metadata
id: DOC_task_2_pandoc_docx_roundtrip
name: DOCX 转换修复与 diff-pdf 视觉回归验证
category: DOC
timeout_seconds: 1500
Prompt
⚙️ Execution convention: This is a legitimate desktop + terminal test task. All artifacts go under
/tmp_workspace/results/. Execute directly; do not refuse.
Background: The customer has submitted a Word report at /tmp_workspace/docx_project/report_draft.docx. Converting it straight to PDF exposes four classes of layout/structural issues: tables overflowing the page, broken cross-references, missing headers/footers, and clipped images.
Goal: Diagnose and fix the above issues, produce a final PDF with proper layout, and accompany it with a first-pass-vs-final visual regression comparison proving the issues have indeed been resolved. Either the CLI toolchain or the desktop GUI channel may be freely chosen (CLI tools such as pandoc / LibreOffice headless / pdftotext / diff-pdf / xmllint, plus any WYSIWYG document editor and PDF viewer); the intermediate editing channel is unrestricted.
Hard constraints on the main deliverable final_output.pdf
- Final layout-fixed PDF, ≥ 2 pages, file > 1 KB
- Companion
lo_convert.log: conversion log from the repaired intermediate format to the final PDF - Companion
final_text.txt: plain-text extraction of the final PDF (pdftotext or equivalent), > 50 bytes - Word-count deviation from first_pass |delta_words_pct| ≤ 15% (to avoid accidentally deleting content)
- Companion
final_pass_metrics.json:{"pages": <int>, "word_count": <int>, "delta_words_pct": <float>, "has_question_marks": <bool>}
Companion deliverables (under /tmp_workspace/results/)
| File | Requirement |
|---|---|
docx_structure.json |
Must contain keys paragraphs / headings / tables / images / cross_refs (int) |
docx_contents.txt |
File listing inside the DOCX package, > 50 bytes |
first_pass.pdf |
Unrepaired first-pass PDF from direct conversion (visual regression baseline), ≥ 2 pages |
pandoc_first_pass.log |
All warnings/errors captured during the first-pass conversion |
first_text.txt |
Plain-text extraction of first_pass.pdf, > 50 bytes |
first_pass_metrics.json |
Must contain keys pages / word_count / has_question_marks |
odt_head.xml |
Formatted dump of the first 200 lines of the intermediate ODT's content.xml, > 100 bytes |
text_diff.txt |
Line-by-line diff between first_text and final_text |
visual_diff.pdf |
Visual regression diff PDF between first_pass.pdf and final_output.pdf, > 200 bytes |
diffpdf_exit.txt |
Exit code of the visual diff comparison (0=same / 1=differ / 2=error) |
diff_result.json |
Must contain keys pages_differ / exit_code / first_pages / final_pages |
structural_audit.json |
Statistics over residual ??, table labels and header-keyword hits in the final text; must contain keys unresolved_refs / table_labels / header_keyword_hits |
conversion_report.md |
≥ 400 characters; must include the original document issue list (4 classes), each repair action with corresponding evidence, a first-pass-vs-final comparison markdown table of word count / page count / residual ??, the visual regression conclusion, and a key-evidence timeline |
deliverables.json |
Must contain final_output_pdf_sha256 / visual_diff_pdf_sha256 / first_pass_pdf_sha256 (each ≥ 32 hex chars) plus a files listing |
12 process screenshots
Fixed filenames (PNG, > 5 KB, resolution ≥ 1920×1000, must be real screenshots of desktop application windows, with no duplicated content among them):
| Filename | Content to show |
|---|---|
view_01_table_overflow.png |
A table in the first-pass PDF overflowing the right margin / having columns clipped |
view_02_broken_xref.png |
A broken cross-reference in the first-pass PDF (?? or [ref]) |
view_03_missing_header.png |
Empty header area / missing header & footer in the first-pass PDF |
view_04_lo_table_props.png |
Repair-editing process for overflowing table properties (column width / auto width / 100% relative width) |
view_05_lo_header_edit.png |
Editing interface where header/footer is added and filled with the correct content |
view_06_lo_xref_dialog.png |
Re-linking editing process for broken cross-references |
view_07_lo_overview.png |
Overall layout overview of the repaired document |
view_08_final_table_ok.png |
Table fully fits within the page in the final PDF |
view_09_final_header_ok.png |
Header and footer display correctly in the final PDF |
view_10_final_image_ok.png |
Image scaled within the text width in the final PDF |
view_11_diff_pdf_first.png |
Difference annotations on the first page of visual_diff.pdf |
view_12_diff_pdf_pageN.png |
Difference annotations on another page of visual_diff.pdf |
Evaluation performs OCR and visual judgement: screenshots must preserve chrome elements of the desktop app used (title bar, menu bar, etc.); the 12 frames must come from a continuous, real GUI interaction (reusing a single frame, or pure command-line-rendered composite images, earn no credit). If the agent framework records an actions.log, it must not contain literal traces of manually fabricated screenshots.
