mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-17 08:31:59 +00:00
数据目录 F3 一致性:助手上下文里「现有工作流用到的表」改用 tool_queries 认节点(以 sqlcheck.tool_query 为准,换了类型的节点残留的查询不再算);正式运行的目录变化事件改用 EventType.CATALOG_DRIFT;写明关系编号区分方向、有意不改
This commit is contained in:
@@ -212,35 +212,28 @@ def _modes(sources: list[Any]) -> list[Literal["all", "pick"]]:
|
||||
def graph_tables(graph: Mapping[str, Any] | None, sources: Iterable[Any]) -> dict[str, list[str]]:
|
||||
"""现有工作流里查库节点的 SQL 用到的表:{源名: [schema_cache 的键]},按出现顺序。
|
||||
|
||||
只认调用工具节点上写死的 SQL(db_query__<源>);Agent 运行时自己写的 SQL 图里看不到。节点认两种写法:
|
||||
画布的 {data: {config}} 和 GraphSpec 的 {config}。SQL 解析不了的跳过。
|
||||
只认调用工具节点上写死的 SQL(db_query__<源>);Agent 运行时自己写的 SQL 图里看不到。哪些节点算写死的查询
|
||||
以 sqlcheck.tool_query 为准(catalog_impact.tool_queries,和发布记录、SQL 检查同一份):换成别的类型的节点
|
||||
配置里残留的 tool、args 不算。节点认两种写法:画布的 {data: {config}} 和 GraphSpec 的 {config}。SQL 解析
|
||||
不了的跳过。
|
||||
"""
|
||||
from app.data.catalog_impact import tool_queries
|
||||
from app.engine.evidence import sql_tables
|
||||
from app.tools.datasource import QUERY_PREFIX
|
||||
|
||||
by_name = {getattr(s, "name", None): s for s in sources}
|
||||
out: dict[str, list[str]] = {}
|
||||
for node in (graph or {}).get("nodes") or []:
|
||||
if not isinstance(node, dict):
|
||||
continue
|
||||
config = node.get("config")
|
||||
if not isinstance(config, dict):
|
||||
config = (node.get("data") or {}).get("config") if isinstance(node.get("data"), dict) else None
|
||||
if not isinstance(config, dict):
|
||||
continue
|
||||
tool = str(config.get("tool") or "")
|
||||
source = by_name.get(tool[len(QUERY_PREFIX):]) if tool.startswith(QUERY_PREFIX) else None
|
||||
args = config.get("args")
|
||||
sql = args.get("sql") if isinstance(args, dict) else None
|
||||
if source is None or not isinstance(sql, str):
|
||||
nodes = (graph or {}).get("nodes") or []
|
||||
for _, source_name, sql in tool_queries(n for n in nodes if isinstance(n, dict)):
|
||||
source = by_name.get(source_name)
|
||||
if source is None:
|
||||
continue
|
||||
try:
|
||||
names = sql_tables(sql)
|
||||
except Exception: # noqa: BLE001 - 认不出表就当没用到,只少一个提示
|
||||
continue
|
||||
found = out.setdefault(source.name, [])
|
||||
for name in names:
|
||||
key = catalog.resolve_table_name(source.schema_cache, name)
|
||||
for table in names:
|
||||
key = catalog.resolve_table_name(source.schema_cache, table)
|
||||
if key and key not in found:
|
||||
found.append(key)
|
||||
return {k: v for k, v in out.items() if v}
|
||||
@@ -604,7 +597,13 @@ def _edge_line(edge: JoinEdge, qualified) -> str:
|
||||
|
||||
|
||||
def _edges_within(graph: Mapping[str, list[JoinEdge]], tables: list[str]) -> list[JoinEdge]:
|
||||
"""两端都在 tables 里的关系,每条一次;一对多的翻过来,从「多」的一端写起(外键所在的表在左)。"""
|
||||
"""两端都在 tables 里的关系,每条一次;一对多的翻过来,从「多」的一端写起(外键所在的表在左)。
|
||||
|
||||
