return AIMessage(content="查一下口径。",
|
||
tool_calls=[{"name": querycore.CALIBER_QUERY, "args": args, "id": "c1"}])
|
||
return AIMessage(content="查完了：gmv 与 peak 都取到了。")
|
||
|
||
monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
|
||
run = await run_manager.start(graph=team_graph(asset["workflow_id"]), input_payload={})
|
||
row = await settled(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
assert any(text.startswith("【证据 Q1、Q2】") for text in seen), "成员拿到的结果前面要写全这一次的编号"
|
||
|
||
from app.engine.evidence_reader import read_sealed
|
||
|
||
team = (await read_sealed(row.id)).payload_of(["team"])
|
||
assert isinstance(team, dict)
|
||
[record] = team["transcript"][0]["tool_calls"]
|
||
traced = record["queries"]
|
||
assert [q["alias"] for q in traced] == ["Q1", "Q2"]
|
||
assert {q["query_id"] for q in traced} == {f"{asset['workflow_id']}:sales", f"{asset['workflow_id']}:shared"}
|
||
# 执行身份指的是真的执行那次 SQL 的子运行；父子关系在子运行的来源里
|
||
child_id = str(traced[0]["execution_id"]).split(":", 1)[0]
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, child_id)
|
||
assert child is not None and (child.origin or {}).get("parent_run_id") == row.id
|
||
assert {str(q["execution_id"]).split(":", 1)[0] for q in traced} == {child_id}, "两条查询在同一次执行里"
|
||
assert [q["attempt"] for q in traced] == [1, 1]
|
||
|
||
|
||
#: 成员这一轮查的是另一个窗口：和上面调用工具节点那条的 SQL 不一样，两条的查询快照各是各的（好断言归属）
|
||
MEMBER_INPUTS = {"start": "2026-08-01", "end": "2026-12-01"}
|
||
|
||
|
||
def team_report_graph(workflow_id: str) -> dict[str, Any]:
|
||
"""入口 → 调用工具（口径查询）→ 协作团队（成员也绑口径查询）→ 报告 → 出口。
|
||
|
||
两处口径查询各交回两条 SQL，报告目录只收调用工具节点那两条：成员的查询是**附加**在封存证据上的。"""
|
||
agents = [
|
||
{"name": "查数", "description": "查口径", "system": "你会查口径", "max_steps": 1,
|
||
"tools": [querycore.CALIBER_QUERY, "analysis_comparable"]},
|
||
{"name": "复核", "description": "复核别人算的数", "system": "你复核", "max_steps": 1, "tools": []},
|
||
]
|
||
return S.chain(
|
||
S.node("start", "input"),
|
||
S.node("look", "tool", tool=querycore.CALIBER_QUERY, args={
|
||
"workflow": workflow_id, "card": CARD, "metric_ids": ["gmv", "peak"], "inputs": dict(INPUTS)}),
|
||
S.node("team", "supervisor", goal="再按全年口径核一遍", agents=agents, max_rounds=2, max_parallel=1,
|
||
approval="never"),
|
||
S.node("write", "report", instructions="写一段说明"),
|
||
S.node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}]),
|
||
)
|
||
|
||
|
||
async def test_member_queries_join_the_sealed_evidence_without_renumbering_the_report(client, monkeypatch):
|
||
"""成员查到的数进父 Run 的封存证据（带成员前缀、来源、执行身份的 query 条目），报告目录的全局 Q 编号一个字不变。
|
||
|
||
报告目录按祖先收证据（report.report_catalog 的 allowed），成员条目的节点 id 是「协作节点 id:成员名」，既不像子工作流
|
||
转发的「协作节点 id/成员名」那样被 engine/compiler 的 _own_evidence 滤掉（那会让条目根本进不了封存），也不落在报告的
|
||
收证据范围里——成员查到的数是附加证据：封存读取（sealed.queries / query_records / entries）、导出包的来源与遮罩判断、
|
||
运行对比读得到，报告里的 Qn 还是原来那几条、面板按报告引用的那一份也不变。"""
|
||
from app.core import artifact_store
|
||
from app.engine.evidence_reader import read_sealed
|
||
from app.providers import mock_model
|
||
|
||
asset = await published(client)
|
||
args = {"workflow": asset["workflow_id"], "card": CARD, "metric_ids": ["gmv", "peak"],
|
||
"inputs": dict(MEMBER_INPUTS)}
|
||
decisions = [
|
||
{"assignments": [{"agent": "查数", "instruction": "按全年口径核一遍"}], "done": False, "reason": "先查"},
|
||
{"assignments": [], "done": True, "reason": "够了"},
|
||
]
|
||
turns = {"n": 0}
|
||
|
||
def _decide(self, messages):
|
||
if self.response_format:
|
||
i = turns["n"]
|
||
turns["n"] += 1
|
||
return AIMessage(content=json.dumps(decisions[i] if i < len(decisions) else decisions[-1],
|
||
ensure_ascii=False))
|
||
if self.tools and not any(isinstance(m, ToolMessage) for m in messages):
|
||
return AIMessage(content="查一下口径。",
|
||
tool_calls=[{"name": querycore.CALIBER_QUERY, "args": args, "id": "c1"}])
|
||
return AIMessage(content="按全年口径核过一遍，与周口径口径一致。")
|
||
|
||
monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
|
||
run = await run_manager.start(graph=team_report_graph(asset["workflow_id"]), input_payload={})
|
||
row = await settled(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
|
||
sealed = await read_sealed(row.id)
|
||
# 1) 调用工具节点那两条（tools.py 记的）与成员那两条（这次新加的）各是各的
|
||
look = [e for e in await finished_evidence(row.id, "look") if e.get("kind") == "query"]
|
||
member = [e for e in await finished_evidence(row.id, "team") if e.get("kind") == "query"]
|
||
assert [e["member"] for e in member] == ["查数", "查数"], "一次调用两条查询，各记一条（不重复）"
|
||
assert {e["node_id"] for e in member} == {"team:查数"}, "条目带成员归属"
|
||
assert all(e["exec"] == 1 and e["call_id"] == "c1" and e["tool"] == querycore.CALIBER_QUERY for e in member)
|
||
assert all(e["source"] == S.CURRENT["source"] and e["columns"] and e["rows"] >= 1
|
||
and e["truncated"] is False for e in member)
|
||
assert all(e["query_id"] and e["execution_id"] and e["attempt"] == 1 for e in member)
