mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 04:31:58 +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。
655 lines
39 KiB
Python
655 lines
39 KiB
Python
"""组合式问数：use_caliber 选用已发布口径，服务端按那一版确定性内联实例化（next-round 02 做法 4，ledger/compose.py）。
|
||
|
||
- 实例化是确定的：同一版、同一个操作得到一模一样的节点；取数节点照抄、id 加前缀、模板按参数渲染成字面 SQL，口径卡 id
|
||
就是操作 id、caliber_from 带改名映射；取数首尾相接再接口径卡，入口连到第一个取数节点；
|
||
- 参数按口径入口的字段校验（未知字段、缺必填、类型、引号里的值带引号、裸写的值不是数），不合规拒收、原因交回模型；
|
||
只能选这一轮查看到的那一版、能组合的卡、卡上的指标，id 不能撞；
|
||
- 问数据的查看对象多了口径：给入口字段和「能不能组合」，不给 SQL；「什么时候该看」换成问数据那一份；使用者不给；
|
||
- K08 S4：op.metrics 按**源卡顺序**归一（Composed.metrics、口径卡上记的 metrics、selection_hash 与执行数组都按那一份，
|
||
调用方传的顺序与重复项都不影响）；选用选项按**每个指标**判能不能组合（B 独有那条查询坏了只让 B 不可选，A 照旧给组合
|
||
写法），真正的实例化仍按同一条选择逐项硬校验；
|
||
- 走一遍问数据的生成（剧本模型）再真跑出来：比对器把这些数标成 composed「助手选用」，整题 how=composed。
|
||
数据是虚构店铺库。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import hashlib
|
||
import json
|
||
|
||
import pytest
|
||
from httpx import ASGITransport, AsyncClient
|
||
|
||
from app.api import copilot_context
|
||
from app.api.copilot_gen import GraphApplier, GraphChecker
|
||
from app.db.base import SessionLocal
|
||
from app.engine.runner import run_manager
|
||
from app.engine.schema import GraphSpec, validate_graph
|
||
from app.ledger import compose
|
||
from app.ledger import match as ledger_match
|
||
from app.ledger import params, store
|
||
from app.look.calibers import CaliberKind
|
||
from app.main import app
|
||
from tests import ledger_samples as S
|
||
|
||
|
||
@pytest.fixture(autouse=True)
|
||
async def engine_up(tmp_path):
|
||
await run_manager.setup()
|
||
await S.fresh_source(tmp_path / "shop.db")
|
||
yield
|
||
await ledger_match.drain()
|
||
await run_manager.shutdown()
|
||
|
||
|
||
@pytest.fixture
|
||
async def client():
|
||
async with AsyncClient(transport=ASGITransport(app=app), base_url="http://test") as c:
|
||
yield c
|
||
|
||
|
||
def use(wf: str, **extra) -> dict:
|
||
return {"op": "use_caliber", "id": "c1", "workflow_id": wf, "version": 1, "metrics": ["gmv"],
|
||
"inputs": {"start": "2026-09-01", "end": "2026-09-08"}, **extra}
|
||
|
||
|
||
async def composer() -> compose.Composer:
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
return compose.Composer(list(kind.options))
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 实例化
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
async def test_instantiation_is_deterministic_and_copies_the_pinned_fetch():
|
||
wf = await S.publish()
|
||
made = (await composer()).compose(use(wf), taken={"start"})
|
||
again = (await composer()).compose(use(wf), taken={"start"})
|
||
assert made.nodes == again.nodes and made.edges == again.edges
|
||
fetch, card = made.nodes
|
||
assert fetch["id"] == "c1_fetch" and fetch["type"] == "tool"
|
||
assert fetch["data"]["config"]["tool"] == S.tool()
|
||
assert fetch["data"]["config"]["args"]["sql"] == S.TEMPLATE.replace("{{ input.start }}", "2026-09-01").replace(
|
||
"{{ input.end }}", "2026-09-08")
|
||
# 值渲染进去的那一次已经发生过：节点上记一份参数契约（ledger/params 的 marker），运行时的模板渲染不再碰 args——
|
||
# 不然值里的 {{ }} 会被再解释一遍（K01 的二次渲染反例）
|
||
marker = fetch["data"]["config"]["parameters"]
|
||
assert marker["mode"] == params.LITERAL_ARGS and marker["adapter"] == params.ADAPTER == "inline_text"
|
||
# 源模板正文不写在节点上（engine/variables 会扫节点配置里的 {{ }}，把占位符报成没人产出）：记哈希 + 出处
|
||
assert "template" not in marker
|
||
assert marker["template_hash"] == hashlib.sha256(S.TEMPLATE.encode("utf-8")).hexdigest()
|
||
# 契约与它注解的 args 绑定：哈希对得上才算（节点被复制、args 改过就作废，运行时回退正常渲染）
|
||
assert marker["rendered_hash"] == params.rendered_hash(fetch["data"]["config"]["args"])
|
||
assert params.literal_keys(fetch["data"]["config"], "tool") == ("args",)
|
||
assert marker["parameters"] == [{"name": "start", "kind": "text", "declared": "string", "slot": "quoted"},
|
||
{"name": "end", "kind": "text", "declared": "string", "slot": "quoted"}]
|
||
assert [d["name"] for d in marker["declared"]] == ["start", "end", "store"]
|
||
assert marker["source"]["workflow_id"] == wf and marker["source"]["node_id"] == "fetch"
|
||
assert card["id"] == "c1" and card["type"] == "metrics" and card["data"]["label"] == "card"
|
||
# 这张卡这次算哪个指标也记在 口径卡来源 上（K08 S2：metrics 是这次的选择；整卡时这一项不写）
|
||
assert card["data"]["config"]["caliber_from"] == {"workflow_id": wf, "workflow_version": 1, "node_id": "card",
