mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 05:21:57 +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。
339 lines
20 KiB
Python
339 lines
20 KiB
Python
"""已接管草稿按图读、老草稿如实说没有来源基线（K03/K05 的前置边界）。
|
||
|
||
试跑（POST /api/ledger/drafts/{id}/trial）对 docattach 那种已接管草稿（tags 有 doc_draft、origin 里没有 kind）读的是
|
||
**当前图版本**和**当前图里的入口字段**，不是 payload 里起草/接上那一刻的快照：
|
||
|
||
- 图上有、payload 里没有的字段能填；payload 里的旧字段名现在是「不认识的入口字段」（422）；
|
||
- 必填字段没给值 → needs_input，**一条 SQL 都不发**（运行表里不多一行）；
|
||
- 数来自当前版本（v2）的 SQL，不是 payload 里写的那一版；
|
||
- 跑完试跑不判 stale（它记的就是当前这一版）。
|
||
|
||
来源基线（K05 的读取侧）：草稿详情对没有可还原快照的老草稿如实给 baseline_unknown（不拿它现在的内容冒充「上次被
|
||
接受的导入值」）；payload 里已经有的快照和基线照原样交回，不重算、不另补。
|
||
|
||
真引擎跑一次（run_manager.setup + tests/ledger_samples 的真实 SQLite 店铺库）：探针图里没有报告节点，不发模型请求。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import copy
|
||
import uuid
|
||
|
||
import pytest
|
||
from sqlalchemy import delete, func, select
|
||
|
||
from app.core.artifact_store import graph_hash
|
||
from app.db.base import SessionLocal
|
||
from app.db.models import CaliberDraft, Run, Workflow, WorkflowVersion
|
||
from app.engine.runner import run_manager
|
||
from tests import ledger_samples as S
|
||
from tests.test_ledger_docdraft_api import client # noqa: F401 - fixture 和起草用例共用
|
||
|
||
#: 这个文件自己建的工作流（连同版本、草稿）跑完删掉
|
||
OWN_PREFIX = "K03 读真源 "
|
||
|
||
|
||
@pytest.fixture(autouse=True)
|
||
async def cleanup_own_rows():
|
||
yield
|
||
async with SessionLocal() as session:
|
||
workflow_ids = list((await session.execute(
|
||
select(Workflow.id).where(Workflow.name.like(f"{OWN_PREFIX}%")))).scalars())
|
||
if workflow_ids:
|
||
for model in (WorkflowVersion, CaliberDraft):
|
||
await session.execute(delete(model).where(model.workflow_id.in_(workflow_ids)))
|
||
await session.execute(delete(Workflow).where(Workflow.id.in_(workflow_ids)))
|
||
await session.commit()
|
||
|
||
|
||
#: 取数 SQL 里的那个加数：v2 比 v1 多 1000，跑出来的数一眼看得出是哪一版
|
||
BUMP = 1000.0
|
||
|
||
|
||
def graph(sql: str, *, fields: list[dict] | None = None) -> dict:
|
||
"""一份能派生探针的口径图：入口 → 取数 → 口径卡（入口字段照给，缺省是 start / end 两个必填）。"""
|
||
return S.chain(
|
||
S.node("start", "input", fields=fields if fields is not None else [
|
||
{"name": "start", "required": True, "description": "开始日期"},
|
||
{"name": "end", "required": True, "description": "结束日期"}]),
|
||
S.node("fetch", "tool", tool=S.tool(), args={"sql": sql}),
|
||
S.node("card", "metrics", caliber="周销售", caliber_version="v1",
|
||
metrics=[{"id": "gmv", "name": "销售额", "unit": "元", "decimals": 2,
|
||
"expression": "cell(nodes.fetch, 0, 'gmv')"}]),
|
||
)
|
||
|
||
|
||
async def attached_draft(*, graph: dict, payload: dict) -> dict:
|
||
"""一份像 docattach 产物那样已接管、但 payload 还停在旧快照上的草稿（tags 有 doc_draft、origin 没有 kind）。"""
|
||
workflow_id = f"wf-{uuid.uuid4().hex[:12]}"
|
||
draft_id = uuid.uuid4().hex[:32]
|
||
async with SessionLocal() as session:
|
||
session.add(Workflow(id=workflow_id, name=f"{OWN_PREFIX}{uuid.uuid4().hex[:6]}", status="draft", graph=graph,
|
||
version=1, tags=["doc_draft"],
|
||
origin={"run_id": "run-1", "freeze": True, "sources": [{"run_id": "run-1"}]}))
|
||
session.add(WorkflowVersion(workflow_id=workflow_id, version=1, graph=graph, graph_hash=graph_hash(graph)))
|
||
await session.flush() # 草稿的外键指着工作流：先落工作流那一行
|
||
session.add(CaliberDraft(id=draft_id, title="接上来的口径", status="ready", workflow_id=workflow_id,