|
||
assert all(e["origin"]["workflow_id"] == asset["workflow_id"] and e["parameters"]["inputs_hash"]
|
||
for e in member), "复算要的出处与材料跟着条目走"
|
||
assert not ({e["artifact"] for e in member} & {e["artifact"] for e in look}), "两处查的是不同窗口，工件各是各的"
|
||
|
||
# 2) 封存证据（证据面板、导出包从这里认）：成员那两条在 sealed.queries / query_records / entries 里
|
||
member_artifacts = {e["artifact"] for e in member}
|
||
assert member_artifacts <= set(sealed.queries)
|
||
assert all(sealed.queries[a]["tool"] == querycore.CALIBER_QUERY for a in member_artifacts)
|
||
records = sealed.query_records_of("team")
|
||
assert {r["execution_id"] for r in records} == {e["execution_id"] for e in member}
|
||
assert [e.get("member") for e in sealed.entries if e.get("artifact") in member_artifacts] == ["查数", "查数"]
|
||
|
||
# 3) 报告目录的全局 Q 编号一个字不变：只有调用工具节点那两条，成员的查数不在目录里
|
||
write = sealed.payload_of(["write"])
|
||
doc = artifact_store.load(str(write["doc_artifact"]))
|
||
aliases = {a: e.get("artifact") for a, e in (doc.get("catalog") or {}).items() if e.get("kind") == "query"}
|
||
assert set(aliases) == {"Q1", "Q2"}, "成员的查询不重编报告里的编号"
|
||
assert set(aliases.values()) == {e["artifact"] for e in look}
|
||
assert not (member_artifacts & set(aliases.values()))
|
||
|
||
# 4) 导出包（可验证报告）也把成员查到的快照算进这次运行的查询：以前成员那条路连工具快照都不落，只能看事件里的预览，
|
||
# 长一点的返回被截断就认不出来（打包时按「认不出」宁可多遮），来源也就认不出
|
||
from app.bundle.build import _query_artifacts, _query_sources
|
||
|
||
events = list(sealed.events)
|
||
assert member_artifacts <= set(_query_artifacts(events))
|
||
assert all(_query_sources(events).get(a) == S.CURRENT["source"] for a in member_artifacts)
|
||
page = await client.get(f"/api/runs/{row.id}/bundle.html")
|
||
assert page.status_code == 200, page.text
|
||
|
||
|
||
async def test_a_member_without_queries_leaves_the_team_output_unchanged(client, monkeypatch):
|
||
"""普通成员（没查库）的产出与以前一字不差：团队产出不带 evidence 键，transcript 里也不多东西。"""
|
||
from app.providers import mock_model
|
||
|
||
asset = await published(client)
|
||
agents = [{"name": "撰稿", "description": "写一段话", "system": "你撰稿", "max_steps": 1, "tools": []}]
|
||
graph = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("team", "supervisor", goal="写一句", agents=agents, max_rounds=1, max_parallel=1, approval="never"),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.team.text }}"}]),
|
||
)
|
||
|
||
turns = {"n": 0}
|
||
|
||
def _decide(self, messages):
|
||
if self.response_format:
|
||
turns["n"] += 1
|
||
decided = ({"assignments": [{"agent": "撰稿", "instruction": "写一句"}], "done": False, "reason": "派活"}
|
||
if turns["n"] == 1 else
|
||
{"assignments": [], "done": True, "reason": "够了"})
|
||
return AIMessage(content=json.dumps(decided, ensure_ascii=False))
|
||
return AIMessage(content="写好了。")
|
||
|
||
monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
row = await settled(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
assert await finished_evidence(row.id, "team") == [], "成员没查库，一条台账条目都不记"
|
||
from app.engine.evidence_reader import read_sealed
|
||
|
||
team = (await read_sealed(row.id)).payload_of(["team"])
|
||
[entry] = team["transcript"]
|
||
assert set(entry) == {"round", "agent", "instruction", "text", "tool_calls", "duration_ms"}, entry
|
||
assert "queries" not in team # 团队产出里只有 transcript 那一条，没有别的证据键
|
||
assert asset["workflow_id"]
|
||
|
||
|
||
async def test_a_member_without_an_analysis_tool_still_gets_its_queries_into_the_evidence(client, monkeypatch):
|
||
"""成员只绑口径查询（没绑分析工具）：交给成员的内容一个字不改（没有容器可放编号），查到的数照样进封存证据。"""
|
||
from app.engine.evidence_reader import read_sealed
|
||
from app.providers import mock_model
|
||
|
||
asset = await published(client)
|
||
args = {"workflow": asset["workflow_id"], "card": CARD, "metric_ids": ["gmv", "peak"],
|
||
"inputs": dict(MEMBER_INPUTS)}
|
||
agents = [{"name": "查数", "description": "查口径", "system": "你会查口径", "max_steps": 1,
|
||
"tools": [querycore.CALIBER_QUERY]},
|
||
{"name": "复核", "description": "复核别人算的数", "system": "你复核", "max_steps": 1, "tools": []}]
|
||
graph = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("team", "supervisor", goal="算一下销售额", agents=agents, max_rounds=2, max_parallel=1,
|
||
approval="never"),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.team.text }}"}]),
|
||
)
|
||
decisions = [
|
||
{"assignments": [{"agent": "查数", "instruction": "算 gmv 与 peak"}], "done": False, "reason": "先查"},
|
||
{"assignments": [], "done": True, "reason": "够了"},
|
||
]
|
||
turns = {"n": 0}
|
||
seen: list[str] = []
|
||
|
||
def _decide(self, messages):
|
||
seen.extend(str(m.content) for m in messages if isinstance(m, ToolMessage))
|
||
if self.response_format:
|
||
i = turns["n"]
|
||
turns["n"] += 1
|
||
return AIMessage(content=json.dumps(decisions[i] if i < len(decisions) else decisions[-1],
|
||
ensure_ascii=False))
|
||
if self.tools and not any(isinstance(m, ToolMessage) for m in messages):