|
||
"nodes": {"fetch": "c1_fetch"}, "metrics": ["gmv"]}
|
||
assert made.edges == [{"source": "c1_fetch", "target": "c1"}]
|
||
# 这次要执行的那条取数只用到 start/end：没用到的 store 不进这次的值（也不填它的缺省值）
|
||
assert made.inputs == {"start": "2026-09-01", "end": "2026-09-08"}
|
||
result = made.result()
|
||
assert result["status"] == "ok" and result["card_id"] == "c1" and result["fetches"] == ["c1_fetch"]
|
||
assert result["level_text"] == "受管" and result["metrics"] == [{"id": "gmv", "name": "销售额"}]
|
||
assert "sql" not in json.dumps(result).lower()
|
||
|
||
|
||
@pytest.mark.parametrize(("change", "words"), [
|
||
({"workflow_id": "nope"}, "只能选用这一轮查看到的已发布口径"),
|
||
({"version": 2}, "能选用的是 v1"),
|
||
({"card": "other"}, "没有口径卡「other」"),
|
||
({"id": "C 1"}, "简短的英文小写"),
|
||
({"metrics": []}, "metrics 要写"),
|
||
({"metrics": ["profit"]}, "没有指标「profit」"),
|
||
({"inputs": {"start": "2026-09-01", "end": "2026-09-08", "city": "x"}}, "入口没有字段「city」"),
|
||
# 这一版入口声明过、只是不在这次选中的指标的范围里：说法不一样（K08 复核），照样拒
|
||
({"inputs": {"start": "2026-09-01", "end": "2026-09-08", "store": "east"}}, "不在这次选用的范围里"),
|
||
({"inputs": {"start": "2026-09-01"}}, "请填写「end」"),
|
||
({"inputs": {"start": "2026-09-01' OR '1'='1", "end": "2026-09-08"}}, "引号里，值里不能有引号"),
|
||
({"inputs": "start=2026-09-01"}, "inputs 要写成"),
|
||
])
|
||
async def test_bad_choices_are_refused_with_a_reason(change, words):
|
||
wf = await S.publish()
|
||
op = use(wf, **change) if change.get("workflow_id") != "nope" else {**use(wf), **change}
|
||
with pytest.raises(compose.ComposeRefused) as e:
|
||
(await composer()).compose(op, taken={"start"})
|
||
assert words in e.value.message, e.value.message
|
||
|
||
|
||
async def test_bare_slots_take_numbers_only_and_ids_must_not_clash():
|
||
graph = S.caliber_graph(sql="SELECT SUM(amount) AS gmv FROM orders WHERE status = 'paid' AND id > {{ input.after }}")
|
||
# 字段没写类型（按文字收）：表单校验放行，裸写进 SQL 的只能是数
|
||
graph["nodes"][0]["data"]["config"]["fields"] = [{"name": "after", "default": "0"}]
|
||
wf = await S.publish(graph=graph)
|
||
pool = await composer()
|
||
with pytest.raises(compose.ComposeRefused, match="只能是数"):
|
||
pool.compose({**use(wf), "inputs": {"after": "1 OR 1=1"}}, taken=set())
|
||
made = pool.compose({**use(wf), "inputs": {"after": "3"}}, taken=set())
|
||
fetch = made.nodes[0]["data"]["config"]
|
||
assert fetch["args"]["sql"].endswith("id > 3")
|
||
assert fetch["parameters"]["parameters"] == [{"name": "after", "kind": "text", "declared": "string", "slot": "bare"}]
|
||
with pytest.raises(compose.ComposeRefused, match="撞了"):
|
||
pool.compose({**use(wf), "inputs": {"after": "3"}}, taken={"c1"})
|
||
|
||
|
||
async def test_a_card_that_reads_model_output_is_listed_but_not_composable():
|
||
graph = S.chain(S.node("start", "input"), S.node("ask", "agent", tools=[S.tool()], assign_to="kpi"),
|
||
S.node("card", "metrics", caliber="周销售", metrics=[{"id": "gmv", "expression": "vars.kpi.gmv"}]),
|
||
S.node("out", "output", fields=[{"name": "r", "value": "{{ nodes.card.text }}"}]))
|
||
wf = await S.publish(graph=graph)
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
[option] = [o for o in kind.options if o.workflow_id == wf]
|
||
# 读模型写的数：这个指标一个也不可选（原因照实给，不是「没找到」）
|
||
assert not option.composable and "输入不全是查库结果里的格" in option.reason
|
||
assert [m["metric_id"] for m in option.unavailable_metrics] == ["gmv"]
|
||
assert option.unavailable_metrics[0]["reason"] in option.reason
|
||
text, _ = await kind.detail([(i, "named") for i in kind.items if i.workflow_id == wf], 4000)
|
||
assert "不能组合" in text and "输入不全是查库结果里的格" in text
|
||
with pytest.raises(compose.ComposeRefused, match="不能组合"):
|
||
compose.Composer(list(kind.options)).compose(use(wf), taken=set())
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 查看：入口字段、能不能组合，不给 SQL
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
async def test_the_answer_look_shows_fields_and_how_to_compose_but_no_sql():
|
||
wf = await S.publish()
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
plain = await CaliberKind.load(session)
|
||
objects = [(i, "named") for i in kind.items if i.workflow_id == wf]
|
||
text, [record] = await kind.detail(objects, 4000)
|
||
# 入口字段只列这次可选指标要用的那几个（store 没有指标用得到，不列）
|
||
