mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-15 07:02:00 +00:00
ThearchyHelios
265ba0c91d
两件事，一条心：把「网页上点数字看得到的效果」带出系统外。
① 报告导出：新增 GET /api/runs/{run_id}/report.html（app/report_export：build + 模板 + 浏览器端脚本）。
正文由服务端按报告文档的块、句、片段渲染成 HTML——不执行脚本也读得到整份报告（附件预览常常不执行脚本）；
执行脚本后，带出处的数字、无证据的数字、可疑的表名字段名都点得开，面板是网页上证据面板同一个拼法的
出处链（api/evidence._chain）：指标步骤带口径卡名与版本、原式、代入式、每个输入取自哪次查询的哪一格、
复算、双盲复算；查询步骤带 SQL 和被引用的行；数据源遮罩列照网页上的规矩写「已遮罩」。封存状态逐项标明；
没封存的运行照常导出（和网页上读报告一个道理）。同样的引用共用一条链，长报告不把整份快照抄几十遍。
入口在问数据答案和运行详情成果上的「导出」（带证据的报告给 HTML，Markdown 留一个小的；旧运行照旧只有
Markdown）——api.reportExport + AssistantStream 的 AnswerActions。
② 可验证报告（bundle）的数字详情补齐：点开数字原来只有出处和值，现在多一节口径卡——指标（名 + id）、
单位、原式、代入式、输入（取值 ← 路径、取自哪次查询的哪一行哪一列）、复算；输入指向的查询快照也随文件
附带（遮罩规则同嵌入那一套，build._card_input_queries），④ 引用链按「已核对的口径卡的输入」认它们
（哈希对不上的卡不认，免得替被改过的卡圆）。收件人不连服务也能看到每个数怎么算出来的。
测试：后端 tests/test_report_export.py（真跑运行导出，正文、出处链、遮罩、只读、404/409、共用链）；
bundle_samples 的口径卡换成带原式/代入式/输入/复算的真实形状，另加 metric-only 样本页（只引用指标的
报告，快照从口径卡的输入这一条路附带）；check-bundle 断言数字详情和这条引用链，加上篡改；新增
scripts/check-report-export.mjs（真导出断网点开、不执行脚本也读得到、成果上的导出按钮、旧答案没有
HTML 按钮、409 提示），挂进 check-all。README 补「导出两份」一节。全量 pytest 通过；check-bundle、
check-report-export、check-stream、check-evidence、check-chat 全部通过（check-runs 5 项失败为基线
上就有的沙箱数据形态问题，与本次无关）。
288 lines
18 KiB
Python
288 lines
18 KiB
Python
"""可验证报告的虚构样本页：给 scripts/check-bundle.mjs 在浏览器里核对（pytest 不执行页面上的核对脚本）。
|
||
|
||
python -m tests.bundle_samples <目录>
|
||
|
||
在目录里写下几份页面（<名字>.html）。每份都是 build_bundle + render_html 真实生成的，封存哈希按 runner 的算法
|
||
（artifact_store.manifest_hash）对事件真实算出，报告文档用 compose_doc 按证据目录真实组装（标题、段落、列表、
|
||
表格、粗体、斜体，数字和取值的引用都和真实运行一样）。check-bundle 断网打开它们，逐项断言核对结果和原因，再在
|
||
这几份上做篡改。这样遮罩快照、多份报告、早期运行、服务端缺文件这些分支，不论真数据里有没有，都有人在浏览器里核过。
|
||
|
||
数据全是虚构的：门店、会员、金额都是编的。
|
||
|
||
- base：一份报告，引用口径卡、门店流水（没有遮罩）、会员消费（有遮罩）各一处，一个没有出处的数，一张门店流水整表
|
||
（数字列右对齐，数字与取值共 21 个，超过折叠的条数）；一条封存之后追加的按需裁判
|
||
- two-reports：另有一份月报，两份报告各自的引用链
|
||
- metric-only：只引用口径卡指标的报告——正文没有一处 [[v:]]，门店流水的快照从口径卡的输入这一条路随文件附带，
|
||
引用链（④）按「已核对的口径卡的输入」认它；点开数字看得到指标、原式、代入式、输入和复算
|
||
- legacy：早期运行，封存范围按结束事件推算（运行结束事件在范围外，运行结果无法核对）
|
||
- failed：失败结束的运行（没有出具判定），封存说明多一句「继续运行会重新封存」
|
||
- continued：早期运行先失败、经继续运行跑到成功：范围里有那条运行失败，之后运行又继续了，成功结束的那条在范围外
|
||
- degraded：两条出具判定，先完整出具、后降档出具，认最后一条
|
||
- missing-doc / missing-card：导出时服务端已没有报告文档 / 口径卡的文件
|
||
- masked-card：口径卡上一个指标的输入取自会员消费的号码列（设了遮罩），卡上记着号码原值：整张卡只给 id，引用它的
|
||
数字无法核对
|
||
- masked-source：会员库设了号码遮罩（查询当时记下了，数据源现在也设着）。报告引用会员按地区汇总的查询（没碰遮罩列：
|
||
快照随文件附带，数字核对得过）和选出了号码的那份会员消费（只给哈希）
|
||
- nan：运行开始事件和门店流水快照里有 json.dumps 写出的裸 NaN（快照里那一列报告没有引用）
|
||
- stray：报告自查记下的文档是另一次运行写的
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
import sys
|
||
from pathlib import Path
|
||
from typing import Any
|
||
|
||