|
||
payload={"version": 1, "workflow_id": workflow_id, **copy.deepcopy(payload)}))
|
||
await session.commit()
|
||
return {"id": draft_id, "workflow_id": workflow_id, "revision": 1, "card_id": "card"}
|
||
|
||
|
||
async def runs_of(workflow_id: str) -> int:
|
||
async with SessionLocal() as session:
|
||
return int(await session.scalar(select(func.count(Run.id)).where(Run.workflow_id == workflow_id)) or 0)
|
||
|
||
|
||
# ==========================================================================
|
||
# 试跑：当前图版本、当前入口字段、缺参不发 SQL
|
||
# ==========================================================================
|
||
|
||
async def test_the_trial_of_a_taken_over_draft_reads_the_current_graph_not_the_payload(client, tmp_path):
|
||
"""docattach 那种已接管草稿：试跑按当前图版本和当前入口字段跑，payload 里的旧快照一个字都不算数。"""
|
||
await run_manager.setup()
|
||
try:
|
||
await S.fresh_source(tmp_path / "shop.db")
|
||
# 起草/接上那一刻的 payload：入口字段是 store，版本停在 1，SQL 是另一句（跑到它身上一眼看得出）
|
||
body = await attached_draft(
|
||
graph=graph(S.TEMPLATE), payload={
|
||
"sql": "SELECT 1 AS gmv", "entry_fields": [{"name": "store", "description": "旧快照里的字段"}],
|
||
"metrics": [{"id": "m1", "name": "旧指标", "expression": "cell(nodes.fetch, 0, 'gmv')"}]})
|
||
|
||
# 画布上改了取数（加了 1000）：新的图版本就是当前这一版
|
||
saved = await client.patch(f"/api/workflows/{body['workflow_id']}", json={
|
||
"graph": graph(S.TEMPLATE.replace("SELECT SUM(amount) AS gmv", f"SELECT SUM(amount) + {BUMP} AS gmv")),
|
||
"expected_revision": body["revision"]})
|
||
assert saved.status_code == 200, saved.text
|
||
assert saved.json()["version"] == 2
|
||
|
||
# 入口字段按图里的：payload 里那个旧字段现在是不认识的字段（不回退旧快照）
|
||
stale_field = await client.post(f"/api/ledger/drafts/{body['id']}/trial",
|
||
json={"inputs": {"store": "east", "start": "2026-09-01", "end": "2026-09-08"}})
|
||
assert stale_field.status_code == 422, stale_field.text
|
||
assert "不认识的入口字段" in stale_field.text and "store" in stale_field.text
|
||
|
||
# 必填没给：表单校验就拒（422，和别的入口同一套），一条 SQL 都不发（运行表里不多行）
|
||
before = await runs_of(body["workflow_id"])
|
||
missing = await client.post(f"/api/ledger/drafts/{body['id']}/trial", json={"inputs": {}})
|
||
assert missing.status_code == 422, missing.text
|
||
assert "start" in missing.text and "end" in missing.text
|
||
assert await runs_of(body["workflow_id"]) == before
|
||
|
||
# 补齐入口：真的跑一次，数来自当前版本（170.5 + 1000）
|
||
done = await client.post(f"/api/ledger/drafts/{body['id']}/trial",
|
||
json={"inputs": {"start": "2026-09-01", "end": "2026-09-08"}})
|
||
assert done.status_code == 200, done.text
|
||
trial = done.json()
|
||
assert trial["status"] == "succeeded", trial.get("error")
|
||
assert trial["version"] == 2 and trial["error"] == ""
|
||
assert float(trial["metrics"][0]["value"]) == 1170.5
|
||
assert [query["node_id"] for query in trial["queries"]] == ["fetch"]
|
||
|
||
# 草稿详情和试跑都认这一版：图里的入口字段和 SQL 就是现在这一版，试跑不算 stale
|
||
made = (await client.get(f"/api/ledger/drafts/{body['id']}")).json()
|
||
assert [field["name"] for field in made["entry_fields"]] == ["start", "end"]
|
||
assert "1000" in made["sql"] and made["version"] == 2
|
||
assert made["trial"]["status"] == "succeeded"
|
||
assert (await client.get(f"/api/ledger/drafts/{body['id']}/trial")).json()["status"] == "succeeded"
|
||
finally:
|
||
await run_manager.shutdown()
|
||
|
||
|
||
async def test_the_trial_of_a_taken_over_draft_reports_a_derivation_problem_without_running(client, tmp_path):