assert "入口字段：start（string，必填）；end（string，必填）" in text and "store" not in text
|
||
assert "能选用的指标：销售额（gmv，要 start、end）" in text and "用不了的指标" not in text
|
||
assert '"op": "use_caliber"' in text and f'"workflow_id": "{wf}"' in text and '"card": "card"' in text
|
||
assert "SELECT" not in text.upper() and "orders" not in text
|
||
assert record["kind"] == "caliber" and record["named"]
|
||
# 蓝图助手那一支（不组合）照旧给复用写法，没有组合写法
|
||
plain_text, _ = await plain.detail([(i, "named") for i in plain.items if i.workflow_id == wf], 4000)
|
||
assert "复用写法" in plain_text and "use_caliber" not in plain_text
|
||
|
||
|
||
async def test_the_answer_prompt_switches_only_when_calibers_can_be_composed():
|
||
from app.look.service import LookService
|
||
|
||
assert copilot_context.LOOK_WHEN_ANSWER_PROMPT != copilot_context.LOOK_WHEN_PROMPT
|
||
assert "use_caliber" in copilot_context.LOOK_WHEN_ANSWER_PROMPT and "含退款" in copilot_context.LOOK_WHEN_ANSWER_PROMPT
|
||
async with SessionLocal() as session:
|
||
composing = LookService([await CaliberKind.load(session, compose=True)])
|
||
plain = LookService([await CaliberKind.load(session)])
|
||
assert copilot_context.look_when(composing) == copilot_context.LOOK_WHEN_ANSWER_PROMPT
|
||
assert copilot_context.look_when(plain) == copilot_context.LOOK_WHEN_PROMPT
|
||
assert copilot_context.composer(plain) is None and copilot_context.composer(composing) is not None
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 落地：GraphApplier
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
async def test_the_applier_lands_the_nodes_wires_the_entry_and_reports_refusals():
|
||
wf = await S.publish()
|
||
nodes: dict = {}
|
||
edges: list = []
|
||
applier = GraphApplier(nodes, edges, composer=await composer())
|
||
op = use(wf)
|
||
applied = applier.apply(op)
|
||
assert applied.changed and applied.forward and op["result"]["status"] == "ok"
|
||
assert set(nodes) == {"c1_fetch", "c1"} and edges == [{"source": "c1_fetch", "target": "c1", "sourceHandle": None}]
|
||
# 入口后加的：加上时接到第一个取数节点；模型再发一条同样的连线不重复
|
||
applier.apply({"op": "add_node", "node": {"id": "start", "type": "input", "label": "入口", "config": {}}})
|
||
assert {"source": "start", "target": "c1_fetch", "sourceHandle": None} in edges
|
||
assert not applier.apply({"op": "add_edge", "edge": {"source": "start", "target": "c1_fetch"}}).changed
|
||
assert len([e for e in edges if e["target"] == "c1_fetch"]) == 1
|
||
# 接着写时重发同一个选用：不再落地
|
||
applier.continuing()
|
||
assert not applier.apply(use(wf)).changed
|
||
assert "c1" in applier.landed()
|
||
|
||
bad = use(wf, id="c2", inputs={"start": "2026-09-01"})
|
||
refused = applier.apply(bad)
|
||
assert not refused.changed and refused.forward and bad["result"]["status"] == "refused"
|
||
checker = GraphChecker(applier)
|
||
first = await checker.issues()
|
||
assert [i["code"] for i in first if i["code"] == compose.REFUSED] == [compose.REFUSED]
|
||
assert "请填写「end」" in first[0]["message"] and "重新输出 use_caliber" in first[0]["for_model"]
|
||
# 交回过一次就不再交
|
||
assert not [i for i in await checker.issues() if i["code"] == compose.REFUSED]
|
||
|
||
# 兜底重放到上一轮的图上：拷进来的节点和接线照样在
|
||
applier.replay_onto({"nodes": [{"id": "start", "type": "input", "data": {"config": {}}}], "edges": []})
|
||
assert {"c1_fetch", "c1"} <= set(nodes) and any(e["target"] == "c1_fetch" for e in edges)
|
||
|
||
|
||
async def test_without_a_composer_use_caliber_is_dropped_like_any_unknown_op():
|
||
nodes: dict = {}
|
||
applier = GraphApplier(nodes, [])
|
||
applied = applier.apply({"op": "use_caliber", "id": "c1", "workflow_id": "w", "version": 1})
|
||
assert not applied.changed and not applied.forward and nodes == {}
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 问数据走一遍：生成 → 运行 → 比对标 composed
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def _answer_script(wf: str) -> list[dict]:
|
||
from tests.test_copilot_answer_ops import _edge, _node
|
||
|
||
return [
|
||
{"op": "plan", "summary": "选用周销售口径"},
|
||
_node("start", "input", {}),
|
||
use(wf),
|
||
_node("write", "report", {"instructions": "回答：{{ input.question }}"}),
|
||
_edge("c1", "write"),
|
||
_node("out", "output", {"fields": [{"name": "答案", "value": "{{ nodes.write.text }}"}]}),
|
||
_edge("write", "out"),
|
||
{"op": "done", "explanation": "按口径周销售回答", "run": True},
|
||
]
|
