diff --git a/backend/app/api/ledger.py b/backend/app/api/ledger.py index f349c1d1..d1bdfa06 100644 --- a/backend/app/api/ledger.py +++ b/backend/app/api/ledger.py @@ -330,6 +330,8 @@ async def trial_caliber(workflow_id: str, version: int, node_id: str, payload: T pinned = await load_pinned(session, workflow_id, int(version), node_id) except NotDerivable as e: raise HTTPException(409, e.message) from None + if not pinned.published or pinned.retired: + raise HTTPException(409, "这一版没有发布或已退役，不能试运行") fields = [dict(f) for f in pinned.entry_fields] known = {str(f.get("name")) for f in fields} unknown = sorted(str(k) for k in payload.inputs if str(k) not in known) @@ -343,7 +345,7 @@ async def trial_caliber(workflow_id: str, version: int, node_id: str, payload: T graph = caliber_probe(pinned).graph principal = current_principal() run = await run_manager.start( - graph=graph, input_payload=values, run_class="exploratory", origin_kind="trial", + graph=graph, input_payload=values, run_class="exploratory", origin_kind="caliber_trial", origin={"kind": "caliber_trial", "workflow_id": workflow_id, "version": int(version), "node_id": node_id, "card": pinned.node_id}, started_by=principal.username) diff --git a/backend/app/data/caliber_lint.py b/backend/app/data/caliber_lint.py index 2e3614ee..98e6ed1d 100644 --- a/backend/app/data/caliber_lint.py +++ b/backend/app/data/caliber_lint.py @@ -15,7 +15,7 @@ ROUND_IN_SOURCE = "caliber.round_in_source" ZERO_DEFAULT = "caliber.zero_default" ROUND_MESSAGE = ( - "取数 SQL 里写了 ROUND(...)：取整交给口径卡的小数位和展示；取数 SQL 里写 ROUND 会让未取整的查询" + "取数 SQL 里写了取整函数：取整交给口径卡的小数位和展示；取数 SQL 里写取整会让未取整的查询" "永远认不成等价（口径账本只认规范化后完全相同的结果）" ) ZERO_MESSAGE = ( diff --git a/backend/app/data/rule_compile.py b/backend/app/data/rule_compile.py index 36298947..90b6467b 100644 --- a/backend/app/data/rule_compile.py +++ b/backend/app/data/rule_compile.py @@ -108,8 +108,11 @@ def run_fixtures(rule: Mapping[str, Any], table: str, fixtures: Iterable[Mapping columns: dict[str, str] = {} for item in items: for row in item.get("rows") or []: + if not isinstance(row, Mapping): + raise RuleInvalid("fixture 的 rows 需要是对象") for name, value in row.items(): - columns.setdefault(str(name), "INTEGER" if isinstance(value, int) and not isinstance(value, bool) + key = _ident(name, "fixture 列名") + columns.setdefault(key, "INTEGER" if isinstance(value, int) and not isinstance(value, bool) else "REAL" if isinstance(value, float) else "TEXT") if not columns: raise RuleInvalid("fixture 需要至少一行 rows") diff --git a/backend/app/engine/nodes/dual_check.py b/backend/app/engine/nodes/dual_check.py index 9ce7e610..4fa40b83 100644 --- a/backend/app/engine/nodes/dual_check.py +++ b/backend/app/engine/nodes/dual_check.py @@ -512,6 +512,9 @@ async def _plan_metric(state: GraphState, metric: dict[str, Any], options: _Opti "status": "pending", "original": {"value": metric.get("value"), "rendered": metric.get("rendered")}, "blind": {"value": None, "rendered": None}, "delta_pct": None, "inputs": [], "differing": [], } + if metric.get("rule") is not None or metric.get("assessment") is not None or metric.get("external") is not None: + row.update(status=UNCHECKED, reason="这个指标不是按表达式算的，不参与双盲复算") + return row, {} tops = _tops(metric) # masks 是每份用到的快照要遮的列（cell() 的、按位置取的格落在的），queries 只是 cell() 用到的那几份：重算时容器 # 输入按内容对快照只对这几份。按位置取的格落在的快照算遮罩只为认这一格，不当成指标用到的结果——放进去的话，同一个 diff --git a/backend/app/engine/nodes/metrics.py b/backend/app/engine/nodes/metrics.py index 3dd97c88..1bcb3227 100644 --- a/backend/app/engine/nodes/metrics.py +++ b/backend/app/engine/nodes/metrics.py @@ -120,18 +120,31 @@ async def run_metrics(state: GraphState, ctx: NodeContext) -> dict[str, Any]: value: Any = None details: list[dict[str, Any]] = [] level: str | None = None + out_of_range: list[str] = [] if not missing: total = sum(float(d.get("weight", 1)) for d in dimensions) points = 0.0 + bad: str | None = None for d in dimensions: low, high = float(d.get("min", 0)), float(d.get("max", 100)) - raw = float(values_by_name[str(d["name"])]) + try: + raw = float(values_by_name[str(d["name"])]) + except (TypeError, ValueError): + bad = f"维度「{d['name']}」的输入不是数（{values_by_name[str(d['name'])]!r}）" + break normalized = max(0.0, min(100.0, (raw - low) / (high - low) * 100)) weight = float(d.get("weight", 1)) points += normalized * weight - details.append({"name": d["name"], "value": raw, "normalized": normalized, "weight": weight}) - value = round_metric(points / total, display["decimals"]) - level = next((str(l["name"]) for l in assessment["levels"] if value >= float(l["min"])), None) + if raw < low or raw > high: + out_of_range.append(str(d["name"])) + details.append({"name": d["name"], "value": raw, "normalized": normalized, "weight": weight, + "out_of_range": raw < low or raw > high}) + if bad: + errors.append(f"指标「{metric_id}」：{bad}") + continue + raw_score = points / total + value = round_metric(raw_score, display["decimals"]) + level = next((str(l["name"]) for l in assessment["levels"] if raw_score >= float(l["min"])), None) inputs = [_input_of(str(d["input"]), values_by_name[str(d["name"])], state, ctx, tctx) for d in dimensions] metrics.append({ @@ -145,10 +158,14 @@ async def run_metrics(state: GraphState, ctx: NodeContext) -> dict[str, Any]: "rendered": _rendered(value, definition, display), "inputs": inputs, "substituted": "加权评分", - "recompute_ok": True, + "recompute_ok": None, + "recompute_reason": "这个指标不按表达式复算", "status": "ok" if value is not None else "missing_input", "assessment": {"score": value, "level": level, "dimensions": details, - "levels": assessment["levels"], **({"missing": missing} if missing else {})}, + "levels": assessment["levels"], + **({"missing": missing} if missing else {}), + **({"out_of_range": out_of_range} if out_of_range else {}), + **({"unlevelled": True} if value is not None and level is None else {})}, }) continue if str(definition.get("kind") or "formula") == "rule": @@ -179,7 +196,8 @@ async def run_metrics(state: GraphState, ctx: NodeContext) -> dict[str, Any]: "rendered": _rendered(value, definition, display), "inputs": inputs, "substituted": _rule_text(rule), - "recompute_ok": True, + "recompute_ok": None, + "recompute_reason": "这个指标不按表达式复算", "status": "ok" if value is not None else "missing_input", "rule": rule, }) @@ -202,7 +220,8 @@ async def run_metrics(state: GraphState, ctx: NodeContext) -> dict[str, Any]: "rendered": _rendered(value, definition, display), "inputs": [], "substituted": str(value), - "recompute_ok": True, + "recompute_ok": None, + "recompute_reason": "这个指标不按表达式复算", "status": "ok", "external": external, }) @@ -605,10 +624,18 @@ def _rule_matches(values: dict[str, Any], rule: dict[str, Any]) -> bool: return False continue b = cond.get("value") - try: - ok = {"eq": a == b, "ne": a != b, "gt": a > b, "gte": a >= b, "lt": a < b, "lte": a <= b}.get(op, False) - except TypeError: - ok = False + # 类型对不上（字符串和数）按条件不成立，不静默走 ne；大小比较只对数 + if op in ("eq", "ne"): + comparable = (isinstance(a, (int, float)) and not isinstance(a, bool) + and isinstance(b, (int, float)) and not isinstance(b, bool)) or type(a) is type(b) + if not comparable: + return False + equals = a == b + ok = equals if op == "eq" else not equals + else: + if not all(isinstance(v, (int, float)) and not isinstance(v, bool) for v in (a, b)): + return False + ok = {"gt": a > b, "gte": a >= b, "lt": a < b, "lte": a <= b}.get(op, False) if not ok: return False return True diff --git a/backend/app/engine/origin.py b/backend/app/engine/origin.py index 25081ff7..ffb8fa17 100644 --- a/backend/app/engine/origin.py +++ b/backend/app/engine/origin.py @@ -29,6 +29,7 @@ ORIGIN_KINDS = ( "ledger_issue", # 问数据的答案按已治理口径正式出具（正式运行） "reconcile", # 口径对账：这次的取数和口径在同一份数据上各算一次（探索运行） "recall", # 口径召回：新旧两版口径在同一份数据上各算一次（探索运行） + "caliber_trial", # 口径目录里的试跑（探索运行，不是问数回答，也不进统计/召回） ) #: 客户端在 POST /api/runs 里能声明的来源 @@ -38,7 +39,7 @@ CLIENT_ORIGINS = ("chat", "canvas", "launcher", "api") #: 不是有人在用。它们是完整的运行（证据、封存都在、点得开），但不进使用统计——探索问题聚类、工具白名单衰减、数据目录的 #: 使用次数和「被哪些运行用过」、审阅队列的使用权重、以后的账本统计——否则一跑评测，治理信号和目录的使用热度就失真了。 #: 记录页默认也不列（筛选里能选，列表照实写有几次没列出）。ledger_issue 是替用户正式出具，照常计入 -SYSTEM_ORIGINS = ("eval", "reconcile", "recall") +SYSTEM_ORIGINS = ("eval", "reconcile", "recall", "caliber_trial") def counted_runs() -> Any: diff --git a/backend/app/engine/schema.py b/backend/app/engine/schema.py index 26fc5de5..36bcaa21 100644 --- a/backend/app/engine/schema.py +++ b/backend/app/engine/schema.py @@ -1086,8 +1086,8 @@ def _check_challenge(node: GraphNode, spec: GraphSpec, result: ValidationResult) field="model", code="challenge.same_model") -#: 指标定义的类型。现在只有公式（表达式）这一种在运行时能算；规则、评估、外部常量还没实现， -#: 先在校验里挡住，免得写进图里以后被当成能算的指标 +#: 指标定义的类型：公式（表达式）、外部常量、单元格条件规则、评估评分。四种都在运行时能算； +#: 不认识的类型仍在校验里挡住，免得被当成能算的指标 METRIC_KINDS = ("formula", "external", "rule", "assessment") #: 定义卡参数的标量类型（和入口表单的字段类型同一套子集） METRIC_PARAM_TYPES = ("string", "number", "integer", "boolean") @@ -1110,8 +1110,8 @@ def metric_definition_issues(node: GraphNode, metric: Any, index: int) -> list[V add(f"指标定义缺少 ID：{field}", code="metrics.id_missing") kind = str(metric.get("kind") or "formula") if kind not in METRIC_KINDS: - add(f"指标「{metric.get('id') or index + 1}」的类型「{kind}」还没有实现：现在支持「公式」（formula）和" - "「外部常量」（external）。规则、评估请先在定义卡里写成说明，等对应的执行方式上线后再发布", + add(f"指标「{metric.get('id') or index + 1}」的类型「{kind}」不认识：现在支持「公式」（formula）、" + "「外部常量」（external）、「规则」（rule）、「评估」（assessment）", f"{field}.kind", code="metrics.kind_unsupported") if kind == "external": external = metric.get("external") @@ -1151,6 +1151,10 @@ def metric_definition_issues(node: GraphNode, metric: Any, index: int) -> list[V if not isinstance(name, str) or not name.strip() or not isinstance(path, str) \ or not path.strip(): add("规则的每个输入都要有名字和取值路径", f"{field}.rule.inputs", code="metrics.rule_invalid") + elif not any(token in path for token in ("cell(", "nodes.", "vars.", "input.")): + add(f"规则输入「{name}」的取值路径要写成 cell(nodes.