|
||
"""图本身派生不出来（入口字段没声明、模板引用了它）时照实说，也不发起运行。"""
|
||
await run_manager.setup()
|
||
try:
|
||
await S.fresh_source(tmp_path / "shop.db")
|
||
sql = S.TEMPLATE.replace("{{ input.end }}", "{{ input.undefined_param }}")
|
||
body = await attached_draft(graph=graph(sql), payload={"entry_fields": [], "sql": "SELECT 1 AS gmv"})
|
||
before = await runs_of(body["workflow_id"])
|
||
res = await client.post(f"/api/ledger/drafts/{body['id']}/trial",
|
||
json={"inputs": {"start": "2026-09-01", "end": "2026-09-08"}})
|
||
assert res.status_code == 200, res.text
|
||
assert res.json()["status"] == "failed" and "undefined_param" in res.json()["error"]
|
||
assert await runs_of(body["workflow_id"]) == before
|
||
finally:
|
||
await run_manager.shutdown()
|
||
|
||
|
||
async def test_a_parameter_with_no_value_stops_before_any_sql(client, tmp_path):
|
||
"""可选、但没声明缺省值的入口被 SQL 用着：这次没给就是缺参数（needs_input），一条 SQL 都不发。
|
||
|
||
和 K01 的参数完整性同一套：声明了 `""` 缺省值的（「空值不过滤」那种写法）照缺省值渲染，不算缺参。
|
||
"""
|
||
await run_manager.setup()
|
||
try:
|
||
await S.fresh_source(tmp_path / "shop.db")
|
||
fields = [{"name": "start", "required": True}, {"name": "end", "required": True},
|
||
{"name": "region", "type": "string", "description": "区域（可选；没声明缺省值）"}]
|
||
sql = ("SELECT SUM(amount) AS gmv FROM orders WHERE status = 'paid' "
|
||
"AND paid_at >= '{{ input.start }}' AND paid_at < '{{ input.end }}' "
|
||
"AND ('{{ input.region }}' = '' OR region = '{{ input.region }}')")
|
||
body = await attached_draft(graph=graph(sql, fields=fields), payload={"sql": "SELECT 1 AS gmv"})
|
||
before = await runs_of(body["workflow_id"])
|
||
res = await client.post(f"/api/ledger/drafts/{body['id']}/trial",
|
||
json={"inputs": {"start": "2026-09-01", "end": "2026-09-08"}})
|
||
assert res.status_code == 200, res.text
|
||
assert res.json()["status"] == "needs_input" and "region" in res.json()["error"]
|
||
assert await runs_of(body["workflow_id"]) == before
|
||
|
||
# 声明 `""` 缺省值的可选入口（空值不过滤）：不给也能跑，SQL 照缺省值渲染
|
||
open_graph = graph(sql, fields=[{**fields[0]}, {**fields[1]}, {**fields[2], "default": ""}])
|
||
async with SessionLocal() as session:
|
||
workflow = await session.get(Workflow, body["workflow_id"])
|
||
workflow.graph, workflow.version = open_graph, 2
|
||
session.add(WorkflowVersion(workflow_id=body["workflow_id"], version=2, graph=open_graph,
|
||
graph_hash=graph_hash(open_graph)))
|
||
await session.commit()
|
||
done = await client.post(f"/api/ledger/drafts/{body['id']}/trial",
|
||
json={"inputs": {"start": "2026-09-01", "end": "2026-09-08"}})
|
||
assert done.status_code == 200, done.text
|
||
assert done.json()["status"] == "succeeded" and float(done.json()["metrics"][0]["value"]) == 170.5
|
||
finally:
|
||
await run_manager.shutdown()
|
||
|
||
|
||
async def test_a_really_attached_draft_runs_its_trial_from_the_current_graph(client, monkeypatch, tmp_path):
|
||
"""真走一次 docattach：接上来的草稿再试跑，读的是接上之后图上又改过的那一版，payload 的旧字段不算数。
|
||
|
||
和上面那条一样的几条断言，但工作流是 `docattach.attach` 真建出来的（问数据 → 比对 → 接上），不是手搭的形状。
|
||
"""
|
||
from app.ledger import docattach
|
||
from tests.test_caliber_docdraft_attach import ask, document, payload as attached
|
||
|
||
await run_manager.setup()
|
||
try:
|
||
await S.fresh_source(tmp_path / "shop.db")
|
||
run = await ask(monkeypatch)
|
||
doc_id, chunk_id = await document()
|
||
draft_id = uuid.uuid4().hex[:32]