||
|
||
|
||
async def test_an_answer_built_with_use_caliber_is_labelled_composed(client, monkeypatch):
|
||
from tests.test_copilot_answer_ops import BODY, _Scripted, _sse, _use
|
||
|
||
wf = await S.publish()
|
||
_use(monkeypatch, _Scripted(_answer_script(wf), []))
|
||
ops = await _sse(client, {**BODY, "instruction": "九月第一周卖了多少"})
|
||
[chosen] = [o for o in ops if o["op"] == "use_caliber"]
|
||
assert chosen["result"]["status"] == "ok" and chosen["result"]["workflow_name"] == "周销售"
|
||
assert not [o for o in ops if o["op"] == "add_node" and o["node"]["id"] in ("c1", "c1_fetch")], \
|
||
"拷进来的节点不另外转给界面：过程里只有 use_caliber 那一行"
|
||
[final] = [o for o in ops if o["op"] == "final"]
|
||
assert final["next"] == "run", final["issues"]
|
||
graph = final["graph"]
|
||
assert {n["id"] for n in graph["nodes"]} == {"start", "c1_fetch", "c1", "write", "out"}
|
||
assert not [i for i in validate_graph(GraphSpec.model_validate(graph)).issues if i.level == "error"]
|
||
# 走过应用器、写进最终图之后，参数契约还认得出它注解的就是这条 args（K01 的哈希绑定不靠内存里的对象）
|
||
fetch = next(n for n in graph["nodes"] if n["id"] == "c1_fetch")["data"]["config"]
|
||
assert params.literal_keys(fetch, "tool") == ("args",)
|
||
assert fetch["parameters"]["rendered_hash"] == params.rendered_hash(fetch["args"])
|
||
|
||
S.script(monkeypatch, "九月第一周销售额 [[m:gmv]]。")
|
||
run = await run_manager.start(graph=graph, input_payload={"question": "九月第一周卖了多少"}, origin_kind="chat")
|
||
row = await S.wait(run.id)
|
||
assert row.status == "succeeded", row.error
|
||
await ledger_match.drain()
|
||
recorded = await store.latest(row.id)
|
||
content = recorded.content
|
||
assert content["algo"] == store.ALGO == "5"
|
||
[query] = content["queries"]
|
||
assert query["status"] == "composed" and query["node_id"] == "c1_fetch"
|
||
[cand] = query["candidates"]
|
||
assert cand["status"] == "proven" and cand["how"] == "composed" and cand["caliber"] == f"{wf}@1"
|
||
[label] = content["reports"]["write"]["labels"].values()
|
||
assert label["kind"] == "caliber" and label["how"] == "composed" and label["queries"] == [query["artifact"]]
|
||
assert content["whole"]["equivalent"] is True and content["whole"]["how"] == "composed"
|
||
assert recorded.content and (await store.latest(row.id)).trigger == "sealed"
|
||
async with SessionLocal() as session:
|
||
from app.db.models import Artifact
|
||
|
||
ref = await session.get(Artifact, (recorded.artifact, row.id))
|
||
assert ref.meta["calibers"] == [f"{wf}@1"] and ref.meta["counts"]["composed"] == 1
|
||
|
||
|
||
async def test_users_get_the_authorized_calibers_but_no_use_caliber(monkeypatch):
|
||
"""使用者（开登录后的 user 角色）：问数据里**看得到**自己有授权的口径（K14），但没有 use_caliber 那条路。
|
||
|
||
看得到的是「按资产授权收窄过的清单 + 入口字段 + 按口径查询工具取数的写法」（look/calibers 的 query 那一档），不是把口径的
|
||
取数 SQL 拷进自己的图（那是搭建者及以上那一档，compose）。"""
|
||
import app.api.copilot as copilot
|
||
from app.api.copilot import GenerateIn
|
||
from app.api import copilot_context
|
||
|
||
seen: dict = {}
|
||
|
||
async def spy(session, payload, sources, **kw):
|
||
seen.update(kw)
|
||
raise RuntimeError("到这里就够了")
|
||
|
||
monkeypatch.setattr(copilot, "_look_service", spy)
|
||
monkeypatch.setattr("app.api.deps.restricted", lambda: True)
|
||
with pytest.raises(RuntimeError):
|
||
await copilot.answer_ops(GenerateIn(instruction="x", intent="answer")).__anext__()
|
||
assert seen["kinds"] == ("table", "capability", "caliber")
|
||
assert seen["caliber"] == copilot_context.CALIBER_QUERY, "使用者这一档只给按口径查询工具取数"
|
||
monkeypatch.setattr("app.api.deps.restricted", lambda: False)
|
||
with pytest.raises(RuntimeError):
|
||
await copilot.answer_ops(GenerateIn(instruction="x", intent="answer")).__anext__()
|
||
assert seen["kinds"] == ("table", "capability", "caliber")
|
||
assert seen["caliber"] == copilot_context.CALIBER_COMPOSE, "搭建者及以上照旧能 use_caliber"
|
||
with pytest.raises(RuntimeError):
|
||
await copilot.answer_ops(GenerateIn(instruction="x", intent="build")).__anext__()
|
||
assert seen["kinds"] == ("table", "capability")
|
||
assert seen["caliber"] == copilot_context.CALIBER_LOOK, "画布那一档根本不看口径"
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# card 填什么：协议写清两个 id，写错了交回的原因能一次改对（校验不放宽）
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
def test_the_protocol_says_card_is_the_cards_node_id_and_id_is_the_name_you_give():