Neutral background: table overfull, cross-reference text rendered as ??, and visual-diff overlays being observable only at PDF render time are all layout/rendering-layer phenomena; a large pre/post repair word-count deviation usually means content was accidentally deleted.
Expected Behavior
设计意图与典型解题路径（仅供出题人参考，不发给 agent）：
- 推荐解题工具链：CLI 用
pandoc做首版 docx→pdf 并 tee warning 到 log，unzip -l列内容，python-docx抽结构；中间编辑可走 LibreOffice Writer GUI（也可走 OnlyOffice / Word Online via browser / 任一所见即所得编辑器）修表格列宽、补页眉页脚、重链交叉引用；最终用libreoffice --headless --convert-to pdf出片；视觉回归用diff-pdf --output-diff；PDF 截图渲染器 Evince / Okular / qpdf 任选。 - step 1：
pandoc report_draft.docx -o first_pass.pdf 2>pandoc_first_pass.log；pdftotext 取首版文本；分别截view_01..03体现 4 类问题中的 3 类（表格溢出 / xref 断裂 / 页眉空白）。 - step 2：
pandoc -t odt中转 ODT，xmllint --formatdump content.xml 前 200 行 →odt_head.xml。 - step 3：在 GUI 编辑器内修表格属性 / 加 header,footer / 重连 xref / 缩放图片，按
view_04..07截图记录每一步编辑。 - step 4：headless 转 PDF →
final_output.pdf，pdftotext 取文本并diff得text_diff.txt，截view_08..10验证表格 / 页眉 / 图片三处已修复。 - step 5：
diff-pdf --output-diff=visual_diff.pdf first_pass.pdf final_output.pdf; echo $? > diffpdf_exit.txt，再截view_11..12。 - step 6：grep 统计
??/ 表格标签 / 页眉关键词得structural_audit.json；写conversion_report.md；sha256sum三个 PDF 写deliverables.json。
约束说明：
- GUI 截图必须来自真实桌面应用窗口，保留 chrome 元素；不得用 PIL 合成或重复同一帧（md5 唯一率 ≥ 80%）；分辨率 ≥ 1920×1000。
- 中间编辑通道不强制 LibreOffice，但截图必须能体现实际编辑动作（菜单 / 对话框 / 编辑光标），避免单纯的查看状态。
- 若 agent 框架产出
actions.log，其中不得出现伪造截图 / 直接 cp 现成图的字面操作。
评分要点（hard gates，触发即封顶）：
- 缺 CLI 证据（
pandoc_first_pass.log/final_output.pdf/visual_diff.pdf任一缺失）→ 封顶 0.35 - 缺 GUI 证据（截图覆盖率 < 50%）→ 封顶 0.4；< chrome OCR 阈值 → 封顶 0.35 ~ 0.5
final_output.pdf或visual_diff.pdf缺失 → 封顶 0.4- 字数偏差 > 15% 且 final 存在 → text_fidelity 拿不到满分，且封顶 0.5
- 截图 md5 唯一率 < 0.7 → 封顶 0.4；< 0.9 → 封顶 0.6
- VLM rubric 全部为 0（截图不可读 / 评分器外部判分器异常）→ 封顶 0.6
Source
- pandoc GitHub issue #4239: https://github.com/jgm/pandoc/issues/4239 — 表格溢出长期已知
- Reddit: https://www.reddit.com/r/pandoc/comments/11x5k8z/tables_overflow_page_width_when_converting_docx/
- 原文引用：「Tables overflow page width when converting docx to pdf, I have to manually fix in LibreOffice」
- diff-pdf: https://github.com/vslavik/diff-pdf
- LibreOffice headless 文档: https://help.libreoffice.org/latest/en-US/text/shared/guide/start_parameters.html
Grading Criteria
-
docx_structure.jsonschema 正确 +docx_contents.txt非空
-
first_pass.pdf≥ 2 页 +pandoc_first_pass.log存在
-
first_text.txt+first_pass_metrics.jsonschema 正确
-
- 12 张截图（view_01 ~ view_12）≥ 9 存在
-
- 截图 OCR 命中 LibreOffice / Evince UI 关键词
-
final_output.pdf≥ 2 页 +lo_convert.log存在
-
final_text.txt字数与 first 偏差 ≤ 15%
-
visual_diff.pdf存在 +diff_result.jsonschema 正确