|
||
return AIMessage(content="查一下口径。",
|
||
tool_calls=[{"name": querycore.CALIBER_QUERY, "args": args, "id": "c1"}])
|
||
return AIMessage(content="查完了。")
|
||
|
||
monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
row = await settled(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
assert seen and all(not t.startswith("【证据") for t in seen), "没绑分析工具：交回成员的内容一个字不改"
|
||
|
||
entries = [e for e in await finished_evidence(row.id, "team") if e.get("kind") == "query"]
|
||
assert len(entries) == 2 and {e["node_id"] for e in entries} == {"team:查数"}
|
||
assert all(e["member"] == "查数" and e["query_id"] and e["execution_id"] and e["attempt"] == 1
|
||
for e in entries)
|
||
sealed = await read_sealed(row.id)
|
||
assert {e["artifact"] for e in entries} <= set(sealed.queries)
|
||
[record] = (sealed.payload_of(["team"]) or {}).get("transcript", [{}])[0].get("tool_calls") or [{}]
|
||
assert [q["artifact"] for q in record["queries"]] == [e["artifact"] for e in entries]
|
||
assert all("alias" not in q for q in record["queries"]), "没有容器就没有本节点内的编号"
|
||
|
||
|
||
# ==========================================================================
|
||
# 6、复核出来的三条：授权范围继承、逐次执行身份、发起时预解析口径的源
|
||
# ==========================================================================
|
||
|
||
#: 记录表的列（和 tests/test_records.py 同一个形状的假数据）
|
||
RECORD_COLUMNS = [{"name": "门店", "label": "门店", "type": "string", "required": True},
|
||
{"name": "数量", "label": "报损数量", "type": "integer", "required": True}]
|
||
#: 记录表的行级范围认的一列：查全部、只看自己的、一行都不给
|
||
RECORD_COUNT = "SELECT count(*) AS n FROM records"
|
||
|
||
|
||
def records_graph(source_name: str) -> dict[str, Any]:
|
||
"""父 Run 的图：查这张记录表（固定的数据版本里带着它的高水位与行级范围）。"""
|
||
return S.chain(
|
||
S.node("start", "input"),
|
||
S.node("q", "tool", tool=f"db_query__{source_name}", args={"sql": RECORD_COUNT}),
|
||
S.node("out", "output", fields=[{"name": "n", "value": "{{ nodes.q }}"}]),
|
||
)
|
||
|
||
|
||
async def records_source(name: str = "门店报损", *, grants: dict[str, str] | None = None) -> tuple[str, str]:
|
||
"""建一张记录表：所属应用 + 逐人的授权。返回 (源 id, 源名)。grants 缺省是给 uma 的一档 use。"""
|
||
from app.db.models import DataSource
|
||
from app.db.models_apps import App, AppGrant
|
||
from app.records.tables import ensure_table
|
||
|
||
async with SessionLocal() as session:
|
||
app = App(name=f"{name}-{uuid.uuid4().hex[:6]}", created_by="建立人")
|
||
session.add(app)
|
||
await session.flush()
|
||
for who, role in ({"uma": "use"} if grants is None else grants).items():
|
||
session.add(AppGrant(app_id=app.id, principal_kind="user", principal=who, role=role))
|
||
table = await ensure_table(session, app, RECORD_COLUMNS)
|
||
await session.commit()
|
||
source = await session.get(DataSource, table.source_id)
|
||
# 一行 uma 提交的记录：行级范围是 all / mine 时看得见这一行，none 看不见
|
||
store.append(path_of(table), columns_of(table), {"门店": "一店", "数量": 1}, version=table.version,
|
||
run_id=None, actor="uma", action_key=uuid.uuid4().hex)
|
||
return source.id, source.name
|
||
|
||
|
||
async def grant_caliber(workflow_id: str, version: int, who: str, role: str) -> str:
|
||
"""给这个人在**这一版口径**上的资产授权：一份应用（当前发布版精确固定着这一版）+ 一条 AppGrant。返回应用 id。
|
||
|
||
普通使用者按口径取数靠的就是它（apps/access.usable_calibers / caliber_role，K14）：没有这一份，QueryCore 在起 Run
|
||
之前就拒（querycore_not_authorized），工具那一层也拿不到数——记录表的行级授权（records_source 的 grants）不管口径。"""
|
||
from app.db.models_apps import App, AppGrant, AppVersion
|
||
|
||
async with SessionLocal() as session:
|
||
app = App(name=f"口径授权-{uuid.uuid4().hex[:6]}", status="active", published_version=1)
|
||
session.add(app)
|
||
await session.flush()
|
||
session.add(AppVersion(app_id=app.id, version=1, spec={}, spec_hash="grant",
|
||
pins=[{"alias": "cal", "workflow_id": workflow_id, "version": int(version)}]))
|
||
session.add(AppGrant(app_id=app.id, principal_kind="user", principal=who, role=role))
|
||
await session.commit()
|
||
return app.id
|
||
|
||
|
||
async def wait_for(run_id: str, statuses: tuple[str, ...], timeout: float = 30.0) -> Run:
|
||
"""等运行到这几个状态之一（settled 把 interrupted 也当终态，审批那一段要另等）。"""
|
||
deadline = asyncio.get_running_loop().time() + timeout
|
||
while asyncio.get_running_loop().time() < deadline:
|
||
async with SessionLocal() as session:
|
||
row = await session.get(Run, run_id)
|
||
if row is not None and row.status in statuses:
|
||
return row
|
||
await asyncio.sleep(0.05)
|
||
raise AssertionError(f"运行 {run_id} 没有进入 {statuses}")
|
||
|
||
|
||
async def test_inherit_scope_keeps_the_parents_authorized_rows():
|
||
"""继承父 Run 的 data_versions：记录表的行级范围原样带着（all/mine/none 都不动），缺 rows 的按最窄的算。"""
|
||
from app.records.views import inherit_scope
|
||
|
||
versions = {
|
||
"a": {"name": "甲", "origin": "records", "table": "t", "version": 1, "seq": 3, "rows": "all"},
|
||
"b": {"name": "乙", "origin": "records", "table": "t", "version": 1, "seq": 4, "rows": "mine",