…)、vars.… 或 input.…；" + "表列名是规则 DSL 编译接口的写法，不能直接写进口径卡", + f"{field}.rule.inputs.{name}", code="metrics.rule_invalid") conditions = rule.get("all") if not isinstance(conditions, list) or not conditions: add("规则需要 all：至少一条条件", f"{field}.rule.all", code="metrics.rule_invalid") @@ -1163,8 +1167,8 @@ def metric_definition_issues(node: GraphNode, metric: Any, index: int) -> list[V if cond.get("op") not in RULE_OPS: add(f"条件运算符只能是 {'、'.join(RULE_OPS)}", f"{key}.op", code="metrics.rule_invalid") if cond.get("op") in ("in", "not_in"): - if not isinstance(cond.get("values"), list): - add("in / not_in 需要 values 列表", f"{key}.values", code="metrics.rule_invalid") + if not isinstance(cond.get("values"), list) or not cond["values"]: + add("in / not_in 需要非空的 values 列表", f"{key}.values", code="metrics.rule_invalid") elif "value" not in cond: add("条件需要 value", f"{key}.value", code="metrics.rule_invalid") for key in ("then", "else"): @@ -1194,6 +1198,9 @@ def metric_definition_issues(node: GraphNode, metric: Any, index: int) -> list[V names.add(name) if not isinstance(dimension.get("input"), str) or not dimension["input"].strip(): add("评估维度需要 input（取值路径）", f"{key}.input", code="metrics.assessment_invalid") + elif not any(token in dimension["input"] for token in ("cell(", "nodes.", "vars.", "input.")): + add("评估维度的 input 要写成 cell(nodes.…)、vars.… 或 input.…", + f"{key}.input", code="metrics.assessment_invalid") weight = dimension.get("weight", 1) if isinstance(weight, bool) or not isinstance(weight, (int, float)) or weight <= 0: add("评估维度的 weight 需要是正数", f"{key}.weight", code="metrics.assessment_invalid") @@ -1220,6 +1227,10 @@ def metric_definition_issues(node: GraphNode, metric: Any, index: int) -> list[V add("评估等级要按 min 从高到低排列", f"{key}.min", code="metrics.assessment_invalid") else: previous = float(minimum) + if levels and isinstance(levels[-1], dict) and isinstance(levels[-1].get("min"), (int, float)) \ + and not isinstance(levels[-1]["min"], bool) and levels[-1]["min"] > 0: + add("最后一个评估等级的 min 要不大于 0，否则低分没有等级可落", + f"{field}.assessment.levels[-1].min", code="metrics.assessment_invalid") elif not metric.get("expression"): add(f"指标定义缺少表达式：{metric.get('id') or '（空）'}", code="metrics.expression_missing") diff --git a/backend/app/ledger/descriptor.py b/backend/app/ledger/descriptor.py index c32bce18..e6e4e48a 100644 --- a/backend/app/ledger/descriptor.py +++ b/backend/app/ledger/descriptor.py @@ -77,6 +77,17 @@ async def _describe(session: AsyncSession, workflow: Workflow, row: Any, node_id cfg = card["config"] level = str(row.level or "") _, reason = card_inputs(graph, node_id) + if reason is None: + from app.ledger.derive import template_problem + + for nid, node in nodes.items(): + tool = str((node.get("config") or {}).get("tool") or "") + sql = ((node.get("config") or {}).get("args") or {}).get("sql") \ + if isinstance((node.get("config") or {}).get("args"), dict) else None + if node.get("type") == "tool" and tool.startswith(QUERY_PREFIX) and isinstance(sql, str) and sql.strip(): + if problem := template_problem(str(nid), node.get("config") or {}): + reason = problem + break key = (workflow.id, int(row.version)) if key not in side_cache: side_cache[key] = await graph_side_effect(session, graph, level=level or None) @@ -86,7 +97,9 @@ async def _describe(session: AsyncSession, workflow: Workflow, row: Any, node_id if level != "governed": reasons.append("这一版是已发布级别，只支持一键出具") if side_effect is not None: - reasons.append(f"这一版{side_effect}，不能替人重跑") + reasons.append(f"这一版{side_effect}，不能替人再次执行") + if not sources: + reasons.append("这一版没有取数节点，账本不能按口径证明等价") return { "workflow_id": workflow.id, "workflow_name": workflow.name, @@ -103,7 +116,7 @@ async def _describe(session: AsyncSession, workflow: Workflow, row: Any, node_id "compose_reason": reason, "dependencies": {"sources": sorted(sources), "tables": sorted(tables), "catalog_versions": row.catalog_versions, "schema_versions": row.schema_versions}, - "issuance": {"level": level, "auto_issue": level == "governed" and side_effect is None, + "issuance": {"level": level, "auto_issue": level == "governed" and side_effect is None and bool(sources), "requires_independent_proof": True, "side_effects": side_effect, "reasons": reasons}, } diff --git a/backend/app/ledger/labels.py b/backend/app/ledger/labels.py index a124d4c8..f339af40 100644 --- a/backend/app/ledger/labels.py +++ b/backend/app/ledger/labels.py @@ -1,8 +1,11 @@ """逐个数标口径（docs/design/next-round/02-口径账本.md 做法 3）：报告里每个引用了查询格、口径卡指标的数，来自哪一份口径。 -标签四种（KINDS）： +标签（KINDS）： - caliber：来自口径「X」vN。how 是怎么认的：independent「等价已证明」（这次 Agent 自己写的 SQL 被证明和口径的取数等价）、 composed「助手选用」（助手 use_caliber 按口径组合出来的取数，证明了和那一版一致：ledger/match 的 composed）。 +- external：外部常量（口径卡上的 kind=external），带来源、生效日期、是否过期。 +- derived：规则型指标（kind=rule），数由结构化规则从输入算出，不是某条查询与口径等价。 +- assessed：评估型指标（kind=assessment），按评分协议算出的评估值，不是测量值。 - related：相关口径——同一个数据源、查了同样的表的已发布口径，但不等价；附差在哪（explain_sql_diff 的条目）。 - reasoning：这次的推理——没有相关的已发布口径。 - unknown：没比对、比对没做完（超时、解析失败）、快照读不出来。 @@ -309,13 +312,26 @@ def metric_label(card: Mapping[str, Any] | None, metric_id: Any, queries: Mappin inputs = [i for i in metric.get("inputs") or [] if isinstance(i, dict)] from_queries = [i for i in inputs if i.get("artifact") and i.get("via") in _SNAPSHOT_VIAS] if metric.get("rule") is not None: - # 规则型指标：数由规则从输入算出，不是某条查询与口径等价；输入仍然逐格可追溯 - return {"kind": "derived", "rule": metric.get("rule"), + # 规则型指标：数由规则从输入算出，不是某条查询与口径等价；输入仍然逐格可追溯。 + # 标签里只给结构（输入名、条件条数、then/else），不给条件里的字面量——那可能是遮罩列的取值 + rule = metric.get("rule") if isinstance(metric.get("rule"), dict) else {} + return {"kind": "derived", + "rule": {"inputs": sorted(str(k) for k in (rule.get("inputs") or {})), + "conditions": len(rule.get("all") or []), + "then": rule.get("then", 1), "else": rule.get("else", 0)}, "queries": [str(i["artifact"]) for i in from_queries], **({"definition_hash": card.get("definition_hash")} if card.get("definition_hash") else {})} if metric.get("assessment") is not None: - # 评估值：按评分协议从输入算出，不是测量值，也不参与自动出具 - return {"kind": "assessed", "assessment": metric.get("assessment"), + # 评估值：按评分协议从输入算出，不是测量值，也不参与自动出具。标签里不带各维度的原始取值 + # （可能是遮罩列的取值），只给得分、等级和维度名/权重 + assessment = metric.get("assessment") if isinstance(metric.get("assessment"), dict) else {} + dimensions = [{"name": d.get("name"), "weight": d.get("weight")} + for d in assessment.get("dimensions") or [] if isinstance(d, dict)] + return {"kind": "assessed", + "assessment": {"score": assessment.get("score"), "level": assessment.get("level"), + "dimensions": dimensions, + "levels": [{"name": l.get("name"), "min": l.get("min")} + for l in assessment.get("levels") or [] if isinstance(l, dict)]}, "queries": [str(i["artifact"]) for i in from_queries], **({"definition_hash": card.get("definition_hash")} if card.get("definition_hash") else {})} if not from_queries: @@ -445,7 +461,11 @@ def whole_of(labels: Mapping[str, Mapping[str, Any]], calibers: Mapping[str, Map keys = {label.get("caliber") for label in labels.values() if label.get("kind") == "caliber"} if counts["caliber"] != len(labels): # 逐格判之后常见「一部分数来自口径、另几个只是相关」：照实说是不一致，不说成这次的推理 - reason = ("有的数没有比对" if counts["unknown"] else "没有数来自口径" if not counts["caliber"] + reason = ("有的数没有比对" if counts["unknown"] + else "有的数是外部常量，不是测量值" if counts["external"] + else "有的数是规则派生的，不是测量值" if counts["derived"] + else "有的数是评估值，不是测量值" if counts["assessed"] + else "没有数来自口径" if not counts["caliber"] else "有的数和相关口径不一致，不来自口径" if counts["related"] else "有的数是这次的推理") return {"equivalent": False, "reason": reason, "counts": counts} if len(keys) != 1: diff --git a/backend/app/ledger/store.py b/backend/app/ledger/store.py index 0d10693f..295bde7b 100644 --- a/backend/app/ledger/store.py +++ b/backend/app/ledger/store.py @@ -34,7 +34,8 @@ FORMAT = "agentlab.caliber_match/1" #: 3：整条比不等价时按报告引用的列投影再比（ledger/match 的 _projected），候选记 projection；没投影的照旧 #: 4：按全部引用列投影仍没证明的查询，被引用的每一列逐格各比一次（queries[].cells、one_row），每个数按它那一格标；老工件 #: 没有 cells，照旧按整条查询读 -ALGO = "4" +#: 5：候选记定义指纹（caliber_hash / definition_hashes），标签新增 external / derived / assessed；老工件没有这些键照旧读 +ALGO = "5" #: 触发方式（见模块说明）。certified（第 3 波）：从问数据固化出的口径发布之后，对来源运行补比的（ledger/certify） TRIGGERS = ("sealed", "backfill", "on_demand", "certified") diff --git a/backend/app/look/calibers.py b/backend/app/look/calibers.py index 8bae0ca2..75f0e394 100644 --- a/backend/app/look/calibers.py +++ b/backend/app/look/calibers.py @@ -106,33 +106,59 @@ def card_inputs(graph: dict[str, Any], card_id: str) -> tuple[list[dict[str, Any return [], CHAINED_REASON found: dict[str, dict[str, Any]] = {} for metric in card["config"].get("metrics") or []: - try: - tree = parse_expression(str((metric or {}).get("expression") or ""))[0] - except ExpressionError: + paths = _metric_paths(metric) + if paths is None: return [], NOT_TOOL_REASON - for path in leaf_refs(tree): - parts = cell_parts(path) - if parts is None: - return [], NOT_TOOL_REASON - chain, row, column = parts - node_id = chain[len("nodes."):] if chain.startswith("nodes.") else "" - tool = str((nodes.get(node_id) or {}).get("config", {}).get("tool") or "") - if (nodes.get(node_id) or {}).get("type") != "tool" or not tool.startswith(QUERY_PREFIX): - return [], NOT_TOOL_REASON - if row.strip() != "0": - return [], ROW_REASON + for path in paths: try: - col = json.loads(column.replace("'", '"')) if column[:1] in ("'", '"') else None - except ValueError: - col = None - if not isinstance(col, str): + tree = parse_expression(path)[0] + except ExpressionError: return [], NOT_TOOL_REASON - entry = found.setdefault(node_id, {"node": node_id, "tool": tool, "columns": []}) - if col not in entry["columns"]: - entry["columns"].append(col) + for leaf in leaf_refs(tree): + parts = cell_parts(leaf) + if parts is None: + return [], NOT_TOOL_REASON + chain, row, column = parts + node_id = chain[len("nodes."):] if chain.startswith("nodes.") else "" + tool = str((nodes.get(node_id) or {}).get("config", {}).get("tool") or "") + if (nodes.get(node_id) or {}).get("type") != "tool" or not tool.startswith(QUERY_PREFIX): + return [], NOT_TOOL_REASON + if row.strip() != "0": + return [], ROW_REASON + try: + col = json.loads(column.replace("'", '"')) if column[:1] in ("'", '"') else None + except ValueError: + col = None + if not isinstance(col, str): + return [], NOT_TOOL_REASON + entry = found.setdefault(node_id, {"node": node_id, "tool": tool, "columns": []}) + if col not in entry["columns"]: + entry["columns"].append(col) return list(found.values()), None +def _metric_paths(metric: Any) -> list[str] | None: + """一个指标读的取值路径：公式的表达式、规则各输入、评估各维度。外部常量没有输入。认不出为 None。""" + if not isinstance(metric, dict): + return None + paths: list[str] = [] + if metric.get("expression"): + paths.append(str(metric["expression"])) + rule = metric.get("rule") + if isinstance(rule, dict): + inputs = rule.get("inputs") + if not isinstance(inputs, dict): + return None + paths.extend(str(path) for path in inputs.values()) + assessment = metric.get("assessment") + if isinstance(assessment, dict): + dimensions = assessment.get("dimensions") + if not isinstance(dimensions, list): + return None + paths.extend(str(d.get("input")) for d in dimensions if isinstance(d, dict) and d.get("input")) + return paths + + def views_of(workflow_id: str, workflow_name: str, version: Any) -> list[CaliberView]: """一个已发布版本里每张口径卡（源头的）的每个指标。""" from app.apps.tree import node_map diff --git a/backend/tests/test_ledger_cells.py b/backend/tests/test_ledger_cells.py index 3f397d81..c4538e40 100644 --- a/backend/tests/test_ledger_cells.py +++ b/backend/tests/test_ledger_cells.py @@ -264,7 +264,7 @@ async def test_last_weeks_sales_labels_only_the_sales_number_as_from_the_caliber row = await S.explore(monkeypatch, ANSWER, origin="chat", text=( "上周销售额 [[v:Q1.r0.销售额]] 元，共 [[v:Q1.r0.订单数]] 单（[[v:Q1.r0.区间首日]] 至 [[v:Q1.r0.区间末日]]）。")) content = (await store.latest(row.id)).content - assert content["algo"] == store.ALGO == "4" + assert content["algo"] == store.ALGO == "5" [query] = content["queries"] assert query["status"] == "related" and set(query["cells"]) == set(COLUMNS) [labels] = [r["labels"] for r in content["reports"].values()] diff --git a/backend/tests/test_ledger_compose.py b/backend/tests/test_ledger_compose.py index 9171b521..6d15fe23 100644 --- a/backend/tests/test_ledger_compose.py +++ b/backend/tests/test_ledger_compose.py @@ -253,7 +253,7 @@ async def test_an_answer_built_with_use_caliber_is_labelled_composed(client, mon await ledger_match.drain() recorded = await store.latest(row.id) content = recorded.content - assert content["algo"] == store.ALGO == "4" + assert content["algo"] == store.ALGO == "5" [query] = content["queries"] assert query["status"] == "composed" and query["node_id"] == "c1_fetch" [cand] = query["candidates"] diff --git a/backend/tests/test_ledger_descriptor.py b/backend/tests/test_ledger_descriptor.py index 5b9c1654..c0541a61 100644 --- a/backend/tests/test_ledger_descriptor.py +++ b/backend/tests/test_ledger_descriptor.py @@ -117,3 +117,17 @@ async def test_metric_definition_and_params_are_described(client): [m] = made["metrics"] assert m["definition"]["text"] == "已支付订单的销售额" and m["definition"]["owner"] == "运营" assert m["params"] == [{"name": "store", "type": "string", "default": "all"}] + + +async def test_an_external_only_card_is_not_promised_automatic_issuance(client): + graph = S.chain( + S.node("start", "input"), + S.node("card", "metrics", caliber="年度目标", caliber_version="v1", metrics=[{ + "id": "target", "name": "年度目标", "unit": "元", "kind": "external", + "external": {"value": 100, "source": "预算", "as_of": "2026-01-01"}}]), + S.node("out", "output", fields=[{"name": "v", "value": "{{ nodes.card.metrics[0].value }}"}])) + wf = await S.publish(name="只有外部常量", graph=graph) + [made] = await descriptors(client, wf) + assert made["composable"] is True + assert made["issuance"]["auto_issue"] is False + assert any("没有取数节点" in r for r in made["issuance"]["reasons"]) diff --git a/backend/tests/test_ledger_trial.py b/backend/tests/test_ledger_trial.py index 51559ea7..39ec69bf 100644 --- a/backend/tests/test_ledger_trial.py +++ b/backend/tests/test_ledger_trial.py @@ -42,8 +42,10 @@ async def test_trial_runs_the_pinned_caliber_and_returns_its_metrics(client): assert body["metrics"][0]["unit"] == "元" async with SessionLocal() as session: run = await session.get(Run, body["run_id"]) - assert run is not None and run.origin_kind == "trial" and run.run_class == "exploratory" + assert run is not None and run.origin_kind == "caliber_trial" and run.run_class == "exploratory" assert run.origin["kind"] == "caliber_trial" and run.origin["workflow_id"] == wf + # 试运行不是问数回答：不参与账本比对、不计入使用统计/召回 + assert not ledger_match.wanted(run.run_class, run.origin_kind) async def test_trial_refuses_missing_required_inputs_and_unknown_fields(client): diff --git a/backend/tests/test_metric_definition.py b/backend/tests/test_metric_definition.py index e3beedd4..d4e02caa 100644 --- a/backend/tests/test_metric_definition.py +++ b/backend/tests/test_metric_definition.py @@ -1,7 +1,6 @@ -"""指标定义卡（definition）与参数（params）的最小结构校验。 +"""指标定义卡（definition）与参数（params）：结构校验、四种 kind（公式/外部常量/规则/评估）的执行与标签。 -执行还是公式这一种；定义文本、参数先作为语义元数据校验、进内容指纹、进 descriptor。规则/评估/外部常量没实现前， -写进图里要按错误挡住，免得被当成能算的指标。 +定义文本、参数进内容指纹和 descriptor；不认识的 kind 写进图里要按错误挡住，免得被当成能算的指标。 """ from __future__ import annotations @@ -30,7 +29,7 @@ def test_a_formula_metric_can_carry_a_definition_and_params(): def test_an_unimplemented_kind_is_refused(): errors = _errors(_graph(kind="normative")) - assert any(i.code == "metrics.kind_unsupported" and "还没有实现" in i.message for i in errors) + assert any(i.code == "metrics.kind_unsupported" and "不认识" in i.message for i in errors) def test_params_must_be_named_and_typed_once_each(): @@ -72,6 +71,7 @@ async def test_an_external_constant_metric_runs_without_a_query(): metric = card["metrics"][0] assert metric["external"]["source"] == "2026 年度预算决议" assert metric["external"]["stale"] is False + assert metric["recompute_ok"] is None and metric["recompute_reason"] finally: await run_manager.shutdown() @@ -147,7 +147,8 @@ def test_a_rule_metric_is_labelled_derived_with_its_input_queries(): }]} label = metric_label(card, "big", {}, {}) assert label["kind"] == "derived" and label["queries"] == ["q1"] - assert label["definition_hash"] == "abc" and label["rule"]["all"][0]["op"] == "eq" + assert label["definition_hash"] == "abc" and label["rule"]["conditions"] == 1 + assert "paid" not in str(label) def test_an_assessment_metric_needs_unique_dimensions_and_descending_levels(): @@ -198,9 +199,12 @@ def test_an_assessment_metric_is_labelled_assessed(): from app.ledger.labels import metric_label card = {"definition_hash": "abc", "metrics": [{ - "id": "quality", "assessment": {"score": 80, "level": "B"}, + "id": "quality", + "assessment": {"score": 80, "level": "B", + "dimensions": [{"name": "手机号", "weight": 1, "value": "19900000501"}]}, "inputs": [{"artifact": "q1", "via": "tool_cell"}], }]} label = metric_label(card, "quality", {}, {}) assert label["kind"] == "assessed" and label["queries"] == ["q1"] assert label["assessment"]["level"] == "B" and label["definition_hash"] == "abc" + assert "19900000501" not in str(label) diff --git a/docs/zh-CN/data-catalog.md b/docs/zh-CN/data-catalog.md index ae8e12da..4cd5eccd 100644 --- a/docs/zh-CN/data-catalog.md +++ b/docs/zh-CN/data-catalog.md @@ -130,7 +130,8 @@ 剖析范围选「取值观察」或定时维护时，还会观察取值：行数、码值分布（包括人工确认过的码值列）、最新业务日期、最近 35 天每天的行数。 - 和同一口径（同样的取样方式）的历史比，比出新码值、比例剧变、断档、缺天、量跌、行数大起大落，记进变化记录（见下文）。 -- 标了「已列全」的码值表出现目录里没有的取值时，系统撤销「已列全」，目录升一版。这是机器唯一会自己写进目录的事；新取值的含义仍要人填。 +- 标了「已列全」的码值表出现目录里没有的取值时，系统撤销「已列全」，目录升一版。新取值的含义仍要人填。 +- 新设遮罩列时，系统会清掉这些列在目录里留下的码值、含义、备注和确认依据，目录升一版。这是机器自己写目录的另一种情况。 - 遮罩列不观察。 ## 版本与提醒 @@ -185,7 +186,7 @@ - 数据目录的「变化」视图按「要处理 / 提醒 / 说明」分组；数据源卡片上的角标是要处理的条数；表详情里有「结构与取值变化」一节。 - 每条写明影响面：说到这些列的目录项和它们的状态、写死了这一列的已发布模板、哪条 SQL 现在会报列不存在。 - 处理动作：「清理目录项」「加入码值」「重新起草」「重新剖析」「知道了」「误报」（误报必须写理由）。动作都落到目录已有的写入上，记下目录版本或建的任务。 -- 机器自己只做一件事：数据证明码值表不再列全时，撤销「已列全」。 +- 机器自己写目录只有两种情况：数据证明码值表不再列全时撤销「已列全」；新设遮罩列时清掉该列的码值、含义、备注和确认依据。 - 有「要处理」的变化时，可以推通知「数据源有变化要处理」，消息里只有源名、条数和链接，不放表名、列名和取值（见[通知](schedules.md#通知)）。也可以作为定时任务的触发事件。 ### 定时维护 diff --git a/docs/zh-CN/evaluation-and-ledger.md b/docs/zh-CN/evaluation-and-ledger.md index 76d81bff..d1028987 100644 --- a/docs/zh-CN/evaluation-and-ledger.md +++ b/docs/zh-CN/evaluation-and-ledger.md @@ -75,9 +75,16 @@ |---|---| | 「等价已证明」 | SQL 被证明和口径等价 | | 「助手选用」 | 助手选用口径拼出来的 | +| 「外部常量」 | 口径卡上的外部常量，带来源和生效日期；过期会降档 | +| 「规则派生」 | 规则型指标从查询格按结构化条件算出，不是某条查询与口径等价 | +| 「评估值」 | 评估型指标按评分协议算出的值，不是测量值 | | 「相关口径」 | 用到同一张表却不等价，写明差在哪 | | 「这次的推理」 | 和哪份口径都不沾边 | +后三种（外部常量、规则派生、评估值）不会让整题变成「等价已证明」，不自动正式出具；证据面板会照实写出处。 + +口径卡上的指标现在有四种：公式（`formula`）、外部常量（`external`）、规则（`rule`）、评估（`assessment`）。定义文本和参数随版本固定，并在「口径」目录页（`/ledger/calibers`）里能看、能按参数试运行；试运行是只读探索运行，不进使用统计，也不是正式出具。 + - 证据面板的出处链最上面多一步「口径」；问数据每一轮底下有一行摘要。比对还没出结果时，写「正在比对口径」并按退避重取。 - 比对结果是一条封存后追加的事件 `ledger.matched`，出现在时间线和可验证报告的「封存后追加」里，不在封存范围内。启动时会补比最近 7 天的运行（跟着 `AGENTLAB_SCHEDULER` 开关）。 - 逐个数判：报告引用了一条查询的几列（例如销售额、订单数、首末日）时，每个数按它那一列单独和口径的指标列比，多查的列不影响这个数。只在去掉别的列不会改变这个数时才这样比；有 DISTINCT、按别的列排序后截断、多出来分组列的，照旧整条比。 @@ -85,7 +92,7 @@ - 整题要每个来自数据的数都被证明来自同一份口径，才自动出具。 - 分析结论里的数（占比、变化、波动范围）是从查询结果派生的，不比对、不标口径，摘要行也不计它们。 -写口径的取数 SQL 时，最好不要取整（`ROUND(SUM(x), 2)`），也不要套 `COALESCE(…, 0)`。