完善口径管理、主动调查与后台任务,修复通用问题

- 完善口径编辑、只读试运行、统一质量验证与执行语义指纹
- 增加可执行指标约束、有预算的主动调查和过程评测
- 提取发布检查、证据读取与起草服务,支持任务停止、重试和检查点恢复
- 增加版本传递影响分析,修复并发覆盖、敏感值遮罩、异常评分及页面状态问题
- 修复 SQLite 复合查询追溯和 microVM 可选依赖缺失处理
- 纳入当前设计文档及部署配置调整

验证:相关主回归 3822 项通过,4 项跳过、8 项预期失败;补充回归通过。
前端构建、文案检查及 git diff --check 通过。
全量检查仍有环境权限、PostgreSQL 可选依赖和既有部署配置、SQLite 文件基线差异导致的未通过项。
This commit is contained in:
ThearchyHelios
2026-11-06 17:14:45 +08:00
parent 5f0b2359ac
commit 9d8b41430e
94 changed files with 9730 additions and 957 deletions
+234 -18
View File
@@ -57,6 +57,16 @@ def plan_reply(chunk_id: str, *, questions=None, sql: str | None = None, metrics
}, ensure_ascii=False)
def groups_reply(*plans: dict) -> str:
"""一次交多组:每组沿用 plan_reply 的 sql / entry_fields / metrics。"""
return json.dumps({
"groups": [{"sql": plan["sql"], "entry_fields": plan["entry_fields"], "metrics": plan["metrics"]}
for plan in plans],
"open_questions": [],
"uncovered": [],
}, ensure_ascii=False)
class FakeModel:
"""起草模型:ainvoke 返回带 .content 的对象;按顺序交回复,交完重复最后一份。"""
@@ -96,16 +106,19 @@ async def fake_trial(monkeypatch):
"""试跑换成写一行 succeeded 的运行:不启动引擎、不查库,接口和 payload 的形状照真的一样。"""
made: list[str] = []
async def start(*, draft_id, workflow_id, graph, values, started_by, entry_fields=()):
async def start(*, draft_id, workflow_id, graph, values, started_by, entry_fields=(), version=1):
run_id = f"run-draft-{uuid.uuid4().hex[:10]}"
async with SessionLocal() as session:
session.add(Run(id=run_id, workflow_id=workflow_id, status="succeeded", graph=graph,
input=values, run_class="exploratory", origin_kind="caliber_trial",
origin={"kind": "caliber_trial", "draft_id": draft_id,
"workflow_id": workflow_id, "version": 1}))
"workflow_id": workflow_id, "version": version}))
await session.commit()
made.append(run_id)
return {"run_id": run_id, "status": "succeeded", "error": ""}
trial = {**api_ledger._trial_binding(version, graph, values),
"run_id": run_id, "status": "succeeded", "error": ""}
await api_ledger._save_trial(draft_id, trial, replace=True)
return trial
monkeypatch.setattr(api_ledger, "_run_trial", start)
yield made
@@ -213,6 +226,10 @@ async def test_extract_produces_a_ready_draft_with_a_graph_and_a_trial(client, d
assert made["alignment"]["new"] == ["已支付退款率"]
assert made["truncated"] is False and made["model_id"] == "fake-model"
assert made["datasource_id"] == source_id
# 发布前检查:五项都从图和试跑状态算,不另起一份判断
assert {check["key"] for check in made["checks"]} == {"graph", "sql", "entries", "trial", "citations"}
assert all(check["ok"] for check in made["checks"] if check["key"] != "trial"), made["checks"]
assert next(check for check in made["checks"] if check["key"] == "trial")["status"] == "unknown"
# 工作流和 v1 已建:合法的草稿图,取数指向这个数据源
async with SessionLocal() as session:
@@ -225,10 +242,27 @@ async def test_extract_produces_a_ready_draft_with_a_graph_and_a_trial(client, d
assert [n["id"] for n in workflow.graph["nodes"]] == ["start", "fetch", "card", "write", "out"]
assert validate_graph(GraphSpec.model_validate(workflow.graph)).ok
# 试跑只读接口:契约里的形状
# 试跑只读接口:契约里的形状(queries 是取数节点交回的查询快照;这次是假试跑,没有快照)
trial = (await client.get(f"/api/ledger/drafts/{body['draft_id']}/trial")).json()
assert set(trial) == {"run_id", "status", "metrics", "error"}
assert set(trial) == {"run_id", "status", "metrics", "queries", "error", "version", "graph_hash", "execution_hash", "inputs_hash", "validation"}
assert trial["run_id"] == made["trial"]["run_id"] and trial["status"] == "succeeded"
assert trial["queries"] == []