「每条一次」按 relation_id 去重,去掉的是同一条关系在关系图里的正反两条边。注意 catalog.relation_id 区分方向:
|
||||
A 表目录里记的 A→B 和 B 表目录里记的 B→A 是两个编号,两边都记了的话这里会各出一行(内容相同、写法相反)。
|
||||
这是有意不改的:编号存在每条关系上,也是审阅路径 relations.<编号> 的一部分,改成不分方向会让已有目录里的
|
||||
编号全部对不上。
|
||||
"""
|
||||
wanted = set(tables)
|
||||
seen: set[str] = set()
|
||||
out = []
|
||||
|
||||
@@ -17,6 +17,7 @@ from sqlalchemy import select as _select
|
||||
from app.api.coded import CodedHTTPException
|
||||
from app.core.bus import bus
|
||||
from app.core import artifact_store
|
||||
from app.core.events import EventType
|
||||
from app.data.catalog_impact import catalog_drift
|
||||
from app.db.base import SessionLocal, get_session
|
||||
from app.db.models import Approval, Artifact, Run, RunEvent, Workflow, WorkflowVersion
|
||||
@@ -188,7 +189,7 @@ async def start_run(
|
||||
if catalog_changes:
|
||||
# 落在封存范围内:之后回看这次正式出具,能知道当时的目录和发布时不一样、差在哪几张表。
|
||||
# 每张表带 impact:direct 写死的 SQL 用到的,possible Agent 可能查询的(界面分开说)
|
||||
await run_manager.note(run.id, "catalog.drift", version=version, count=len(catalog_changes),
|
||||
await run_manager.note(run.id, EventType.CATALOG_DRIFT, version=version, count=len(catalog_changes),
|
||||
tables=catalog_changes)
|
||||
# 已声明策略的升版记入事件流,出具物上能看到"这期换口径了、怎么处置的"
|
||||
for ev in upgrade_events:
|
||||
|
||||
@@ -183,3 +183,25 @@ def test_graph_sql_tables_uses_the_evidence_rules():
|
||||
))
|
||||
tables = {"shop": {"orders": {}, "order_items": {}, "customers": {}}, "other": {"orders": {}}}
|
||||
assert graph_sql_tables(spec, tables) == {"shop": {"orders", "order_items", "customers"}, "other": {"orders"}}
|
||||
|
||||
|
||||
def test_assistant_seeds_read_tool_nodes_the_same_way():
|
||||
"""助手给全字段的「现有工作流用到的表」和发布记录、SQL 检查认的是同一种节点(sqlcheck.tool_query)。
|
||||
|
||||
画布上把调用工具节点换成别的类型,配置里的 tool、args 还留着:那已经不是一条会执行的查询,不能当成用到的表。
|
||||
"""
|
||||
from app.api.copilot_context import graph_tables
|
||||
|
||||
shop = SimpleNamespace(name="shop", schema_cache={"tables": {
|
||||
"orders": {"qualified": "orders"}, "customers": {"qualified": "customers"}, "stores": {"qualified": "stores"}}})
|
||||
graph = _graph(
|
||||
_query("q1", "shop", "SELECT * FROM customers c JOIN orders o ON o.customer_id = c.id"),
|
||||
# 换成了 Agent 的节点:残留的写死查询不算
|
||||
_node("ask", "agent", tool="db_query__shop", args={"sql": "SELECT * FROM stores"}, tools=["db_query__shop"]),
|
||||
# GraphSpec 的写法(config 在节点上)照样认
|
||||
{"id": "q2", "type": "tool", "config": {"tool": "db_query__shop", "args": {"sql": "SELECT * FROM stores"}}},
|
||||
_query("q3", "shop", " "),
|
||||
)
|
||||
assert graph_tables(graph, [shop]) == {"shop": ["customers", "orders", "stores"]}
|
||||
graph["nodes"] = [n for n in graph["nodes"] if n["id"] != "q2"]
|
||||
assert graph_tables(graph, [shop]) == {"shop": ["customers", "orders"]}
|
||||
|
||||
Reference in New Issue
Block a user