取整和缺省值交给口径卡的小数位和展示；否则问数据里没取整的写法永远认不成等价（差在哪会照实写出来）。 +写口径的取数 SQL 时，不要取整（`ROUND(SUM(x), 2)`），也不要套 `COALESCE(…, 0)`。取整和缺省值交给口径卡的小数位和展示；否则问数据里没取整的写法永远认不成等价（差在哪会照实写出来）。受管级别发布时，这两类写法会直接挡住发布；已发布级别只提示。 ### 按口径正式出具 diff --git a/frontend/src/components/CommandPalette.tsx b/frontend/src/components/CommandPalette.tsx index aa5b783c..b811a8fe 100644 --- a/frontend/src/components/CommandPalette.tsx +++ b/frontend/src/components/CommandPalette.tsx @@ -67,7 +67,7 @@ export const PAGES: PageDef[] = [ // 评测（评测-D8）：搭建者以上可见（使用者的页面表在 visiblePages 里收窄，用不着另判），同样排在设置前面；快捷键用最后一个空着的 ⌥0 { to: '/evals', label: EVAL_TEXT.nav, hint: EVAL_TEXT.hint, icon: Gauge, shortcut: 'Alt+0', keywords: 'evals eval evaluation 评测 回归 题集 准确率 A/B 口径命中 重复度' }, // 口径目录：已发布口径的入口、依赖和试跑。后加的页，快捷键用没占用的 Alt+L - { to: '/ledger/calibers', label: '口径', hint: '查看已发布口径的入口、依赖，并按参数试跑', icon: Layers, shortcut: 'Alt+L', keywords: 'caliber ledger 口径 指标 试跑 入口' }, + { to: '/ledger/calibers', label: '口径', hint: '查看已发布口径的入口、依赖，并按参数试运行', icon: Layers, shortcut: 'Alt+L', keywords: 'caliber ledger 口径 指标 试运行 入口' }, { to: '/settings', label: '设置', hint: '模型接入、偏好、运行环境', icon: Settings, shortcut: 'Alt+7', keywords: 'settings provider 模型 署名 主题 偏好' }, ] diff --git a/frontend/src/lib/terms.ts b/frontend/src/lib/terms.ts index e672c20a..cc153e90 100644 --- a/frontend/src/lib/terms.ts +++ b/frontend/src/lib/terms.ts @@ -5482,6 +5482,7 @@ export const ORIGIN_LABEL: Record = { ledger_issue: '按口径正式出具', reconcile: '口径对账', recall: '口径召回', + caliber_trial: '口径试运行', } export const ORIGIN_TEXT = { diff --git a/frontend/src/lib/termsLedger.ts b/frontend/src/lib/termsLedger.ts index 29165b0b..81e68ac3 100644 --- a/frontend/src/lib/termsLedger.ts +++ b/frontend/src/lib/termsLedger.ts @@ -153,12 +153,16 @@ export const LEDGER_UI = { pending: '正在比对口径…', timeout: '比对口径还没有结果', /** 一行摘要：几个数来自哪份口径、几个是这次的推理、几个没有比对 */ - summary: (parts: { caliber: number; calibers: string[]; reasoning: number; related: number; unknown: number }) => { + summary: (parts: { caliber: number; calibers: string[]; reasoning: number; related: number; unknown: number; + external?: number; derived?: number; assessed?: number }) => { const out: string[] = [] if (parts.caliber) { out.push(parts.calibers.length === 1 ? `${formatNumber(parts.caliber)} 个数来自口径${parts.calibers[0]}` : `${formatNumber(parts.caliber)} 个数来自 ${formatNumber(parts.calibers.length)} 份口径`) } + if (parts.external) out.push(`${formatNumber(parts.external)} 个数是外部常量`) + if (parts.derived) out.push(`${formatNumber(parts.derived)} 个数是规则派生`) + if (parts.assessed) out.push(`${formatNumber(parts.assessed)} 个数是评估值`) const reasoning = parts.reasoning + parts.related if (reasoning) { out.push(`${formatNumber(reasoning)} 个数是这次的推理${parts.related ? `（其中 ${formatNumber(parts.related)} 个有相关口径）` : ''}`) diff --git a/frontend/src/pages/ledger/CaliberCatalog.tsx b/frontend/src/pages/ledger/CaliberCatalog.tsx index 338db22e..1a037449 100644 --- a/frontend/src/pages/ledger/CaliberCatalog.tsx +++ b/frontend/src/pages/ledger/CaliberCatalog.tsx @@ -65,9 +65,9 @@ export function CaliberCatalog() { try { const out = await api.caliberLedger.trial(d.workflow_id, d.version, d.card.node_id, values[key] ?? {}) setTrials((t) => ({ ...t, [key]: out })) - if (out.status === 'succeeded') toast.ok('试跑完成') - else if (out.status === 'running') toast.warn('试跑还在运行，打开运行记录看结果') - else toast.error(out.error || '试跑没有成功') + if (out.status === 'succeeded') toast.ok('试运行完成') + else if (out.status === 'running') toast.warn('试运行还在进行，打开运行记录看结果') + else toast.error(out.error || '试运行没有成功') } catch (e) { toast.error(e) } finally { @@ -77,7 +77,7 @@ export function CaliberCatalog() { return (
- } title="口径" subtitle="已发布口径的入口、指标、依赖与试跑" + } title="口径" subtitle="已发布口径的入口、指标、依赖与试运行" actions={} /> @@ -164,17 +164,24 @@ export function CaliberCatalog() {

入口字段

{d.entry_schema.properties && Object.keys(d.entry_schema.properties).length > 0 ? ( setValues((all) => ({ ...all, [key]: v }))} + onChange={(v) => { + setValues((all) => ({ ...all, [key]: v })) + setTrials((all) => { + const next = { ...all } + delete next[key] + return next + }) + }} idPrefix={`caliber-${key}`} /> ) :

这份口径不需要参数。

}
打开工作流 - 试跑是只读探索运行，不是正式出具 + 试运行是只读探索运行，不是正式出具
{result && (
@@ -189,7 +196,7 @@ export function CaliberCatalog() { ) : ( - {result.status === 'running' ? '还在运行：' : '试跑没有成功：'} + {result.status === 'running' ? '还在运行：' : '试运行没有成功：'} {result.error || result.status} {' '}看运行 diff --git a/frontend/src/run/LedgerStep.tsx b/frontend/src/run/LedgerStep.tsx index f7b2bbf4..e8710682 100644 --- a/frontend/src/run/LedgerStep.tsx +++ b/frontend/src/run/LedgerStep.tsx @@ -364,7 +364,7 @@ export function LedgerPart({ runId, runClass, report, segId, segText, segLedger,

数由规则从查询结果算出，不是某条查询与口径等价；输入仍然逐格可追溯。

- {label.rule?.all?.length ?

条件：{label.rule.all.length} 条

: null} + {label.rule?.conditions ?