async def test_publish_checks_point_at_an_entry_field_the_sql_never_uses(client, documents, datasource, monkeypatch):
"""发布前检查照实说:声明了入口但 SQL 没用,也算不一致;其他项不跟着判失败。"""
doc = await documents("dd-checks")
source_id = await datasource()
plan = json.loads(plan_reply(f"{doc}-2", sql="SELECT COUNT(*) * 1.0 AS refund_rate FROM orders "
"WHERE status = 'paid'"))
fake_model(monkeypatch, json.dumps(plan, ensure_ascii=False))
draft_id = (await extract(client, doc, source_id))["draft_id"]
made = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
by_key = {check["key"]: check for check in made["checks"]}
assert by_key["entries"]["ok"] is False and "store" in by_key["entries"]["detail"]
assert by_key["graph"]["ok"] and by_key["sql"]["ok"] and by_key["citations"]["ok"]
assert by_key["trial"]["status"] == "unknown"
async def test_open_questions_keep_the_draft_in_questions_open_but_still_run_it(
@@ -240,6 +274,7 @@ async def test_open_questions_keep_the_draft_in_questions_open_but_still_run_it(
body = await extract(client, doc, source_id)
made = (await client.get(f"/api/ledger/drafts/{body['draft_id']}")).json()
assert made["status"] == "questions_open"
assert made["stage"] == "review"
assert made["open_questions"] == ["退款算在哪一天"]
assert made["workflow_id"] and made["sql"]
assert made["trial"]["status"] == "succeeded"
@@ -318,6 +353,15 @@ async def test_patch_replaces_metrics_and_clears_resolved_questions(client, docu
assert [m["name"] for m in made["metrics"]] == ["已支付退款率(按支付日)"]
assert made["open_questions"] == [] and made["question_count"] == 0
# PATCH 和画布一样新增版本,读取新版本而不是覆盖 v1
async with SessionLocal() as session:
draft_row = await session.get(CaliberDraft, draft_id)
version_row = (await session.execute(select(WorkflowVersion).where(
WorkflowVersion.workflow_id == draft_row.workflow_id,
WorkflowVersion.version == made["version"]))).scalars().one()
card = next(n for n in version_row.graph["nodes"] if n["type"] == "metrics")
assert [m["name"] for m in card["data"]["config"]["metrics"]] == ["已支付退款率(按支付日)"]
# 结构不对的指标 422,草稿不动
bad = await client.patch(f"/api/ledger/drafts/{draft_id}", json={"metrics": [{"synonyms": ["x"]}]})
assert bad.status_code == 422
@@ -330,6 +374,94 @@ async def test_patch_replaces_metrics_and_clears_resolved_questions(client, docu
assert res.json()["status"] == "questions_open" and res.json()["question_count"] == 1
async def test_draft_view_reads_the_graph_so_canvas_edits_show_up(client, documents, datasource, monkeypatch):
"""画布改了指标名和取数 SQL 之后,草稿页从图里读最新的那几项,不再显示起草时的 payload 快照。"""
doc = await documents("dd-live")
source_id = await datasource()
fake_model(monkeypatch, plan_reply(f"{doc}-2"))
draft_id = (await extract(client, doc, source_id))["draft_id"]
# 必须走画布的真实保存接口:它新增 v2,不能直接覆盖 v1 来模拟
async with SessionLocal() as session:
draft = await session.get(CaliberDraft, draft_id)
row = (await session.execute(select(WorkflowVersion).where(
WorkflowVersion.workflow_id == draft.workflow_id,
WorkflowVersion.version == 1))).scalars().one()
graph = copy.deepcopy(row.graph)
for node in graph["nodes"]:
if node["type"] == "metrics":
node["data"]["config"]["metrics"][0]["name"] = "已支付退款率(画布改过)"
if node["type"] == "tool" and (node["data"]["config"].get("args") or {}).get("sql"):
node["data"]["config"]["args"]["sql"] = "SELECT 1 AS refund_rate"
workflow_id = draft.workflow_id
saved = await client.patch(f"/api/workflows/{workflow_id}", json={"graph": graph})
assert saved.status_code == 200 and saved.json()["version"] == 2, saved.text
made = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
assert made["metrics"][0]["name"] == "已支付退款率(画布改过)"