|
||
"viewer": "uma"},
|
||
"c": {"name": "丙", "origin": "records", "table": "t", "version": 1, "seq": 5, "rows": "none"},
|
||
"d": {"name": "丁", "origin": "records", "table": "t", "version": 1, "seq": 6}, # 老运行：没有范围
|
||
"e": {"snapshot": "s1", "name": "上传表格"}, # 不是记录表
|
||
}
|
||
inherit_scope(versions)
|
||
assert versions["a"]["rows"] == "all"
|
||
assert versions["b"]["rows"] == "mine" and versions["b"]["viewer"] == "uma"
|
||
assert versions["c"]["rows"] == "none"
|
||
assert versions["d"]["rows"] == "none", "没有范围的老项按最窄的算，绝不当成 all"
|
||
assert versions["e"] == {"snapshot": "s1", "name": "上传表格"}, "上传源不碰"
|
||
|
||
|
||
async def test_a_derived_child_inherits_the_parents_authorized_scope(client):
|
||
"""正式父 Run 的行级范围是 all；派生的子运行是 exploratory，也不能按发起人重算成 mine。"""
|
||
from app.engine.runner import _data_versions # noqa: F401
|
||
from tests.fakes import auth_env
|
||
|
||
await auth_env.turn_on_local()
|
||
try:
|
||
await auth_env.make_user("uma", "user")
|
||
source_id, name = await records_source()
|
||
graph = records_graph(name)
|
||
# 父 Run：正式运行 → 看全部
|
||
parent = await run_manager.start(graph=graph, input_payload={}, run_class="formal", started_by="uma")
|
||
parent = await settled(parent.id)
|
||
assert parent.status == "succeeded", parent.error
|
||
assert parent.data_versions[source_id]["rows"] == "all"
|
||
# 反例：同一个起点自己发起（不继承父 Run）→ 按发起人的 use 档收成 mine。没有这一条，上面那句断言说明不了什么
|
||
own = await settled((await run_manager.start(graph=graph, input_payload={}, run_class="exploratory",
|
||
started_by="uma")).id)
|
||
assert own.data_versions[source_id]["rows"] == "mine"
|
||
assert own.data_versions[source_id]["viewer"] == "uma"
|
||
# 反例（同一条规则直接跑一遍）：拿父 Run 那一份按子运行的类别与发起人重算，all 会收成 mine
|
||
from app.records.views import scope_entries
|
||
|
||
recomputed = copy.deepcopy(parent.data_versions)
|
||
async with SessionLocal() as session:
|
||
await scope_entries(session, recomputed, run_class="exploratory", started_by="uma")
|
||
assert recomputed[source_id]["rows"] == "mine"
|
||
# 派生：照父 Run 那一份原样，不收窄、也不添 viewer
|
||
child = await settled((await run_manager.start(graph=graph, input_payload={}, run_class="exploratory",
|
||
started_by="uma", data_versions=parent.data_versions)).id)
|
||
assert child.status == "succeeded", child.error
|
||
assert child.data_versions[source_id]["rows"] == "all", "父已授权 all，子运行不能收成 mine"
|
||
assert "viewer" not in child.data_versions[source_id]
|
||
assert child.data_versions[source_id]["seq"] == parent.data_versions[source_id]["seq"]
|
||
finally:
|
||
await auth_env.reset()
|
||
|
||
|
||
async def test_a_compose_parent_keeps_its_scope_after_the_user_is_disabled(client):
|
||
"""发起之后人停用了：同一次运行（派生出来的那一段）仍按发起时定下的那一档看，不改判成 none。"""
|
||
from app.db.models import User
|
||
from tests.fakes import auth_env
|
||
|
||
await auth_env.turn_on_local()
|
||
try:
|
||
await auth_env.make_user("uma", "user")
|
||
source_id, name = await records_source()
|
||
graph = records_graph(name)
|
||
parent = await settled((await run_manager.start(graph=graph, input_payload={}, started_by="uma")).id)
|
||
assert parent.data_versions[source_id]["rows"] == "mine"
|
||
async with SessionLocal() as session:
|
||
user = (await session.execute(select(User).where(User.username == "uma"))).scalar_one()
|
||
user.disabled = True
|
||
await session.commit()
|
||
# 现在自己发起一次：人不在（停用）→ 一行都不给
|
||
fresh = await settled((await run_manager.start(graph=graph, input_payload={}, started_by="uma")).id)
|
||
assert fresh.data_versions[source_id]["rows"] == "none"
|
||
# 派生：照父 Run 那一份（mine + 看的人），不因为人后来停用而改判
|
||
child = await settled((await run_manager.start(graph=graph, input_payload={}, run_class="exploratory",
|
||
started_by="uma", data_versions=parent.data_versions)).id)
|
||
assert child.data_versions[source_id]["rows"] == "mine"
|
||
assert child.data_versions[source_id]["viewer"] == "uma"
|
||
finally:
|
||
await auth_env.reset()
|
||
|
||
|
||
# ---- 2) 同一份工件对应两条不同的逻辑查询 ----------------------------------
|
||
|
||
|
||
def same_duration(monkeypatch) -> None:
|
||
"""把每条查询的耗时记成 0：同一条 SQL、同一批数据、同一份参数的两条查询就得到**同一个**工件（内容寻址）。
|
||
|
||
耗时是快照内容的一部分（每条查询各记各的、几毫秒上下的抖动）：这一条测的是「同一份工件对应两条不同的逻辑查询」
|
||
时证据索引的取法，所以先把那个抖动按平，让两份内容真的逐字相同。"""
|
||
import dataclasses
|
||
|
||
from app.tools import datasource
|
||
|
||
real = datasource.run_query
|
||
|
||
async def zeroed(source: Any, sql: str, **kw: Any) -> Any:
|
||
return dataclasses.replace(await real(source, sql, **kw), elapsed_ms=0)
|
||
|
||
monkeypatch.setattr(datasource, "run_query", zeroed)
|
||
|
||
|
||
async def test_two_logical_queries_sharing_one_artifact_are_both_reported(client, monkeypatch):
|
||
"""两个取数节点写同一条 SQL、读同一批数据：内容寻址下是**同一个**工件，但它们是两条不同的逻辑查询。
|
||
|
||
按工件 id 去重的证据索引会把它们并成一条，早先那条就找不到自己那一次——所以逐次记下的执行身份才是取法。"""
|
||
same_duration(monkeypatch)
|
||