|
||
text = copilot_context.USE_CALIBER_PROMPT
|
||
assert '"card":"card"' in text and '"card":"口径卡节点 id"' not in text
|
||
assert "card 填那一版工作流图里口径卡的节点 id" in text and "不是你起的名字" in text
|
||
assert "id 是你给这次选用起的名字" in text and "workflow_id、version、card 原样照抄" in text
|
||
|
||
|
||
def test_a_user_round_refuses_use_caliber_and_points_at_the_query_tool():
|
||
"""使用者那一轮看得见口径但不能选用：use_caliber 照实拒收、自查交回模型改用口径查询工具，不静默丢掉。
|
||
|
||
没有口径这一档的入口（画布、发布前修复、升级）照旧：use_caliber 是认不出的操作，不进拒收。"""
|
||
from app.api.copilot_gen import NO_COMPOSE_TEXT, GraphApplier, _refused_issues
|
||
|
||
op = {"op": "use_caliber", "id": "c1", "workflow_id": "wf", "version": 1, "card": "card",
|
||
"metrics": ["gmv"], "inputs": {}}
|
||
user = GraphApplier({}, [], composer=None, no_compose=True)
|
||
assert user.apply(dict(op)).changed is False
|
||
[issue] = _refused_issues(user)
|
||
assert issue["level"] == "error" and NO_COMPOSE_TEXT in issue["message"]
|
||
assert "use_caliber" in issue["for_model"] and "caliber_query" in issue["for_model"]
|
||
assert user.nodes == {} and user.edges == [], "拒收的选用一个节点、一条连线都不落地"
|
||
|
||
plain = GraphApplier({}, [])
|
||
plain.apply(dict(op))
|
||
assert _refused_issues(plain) == []
|
||
|
||
|
||
async def test_writing_your_own_id_as_the_card_is_refused_with_what_to_write_instead():
|
||
wf = await S.publish()
|
||
pool = await composer()
|
||
with pytest.raises(compose.ComposeRefused) as e:
|
||
pool.compose(use(wf, card="c1"), taken={"start"})
|
||
message = e.value.message
|
||
assert "没有口径卡「c1」" in message and "card 要填那一版工作流图里口径卡的节点 id" in message
|
||
assert "可填的有 「card」（口径卡「card」）" in message, message
|
||
assert "「c1」是你给这次选用起的 id，不是口径卡的节点 id" in message
|
||
# 别的写错（不是自己起的 id）不多那一句
|
||
with pytest.raises(compose.ComposeRefused) as other:
|
||
pool.compose(use(wf, card="other"), taken={"start"})
|
||
assert "是你给这次选用起的 id" not in other.value.message
|
||
# 校验不放宽：照着交回的原因改成节点 id 就过
|
||
made = pool.compose(use(wf, card="card"), taken={"start"})
|
||
assert made.op_id == "c1"
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 只算选中的指标（K08 S2）：op.metrics 贯穿参数校验与拷贝
|
||
# --------------------------------------------------------------------------
|
||
|
||
#: 只有 B 那条查询用的字段（region）：只选 A 时它不该被要求
|
||
FIELDS_AB = [{"name": "start", "required": True}, {"name": "end", "required": True},
|
||
{"name": "buyer", "required": True}]
|
||
SQL_A = ("SELECT SUM(amount) AS gmv FROM orders WHERE status = 'paid' "
|
||
"AND paid_at >= '{{ input.start }}' AND paid_at < '{{ input.end }}'")
|
||
SQL_B = SQL_A + " AND region = '{{ input.buyer }}'"
|
||
|
||
|
||
def two_metric_graph(*, on_missing: bool = False, a_sql: str = SQL_A, b_sql: str = SQL_B) -> dict:
|
||
"""一张卡两个指标、两条各自独立的查询；buyer 只有 B 那条用（A 不需要它）。"""
|
||
card: dict = {"caliber": "双指标", "caliber_version": "v1",
|
||
"metrics": [{"id": "gmv", "name": "销售额", "expression": "cell(nodes.fetch_a, 0, 'gmv')"},
|
||
{"id": "n", "name": "单量", "expression": "cell(nodes.fetch_b, 0, 'n')"}]}
|
||
if on_missing:
|
||
card["on_missing"] = "null"
|
||
return S.chain(S.node("start", "input", fields=FIELDS_AB),
|
||
S.node("fetch_a", "tool", tool=S.tool(), args={"sql": a_sql}),
|
||
S.node("fetch_b", "tool", tool=S.tool(), args={"sql": b_sql}),
|
||
S.node("card", "metrics", **card),
|
||
S.node("out", "output", fields=[{"name": "报告", "value": "{{ nodes.write.text }}"}]),
|
||
S.node("write", "report", instructions="写周销售"))
|
||
|
||
|
||
def shared_graph() -> dict:
|
||
"""两个指标读同一条取数（共享 SQL）：选一个时那条 SQL 保留，还在算另一个。"""
|
||
return S.chain(S.node("start", "input", fields=FIELDS_AB[:2]),
|
||
S.node("fetch", "tool", tool=S.tool(), args={"sql": SQL_A}),
|
||
S.node("card", "metrics", caliber="共享取数", caliber_version="v1",
|
||
metrics=[{"id": "gmv", "name": "销售额", "expression": "cell(nodes.fetch, 0, 'gmv')"},
|
||
{"id": "n", "name": "单量", "expression": "cell(nodes.fetch, 0, 'n')"}]),
|
||
S.node("out", "output"))
|
||
|
||
|
||
def select(wf: str, metrics: list[str], inputs: dict | None = None, *, id: str = "c1") -> dict:
|
||
return {"op": "use_caliber", "id": id, "workflow_id": wf, "version": 1, "card": "card",
|
||
"metrics": metrics, "inputs": inputs if inputs is not None else
|
||
{"start": "2026-09-01", "end": "2026-09-08"}}
|
||
|
||
|
||