条件：{label.rule.conditions} 条

: null}
)} {label.kind === 'assessed' && ( diff --git a/frontend/src/run/LedgerTurn.tsx b/frontend/src/run/LedgerTurn.tsx index 9ce300f9..1c3a549b 100644 --- a/frontend/src/run/LedgerTurn.tsx +++ b/frontend/src/run/LedgerTurn.tsx @@ -21,8 +21,8 @@ import { ReconcileDialog, ReconcileResults } from './LedgerReconcile' * 能不能自动、不能的原因都是服务端给的；摘要只是把主报告里服务端标好的数按标签数一数、写出是哪份口径。 */ -/** 摘要认的四种标签。服务端不给别的标签（分析结论的数不和口径比对，00-总方案 D11），以后多了也不计、不出 NaN */ -const COUNTED_KINDS = ['caliber', 'related', 'reasoning', 'unknown'] as const +/** 摘要认的标签。分析结论的数不和口径比对（00-总方案 D11），但外部常量、规则派生、评估值要照实计数 */ +const COUNTED_KINDS = ['caliber', 'external', 'derived', 'assessed', 'related', 'reasoning', 'unknown'] as const type CountedKind = (typeof COUNTED_KINDS)[number] const counted = (kind: unknown): kind is CountedKind => (COUNTED_KINDS as readonly unknown[]).includes(kind) @@ -33,7 +33,9 @@ const counted = (kind: unknown): kind is CountedKind => (COUNTED_KINDS as readon function summaryOf(data: RunLedger) { const labels = Object.values((data.primary ? data.reports?.[data.primary]?.labels : undefined) ?? {}) .filter((l) => !!l && typeof l === 'object' && counted(l.kind)) - const counts: Record = { caliber: 0, related: 0, reasoning: 0, unknown: 0 } + const counts: Record = { + caliber: 0, external: 0, derived: 0, assessed: 0, related: 0, reasoning: 0, unknown: 0, + } const calibers: string[] = [] for (const l of labels) { counts[l.kind as CountedKind] += 1 diff --git a/frontend/src/typesApp.ts b/frontend/src/typesApp.ts index a5fa0217..9a425c12 100644 --- a/frontend/src/typesApp.ts +++ b/frontend/src/typesApp.ts @@ -822,7 +822,7 @@ export interface NotifyEventRow { export type OriginKind = | 'chat' | 'canvas' | 'launcher' | 'schedule' | 'app_entry' | 'app_view' | 'app_backfill' | 'hook' | 'api' | 'trial' | 'personal_app' - | 'eval' | 'ledger_issue' | 'reconcile' | 'recall' + | 'eval' | 'ledger_issue' | 'reconcile' | 'recall' | 'caliber_trial' /** * GET /api/runs/unlisted：同样的筛选下，记录页因为「系统派生的运行默认不列」而没列出的次数。系统派生的是评测、口径对账、 diff --git a/frontend/src/typesLedger.ts b/frontend/src/typesLedger.ts index 166dd9c8..77ce9df4 100644 --- a/frontend/src/typesLedger.ts +++ b/frontend/src/typesLedger.ts @@ -58,8 +58,13 @@ export interface LedgerLabel { source?: string as_of?: string stale?: boolean - rule?: { inputs?: Record; all?: unknown[]; then?: number; else?: number } - assessment?: { score?: number | null; level?: string | null; dimensions?: unknown[]; levels?: unknown[] } + rule?: { inputs?: string[]; conditions?: number; then?: number; else?: number } + assessment?: { + score?: number | null + level?: string | null + dimensions?: { name?: string; weight?: number }[] + levels?: { name?: string; min?: number }[] + } caliber?: string workflow_id?: string version?: number @@ -385,7 +390,7 @@ export interface CaliberMetric { unit?: string decimals?: number format?: string - /** 公式（formula）；规则、评估、外部常量还没实现 */ + /** 公式（formula）、外部常量（external）、单元格条件规则（rule）、评估（assessment） */ kind?: string /** 定义卡：文字定义、范围、排除项、负责人、生效日期、引用 */ definition?: { text?: string; scope?: string; exclusions?: string[]; owner?: string; effective_from?: string; citations?: unknown[] } diff --git a/scripts/check-shell.mjs b/scripts/check-shell.mjs index 59ca52fd..b43b6e92 100644 --- a/scripts/check-shell.mjs +++ b/scripts/check-shell.mjs @@ -147,11 +147,11 @@ await section('导航', '=== 导航 ===', async () => { const labels = await nav(page).locator('li > a[aria-keyshortcuts] > span:not([aria-hidden])').allInnerTexts() // 定时任务是后加的，排在设置前面；设置留在最底下，快捷键还是 ⌥7（定时任务用 ⌥8）。应用（⌥9）排在编排后面。 // 评测（next-round 评测-D8，搭建者以上）也排在设置前面，用最后一个空着的 ⌥0 - check('导航十项：问数据、编排、应用、记录、数据、工具、知识、定时任务、评测、设置', - labels.join(',') === '问数据,编排,应用,记录,数据,工具,知识,定时任务,评测,设置', labels.join(',')) + check('导航十一项：问数据、编排、应用、记录、数据、工具、知识、定时任务、评测、口径、设置', + labels.join(',') === '问数据,编排,应用,记录,数据,工具,知识,定时任务,评测,口径,设置', labels.join(',')) const keys = await nav(page).locator('li > a[aria-keyshortcuts]').evaluateAll((els) => els.map((el) => el.getAttribute('aria-keyshortcuts'))) - check('原有快捷键都没动：问数据 ⌥1、编排 ⌥2、记录 ⌥3 … 设置 ⌥7、定时任务 ⌥8；应用是 ⌥9、评测是 ⌥0', - keys.join(',') === 'Alt+1,Alt+2,Alt+9,Alt+3,Alt+4,Alt+5,Alt+6,Alt+8,Alt+0,Alt+7', keys.join(',')) + check('原有快捷键都没动：问数据 ⌥1、编排 ⌥2、记录 ⌥3 … 设置 ⌥7、定时任务 ⌥8；应用是 ⌥9、评测是 ⌥0；口径是 ⌥L', + keys.join(',') === 'Alt+1,Alt+2,Alt+9,Alt+3,Alt+4,Alt+5,Alt+6,Alt+8,Alt+0,Alt+L,Alt+7', keys.join(',')) const fontSize = await nav(page).locator('li > a[aria-keyshortcuts] > span:not([aria-hidden])').first() .evaluate((el) => getComputedStyle(el).fontSize) check('导航标签 11px', fontSize === '11px', fontSize)