asset = await make_workflow(graph=twin_graph())
|
||
await publish(client, asset)
|
||
outcome = await run_once(asset, metric_ids=["left", "right"], inputs=dict(INPUTS))
|
||
assert outcome.ok, [p["code"] for p in outcome.problems]
|
||
data = outcome.as_dict()
|
||
assert [q["query_id"] for q in data["queries"]] == [f"{asset['workflow_id']}:one",
|
||
f"{asset['workflow_id']}:two"]
|
||
assert all(q["found"] for q in data["queries"])
|
||
assert len({q["artifact"] for q in data["queries"]}) == 1, "同一条 SQL 同一批数据 → 同一个工件"
|
||
assert len({q["execution_id"] for q in data["queries"]}) == 2, "两条逻辑查询各有各的执行身份"
|
||
assert all(q["row_count"] == 1 and q["columns"] for q in data["queries"])
|
||
assert {m["id"]: m["value"] for m in outcome.metrics} == {"left": WINDOW_SUM, "right": WINDOW_SUM}
|
||
|
||
|
||
async def test_a_later_execution_of_the_same_sql_takes_its_own_record(client):
|
||
"""同一个节点执行两次（同一条 SQL、同一份数据、同一个工件）：引用取**最后一次**那次执行的身份。"""
|
||
from app.engine.evidence_reader import read_sealed
|
||
|
||
asset = await published(client)
|
||
outcome = await run_once(asset, metric_ids=["gmv"], inputs=dict(INPUTS))
|
||
assert outcome.ok
|
||
sealed = await read_sealed(outcome.run_id)
|
||
records = sealed.query_records_of("sales")
|
||
assert len(records) == 1 and records[-1]["artifact"] == outcome.queries[0]["artifact"]
|
||
# 逐次记录与按工件去重的那一份指向同一次执行（认法不同、结论一致）
|
||
assert sealed.queries[records[-1]["artifact"]]["exec"] == records[-1]["exec"]
|
||
assert querycore._query_record_of(sealed, "sales")["execution_id"] == records[-1]["execution_id"]
|
||
# 没有这个节点的逐次记录时（老运行、只回了 tool.end）退回按工件去重的那一份
|
||
assert querycore._query_record_of(sealed, "shared") is None
|
||
|
||
|
||
# ---- 3) 父 Run 发起时预解析口径那一版读的源 -------------------------------
|
||
|
||
|
||
def book(sheets: dict[str, list[list]]) -> bytes:
|
||
import io
|
||
|
||
from openpyxl import Workbook
|
||
|
||
wb = Workbook()
|
||
wb.remove(wb.active)
|
||
for title, rows in sheets.items():
|
||
ws = wb.create_sheet(title)
|
||
for row in rows:
|
||
ws.append(row)
|
||
buf = io.BytesIO()
|
||
wb.save(buf)
|
||
return buf.getvalue()
|
||
|
||
|
||
#: 上传表格第一版：数量合计 3；第二版（重传）：合计 100，另多一张表
|
||
UP_V1 = {"明细": [["区域", "数量"], ["分区甲", 1], ["分区乙", 2]]}
|
||
UP_V2 = {"明细": [["区域", "数量"], ["分区甲", 60], ["分区乙", 40]], "新增": [["时段", "客流"], ["上午", 7]]}
|
||
|
||
|
||
async def upload_source(sheets: dict[str, list[list]], *, name: str | None = None) -> tuple[str, str, str]:
|
||
"""发布一版上传表格，返回 (源 id, 快照 id, 源名)。给了 name 就是**往同一个源上重传**（新的一版）。"""
|
||
from app.data import table_versions
|
||
|
||
name = name or f"up_{uuid.uuid4().hex[:8]}"
|
||
async with SessionLocal() as session:
|
||
row = (await session.execute(select(DataSource).where(DataSource.name == name))).scalar_one_or_none()
|
||
if row is None:
|
||
row = DataSource(name=name, kind="sqlite", readonly=True, origin="upload", enabled=True)
|
||
session.add(row)
|
||
result = await table_versions.publish_upload(session, row, book(sheets), "客流.xlsx",
|
||
header_row=1, mixed="reject", raw_mode=False)
|
||
return row.id, result.snapshot_id, name
|
||
|
||
|
||
def upload_caliber_graph(source_name: str) -> dict[str, Any]:
|
||
"""口径的取数节点查这张上传表格（国产表名照原样写在 SQL 里）。"""
|
||
return S.chain(
|
||
S.node("start", "input"),
|
||
S.node("fetch", "tool", tool=f"db_query__{source_name}",
|
||
args={"sql": 'SELECT SUM("数量") AS total FROM "明细"'}),
|
||
S.node(CARD, "metrics", caliber="上传合计", caliber_version="v1",
|
||
metrics=[{"id": "total", "name": "合计", "unit": "件",
|
||
"expression": "cell(nodes.fetch, 0, 'total')"}]),
|
||
S.node("write", "report", instructions="写周销售"),
|
||
S.node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}]),
|
||
)
|
||
|
||
|
||
async def test_the_parent_pins_the_sources_its_caliber_reads(client):
|
||
"""绑了 caliber_query 的节点读的**不在本图里**：口径那一版图里的源也要在父 Run 发起时固定下来。"""
|
||
from app.engine.schema import GraphSpec
|
||
from app.engine.runner import _data_versions
|
||
|
||
other_id, _, _ = await upload_source(UP_V1) # 别的源：不许被顺手固定
|
||
source_id, snap1, name = await upload_source(UP_V1)
|
||
asset = await make_workflow(graph=upload_caliber_graph(name))
|
||
await publish(client, asset)
|
||
graph = parent_graph(asset["workflow_id"], metric_ids=["total"], inputs={})
|
||
async with SessionLocal() as session:
|
||
pins = await _data_versions(session, GraphSpec.model_validate(graph))
|
||
assert pins[source_id]["snapshot"] == snap1, "口径那一版的取数源要在父 Run 发起时就固定"
|
||
assert pins[source_id]["name"] == name
|
||
assert other_id not in pins, "没用到的源照旧不固定"
|
||
# 认不到工作流（名字写错了）：这一次调用会被 QueryCore 拒，一条数据都不读——不必把全部源都固定上
|
||
broken = parent_graph("nope-not-a-workflow", metric_ids=["total"], inputs={})
|
||
async with SessionLocal() as session:
|
||
assert await _data_versions(session, GraphSpec.model_validate(broken)) == {}
|
||
|
||
|
||
async def test_an_agent_or_templated_caliber_reference_pins_every_source(client):