async def test_selecting_one_metric_no_longer_asks_for_the_other_query_s_parameters():
|
||
"""op.metrics 决定这次真正要执行哪些取数：只选 A 时不再要求 B 的参数，也只拷 A 那条查询。"""
|
||
wf = await S.publish(graph=two_metric_graph())
|
||
pool = await composer()
|
||
made = pool.compose(select(wf, ["gmv"]), taken={"start"})
|
||
assert [n["id"] for n in made.nodes] == ["c1_fetch_a", "c1"]
|
||
card = made.nodes[-1]["data"]["config"]
|
||
assert card["caliber_from"] == {"workflow_id": wf, "workflow_version": 1, "node_id": "card",
|
||
"nodes": {"fetch_a": "c1_fetch_a"}, "metrics": ["gmv"]}
|
||
assert made.inputs == {"start": "2026-09-01", "end": "2026-09-08"}
|
||
assert made.fetch_ids == ["c1_fetch_a"]
|
||
# 整卡（两个指标）时还是要求 B 的参数：不收窄、也不静默放宽
|
||
with pytest.raises(compose.ComposeRefused, match="buyer"):
|
||
pool.compose(select(wf, ["gmv", "n"], id="c2"), taken={"start"})
|
||
# 结果带上这次选择的身份与排除的查询
|
||
result = made.result()
|
||
assert result["selection"]["mode"] == "explicit" and result["selection"]["metric_ids"] == ["gmv"]
|
||
assert result["selection"]["query_ids"] == [f"{wf}:fetch_a"]
|
||
assert result["selection"]["excluded_query_ids"] == [f"{wf}:fetch_b"]
|
||
assert result["selection"]["selection_hash"] and result["selection"]["closure_hash"]
|
||
assert "shared_queries" not in result
|
||
# 拷进来的图能过画布校验（口径卡上多记的选择不影响它）
|
||
nodes = [S.node("start", "input", fields=[{"name": "x"}]), *made.nodes,
|
||
S.node("write", "report", instructions="答"), S.node("out", "output")]
|
||
issues = validate_graph(GraphSpec.model_validate(S.chain(*nodes))).issues
|
||
assert not [i for i in issues if i.level == "error"], issues
|
||
|
||
|
||
async def test_a_shared_query_is_kept_and_declared_as_still_computing_the_unselected_metric():
|
||
"""A 和 B 共用一条 SQL：选 A 也保留那条 SQL，并在结果里声明它还在算 B（不做投影）。"""
|
||
wf = await S.publish(graph=shared_graph())
|
||
made = (await composer()).compose(select(wf, ["n"]), taken={"start"})
|
||
assert [n["id"] for n in made.nodes] == ["c1_fetch", "c1"]
|
||
assert made.nodes[-1]["data"]["config"]["caliber_from"]["metrics"] == ["n"]
|
||
result = made.result()
|
||
assert result["selection"]["metric_ids"] == ["n"] and result["selection"]["excluded_query_ids"] == []
|
||
assert result["shared_queries"] == [{"query_id": f"{wf}:fetch", "node_id": "fetch",
|
||
"computes_unselected": ["gmv"]}]
|
||
|
||
|
||
async def test_a_selection_whose_dependencies_do_not_resolve_is_refused_not_widened():
|
||
"""点名的指标一个取数也不依赖（解析不出闭包）：拒收，不悄悄退回整卡。"""
|
||
graph = S.chain(S.node("start", "input", fields=FIELDS_AB[:2]),
|
||
S.node("fetch", "tool", tool=S.tool(), args={"sql": SQL_A}),
|
||
S.node("card", "metrics", caliber="半个常量", caliber_version="v1",
|
||
metrics=[{"id": "gmv", "name": "销售额", "expression": "cell(nodes.fetch, 0, 'gmv')"},
|
||
{"id": "constant", "name": "常量", "expression": "1 + 1"}]),
|
||
S.node("out", "output"))
|
||
wf = await S.publish(graph=graph)
|
||
pool = await composer()
|
||
# 整卡两张指标都在：gmv 那条查询照拷（常量指标不牵连它）
|
||
ok = pool.compose(select(wf, ["gmv"]), taken={"start"})
|
||
assert [n["id"] for n in ok.nodes] == ["c1_fetch", "c1"]
|
||
with pytest.raises(compose.ComposeRefused) as caught:
|
||
pool.compose(select(wf, ["constant"], id="c2"), taken={"start"})
|
||
assert "no_query_dependency" in caught.value.message and "constant" in caught.value.message
|
||
|
||
|
||
async def test_check_inputs_takes_the_same_selection_by_itself():
|
||
"""check_inputs 也能直接按 metrics 收窄（callers 不必先裁一份）：只选 A 时不要求 B 的参数。"""
|
||
from app.ledger import derive
|
||
|
||
wf = await S.publish(graph=two_metric_graph())
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
[option] = [o for o in kind.options if o.workflow_id == wf]
|
||
pinned = option.pinned
|
||
values = compose.check_inputs(pinned, {"start": "2026-09-01", "end": "2026-09-08"}, metric_ids=["gmv"])
|
||
assert values == {"start": "2026-09-01", "end": "2026-09-08"}
|
||
with pytest.raises(compose.ComposeRefused, match="buyer"):
|
||
compose.check_inputs(pinned, {"start": "2026-09-01", "end": "2026-09-08"})
|
||
# caliber_block 拿同一份选择：改名映射只列保留的那条查询，卡上记下这次算哪个指标
|
||
block = derive.caliber_block(pinned, prefix="c1_", metrics=["gmv"])
|
||
assert [n["id"] for n in block.nodes] == ["c1_fetch_a", "c1_card"]
|
||
assert block.nodes[-1]["data"]["config"]["caliber_from"]["metrics"] == ["gmv"]