assert made["sql"] == "SELECT 1 AS refund_rate"
assert [group["sql"] for group in made["groups"]] == ["SELECT 1 AS refund_rate"]
assert made["version"] == 2 and made["trial"]["status"] == "stale"
async def test_resolving_a_metric_question_writes_it_back_to_the_graph(client, documents, datasource, monkeypatch):
"""答清一条指标自己的未决问题之后,图里那条指标的 open_questions 也要去掉:草稿页从图读,不能又显示出来。"""
doc = await documents("dd-resolve")
source_id = await datasource()
plan = json.loads(plan_reply(f"{doc}-2"))
plan["metrics"][0]["open_questions"] = ["按哪个门店口径"]
fake_model(monkeypatch, json.dumps(plan, ensure_ascii=False))
draft_id = (await extract(client, doc, source_id))["draft_id"]
made = (await client.get(f"/api/ledger/drafts/{draft_id}")).json()
assert made["status"] == "questions_open"
assert made["metrics"][0]["open_questions"] == ["按哪个门店口径"]
res = await client.patch(f"/api/ledger/drafts/{draft_id}", json={"resolved": ["按哪个门店口径"]})
assert res.status_code == 200, res.text
fixed = res.json()
assert fixed["status"] == "ready" and fixed["question_count"] == 0
assert fixed["metrics"][0]["open_questions"] == []
async with SessionLocal() as session:
draft = await session.get(CaliberDraft, draft_id)
row = (await session.execute(select(WorkflowVersion).where(
WorkflowVersion.workflow_id == draft.workflow_id,
WorkflowVersion.version == fixed["version"]))).scalars().one()
card = next(n for n in row.graph["nodes"] if n["type"] == "metrics")
assert not card["data"]["config"]["metrics"][0].get("open_questions")
async def test_trial_validates_against_graph_entry_fields_after_canvas_edit(
client, documents, datasource, monkeypatch):
"""画布上给入口加了必填字段之后,试跑按图里的字段校验;payload 的旧入口字段不能把新字段漏掉。"""
doc = await documents("dd-entry-live")
source_id = await datasource()
fake_model(monkeypatch, plan_reply(f"{doc}-2"))
draft_id = (await extract(client, doc, source_id))["draft_id"]
async with SessionLocal() as session:
draft = await session.get(CaliberDraft, draft_id)
row = (await session.execute(select(WorkflowVersion).where(
WorkflowVersion.workflow_id == draft.workflow_id,
WorkflowVersion.version == 1))).scalars().one()
graph = copy.deepcopy(row.graph)
for node in graph["nodes"]:
if node["type"] == "input":
node["data"]["config"]["fields"].append(
{"name": "region", "type": "string", "required": True, "default": "",
"description": "区域"})
workflow_id = draft.workflow_id
saved = await client.patch(f"/api/workflows/{workflow_id}", json={"graph": graph})
assert saved.status_code == 200 and saved.json()["version"] == 2, saved.text
missing = await client.post(f"/api/ledger/drafts/{draft_id}/trial", json={"inputs": {}})
assert missing.status_code == 422 and "region" in missing.text
ok = await client.post(f"/api/ledger/drafts/{draft_id}/trial", json={"inputs": {"region": "华东"}})
assert ok.status_code == 200, ok.text and ok.text[:200]
assert ok.json()["status"] == "succeeded"
async def test_dismiss_marks_rejected_and_writes_an_audit(client, documents, datasource, monkeypatch):
doc = await documents("dd-7")
source_id = await datasource()
@@ -337,7 +469,7 @@ async def test_dismiss_marks_rejected_and_writes_an_audit(client, documents, dat
draft_id = (await extract(client, doc, source_id))["draft_id"]
res = await client.post(f"/api/ledger/drafts/{draft_id}/dismiss", json={"note": "已经有口径了"})
assert res.status_code == 200 and res.json()["status"] == "rejected"
assert res.status_code == 200 and res.json()["status"] == "rejected" and res.json()["stage"] == "discarded"
async with SessionLocal() as session:
event = (await session.execute(select(AuditEvent).where(
AuditEvent.action == "ledger.docdraft.dismiss", AuditEvent.subject_id == draft_id))).scalars().one()
@@ -358,14 +490,20 @@ async def test_list_counts_orders_by_last_change_and_filters_by_status(client, d
listed = (await client.get("/api/ledger/drafts")).json()["drafts"]
assert [d["id"] for d in listed] == [two, one]
latest, earlier = listed
assert (latest["title"], latest["status"], latest["metric_count"], latest["question_count"]) == \
("二.md", "ready", 1, 0)
assert (earlier["status"], earlier["metric_count"], earlier["question_count"]) == ("questions_open", 1, 1)
assert (latest["title"], latest["status"], latest["stage"], latest["metric_count"], latest["question_count"]) == \
("二.md", "ready", "review", 1, 0)
assert (earlier["status"], earlier["stage"], earlier["metric_count"], earlier["question_count"]) == \
("questions_open", "review", 1, 1)
assert earlier["workflow_id"] and earlier["created_at"] and earlier["error"] is None
only = (await client.get("/api/ledger/drafts", params={"status": "questions_open"})).json()["drafts"]
assert [d["id"] for d in only] == [one]
# 界面按阶段筛:ready 和 questions_open 都是「待审」
review = (await client.get("/api/ledger/drafts", params={"stage": "review"})).json()["drafts"]
assert {d["id"] for d in review} == {one, two}
assert (await client.get("/api/ledger/drafts", params={"stage": "discarded"})).json()["drafts"] == []
assert (await client.get("/api/ledger/drafts", params={"status": "nope"})).status_code == 422
assert (await client.get("/api/ledger/drafts", params={"stage": "nope"})).status_code == 422
assert (await client.get("/api/ledger/drafts/d-nope")).status_code == 404
@@ -382,9 +520,11 @@ async def test_trial_endpoint_returns_metric_values_from_the_run(client, documen
res = await client.get(f"/api/ledger/drafts/{draft_id}/trial")
assert res.status_code == 200, res.text
body = res.json()
assert body == {"run_id": body["run_id"], "status": "succeeded",
assert body == {"run_id": body["run_id"], "status": "succeeded", "version": 1,
"graph_hash": body["graph_hash"], "execution_hash": body["execution_hash"],
"validation": body["validation"], "inputs_hash": body["inputs_hash"],
"metrics": [{"id": "m1", "name": "已支付退款率", "rendered": "3.2%", "value": 0.032}],
"error": ""}
"queries": [], "error": ""}
# 没有试跑的老草稿
async with SessionLocal() as session:
@@ -489,6 +629,14 @@ async def test_a_real_trial_run_feeds_the_trial_endpoint(client, documents, tmp_
assert [m["id"] for m in trial["metrics"]] == ["m1"]
assert trial["metrics"][0]["name"] == "已支付销售额"
assert float(trial["metrics"][0]["value"]) == 156.5
# 试跑结果集:取数节点交回的查询快照,给列和行、不给 SQL
assert [query["node_id"] for query in trial["queries"]] == ["fetch"]
query = trial["queries"][0]
assert set(query) == {"node_id", "tool", "source", "columns", "rows", "total_rows",
"truncated", "masked", "source_truncated", "preview_truncated"}
assert query["columns"] == ["gmv"] and query["rows"] == [[156.5]]
assert query["total_rows"] == 1 and query["truncated"] is False and query["masked"] == []
assert "sql" not in query
async with SessionLocal() as session:
run = await session.get(Run, run_id)
assert run.origin_kind == "caliber_trial" and run.run_class == "exploratory"
@@ -510,6 +658,64 @@ async def test_a_real_trial_run_feeds_the_trial_endpoint(client, documents, tmp_
await session.commit()
async def test_one_reply_can_bring_several_independent_groups(client, documents, datasource, monkeypatch):
"""模型一次交两组:只调一次模型,两组各自成块,不再一组一轮。"""
doc = await documents("dd-groups")
