可点击证据第二期:取数链路打通、问数据的答案逐段可点;发布前自动修复

一、可点击证据第二期

取数链路
- 智能体、工具节点、代码节点、知识检索取数时都记进证据台账;报告目录按台账编出 Q1… / K1…
- 报告可以引用快照里的一格([[v:Q1.r0.gmv]]),也可以由系统从快照生成整张表([[table:Q1 cols=… rows=0-4]]),
  表里每一格都是可点的片段;取值一律以快照为准,取回快照时复验哈希
- 口径卡表达式新增 cell(x, row, column),lineage 精确到单元格
- 智能体开 cite_fields(配合 output_schema)后,循环结束做一次结构化抽取,每个字段标出取自哪次查询的哪一格,
  再到快照里核对:一致记 verified;不一致以快照为准并告警;查不到的记为空值,永远不兜底成 0
- 代码节点新增 evidence_role(source / compute)和 code_sha;沙箱算出来的数不能被报告直接引用,要进口径卡
- 查询快照记下列类型,文本列按文本渲染(年份 "2026" 不再显示成「2,026」)
- 升级前发起的运行(停在审批上的)一律走旧逻辑:新的 task 只追加在循环之后,恢复时不会错号

证据接口与面板
- 片段链接上查询步骤:SQL、被引用的行加前后各 2 行、高亮到格;数据源可设 mask_columns 遮罩列,
  遮罩在查询时记进快照,数据源事后改名或删除也照样遮罩(在有身份体系之前,遮罩只减少暴露,不是安全边界)
- 输入来源写明核对结果(与快照一致 / 模型报的值和快照不一致 / 没查到记为空);口径卡缺输入时也给出完整的链
- 受管出具要在契约里声明 cells 才能引用单元格;受管与否记在所钉版本上,改画布不再让受管版本的正式运行降级
- 口径卡可以用 caliber_from 钉住另一个已发布工作流里的口径卡,上游有新版本时按 upgrade_policy 处置,
  和子工作流同一套规则;子工作流节点也补上了升版处置的配置项

Copilot、校验与报错
- 问数据的答案里有数字时,搭 input → 智能体 → 报告 → 出口,报告直接引用查过的单元格
- 写 SQL 前核对数据源的表结构,SQL 里出现表结构里没有的列会被打回
- 不再搭「智能体文字 → 转换节点解析 JSON」这种图;转换节点的 JSON 报错分清是上游模型写坏了、
  模板里裸插了模型文字(给出 | json 的改法),还是模板本身写错了
- 工具节点遇到「查询失败」「SQL 被拒绝」判为失败;发起运行前检查绑定的工具是否存在,不存在时直接拒绝并给出入口

二、发布前自动修复

- 问题带上机器编号(code)和修复 id;新增只读接口 POST /api/workflows/{id}/publish-check 和 /autofix
- 修复分三类:答案唯一、不降低要求的一键修复;要人拿主意的给选项(不替人选);结构性问题交给 Copilot
- 一律先预览,不自动保存、不自动发布;删节点、删契约、放宽审批、关掉 strict、改级别等降低要求的改动一律拒绝,
  Copilot 的方案也要过这一关;应用后重跑检查,错误必须变少且不能冒出新的
- 发布弹窗打开就先检查;问题面板新增「发布前检查」,平时就能看到、就地修掉
- 门禁修正:契约有 report_from 时不再要求 narrative(模板⑨可以发成受管);全图默认审批为「全部放行」时同样拦下
- 修掉画布保存会抹掉全图默认(defaults)的老问题:以前任何一次保存都会把它清空,审批 always 因此被悄悄放宽