from app.bundle.build import build_bundle, render_html
|
||
from app.core.artifact_store import canonical_json, content_hash, manifest_hash
|
||
from app.engine.evidence import build_catalog, compose_doc
|
||
|
||
RUN_ID = "5a3c0f4b9d2e4e1f8b7a6c5d4e3f2a1b"
|
||
RUN = {"id": RUN_ID, "workflow_id": "wf-sample-weekly", "workflow_name": "门店周报", "version": 3,
|
||
"version_hash": "ab" * 32, "run_class": "formal", "status": "succeeded",
|
||
"started_at": "2026-09-07T01:00:00+00:00", "finished_at": "2026-09-07T01:02:03+00:00"}
|
||
GRAPH = {"nodes": [{"id": n, "type": t, "data": {"label": label}} for n, t, label in [
|
||
("start", "input", "开始"), ("fetch", "tool", "门店流水"), ("members", "tool", "会员消费"),
|
||
("regions", "tool", "会员地区汇总"), ("card", "metrics", "周报口径"), ("write", "report", "写周报"),
|
||
("monthly", "report", "写月报"), ("out", "output", "成果")]]}
|
||
EXPORTED_AT = "2026-10-05T08:00:00+00:00"
|
||
|
||
STORES = [["青石店", 1280.5], ["白沙店", "960.00"], ["临川店", 732], ["望江店", 1105.25], ["枫桥店", 418],
|
||
["云岭店", 2210], ["石桥店", 655.5], ["柳湾店", 980]]
|
||
CRM = {"columns": ["name", "phone", "spend"], "column_types": {"name": "text", "phone": "text", "spend": "number"},
|
||
"rows": [["林一", "000-0001", 320.0], ["周二", "000-0002", 188.5]], "source": "crm",
|
||
"sql": "SELECT name, phone, spend FROM members ORDER BY id", "truncated": False, "mask_columns": ["phone"]}
|
||
#: 会员按地区汇总：同一个设了号码遮罩的会员库，SQL 没碰号码列
|
||
REGIONS = {"columns": ["region", "spend"], "column_types": {"region": "text", "spend": "number"},
|
||
"rows": [["华东", 320.0], ["华南", 188.5]], "source": "crm",
|
||
"sql": "SELECT region, SUM(spend) AS spend FROM members GROUP BY region ORDER BY region", "truncated": False,
|
||
"mask_columns": ["phone"]}
|
||
|
||
|
||
def _num(value: Any) -> str:
|
||
"""代入式里的一个数：照 Python 字面量的样子（整型不带 .0）；DECIMAL 落成的数字文本原样。"""
|
||
if isinstance(value, str):
|
||
return value
|
||
return str(int(value)) if isinstance(value, float) and value.is_integer() else str(value)
|
||
|
||
|
||
def card_of(shop_id: str) -> dict[str, Any]:
|
||
"""周报口径卡：销售额是八家门店流水之和。和真实口径卡一样带着计算方式（原式、代入式）、每个输入
|
||
（取自哪次查询的哪一格）和复算结果——可验证报告点开数字时据此展示，引用链据此附上输入指向的查询快照。"""
|
||
cells = [f"cell(nodes.fetch, {i}, 'amount')" for i in range(len(STORES))]
|
||
gmv = {"id": "gmv", "name": "销售额", "value": round(sum(float(v) for _, v in STORES), 2), "unit": "元",
|
||
"decimals": 2, "format": "thousands", "expression": " + ".join(cells),
|
||
"substituted": " + ".join(_num(v) for _, v in STORES), "recompute_ok": True,
|
||
"inputs": [{"path": f"cell(nodes.fetch, {i}, 'amount')", "value": v, "node_id": "fetch",
|
||
"via": "tool_cell", "artifact": shop_id, "locator": {"row": i, "column": "amount"},
|
||
"status": "verified"}
|
||
