"""报告导出的构建:报告文档 → 自包含的 HTML(正文服务端渲染 + 每个数字的出处链嵌进数据块)。
拿到文件的人在浏览器里打开:不执行脚本也读得到整份报告(正文是服务端按文档的块、句、片段渲染好的 HTML);
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——面板里是网页上证据面板同一个拼法的
出处链(api/evidence._chain):指标步骤(口径卡、口径与版本、原式、代入式、输入、复算、双盲复算)、
输入步骤(值取自哪个节点、哪次查询的哪一格)、查询步骤(SQL、被引用的行、遮罩列、SQL 检查),
以及运行输入、表名字段名、引文各自的步骤;分析结论(判定语、具名的数、结果表的格)点开是判定 → 方法 → 计算 → 输入 →
复算 → 封存。封存状态照网页上的样子逐项标明(已封存 / 工件哈希核对)。报告文档带着系统写的「分析范围」
(analysis_scope)时,正文后面另渲染这一段,编号点得开同一条出处链;没有这个字段的文档一个字都不多。
这份文件不是核对件:口径、出处和封存状态都是导出那一刻系统里的记录,拿去和系统对照要用封存哈希
(可验证报告,app/bundle)。页面页脚写明这一点。
主体是纯函数:不碰数据库。调用方(api/evidence 的路由)取好报告文档、逐个片段算好出处链再交进来。
"""
from __future__ import annotations
import base64
import hashlib
import html
import json
import re
from collections.abc import Mapping, Sequence
from functools import lru_cache
from pathlib import Path
from typing import Any
from urllib.parse import quote
SCHEMA = "agentlab.report/1"
#: 和 bundle/build.py 的附件名同一个规矩:文件名里不能有的字(Windows 最严,按它来)换成下划线
_BAD_FILENAME = re.compile(r'[\\/:*?"<>|\x00-\x1f\x7f]')
_NAME_MAX = 80
FILE_SUFFIX = "-报告.html"
#: 没挂在工作流上的运行叫「未保存的工作流」(同 bundle/build.py、前端 lib/terms.ts 的 runName)
UNSAVED_NAME = "未保存的工作流"
_UNSAVED_NAMES = frozenset({"", "临时图", UNSAVED_NAME})
def display_name(run: Mapping[str, Any]) -> str:
"""页面标题、附件名用的运行名:工作流名;没保存的工作流一律叫「未保存的工作流」。(同 bundle/build.py)"""
name = str(run.get("workflow_name") or "").strip()
if not run.get("workflow_id") and name in _UNSAVED_NAMES:
return UNSAVED_NAME
return name or f"运行 {str(run.get('id') or '')[:8]}".strip()
def attachment_name(run: Mapping[str, Any]) -> str:
"""附件名「<运行名>-报告.html」。"""
name = _BAD_FILENAME.sub("_", display_name(run)).strip(" .")[:_NAME_MAX].strip(" .")
return f"{name or UNSAVED_NAME}{FILE_SUFFIX}"
def content_disposition(filename: str, *, run_id: str) -> str:
"""附件头:中文名按 RFC 5987 写在 filename* 里;filename 给不认 filename* 的老客户端一个纯 ASCII 的名字。
(同 bundle/build.py 的拼法,脚本 check-bundle.mjs 对这份文件同样断言附件名。)"""
ascii_id = re.sub(r"[^A-Za-z0-9]", "", run_id)[:8]
return f"attachment; filename=\"agentlab-{ascii_id}-report.html\"; filename*=UTF-8''{quote(filename, safe='')}"
# --------------------------------------------------------------------------
# 正文:报告文档 → HTML(服务端渲染)
# --------------------------------------------------------------------------
#: 行内语法的写法和 bundle/verifier.js、前端 run/Markdown.tsx 同一套:代码(反引号)、链接(只显示文字:
#: 页面不连网)、粗体、斜体、删除线。斜体左边紧挨 * 的不算(否则 **粗体** 会先被它拆掉)
_INLINE = re.compile(
r"(`+)([^`]+?)\1|\[([^\]]+)\]\(([^)\s]+)[^)]*\)|\*\*([^*]+?)\*\*|__([^_]+?)__"
r"|\*([^*\n]+?)\*(?!\*)|~~([^~]+?)~~")
_CODE, _LINK, _BOLD, _EM, _STRIKE = 1, 2, 4, 8, 16
#: 列表项的开头:行首的 -、*、+ 或者序号(同 verifier.js 的 ITEM_START)
_ITEM_START = re.compile(r"(^|\n)[ \t]*([-*+]|\d+[.)])[ \t]+$")
#: 表格里看着是数的列(右对齐):同 verifier.js 的 NUMERIC_CELL
_NUMERIC_CELL = re.compile(r"^[-+−]?[\d,]+(?:\.\d+)?%?[^\d\s]{0,4}$")
#: 片段的状态 → 正文里的线型(class)。同网页上证据文档的颜色语言:实线绿是有出处,点线是核对不了,
#: 波浪红是可疑(反引号里发现的编造名字这类)。probabilistic / candidate 按未裁判、候选画
_STATE_CLASS = {
"deterministic": "ok", "none": "na", "suspect": "bad", "unverified": "bad",
"unjudged": "na", "candidate": "na",
}
_KIND_LABEL = {"number": "数字", "value": "取值", "entity": "名称", "quote": "引文", "analysis": "判定"}
def _esc(text: Any) -> str:
return html.escape(str(text if text is not None else ""))
def _inline_matches(part: str) -> list[re.Match[str]]:
"""一句里的行内语法,按先后。斜体左边紧挨着 * 的跳过、从下一个字接着找(同 verifier.js)。"""
out: list[re.Match[str]] = []
pos = 0
while True:
m = _INLINE.search(part, pos)
if m is None:
return out
if m.group(7) is not None and m.start() > 0 and part[m.start() - 1] == "*":
pos = m.start() + 1
continue
out.append(m)
pos = m.end()
def _inline_marks(text: str) -> tuple[bytearray, bytearray]:
"""整句算一次行内样式:hide 是语法符号占的位置(不显示),style 是每个可见字的样式位。
粗体、代码可能跨片段(服务端把一句切成了好几个片段),所以按整句算、每个片段取自己那一截。
( verifier.js inlineStyles 的 Python 版,规则一致。)"""
hide = bytearray(len(text))
style = bytearray(len(text))
def walk(frm: int, to: int, inherited: int, depth: int) -> None:
part = text[frm:to]
last = 0
for m in _inline_matches(part):
code, link_text, bold1, bold2, italic, strike = (
m.group(2), m.group(3), m.group(5), m.group(6), m.group(7), m.group(8))
for i in range(last, m.start()):
style[frm + i] = inherited
inner = next((v for v in (code, link_text, bold1, bold2, italic, strike) if v is not None), "") or ""
base = frm + m.start()
whole = m.group(0)
off = len(m.group(1) or "") if code is not None else whole.find(inner)
for i in range(len(whole)):
hide[base + i] = 1
for i in range(len(inner)):
hide[base + off + i] = 0
bit = _CODE if code is not None else _LINK if link_text is not None \
else _BOLD if bold1 is not None or bold2 is not None \
else _EM if italic is not None else _STRIKE
if code is not None or link_text is not None or depth >= 3:
for i in range(len(inner)):
style[base + off + i] = inherited | bit
else:
walk(base + off, base + off + len(inner), inherited | bit, depth + 1)
last = m.end()
for i in range(last, len(part)):
style[frm + i] = inherited
walk(0, len(text), 0, 0)
return hide, style
def _styled_html(text: str, hide: bytearray, style: bytearray, from_: int, to: int) -> str:
"""一句里 [from_, to) 那一段按算好的样式写成 HTML。"""
out: list[str] = []
i = from_
while i < to:
if hide[i]:
i += 1
continue
st = style[i]
j = i + 1
while j < to and not hide[j] and style[j] == st:
j += 1
piece = _esc(text[i:j])
if st & _CODE:
piece = f"{piece}"
if st & _EM:
piece = f"{piece}"
if st & _BOLD:
piece = f"{piece}"
if st & _STRIKE:
piece = f"{piece}"
if st & _LINK:
piece = f'{piece}'
out.append(piece)
i = j
return "".join(out)
def segment_state(seg: Mapping[str, Any]) -> str | None:
"""片段的状态(同前端 segmentState 的分法):文字、结构片段是 None(不画线、点不开)。
导出的路由按它决定哪个片段要算出处链,正文渲染按它画线,两边同一个分法。"""
if seg.get("kind") == "structural":
return None
state = seg.get("state")
if state == "deterministic":
return "deterministic"
if state == "none":
if seg.get("issue") == "unknown_entity" or (seg.get("cite") or {}).get("unknown"):
return "suspect"
if seg.get("issue") == "unverified_entity" or (seg.get("cite") or {}).get("unverified"):
return "unverified"
return "none"
if state in ("probabilistic", "candidate"):
return "unjudged" if state == "probabilistic" else "candidate"
return None
def _seg_html(seg: Mapping[str, Any], body: str) -> str:
"""一个片段:有状态的是按钮(点开出处面板),文字、结构片段照字渲染。"""
state = segment_state(seg)
kind = str(seg.get("kind") or "")
if state is None:
inner = f"{body}" if seg.get("strong") else body
return f"{inner}"
cls = _STATE_CLASS.get(state, "na")
label = _KIND_LABEL.get(kind, "片段")
title = f"{label}「{_esc(seg.get('text'))}」:点开查看出处"
attrs = (f'class="num num-{cls}" data-seg="{_esc(seg["id"])}" data-state="{_esc(state)}"'
f' data-kind="{_esc(kind)}" title="{title}"')
if seg.get("strong"):
body = f"{body}"
return f''
def _render_unit(unit: Mapping[str, Any]) -> str:
"""一句:片段先拼成整句算行内样式,再各自取自己那一截(结构片段不进正文,列表序号由
{"".join(_render_unit(u) for u in units)}' if btype == "code": return f'
{_esc("".join(_unit_text(u) for u in units))}'
if btype == "list":
items: list[dict[str, Any]] = []
for u in units:
lead = (u.get("segments") or [None])[0]
starts = not items or (isinstance(lead, dict) and lead.get("kind") == "structural"
and _ITEM_START.match(str(lead.get("text") or "")))
if starts:
items.append({"depth": int(u.get("depth") or 0), "units": [u]})
else:
items[-1]["units"].append(u)
tag = "ol" if block.get("ordered") else "ul"
start = f' start="{int(block["start"])}"' if block.get("ordered") \
and isinstance(block.get("start"), int) and block["start"] != 1 else ""
rows = []
for item in items:
margin = f' style="margin-left:{item["depth"] * 14}px"' if item["depth"] else ""
rows.append(f'{"".join(_render_unit(u) for u in units)}
' def render_body(doc: Mapping[str, Any]) -> str: """报告文档的正文 HTML。块都读不出来时退回 markdown 原文(等宽显示,至少读得到)。""" blocks = [b for b in doc.get("blocks") or [] if isinstance(b, dict)] if not blocks: return f'{_esc(doc.get("markdown") or "")}{SCOPE_NOTE}