验证(沙箱):tsc 干净,生产构建通过,pytest 2268 passed(两遍),check-all 17 项检查、3915 项全过
(新增 check-publish);关键行为做了变异验证、全部抓到。
This commit is contained in:
ThearchyHelios
2026-09-28 21:20:04 +08:00
parent 62aac187d8
commit fcdcc30d4e
74 changed files with 19372 additions and 380 deletions
+136 -7
View File
@@ -12,12 +12,14 @@ from app.api.runs import actor_of
from app.core.artifact_store import graph_hash
from app.db.base import get_session
from app.db.models import Run, Workflow, WorkflowVersion
from app.engine.governance import lint_for_publish
from app.engine.schema import GraphSpec, validate_graph
from app.engine.governance import publish_issues
from app.engine.schema import GraphSpec, NodeType, validate_graph
router = APIRouter(prefix="/api/workflows", tags=["workflows"])
_GONE = "这个工作流不存在,可能已经被删了"
_LEVELS = ("published", "governed")
_BAD_LEVEL = "发布档位只能选「已发布」或「受管」"
class WorkflowIn(BaseModel):
@@ -181,6 +183,8 @@ class VersionOut(BaseModel):
version: int
note: str
created_at: Any = None
#: 这一版按哪一档发布的(published / governed)。没发布过、或者是记级别之前发布的为 null
level: str | None = None
model_config = {"from_attributes": True}
@@ -238,7 +242,7 @@ async def get_version(
id=snapshot.id, version=snapshot.version, note=snapshot.note,
created_at=snapshot.created_at, workflow_id=workflow_id, graph=graph,
graph_hash=snapshot.graph_hash or graph_hash(graph), input_fields=fields,
published=workflow.published_version == snapshot.version,
published=workflow.published_version == snapshot.version, level=snapshot.level,
)
@@ -297,8 +301,8 @@ async def publish_workflow(
错误会挡住发布——这就是"必经检查点"落地的地方。
返回 200 + ok/issues 而不是抛错,让前端能把问题列表渲染出来。
"""
if payload.level not in ("published", "governed"):
raise HTTPException(400, "发布档位只能选「已发布」或「受管」")
if payload.level not in _LEVELS:
raise HTTPException(400, _BAD_LEVEL)
workflow = await session.get(Workflow, workflow_id)
if not workflow:
raise HTTPException(404, _GONE)
@@ -320,13 +324,14 @@ async def publish_workflow(
except Exception as e: # noqa: BLE001
return {"ok": False, "issues": [{"level": "error", "message": f"工作流的结构读不懂:{graph_error(e)}"}]}
issues = [i.model_dump() for i in validate_graph(spec).issues]
issues += [i.model_dump() for i in lint_for_publish(spec, level=payload.level).issues]
issues = [i.model_dump() for i in publish_issues(spec, level=payload.level)]
if any(i["level"] == "error" for i in issues):
return {"ok": False, "level": payload.level, "version": version, "issues": issues}
if not snapshot.graph_hash:
snapshot.graph_hash = graph_hash(snapshot.graph)
# 级别记在这一版上:之后改画布(status 退回 draft),从这一版发起的正式运行照样按发布时的级别出具
snapshot.level = payload.level
workflow.status = payload.level
workflow.published_version = version
workflow.published_by = actor_of(x_actor)
@@ -341,6 +346,130 @@ async def publish_workflow(
}
# --------------------------------------------------------------------------
# 发布前检查与自动修复:两个接口都只读,不改库、不发布
# --------------------------------------------------------------------------
class PublishCheckIn(BaseModel):
level: str = "published"
#: 画布上还没保存的图。不带就用库里当前的草稿
graph: dict[str, Any] | None = None
class AutofixIn(PublishCheckIn):
#: 要应用的修复 id(publish-check 给的 fixes[].id)
apply: list[str] = Field(default_factory=list)
#: choice 类修复的选择:{fix_id: 值},可以多选的给列表
choices: dict[str, Any] = Field(default_factory=dict)
#: 剩下的 error 交给 Copilot 试着修(结果同样只是预览)
assist: bool = False
provider: str | None = None
model: str | None = None
async def _draft(session: AsyncSession, workflow_id: str, payload: PublishCheckIn) -> dict[str, Any]:
if payload.level not in _LEVELS:
raise HTTPException(400, _BAD_LEVEL)
workflow = await session.get(Workflow, workflow_id)
if not workflow:
raise HTTPException(404, _GONE)
return payload.graph if payload.graph is not None else (workflow.graph or {})
async def _gate_context(session: AsyncSession, spec: GraphSpec) -> tuple[dict[str, Any], dict[str, Any]]:
"""修复要用、但门禁本身不查库的东西:子工作流当前的发布版本,钉在别处的口径卡有哪些指标。"""
versions: dict[str, Any] = {}
cards: dict[str, Any] = {}
for node in spec.nodes:
cfg = node.config
if node.type == NodeType.SUBGRAPH and cfg.get("workflow_id") and str(cfg["workflow_id"]) not in versions:
upstream = await session.get(Workflow, str(cfg["workflow_id"]))
versions[str(cfg["workflow_id"])] = upstream.published_version if upstream else None
elif node.type == NodeType.METRICS and isinstance(cfg.get("caliber_from"), dict):
ref = cfg["caliber_from"]
version = ref.get("workflow_version")
if not ref.get("workflow_id") or not str(version or "").isdigit():
continue
snapshot = (await session.execute(select(WorkflowVersion).where(
WorkflowVersion.workflow_id == str(ref["workflow_id"]),
WorkflowVersion.version == int(version)))).scalar_one_or_none()
card = next((n for n in ((snapshot.graph if snapshot else None) or {}).get("nodes") or []
if isinstance(n, dict) and n.get("id") == ref.get("node_id")), None)
if card is not None:
cards[node.id] = ((card.get("data") or {}).get("config") or {}).get("metrics") or []
return versions, cards
def _unreadable(level: str, e: Exception) -> dict[str, Any]:
return {"level": level, "ok": False, "fixes": [], "issues": [
{"level": "error", "node_id": None, "edge_id": None, "field": None, "code": None, "fix": None,
"message": f"工作流的结构读不懂:{graph_error(e)}"}]}
@router.post("/{workflow_id}/publish-check")
async def publish_check(
workflow_id: str, payload: PublishCheckIn, session: AsyncSession = Depends(get_session)
) -> dict[str, Any]:
"""发布前检查:和真正发布同一套口径(validate + 门禁)列出问题,以及能怎么修。
只读:不改库、不发布。发布弹窗一打开就调它,不用等点了发布才知道被拦。
"""
from app.engine.autofix import check
graph = await _draft(session, workflow_id, payload)
try:
spec = GraphSpec.model_validate(graph)
except Exception as e: # noqa: BLE001
return _unreadable(payload.level, e)
versions, cards = await _gate_context(session, spec)
return check(spec, level=payload.level, versions=versions, cards=cards)
@router.post("/{workflow_id}/autofix")
async def autofix(
workflow_id: str, payload: AutofixIn, session: AsyncSession = Depends(get_session)
) -> dict[str, Any]:
"""应用选中的修复,返回修复后的整张图——只是预览。
不自动保存、不自动发布:人看过逐项变更再确认,前端走现有的保存接口存草稿,再检查、再发布。
每条修复都复核过(降低要求的、错误没变少或者冒出新错误的丢弃并写明原因)。assist 为真时,
剩下的 error 交给 Copilot,同一道复核通过才采纳。
"""
from app.engine.autofix import apply_fixes, diff_changes
graph = await _draft(session, workflow_id, payload)
try:
spec = GraphSpec.model_validate(graph)
except Exception as e: # noqa: BLE001
bad = _unreadable(payload.level, e)
return {"graph": graph, "changes": [], "ops": [], "applied": [], "rejected": [
{"fix_id": fid, "reason": "工作流的结构读不懂,没法应用修复"} for fid in payload.apply],
"remaining": bad["issues"], "fixes": [], "assist": None, "ok": False}
versions, cards = await _gate_context(session, spec)
out = apply_fixes(graph, payload.apply, payload.choices, level=payload.level, versions=versions, cards=cards)
out["assist"] = None
if payload.assist:
if out["ok"]:
out["assist"] = {"ok": True, "summary": "修复之后已经没有挡住发布的错误,没有再交给 Copilot",
"questions": []}
return out
from app.api.copilot import assist_publish_fix
helped = await assist_publish_fix(session, out["graph"], level=payload.level,
provider=payload.provider, model=payload.model)
out["assist"] = {"ok": helped["accepted"], "summary": helped["summary"], "questions": helped["questions"]}
if helped["accepted"]:
fixed = helped["graph"]
out["changes"] += diff_changes(out["graph"], fixed, fix_id="assist", label="Copilot 的修改")
out["ops"] += helped["ops"]
out["applied"].append("assist")
after = apply_fixes(fixed, [], level=payload.level, versions=versions, cards=cards)
out.update(graph=fixed, remaining=after["remaining"], fixes=after["fixes"], ok=after["ok"])
elif helped["reason"]:
out["rejected"].append({"fix_id": "assist", "reason": helped["reason"]})
return out
class ValidateIn(BaseModel):
graph: dict[str, Any]