for i, (_, v) in enumerate(STORES)]}
|
||
return {"kind": "metric_set", "caliber": "周报口径", "caliber_version": "v2", "metrics": [gmv]}
|
||
|
||
|
||
def masked_card(card: dict[str, Any], crm_id: str) -> dict[str, Any]:
|
||
"""口径卡多一个指标：联系电话位数，输入是会员消费第 0 行的号码（phone 设了遮罩），卡上记着号码原值和代入式。"""
|
||
cell = "cell(nodes.members, 0, 'phone')"
|
||
contact = {"id": "contact", "name": "联系电话位数", "value": 8, "unit": "位", "decimals": 0, "format": "thousands",
|
||
"expression": f"len({cell})", "substituted": f"len('{CRM['rows'][0][1]}')",
|
||
"inputs": [{"path": cell, "value": CRM["rows"][0][1], "node_id": "members", "via": "tool_cell",
|
||
"artifact": crm_id, "locator": {"row": 0, "column": "phone"}, "status": "verified"}]}
|
||
return {**card, "metrics": [*card["metrics"], contact]}
|
||
|
||
|
||
WEEKLY = ("## 门店周报\n\n本周销售额 [[m:gmv]]，*按发生周统计*。青石店流水 [[v:Q1.r0.amount]]，白沙店 [[v:Q1.r1.amount]]，"
|
||
"**会员消费** [[v:Q2.r0.spend]]，另有 42 笔退款待入账。[[see:Q1]]\n\n- 云岭店流水最高\n- 枫桥店流水最低\n\n"
|
||
"[[table:Q1 cols=store,amount rows=0-7]]")
|
||
MONTHLY = "## 门店月报\n\n本月销售额 [[m:gmv]]，青石店流水 [[v:Q1.r0.amount]]。[[see:Q1]]"
|
||
#: 会员周报：Q3 是会员按地区汇总（没碰遮罩列），Q2 是选出了号码的会员消费
|
||
MEMBERS = ("## 会员周报\n\n华东会员消费 [[v:Q3.r0.spend]]，华南 [[v:Q3.r1.spend]]。[[see:Q3]]\n\n"
|
||
"会员林一消费 [[v:Q2.r0.spend]]。")
|
||
#: 只引用指标的报告：正文没有一处 [[v:]]，门店流水的快照只能从口径卡的输入这一条路随文件附带（引用链 ④）
|
||
METRIC_ONLY = "## 门店周报\n\n本周销售额 [[m:gmv]]，各门店明细随后补上。"
|
||
|
||
|
||
def stored(obj: Any) -> tuple[str, bytes]:
|
||
"""照工件库的样子落一份：规范 JSON 的字节，id 是它的 sha256。"""
|
||
raw = canonical_json(obj).encode()
|
||
return content_hash(raw), raw
|
||
|
||
|
||
def shop(nan: bool = False) -> dict[str, Any]:
|
||
"""门店流水的查询快照。ratio 那一列报告不引用；nan=True 时第一行是浮点 NaN（json.dumps 写成裸的 NaN）。"""
|
||
rows = [[name, amount, float("nan") if nan and i == 0 else round(0.05 * (i + 1), 2)]
|
||
for i, (name, amount) in enumerate(STORES)]
|
||
return {"columns": ["store", "amount", "ratio"],
|
||
"column_types": {"store": "text", "amount": "number", "ratio": "number"}, "rows": rows, "source": "shop",
|
||
"sql": "SELECT store, amount, ratio FROM orders ORDER BY id", "truncated": False}
|
||
|
||
|
||
class Sample:
|
||
"""一次虚构的已封存运行：工件（原字节）、证据目录、报告文档、事件。masked=True 时口径卡带着取自遮罩列的输入；
|
||
regions=True 时会员库另有一次按地区汇总的查询（目录里排在 Q3）。"""
|
||
|
||
def __init__(self, *, nan: bool = False, masked: bool = False, regions: bool = False) -> None:
|
||
self.files: dict[str, bytes] = {}
|
||
self.shop_id = self.put(shop(nan))
|
||
self.crm_id = self.put(CRM)
|
||
self.card = masked_card(card_of(self.shop_id), self.crm_id) if masked else card_of(self.shop_id)
|
||
self.card_id = self.put(self.card)
|
||
self.entries = {
|
||
"fetch": {"kind": "query", "node_id": "fetch", "exec": 1, "artifact": self.shop_id,