|
||
definition = attached(doc_id, chunk_id)
|
||
async with SessionLocal() as session:
|
||
session.add(CaliberDraft(id=draft_id, title="周销售口径", status="ready", payload=definition))
|
||
await session.commit()
|
||
async with SessionLocal() as session:
|
||
made = await docattach.attach(session, payload=definition, name="周销售口径", run_id=run.id,
|
||
draft_id=draft_id)
|
||
workflow_id = made["workflow_id"]
|
||
|
||
# 接上来之后画布又改过：图上加了入口字段 region，取数按它筛（这一版才是当前版）
|
||
async with SessionLocal() as session:
|
||
workflow = await session.get(Workflow, workflow_id)
|
||
current = copy.deepcopy(workflow.graph)
|
||
draft = await session.get(CaliberDraft, draft_id)
|
||
# payload 停在接上那一刻的旧快照：一个旧字段名、版本 1
|
||
draft.payload = {**draft.payload, "workflow_id": workflow_id, "version": 1,
|
||
"entry_fields": [{"name": "store", "description": "旧快照里的字段"}]}
|
||
draft.workflow_id, draft.status = workflow_id, "attached"
|
||
await session.commit()
|
||
region_sql = ("SELECT SUM(amount) AS gmv FROM orders WHERE status = 'paid' "
|
||
"AND region = '{{ input.region }}'")
|
||
for node in current["nodes"]:
|
||
if node["type"] == "input":
|
||
node["data"]["config"]["fields"] = [{"name": "region", "type": "string", "required": True,
|
||
"description": "地区"}]
|
||
elif node["type"] == "tool":
|
||
node["data"]["config"]["args"]["sql"] = region_sql
|
||
saved = await client.patch(f"/api/workflows/{workflow_id}", json={"graph": current, "expected_revision": 1})
|
||
assert saved.status_code == 200 and saved.json()["version"] == 2, saved.text
|
||
|
||
# 入口字段按图里的：payload 里的旧字段名是不认识的字段
|
||
stale_field = await client.post(f"/api/ledger/drafts/{draft_id}/trial",
|
||
json={"inputs": {"store": "east"}})
|
||
assert stale_field.status_code == 422, stale_field.text
|
||
assert "不认识的入口字段" in stale_field.text and "store" in stale_field.text
|
||
|
||
before = await runs_of(workflow_id)
|
||
missing = await client.post(f"/api/ledger/drafts/{draft_id}/trial", json={"inputs": {}})
|
||
assert missing.status_code == 422 and "region" in missing.text # 必填没给：不发 SQL
|
||
assert await runs_of(workflow_id) == before
|
||
|
||
done = await client.post(f"/api/ledger/drafts/{draft_id}/trial", json={"inputs": {"region": "east"}})
|
||
assert done.status_code == 200, done.text
|
||
assert done.json()["status"] == "succeeded", done.json().get("error")
|
||
assert done.json()["version"] == 2 and float(done.json()["metrics"][0]["value"]) == 156.5
|
||
assert (await client.get(f"/api/ledger/drafts/{draft_id}")).json()["trial"]["status"] == "succeeded"
|
||
finally:
|
||
await run_manager.shutdown()
|
||
|
||
|
||
# ==========================================================================
|
||
# 来源基线：老草稿如实说 unknown，已记的快照照原样
|
||
# ==========================================================================
|
||
|
||
async def test_an_old_draft_says_its_source_baseline_is_unknown(client):
|
||
"""没有可还原快照的老草稿：草稿详情给 baseline_unknown，也不凭空补一条快照。"""
|
||
draft_id = uuid.uuid4().hex[:32]
|
||
async with SessionLocal() as session:
|
||
session.add(CaliberDraft(id=draft_id, title="老草稿", status="path_pending",
|
||
payload={"metrics": [{"id": "m1", "name": "旧指标"}], "sql": "SELECT 1 AS n",
|
||
"version": 1, "workflow_id": None}))
|
||
await session.commit()
|
||
try:
|
||
body = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
|
||
assert body["metrics"][0]["name"] == "旧指标"
|
||