source_id = await datasource()
first = json.loads(plan_reply(f"{doc}-2"))
second = json.loads(plan_reply(f"{doc}-2"))
second["sql"] = ("SELECT COUNT(*) AS orders FROM orders WHERE status = 'paid' "
"AND store = '{{ input.store }}'")
second["metrics"] = [{**second["metrics"][0], "id": "m1", "name": "已支付订单数",
"expression": "cell(nodes.fetch, 0, 'orders')"}]
model = fake_model(monkeypatch, groups_reply(first, second))
body = await extract(client, doc, source_id)
made = (await client.get(f"/api/ledger/drafts/{body['draft_id']}")).json()
assert made["status"] == "ready" and made["blocks"] == 2 and made["metric_count"] == 2
assert [group["sql"] for group in made["groups"]] == [first["sql"], second["sql"]]
assert [m["id"] for m in made["metrics"]] == ["m1", "m2"]
assert made["progress"] == {"round": 1, "covered": 2, "uncovered": 0}
assert len(model.calls) == 1
async def test_one_bad_group_is_repaired_without_failing_the_others(client, documents, datasource, monkeypatch):
"""一组 SQL 不合格只重写这一组;修好后照常落图,好的那组不受影响。"""
doc = await documents("dd-groups-bad")
source_id = await datasource()
good = json.loads(plan_reply(f"{doc}-2"))
bad = json.loads(plan_reply(f"{doc}-2", sql="SELECT * FROM secrets"))
bad["metrics"] = [{**bad["metrics"][0], "id": "m1", "name": "已支付订单数",
"expression": "cell(nodes.fetch, 0, 'orders')"}]
fixed = {**bad, "sql": ("SELECT COUNT(*) AS orders FROM orders WHERE status = 'paid' "
"AND store = '{{ input.store }}'")}
model = fake_model(monkeypatch, groups_reply(good, bad), groups_reply(fixed))
body = await extract(client, doc, source_id)
made = (await client.get(f"/api/ledger/drafts/{body['draft_id']}")).json()
assert made["status"] == "ready" and made["blocks"] == 2 and made["metric_count"] == 2
assert made["workflow_id"] and len(model.calls) == 2
assert [group["sql"] for group in made["groups"]] == [good["sql"], fixed["sql"]]
assert not any("secrets" in item["reason"] for item in made["dropped"])
async def test_a_bad_group_that_still_cannot_be_repaired_is_dropped(client, documents, datasource, monkeypatch):
"""重写也没修好:只丢这一组,问题照实写下来,不把整份草稿判死。"""
doc = await documents("dd-groups-unfixable")
source_id = await datasource()
good = json.loads(plan_reply(f"{doc}-2"))
bad = json.loads(plan_reply(f"{doc}-2", sql="SELECT * FROM secrets"))
bad["metrics"] = [{**bad["metrics"][0], "id": "m1", "name": "已支付订单数",
"expression": "cell(nodes.fetch, 0, 'orders')"}]
model = fake_model(monkeypatch, groups_reply(good, bad))
body = await extract(client, doc, source_id)
made = (await client.get(f"/api/ledger/drafts/{body['draft_id']}")).json()
assert made["status"] == "ready" and made["blocks"] == 1 and made["metric_count"] == 1
assert len(model.calls) == 2
assert any("secrets" in item["reason"] for item in made["dropped"])
async def test_a_big_document_is_drafted_in_several_rounds_onto_one_multi_fetch_graph(
client, documents, datasource, monkeypatch):
"""第一轮交一组、uncovered 还剩一条:续跑第二轮,收成一张两个取数节点、卡上两条指标的图。"""
@@ -531,6 +737,13 @@ async def test_a_big_document_is_drafted_in_several_rounds_onto_one_multi_fetch_
assert made["status"] == "ready"
assert [m["name"] for m in made["metrics"]] == ["已支付退款率", "已支付订单数"]
assert made["blocks"] == 2
# groups 逐组给出 SQL、入口字段和指标 id:页面按组画「这条 SQL 出哪几个指标」,不必从拼接的 sql 里猜边界
assert [set(group) for group in made["groups"]] == [{"sql", "entry_fields", "metric_ids"}] * 2
assert [group["sql"] for group in made["groups"]] == [first["sql"], second["sql"]]
assert [group["metric_ids"] for group in made["groups"]] == [["m1"], ["m2"]]