|
||
"tool": "db_query__shop", "source": "shop", "columns": ["store", "amount", "ratio"],
|
||
"rows": len(STORES), "truncated": False},
|
||
"members": {"kind": "query", "node_id": "members", "exec": 1, "artifact": self.crm_id,
|
||
"tool": "db_query__crm", "source": "crm", "columns": CRM["columns"], "rows": len(CRM["rows"]),
|
||
"truncated": False},
|
||
}
|
||
if regions:
|
||
self.entries["regions"] = {"kind": "query", "node_id": "regions", "exec": 1,
|
||
"artifact": self.put(REGIONS), "tool": "db_query__crm", "source": "crm",
|
||
"columns": REGIONS["columns"], "rows": len(REGIONS["rows"]), "truncated": False}
|
||
self.entries["card"] = {"kind": "metric_set", "node_id": "card", "artifact": self.card_id}
|
||
self.ledger = list(self.entries.values())
|
||
self.catalog = build_catalog(nodes={"card": self.card}, ledger=self.ledger)
|
||
|
||
def put(self, obj: Any) -> str:
|
||
artifact, raw = stored(obj)
|
||
self.files[artifact] = raw
|
||
return artifact
|
||
|
||
def load(self, artifact: str) -> Any:
|
||
raw = self.files.get(artifact)
|
||
return json.loads(raw) if raw is not None else None
|
||
|
||
def doc(self, text: str, node_id: str, run_id: str = RUN_ID) -> str:
|
||
return self.put(compose_doc(text, self.catalog, node_id=node_id, run_id=run_id, loader=self.load))
|
||
|
||
def events(self, reports: dict[str, str], *, ending: str = "finished", tiers: tuple[str, ...] = ("formal",),
|
||
nan: bool = False) -> tuple[list[tuple[int, str, str | None, Any]], int]:
|
||
"""(全部事件, 封存范围)。封存范围到结束事件为止，后面再追加一条按需裁判。"""
|
||
rows: list[tuple[int, str, str | None, Any]] = []
|
||
|
||
def add(etype: str, node_id: str | None, data: dict[str, Any]) -> None:
|
||
rows.append((len(rows) + 1, etype, node_id, data))
|
||
|
||
add("run.started", None, {"nodes": len(GRAPH["nodes"]), "resumed": False, "memory_scope": "default",
|
||
"collection": "default", "approval_default": None, "tool_trust": "formal",
|
||
"agent_limits": None, "data_versions": None, "replay_protocol": 2,
|
||
**({"sample_rate": float("nan")} if nan else {})})
|
||
add("node.started", "start", {"node_type": "input", "label": "开始"})
|
||
add("node.finished", "start", {"duration_ms": 1, "attempt": 1, "preview": {"week": "2026-W37"}})
|
||
add("node.started", "fetch", {"node_type": "tool", "label": "门店流水"})
|
||
add("tool.end", "fetch", {"tool": "db_query__shop", "duration_ms": 12, "query_artifact": self.shop_id,
|
||
"preview": json.dumps({"columns": ["store", "amount"], "rows": STORES[:2]},
|
||
ensure_ascii=False)})
|
||
add("node.finished", "fetch", {"duration_ms": 14, "attempt": 1, "preview": "{\"rows\": …}",
|
||
"evidence": [self.entries["fetch"]]})
|
||