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 按指标判可组合性（K08 S4）：B 用不了不牵连 A；一次选用按源卡顺序
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
async def test_a_broken_query_of_one_metric_leaves_the_other_composable():
|
||
"""B 那条查询引用了入口没声明的字段：B 用不了，A 照旧可选、能组合、能实例化；不再把整张卡提前判死。"""
|
||
wf = await S.publish(graph=two_metric_graph(b_sql=SQL_B + " AND note = '{{ input.nope }}'"))
|
||
pool = await composer()
|
||
[option] = [o for o in pool.options if o.workflow_id == wf]
|
||
assert option.composable and [m["metric_id"] for m in option.selectable_metrics] == ["gmv"]
|
||
assert [m["metric_id"] for m in option.unavailable_metrics] == ["n"]
|
||
assert "undeclared_parameter" in option.unavailable_metrics[0]["reason"]
|
||
# 不点名（整卡）那一份算不出来：原因照实给，但按指标还能用
|
||
assert option.default_closure is None and "undeclared_parameter" in option.reason
|
||
[sel] = option.selectable_metrics
|
||
assert sel["query_ids"] == [f"{wf}:fetch_a"] and sel["required_entry_fields"] == ["start", "end"]
|
||
blob = json.dumps(sel, ensure_ascii=False)
|
||
assert "SUM(" not in blob and "orders" not in blob and "paid_at" not in blob and '"sql"' not in blob
|
||
# 能选的照旧实例化（只拷 A 那条查询）；不能选的逐项硬校验拒收，不靠预筛放宽
|
||
made = pool.compose(select(wf, ["gmv"]), taken={"start"})
|
||
assert [n["id"] for n in made.nodes] == ["c1_fetch_a", "c1"]
|
||
with pytest.raises(compose.ComposeRefused) as caught:
|
||
pool.compose(select(wf, ["n"], id="c2"), taken={"start"})
|
||
assert "undeclared_parameter" in caught.value.message
|
||
# 查看结果：能选的写要哪些字段，用不了的写原因；整段没有 SQL
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
text, _ = await kind.detail([(i, "named") for i in kind.items if i.workflow_id == wf], 4000)
|
||
assert "能选用的指标：销售额（gmv，要 start、end）" in text
|
||
assert "用不了的指标：单量（n）：" in text and "undeclared_parameter" in text
|
||
assert "SELECT" not in text.upper() and "orders" not in text
|
||
|
||
|
||
async def test_a_metric_whose_binding_is_out_of_step_does_not_close_the_whole_card():
|
||
"""B 声明了业务字段（统计时间口径）但那条取数没有受控绑定：B 用不了，A 照旧可选、能实例化（K08 S4）。"""
|
||
graph = two_metric_graph()
|
||
card = next(n for n in graph["nodes"] if n["type"] == "metrics")
|
||
card["data"]["config"]["metrics"][1]["time_role"] = "payment_time" # B 声明了，可是 fetch_b 上没有绑定
|
||
wf = await S.publish(graph=graph)
|
||
pool = await composer()
|
||
[option] = [o for o in pool.options if o.workflow_id == wf]
|
||
assert option.composable and [m["metric_id"] for m in option.selectable_metrics] == ["gmv"]
|
||
[bad] = option.unavailable_metrics
|
||
assert bad["metric_id"] == "n" and "受控绑定" in bad["reason"]
|
||
assert option.default_closure is None and "受控绑定" in option.reason
|
||
made = pool.compose(select(wf, ["gmv"]), taken={"start"})
|
||
assert [n["id"] for n in made.nodes] == ["c1_fetch_a", "c1"]
|
||
with pytest.raises(compose.ComposeRefused) as caught:
|
||
pool.compose(select(wf, ["n"], id="c2"), taken={"start"})
|
||
assert "binding_unresolved" in caught.value.message
|
||
|
||
|
||
async def test_a_card_whose_metrics_all_fail_is_not_offered_for_composition():
|
||
"""两个指标都用不了：不提供组合写法、也不实例化，原因与每个指标的原因照实给。"""
|
||
graph = two_metric_graph(a_sql=SQL_A + " AND note = '{{ input.nope }}'",
|
||
b_sql=SQL_B + " AND note = '{{ input.nope }}'")
|
||
wf = await S.publish(graph=graph)
|
||
pool = await composer()
|
||
[option] = [o for o in pool.options if o.workflow_id == wf]
|
||
assert not option.composable and option.pinned is None and option.selectable_metrics == ()
|
||
assert [m["metric_id"] for m in option.unavailable_metrics] == ["gmv", "n"]
|
||
assert "undeclared_parameter" in option.reason
|
||
with pytest.raises(compose.ComposeRefused, match="不能组合"):
|
||
pool.compose(select(wf, ["gmv"]), taken={"start"})
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
text, _ = await kind.detail([(i, "named") for i in kind.items if i.workflow_id == wf], 4000)
|
||
assert "不能组合" in text and "undeclared_parameter" in text and "SELECT" not in text.upper()
|
||
|
||
|
||
async def test_a_selection_is_normalized_to_the_source_card_order():
|
||
"""一次选用里的指标一律按源卡顺序：Composed.metrics、口径卡上记的 metrics、selection_hash 与执行数组都不随
|
||
调用方传的顺序、重复项变（K08 S4）。"""
|
||
wf = await S.publish(graph=two_metric_graph(on_missing=True))
|
||
pool = await composer()
|
||
both = {"start": "2026-09-01", "end": "2026-09-08", "buyer": "east"}
|
||