|
||
"""口径是 Agent 挑的、或者写的是模板：发起时说不准读哪个源，把启用的上传源全部固定上。"""
|
||
from app.engine.schema import GraphSpec
|
||
from app.engine.runner import _data_versions
|
||
|
||
source_id, snap1, _ = await upload_source(UP_V1)
|
||
agent_graph = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("bot", "agent", prompt="算一下", tools=[querycore.CALIBER_QUERY], approval="never"),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.bot.text }}"}]),
|
||
)
|
||
async with SessionLocal() as session:
|
||
pins = await _data_versions(session, GraphSpec.model_validate(agent_graph))
|
||
assert pins.get(source_id, {}).get("snapshot") == snap1
|
||
templated = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("look", "tool", tool=querycore.CALIBER_QUERY, args={
|
||
"workflow": "{{ vars.w }}", "card": CARD, "metric_ids": ["total"], "inputs": {}}),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.look.ok }}"}]),
|
||
)
|
||
async with SessionLocal() as session:
|
||
pins = await _data_versions(session, GraphSpec.model_validate(templated))
|
||
assert pins.get(source_id, {}).get("snapshot") == snap1
|
||
|
||
|
||
async def test_a_reupload_after_launch_still_reads_the_pinned_version(client):
|
||
"""父 Run 发起之后重传：这次运行里口径读的还是发起时那一版（run.started 里记的就是它）。"""
|
||
source_id, s1, name = await upload_source(UP_V1)
|
||
asset = await make_workflow(graph=upload_caliber_graph(name))
|
||
await publish(client, asset)
|
||
graph = parent_graph(asset["workflow_id"], metric_ids=["total"], inputs={}, gate=True)
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
assert (await wait_for(run.id, ("interrupted",))).status == "interrupted"
|
||
|
||
_, s2, _ = await upload_source(UP_V2, name=name) # 停在确认上时重传：当前版本换成合计 100 的那一版
|
||
assert s2 != s1
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
done = await settled(run.id)
|
||
assert done.status == "succeeded", done.error
|
||
|
||
assert done.data_versions[source_id]["snapshot"] == s1
|
||
started = await events_of(run.id, "run.started")
|
||
assert all(e["data_versions"][source_id]["snapshot"] == s1 for e in started), "run.started 里可见实际固定的版本"
|
||
tool = await tool_result(done.id)
|
||
assert tool["ok"], tool.get("problems")
|
||
assert tool["metrics"][0]["value"] == 3, "重传之后这次运行读的还是发起时那一版"
|
||
assert tool["data"]["data_versions"] == {name: s1}
|
||
|
||
|
||
async def test_two_calls_in_one_parent_run_read_the_same_version(client):
|
||
"""父 Run 的两次调用（中间重传过）读的还是同一版：不是各自在子运行里补固定当前版本。"""
|
||
source_id, s1, name = await upload_source(UP_V1)
|
||
asset = await make_workflow(graph=upload_caliber_graph(name))
|
||
await publish(client, asset)
|
||
args = {"workflow": asset["workflow_id"], "card": CARD, "metric_ids": ["total"], "inputs": {}}
|
||
graph = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("one", "tool", tool=querycore.CALIBER_QUERY, args=dict(args)),
|
||
S.node("gate", "human", mode="approve", title="看一眼"),
|
||
S.node("two", "tool", tool=querycore.CALIBER_QUERY, args=dict(args)),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "{{ nodes.two.ok }}"}]),
|
||
)
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
assert (await wait_for(run.id, ("interrupted",))).status == "interrupted"
|
||
|
||
_, s2, _ = await upload_source(UP_V2, name=name) # 第一次调用之后重传（同一个源的新一版）
|
||
assert s2 != s1
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
done = await settled(run.id)
|
||
assert done.status == "succeeded", done.error
|
||
|
||
first, second = await tool_result(done.id, "one"), await tool_result(done.id, "two")
|
||
assert first["ok"] and second["ok"], (first.get("problems"), second.get("problems"))
|
||
assert first["metrics"][0]["value"] == 3 and second["metrics"][0]["value"] == 3, \
|
||
"同一次运行里的两次调用要读同一版数据"
|
||
assert done.data_versions[source_id]["snapshot"] == s1
|
||
for result in (first, second):
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, result["run_id"])
|
||
assert child.data_versions[source_id]["snapshot"] == s1, "子运行照父 Run 固定的那一版，不各自补固定"
|
||
|
||
|
||
# ==========================================================================
|
||
# 7、晚解析的源固定在**父 Run** 上（发布得晚 / 推荐版变了 / 动态引用）
|
||
# ==========================================================================
|
||
|
||
|
||
def records_caliber_graph(source_name: str) -> dict[str, Any]:
|
||
"""口径的取数节点查这张记录表（读哪些行由这次运行定下的行级范围决定）。"""
|
||
return S.chain(
|
||
S.node("start", "input"),
|
||
S.node("fetch", "tool", tool=f"db_query__{source_name}", args={"sql": RECORD_COUNT}),
|
||
S.node(CARD, "metrics", caliber="记录合计", caliber_version="v1",
|
||
metrics=[{"id": "n", "name": "条数", "unit": "条",
|
||
"expression": "cell(nodes.fetch, 0, 'n')"}]),
|
||
S.node("write", "report", instructions="写周销售"),
|
||
S.node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}]),
|
||
)
|
||
|
||
|
||
def gated_parent_graph(workflow_id: str, *, metric_ids: list[str], card: str = CARD,