add("tool.end", "members", {"tool": "db_query__crm", "duration_ms": 9, "query_artifact": self.crm_id,
|
||
"preview": json.dumps(CRM["rows"][:1], ensure_ascii=False)})
|
||
add("node.finished", "members", {"duration_ms": 10, "attempt": 1, "evidence": [self.entries["members"]]})
|
||
if "regions" in self.entries:
|
||
add("tool.end", "regions", {"tool": "db_query__crm", "duration_ms": 7,
|
||
"query_artifact": self.entries["regions"]["artifact"],
|
||
"preview": json.dumps(REGIONS["rows"], ensure_ascii=False)})
|
||
add("node.finished", "regions", {"duration_ms": 8, "attempt": 1, "evidence": [self.entries["regions"]]})
|
||
add("node.finished", "card", {"duration_ms": 3, "attempt": 1, "preview": {"kind": "metric_set"},
|
||
"evidence": [self.entries["card"]]})
|
||
for node_id, doc_id in reports.items():
|
||
add("llm.end", node_id, {"model": "mock", "output": "（写作草稿）"})
|
||
add("report.checked", node_id, {"doc_artifact": doc_id, "ok": True})
|
||
add("node.finished", node_id, {"duration_ms": 20, "attempt": 1,
|
||
"preview": {"text": "（报告正文）", "doc_artifact": doc_id}})
|
||
if ending == "finished":
|
||
for tier in tiers:
|
||
add("issuance", "out", {"tier": tier, "mode": "citations"})
|
||
add("node.finished", "out", {"duration_ms": 2, "attempt": 1, "preview": {"周报": "（报告正文）"}})
|
||
add("run.finished", None, {"output": {"周报": "（报告正文）"}})
|
||
elif ending == "continued":
|
||
# 先失败，再从成果节点继续运行到成功（runner.continue_failed 记一条 run.resumed）
|
||
add("node.failed", "out", {"error": "出口复核没有通过"})
|
||
add("run.failed", None, {"error": "出口复核没有通过"})
|
||
add("run.resumed", None, {"from": ["out"], "message": "从「成果」继续运行，保留前面 6 个节点的结果",
|
||
"actor": "值班分析员"})
|
||
for tier in tiers:
|
||
add("issuance", "out", {"tier": tier, "mode": "citations"})
|
||
add("node.finished", "out", {"duration_ms": 2, "attempt": 1, "preview": {"周报": "（报告正文）"}})
|
||
add("run.finished", None, {"output": {"周报": "（报告正文）"}})
|
||
else:
|
||
add("node.failed", "out", {"error": "出口复核没有通过"})
|
||
add("run.failed", None, {"error": "出口复核没有通过"})
|
||
bound = len(rows)
|
||
add("evidence.judged", next(iter(reports), None),
|
||
{"verdicts": {"u1": {"verdict": "supported"}}, "model": "judge-x", "post_seal": True})
|
||
return rows, bound
|
||
|
||
|
||
def render(sample: Sample, events: list[tuple[int, str, str | None, Any]], bound: int, *, run: dict[str, Any] | None = None,
|
||
legacy: bool = False, drop: tuple[str, ...] = (), source_masks: dict[str, list[str]] | None = None) -> str:
|
||
"""真实封存哈希的页面。drop：导出时服务端已没有文件的工件；source_masks：数据源现在设的遮罩。"""
|
||
sealed = [e for e in events if e[0] <= bound]
|
||
files = {k: v for k, v in sample.files.items() if k not in drop}
|
||
bundle = build_bundle(run or RUN, events, manifest_seq=bound, manifest_hash=manifest_hash(sealed), read=files.get,