forward = pool.compose(select(wf, ["gmv", "n"], both), taken=set())
|
||
backward = pool.compose(select(wf, ["n", "gmv", "n"], both), taken=set())
|
||
assert [m["id"] for m in forward.metrics] == [m["id"] for m in backward.metrics] == ["gmv", "n"]
|
||
assert forward.result()["selection"]["metric_ids"] == ["gmv", "n"]
|
||
assert forward.nodes[-1]["data"]["config"]["caliber_from"]["metrics"] == ["gmv", "n"]
|
||
assert forward.nodes == backward.nodes and forward.edges == backward.edges
|
||
assert forward.result()["selection"]["selection_hash"] == backward.result()["selection"]["selection_hash"]
|
||
assert forward.result()["selection"]["closure_hash"] == backward.result()["selection"]["closure_hash"]
|
||
|
||
|
||
async def test_the_offered_option_follows_the_recommended_version():
|
||
"""K07：问数据里看到的是推荐发布版（Workflow.published_version），不是机械取最高已发布版；按指标的可选性也按那一版。"""
|
||
from app.db.models import Workflow
|
||
|
||
wf = await S.publish(graph=two_metric_graph()) # v1：两个指标都读得出来
|
||
await S.publish(name="周销售", version=2, workflow_id=wf,
|
||
graph=two_metric_graph(b_sql=SQL_B + " AND note = '{{ input.nope }}'")) # v2：B 坏了
|
||
async with SessionLocal() as session:
|
||
row = await session.get(Workflow, wf)
|
||
row.published_version = 1 # 重新推回 v1
|
||
await session.commit()
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
assert {i.version for i in kind.items if i.workflow_id == wf} == {1}
|
||
[option] = [o for o in kind.options if o.workflow_id == wf]
|
||
assert option.version == 1 and option.composable and option.unavailable_metrics == ()
|
||
assert [m["metric_id"] for m in option.selectable_metrics] == ["gmv", "n"]
|
||
assert option.default_closure["query_ids"] == [f"{wf}:fetch_a", f"{wf}:fetch_b"]
|
||
|
||
|
||
# --------------------------------------------------------------------------
|
||
# 复核：重复的指标 id、字段「不在本次选择范围」与「入口没这个字段」
|
||
# --------------------------------------------------------------------------
|
||
|
||
|
||
async def test_a_card_with_a_duplicated_metric_id_is_never_offered_for_composition():
|
||
"""同 id 指向两条不同查询的卡：look 一个指标也不提供（理由照实）、实例化也不放行（K08 复核）。"""
|
||
graph = two_metric_graph()
|
||
card = next(n for n in graph["nodes"] if n["type"] == "metrics")
|
||
card["data"]["config"]["metrics"][1]["id"] = "gmv" # 两个 gmv：一个读 fetch_a、一个读 fetch_b
|
||
wf = await S.publish(graph=graph)
|
||
pool = await composer()
|
||
[option] = [o for o in pool.options if o.workflow_id == wf]
|
||
assert not option.composable and option.pinned is None and option.selectable_metrics == ()
|
||
assert [m["metric_id"] for m in option.unavailable_metrics] == ["gmv", "gmv"]
|
||
assert all("duplicate_metric_id" in m["reason"] for m in option.unavailable_metrics)
|
||
assert "duplicate_metric_id" in (option.reason or "")
|
||
with pytest.raises(compose.ComposeRefused, match="不能组合"):
|
||
pool.compose(select(wf, ["gmv"]), taken={"start"})
|
||
async with SessionLocal() as session:
|
||
kind = await CaliberKind.load(session, compose=True)
|
||
text, _ = await kind.detail([(i, "named") for i in kind.items if i.workflow_id == wf], 4000)
|
||
assert "不能组合" in text and "duplicate_metric_id" in text and "SELECT" not in text.upper()
|
||
|
||
|
||
async def test_a_field_of_another_metric_is_out_of_scope_not_missing():
|
||
"""这一版入口声明过、只是不在这次选中的指标的范围里：说法是「不在这次选用的范围里」，不是「入口没有字段」；
|
||
真的没声明过的字段照旧说「入口没有字段」（K08 复核）。两种都拒，行为不变。"""
|
||
wf = await S.publish(graph=two_metric_graph(on_missing=True))
|
||
pool = await composer()
|
||
both_fields = {"start": "2026-09-01", "end": "2026-09-08", "buyer": "east"}
|
||
with pytest.raises(compose.ComposeRefused) as caught:
|
||
pool.compose(select(wf, ["gmv"], both_fields), taken=set())
|
||
message = caught.value.message
|
||
assert "「buyer」这一版入口声明过，但不在这次选用的范围里" in message
|
||
assert "这次算 gmv，只要 start、end" in message and "入口没有字段" not in message
|
||
with pytest.raises(compose.ComposeRefused) as unknown:
|
||
pool.compose(select(wf, ["gmv"], {"start": "2026-09-01", "end": "2026-09-08", "买了没": "x"}, id="c2"),
|
||
taken=set())
|
||
assert "入口没有字段「买了没」" in unknown.value.message and "不在这次选用的范围里" not in unknown.value.message
|
||
# 两张指标都选上时 buyer 就在这次的范围里：照旧收（整卡老行为不变）
|
||
made = pool.compose(select(wf, ["gmv", "n"], both_fields, id="c3"), taken=set())
|
||
assert made.inputs == both_fields
|