assert "source_snapshot" not in body
|
||
assert body["source_baseline"]["status"] == "baseline_unknown"
|
||
assert "无法还原" in body["source_baseline"]["reason"] and "artifact_id" not in body["source_baseline"]
|
||
finally:
|
||
async with SessionLocal() as session:
|
||
row = await session.get(CaliberDraft, draft_id)
|
||
if row is not None:
|
||
await session.delete(row)
|
||
await session.commit()
|
||
|
||
|
||
async def test_a_baseline_is_recomputed_from_the_artifact_and_never_read_from_the_payload(client):
|
||
"""反例（复核）：payload 里写着 known、工件却不在（或被人改过）时，读草稿要照实说 baseline_unknown。
|
||
|
||
来源基线是**读工件现算**的：工件读得回来才 known（readable=true，段落以工件为准），读不回来就是 unknown
|
||
（readable=false，带上那个工件 id 和原因），不照抄 payload 里存过的状态，也不抛出去把页面弄崩。"""
|
||
from app.core.artifact_store import put_json
|
||
from app.core.config import settings
|
||
from app.ledger import docdraft
|
||
|
||
draft_id = uuid.uuid4().hex[:32]
|
||
snapshot = docdraft.source_snapshot(documents=[{"document_id": "d1", "content": "退款率按已支付订单统计"}],
|
||
chunks={"d1-0": {"document_id": "d1", "ordinal": 0,
|
||
"content": "退款率按已支付订单统计"}})
|
||
artifact_id = await put_json(snapshot, kind=docdraft.SNAPSHOT_KIND)
|
||
ref = docdraft.snapshot_ref(artifact_id, snapshot)
|
||
# payload 里存着的那份状态是「当时算出来的」，工件真伪由读的时候现验
|
||
stale = {"status": "known", "artifact_id": artifact_id, "version": 1, "chunk_ids": ["d1-0"], "note": "起草时写的"}
|
||
async with SessionLocal() as session:
|
||
session.add(CaliberDraft(id=draft_id, title="有基线的草稿", status="ready",
|
||
payload={"metrics": [{"id": "m1", "name": "退款率"}],
|
||
"source_snapshot": copy.deepcopy(ref),
|
||
"source_baseline": copy.deepcopy(stale)}))
|
||
await session.commit()
|
||
path = settings.data_dir / "artifacts" / artifact_id[:2] / f"{artifact_id}.json"
|
||
original = path.read_bytes()
|
||
try:
|
||
body = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
|
||
assert body["source_snapshot"] == ref
|
||
# 工件读得回来：known（readable=true），段落以工件为准，不照抄 payload 里那份（note 不再出现）
|
||
assert body["source_baseline"] == {"status": "known", "readable": True, "artifact_id": artifact_id,
|
||
"version": docdraft.SNAPSHOT_VERSION, "chunk_ids": ["d1-0"]}
|
||
assert docdraft.source_baseline(body) == body["source_baseline"]
|
||
|
||
path.write_text('{"kind": "caliber_source_snapshot", "tampered": true}', encoding="utf-8")
|
||
tampered = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
|
||
assert tampered["source_baseline"]["status"] == "baseline_unknown"
|
||
assert tampered["source_baseline"]["readable"] is False
|
||
assert tampered["source_baseline"]["artifact_id"] == artifact_id # 说得出是哪一个工件读不回来
|
||
assert "读不回来" in tampered["source_baseline"]["reason"]
|
||
assert (await client.get(f"/api/ledger/drafts/{draft_id}/trial")).status_code in (200, 404)
|
||
assert (await client.get(f"/api/ledger/drafts/{draft_id}/validations")).status_code == 200
|
||
|
||
path.unlink()
|
||
gone = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
|
||
assert gone["source_baseline"]["status"] == "baseline_unknown"
|
||
assert gone["source_baseline"]["readable"] is False
|
||
finally:
|
||
path.write_bytes(original)
|
||
async with SessionLocal() as session:
|
||
row = await session.get(CaliberDraft, draft_id)
|
||
if row is not None:
|
||
await session.delete(row)
|
||
await session.commit()
|
||
assert (await client.get(f"/api/ledger/drafts/{draft_id}")).status_code == 404
|