|
||
exported_at=EXPORTED_AT, legacy=legacy, graph=GRAPH, source_masks=source_masks or {})
|
||
return render_html(bundle)
|
||
|
||
|
||
def pages() -> dict[str, str]:
|
||
"""{页面名: HTML}"""
|
||
out: dict[str, str] = {}
|
||
s = Sample()
|
||
weekly = s.doc(WEEKLY, "write")
|
||
events, bound = s.events({"write": weekly})
|
||
out["base"] = render(s, events, bound)
|
||
out["missing-doc"] = render(s, events, bound, drop=(weekly,))
|
||
out["missing-card"] = render(s, events, bound, drop=(s.card_id,))
|
||
# 早期运行：运行结束事件落库之前封存，范围到它前一条为止（runner.verify_manifest 的老口径）
|
||
finished = next(seq for seq, etype, _, _ in events if etype == "run.finished")
|
||
out["legacy"] = render(s, events, finished - 1, legacy=True)
|
||
|
||
monthly = s.doc(MONTHLY, "monthly")
|
||
events, bound = s.events({"write": weekly, "monthly": monthly})
|
||
out["two-reports"] = render(s, events, bound)
|
||
|
||
events, bound = s.events({"write": weekly}, ending="failed")
|
||
out["failed"] = render(s, events, bound, run={**RUN, "status": "failed"})
|
||
|
||
# 早期运行先失败、经继续运行跑到成功：封存范围同样按结束事件推算，到成功结束的前一条为止
|
||
events, _ = s.events({"write": weekly}, ending="continued")
|
||
finished = next(seq for seq, etype, _, _ in events if etype == "run.finished")
|
||
out["continued"] = render(s, events, finished - 1, legacy=True)
|
||
|
||
events, bound = s.events({"write": weekly}, tiers=("formal", "degraded"))
|
||
out["degraded"] = render(s, events, bound)
|
||
|
||
stray = s.doc(WEEKLY, "write", run_id="another-run-0000")
|
||
events, bound = s.events({"write": stray})
|
||
out["stray"] = render(s, events, bound)
|
||
|
||
n = Sample(nan=True)
|
||
events, bound = n.events({"write": n.doc(WEEKLY, "write")}, nan=True)
|
||
out["nan"] = render(n, events, bound)
|
||
|
||
# 只引用指标的报告：门店流水的快照没被正文引用，从口径卡的输入这一条路随文件附带、引用链认得它
|
||
out["metric-only"] = render(s, *s.events({"write": s.doc(METRIC_ONLY, "write")}))
|
||
|
||
m = Sample(masked=True)
|
||
events, bound = m.events({"write": m.doc(WEEKLY, "write")})
|
||
out["masked-card"] = render(m, events, bound)
|
||
|
||
r = Sample(regions=True)
|
||
events, bound = r.events({"write": r.doc(MEMBERS, "write")})
|
||
out["masked-source"] = render(r, events, bound, source_masks={"crm": ["phone"]})
|
||
return out
|
||
|
||
|
||
def main(argv: list[str]) -> int:
|
||
if len(argv) != 2:
|
||
print("用法：python -m tests.bundle_samples <目录>", file=sys.stderr)
|
||
return 2
|
||
target = Path(argv[1])
|
||
target.mkdir(parents=True, exist_ok=True)
|
||
for name, page in pages().items():
|
||
(target / f"{name}.html").write_text(page, encoding="utf-8")
|
||
return 0
|
||
|
||
|
||
if __name__ == "__main__":
|
||
sys.exit(main(sys.argv))
|