|
||
calls: tuple[str, ...] = ("look",)) -> dict[str, Any]:
|
||
"""父 Run 的图：入口 → 人工确认 →（一到两处）调用工具（caliber_query）→ 出口。
|
||
|
||
审批挡在调用之前：测试要在「发起」与「第一次调用」之间做事（发布口径、重传、换推荐版）。"""
|
||
parts = [S.node("start", "input"), S.node("gate", "human", mode="approve", title="看一眼")]
|
||
for name in calls:
|
||
parts.append(S.node(name, "tool", tool=querycore.CALIBER_QUERY, args={
|
||
"workflow": workflow_id, "card": card, "metric_ids": metric_ids, "inputs": {}}))
|
||
parts.append(S.node("out", "output", fields=[{"name": "答", "value": "运行结束"}]))
|
||
return S.chain(*parts)
|
||
|
||
|
||
async def pin_events(run_id: str) -> list[dict[str, Any]]:
|
||
"""这次运行里补固定数据版本的那些事件（tools/datasource.LATE_PIN_CODE）。"""
|
||
from app.tools.datasource import LATE_PIN_CODE
|
||
|
||
return [e for e in await events_of(run_id, "log") if e.get("code") == LATE_PIN_CODE]
|
||
|
||
|
||
@pytest.mark.parametrize("run_class, grant, rows, viewer", [
|
||
("formal", "use", "all", None), # 正式父 Run：看全部
|
||
("exploratory", "use", "mine", "uma"), # 交互：use 档只看自己的
|
||
("exploratory", "observe", "all", None), # 交互：observe 档看全部
|
||
("exploratory", None, "none", None), # 交互：没有口径授权，这一次调用在起 Run 之前就被拒
|
||
])
|
||
async def test_a_caliber_published_after_launch_is_pinned_on_the_parent(client, run_class, grant, rows, viewer):
|
||
"""新发布：口径在父 Run 发起时还没有推荐版（发起时算不到），第一次调用时按**父 Run 的**类别与授权补固定。
|
||
|
||
固定记在父 Run 上、追加一条补固定事件（run.started 已经落库），子运行照父 Run 那一份继承——不能落到子运行里
|
||
按 exploratory 重算（那会把正式父 Run 的 all 收成 mine/none）。
|
||
|
||
exploratory 这一档（普通使用者按口径取数）另有一道 K14 的资产授权：发布之后另给 uma 一份与这里那一档相同的授权
|
||
（grant_caliber）；**没有**口径授权的那一档调用在起 Run 之前就被拒，一行都不读、也不补固定。"""
|
||
from app.engine.runner import _data_versions
|
||
from app.engine.schema import GraphSpec
|
||
from tests.fakes import auth_env
|
||
|
||
admin = await auth_env.turn_on_local()
|
||
try:
|
||
await auth_env.make_user("uma", "user")
|
||
source_id, name = await records_source(grants={} if grant is None else {"uma": grant})
|
||
asset = await make_workflow(graph=records_caliber_graph(name)) # 建了，但**还没发布**
|
||
graph = gated_parent_graph(asset["workflow_id"], metric_ids=["n"])
|
||
async with SessionLocal() as session:
|
||
assert await _data_versions(session, GraphSpec.model_validate(graph)) == {}, "还没发布：发起时算不到"
|
||
run = await run_manager.start(graph=graph, input_payload={}, run_class=run_class, started_by="uma")
|
||
assert (await wait_for(run.id, ("interrupted",))).status == "interrupted"
|
||
|
||
if run_class == "formal": # 正式父 Run 走 formal judge：依据要齐
|
||
await confirm(asset, await add_record(asset, metric_ids=("n",)))
|
||
await publish(admin, asset) # 停在确认上时发布：推荐版出现了
|
||
if grant is not None:
|
||
await grant_caliber(asset["workflow_id"], 1, "uma", grant)
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
done = await settled(run.id)
|
||
if grant is None:
|
||
assert done.status == "failed" and querycore.NOT_AUTHORIZED in str(done.error or ""), done.error
|
||
assert source_id not in (done.data_versions or {}), "被拒的这一次没有固定任何数据版本"
|
||
assert await pin_events(done.id) == [], "被拒的这一次连一条补固定事件都不该有"
|
||
return
|
||
assert done.status == "succeeded", done.error
|
||
|
||
entry = done.data_versions[source_id]
|
||
assert entry["rows"] == rows, "行级范围取父 Run 已授权的那一档，不按子运行的类别重算"
|
||
assert entry.get("viewer") == viewer
|
||
assert entry["late"] is True and entry["caliber"] == f"{asset['workflow_id']}@1:{CARD}"
|
||
events = await pin_events(done.id)
|
||
assert len(events) == 1 and events[0]["source"] == name and events[0]["rows"] == rows
|
||
assert events[0]["caliber"] == f"{asset['workflow_id']}@1:{CARD}" and events[0]["late"] is True
|
||
# 子运行照父 Run 那一份继承：不是它自己按 exploratory 另算一份
|
||
tool = await tool_result(done.id)
|
||
assert tool["ok"], tool.get("problems")
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, tool["run_id"])
|
||
assert child.data_versions[source_id]["rows"] == rows
|
||
assert child.data_versions[source_id]["caliber"] == f"{asset['workflow_id']}@1:{CARD}"
|
||
assert tool["metrics"][0]["value"] == (1 if rows != "none" else 0)
|
||
finally:
|
||
await admin.aclose()
|
||
await auth_env.reset()
|
||
|
||
|
||
async def test_later_calls_reuse_the_first_pin(client):
|
||
"""同父 Run 的两次调用：第一次把源固定在父 Run 上，第二次照旧复用那一版，不重新固定、不再补一条事件。"""
|
||
from app.engine.runner import _data_versions
|
||
from app.engine.schema import GraphSpec
|
||
from tests.fakes import auth_env
|
||
|
||
admin = await auth_env.turn_on_local()
|
||
try:
|
||
await auth_env.make_user("uma", "user")
|
||
source_id, _ = await records_source()
|
||
asset = await make_workflow(graph=records_caliber_graph(await _records_name(source_id)))
|
||
graph = gated_parent_graph(asset["workflow_id"], metric_ids=["n"], calls=("one", "two"))
|
||
async with SessionLocal() as session:
|
||
assert await _data_versions(session, GraphSpec.model_validate(graph)) == {}
|