# 各组都从 m1 起草,拼图前重排:payload 的指标 id 和 groups 指同一组指标,页面按 id 配试跑值才不会张冠李戴
assert [m["id"] for m in made["metrics"]] == ["m1", "m2"]
assert [f["name"] for f in made["groups"][0]["entry_fields"]] == ["store"]
assert made["progress"] == {"round": 2, "covered": 2, "uncovered": 0}
assert made["truncated"] is False and made["doc_truncated"] is False and made["catalog_truncated"] is False
# 第二轮提示词写明上一轮起草过什么:模型据此不再重复起草
@@ -641,6 +854,8 @@ async def test_a_real_trial_run_reads_values_from_both_fetch_nodes(client, docum
trial = (await client.get(f"/api/ledger/drafts/{draft_id}/trial")).json()
assert trial["status"] == "succeeded" and trial["error"] == ""
assert {m["id"]: float(m["value"]) for m in trial["metrics"]} == {"m1": 156.5, "m2": 4.0}
assert [query["node_id"] for query in trial["queries"]] == ["fetch_1", "fetch_2"]
assert [query["columns"] for query in trial["queries"]] == [["gmv"], ["orders"]]
async with SessionLocal() as session:
run = await session.get(Run, run_id)
tools = [n["id"] for n in (run.graph or {}).get("nodes") or [] if n.get("type") == "tool"]
@@ -712,25 +927,26 @@ async def test_running_the_trial_with_inputs_typed_on_the_page(client, documents
"AND paid_at >= '{{ input.start }}'")
fake_model(monkeypatch, json.dumps(reply, ensure_ascii=False))
async def waiting(*, draft_id, workflow_id, graph, values, started_by, entry_fields=()):
return {"run_id": "", "status": "needs_input", "error": "缺少必填输入:start"}
async def waiting(*, draft_id, workflow_id, graph, values, started_by, entry_fields=(), version=1):
return {**api_ledger._trial_binding(version, graph, values),
"run_id": "", "status": "needs_input", "error": "缺少必填输入:start"}
monkeypatch.setattr(api_ledger, "_run_trial", waiting)
body = await extract(client, doc, source_id)
made = (await client.get(f"/api/ledger/drafts/{body['draft_id']}")).json()
assert made["trial"]["status"] == "needs_input"
trial = (await client.get(f"/api/ledger/drafts/{body['draft_id']}/trial")).json()
assert trial["status"] == "needs_input" and trial["metrics"] == []
assert trial["status"] == "needs_input" and trial["metrics"] == [] and trial["queries"] == []
missing = await client.post(f"/api/ledger/drafts/{body['draft_id']}/trial", json={"inputs": {}})
assert missing.status_code == 422
seen: dict[str, object] = {}
async def started(*, draft_id, workflow_id, graph, values, started_by, entry_fields=()):
async def started(*, draft_id, workflow_id, graph, values, started_by, entry_fields=(), version=1):
seen.clear()
seen.update(values)
return {"run_id": "", "status": "running", "error": ""}
return {**api_ledger._trial_binding(version, graph, values), "run_id": "", "status": "running", "error": ""}
monkeypatch.setattr(api_ledger, "_run_trial", started)
res = await client.post(f"/api/ledger/drafts/{body['draft_id']}/trial", json={"inputs": {"start": "2026-01-01"}})
@@ -746,13 +962,13 @@ async def test_extract_hands_the_fixed_filters_to_every_round(client, documents,
source_id = await datasource()
fake_model(monkeypatch, plan_reply(f"{doc}-2"))
seen: dict[str, object] = {}
real_plan = draftplan.plan
real_plan = draftplan.plan_groups
async def spy(session, **kwargs):
seen.update(kwargs)
return await real_plan(session, **kwargs)
monkeypatch.setattr(draftplan, "plan", spy)
monkeypatch.setattr(draftplan, "plan_groups", spy)
body = await client.post("/api/ledger/drafts/extract", json={
"documents": [doc], "datasource_id": source_id, "fixed_filters": "line_code = 'FOAM01'"})
made = (await client.get(f"/api/ledger/drafts/{body.json()['draft_id']}")).json()