||
run = await run_manager.start(graph=graph, input_payload={}, run_class="formal", started_by="uma")
|
||
assert (await wait_for(run.id, ("interrupted",))).status == "interrupted"
|
||
await confirm(asset, await add_record(asset, metric_ids=("n",))) # 正式父 Run：依据要齐
|
||
await publish(admin, asset)
|
||
# 正式运行也要这一版口径的资产授权（K14）：发布之后按同一档补上
|
||
await grant_caliber(asset["workflow_id"], 1, "uma", "use")
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
done = await settled(run.id)
|
||
assert done.status == "succeeded", done.error
|
||
|
||
entry = done.data_versions[source_id]
|
||
assert entry["rows"] == "all" and entry["late"] is True
|
||
assert len(await pin_events(done.id)) == 1, "两次调用只补固定一次：后面那次复用第一版"
|
||
first, second = await tool_result(done.id, "one"), await tool_result(done.id, "two")
|
||
assert first["ok"] and second["ok"], (first.get("problems"), second.get("problems"))
|
||
assert first["data"]["data_versions"] == second["data"]["data_versions"]
|
||
for tool in (first, second):
|
||
async with SessionLocal() as session:
|
||
child = await session.get(Run, tool["run_id"])
|
||
assert child.data_versions[source_id]["seq"] == entry["seq"]
|
||
assert child.data_versions[source_id]["rows"] == "all"
|
||
finally:
|
||
await admin.aclose()
|
||
await auth_env.reset()
|
||
|
||
|
||
async def _records_name(source_id: str) -> str:
|
||
async with SessionLocal() as session:
|
||
return str((await session.get(DataSource, source_id)).name)
|
||
|
||
|
||
async def test_a_new_recommended_version_with_a_moved_source_is_refused(client):
|
||
"""审批期间推荐版换了一版、源也重传了：这一次要执行的口径引用的是另一版——明确说冲突，不静默跨版。"""
|
||
source_id, s1, name = await upload_source(UP_V1)
|
||
asset = await make_workflow(graph=upload_caliber_graph(name))
|
||
await publish(client, asset) # v1 是推荐版
|
||
graph = gated_parent_graph(asset["workflow_id"], metric_ids=["total"])
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
row = await wait_for(run.id, ("interrupted",))
|
||
assert row.status == "interrupted"
|
||
assert row.data_versions[source_id]["snapshot"] == s1 # 发起时固定在这一版
|
||
assert row.data_versions[source_id]["caliber"] == f"{asset['workflow_id']}@1:{CARD}"
|
||
|
||
_, s2, _ = await upload_source(UP_V2, name=name) # 重传：当前版本换成合计 100 的那一版
|
||
changed = upload_caliber_graph(name)
|
||
next(metrics for metrics in changed["nodes"] if metrics["id"] == CARD)["data"]["config"]["metrics"][0]["label"] = "合计2"
|
||
asset = await add_version(asset, changed, version=2)
|
||
await publish(client, asset, version=2) # 推荐版换成 v2（引用的是重传后的那一版）
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
done = await settled(run.id)
|
||
|
||
assert done.status == "failed", done.error
|
||
error = done.error or ""
|
||
assert querycore.DATA_VERSION_CONFLICT in error, error
|
||
assert name in error and s1[:12] in error and s2[:12] in error, error
|
||
async with SessionLocal() as session:
|
||
children = (await session.execute(select(Run).where(Run.origin.is_not(None)))).scalars().all()
|
||
assert [c for c in children if (c.origin or {}).get("parent_run_id") == done.id] == [], "冲突就不发起子运行"
|
||
|
||
|
||
async def test_two_calls_after_a_late_pin_read_the_same_version(client):
|
||
"""新发布的口径：第一次调用把源固定在父 Run 上，两次调用之间重传，第二次读的还是那一版（同一次运行不跨版）。"""
|
||
source_id, s1, name = await upload_source(UP_V1)
|
||
asset = await make_workflow(graph=upload_caliber_graph(name)) # 还没发布
|
||
args = {"workflow": asset["workflow_id"], "card": CARD, "metric_ids": ["total"], "inputs": {}}
|
||
graph = S.chain(
|
||
S.node("start", "input"),
|
||
S.node("gate1", "human", mode="approve", title="看一眼"),
|
||
S.node("one", "tool", tool=querycore.CALIBER_QUERY, args=dict(args)),
|
||
S.node("gate2", "human", mode="approve", title="再确认"),
|
||
S.node("two", "tool", tool=querycore.CALIBER_QUERY, args=dict(args)),
|
||
S.node("out", "output", fields=[{"name": "答", "value": "结束"}]),
|
||
)
|
||
run = await run_manager.start(graph=graph, input_payload={})
|
||
assert (await wait_for(run.id, ("interrupted",))).status == "interrupted"
|
||
|
||
await publish(client, asset) # 新发布：发起时算不到
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
assert (await wait_for(run.id, ("interrupted",))).status == "interrupted"
|
||
|
||
_, s2, _ = await upload_source(UP_V2, name=name) # 两次调用之间重传
|
||
assert s2 != s1
|
||
await run_manager.resume(run.id, {"approved": True})
|
||
done = await settled(run.id)
|
||
assert done.status == "succeeded", done.error
|
||
|
||
assert done.data_versions[source_id]["snapshot"] == s1
|
||
assert len(await pin_events(done.id)) == 1, "第一次调用补固定一次，第二次复用"
|
||
first, second = await tool_result(done.id, "one"), await tool_result(done.id, "two")
|
||
assert first["ok"] and second["ok"], (first.get("problems"), second.get("problems"))
|
||
assert [first["metrics"][0]["value"], second["metrics"][0]["value"]] == [3, 3], \
|
||
"同一次运行里的两次调用读同一版数据"
|