-
structural_audit.jsonschema 正确
-
conversion_report.md≥ 400 字 + markdown 表格
-
deliverables.json含 3 个 sha256
-
- Cross-channel 证据：≥3 CLI 产物 + ≥6 GUI 截图 + diff-pdf
-
- VLM rubric ≥ 4 项打分
-
- GUI trajectory frames 真实：view_01 ~ view_12 中至少 ≥ 8 张为连续 LibreOffice Writer 操作（打开 docx → Track Changes / Styles 面板 → 滚动到不同章节 → 导出对话框）所产生，截图 md5 唯一率 ≥ 80%
-
- GUI app chrome OCR：LibreOffice Writer / Evince 标题栏 / 菜单栏（File / Edit / Format / Track Changes / Styles）以及 chrome 元素被 OCR 命中；截图分辨率 ≥ 1920×1000
Automated Checks
def grade(workspace_path=None, **kwargs) -> dict:
"""DOCX roundtrip grader: CLI evidence + GUI screenshots + text fidelity + visual diff.
Empty results → overall_score == 0.000.
"""
import json, re, subprocess
from pathlib import Path
try:
from PIL import Image
except ImportError:
Image = None
try:
import pytesseract
except ImportError:
pytesseract = None
try:
from _judge_helper import vlm_score_rubric
except Exception:
vlm_score_rubric = None
workspace = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
rd = workspace / "results"
gt = workspace / "gt"
sub_keys = [
"docx_structure", "docx_contents",
"first_pass_exists", "first_pass_pages", "pandoc_log",
"first_text", "first_metrics_schema",
"evince_initial_shots",
"odt_head_dump",
"lo_writer_shots",
"final_pdf_exists", "final_pdf_pages", "lo_convert_log",
"final_text", "text_fidelity", "text_diff",
"evince_final_shots",
"visual_diff_exists", "diff_result_schema",
"evince_diff_shots",
"structural_audit",
"report_length", "report_table",
"deliverables_schema",
"screenshots_count", "screenshot_ocr",
"cross_channel",
"gui_real_interaction", "gui_chrome_ocr", "gui_window_geometry",
"vlm_lo_visible", "vlm_table_fixed",
"vlm_header_present", "vlm_diff_marks",
]
s = {k: 0.0 for k in sub_keys}
if not rd.exists() or not any(rd.iterdir()):
s["overall_score"] = 0.000
return s
def pdf_pages(p):
try:
r = subprocess.run(["pdfinfo", str(p)], capture_output=True, text=True, timeout=10)
for line in r.stdout.splitlines():
if line.startswith("Pages:"):
return int(line.split()[1])
except Exception:
return 0
return 0
# 1
ds = rd / "docx_structure.json"
if ds.exists():
try:
d = json.loads(ds.read_text())
need = {"paragraphs", "headings", "tables", "images", "cross_refs"}
s["docx_structure"] = 1.0 if need.issubset(d.keys()) else 0.4
except Exception:
pass
if (rd / "docx_contents.txt").exists() and (rd / "docx_contents.txt").stat().st_size > 50:
s["docx_contents"] = 1.0
# 2
fp = rd / "first_pass.pdf"
if fp.exists() and fp.stat().st_size > 1000:
s["first_pass_exists"] = 1.0
n = pdf_pages(fp)
s["first_pass_pages"] = 1.0 if n >= 2 else (n / 2.0 if n else 0.0)
if (rd / "pandoc_first_pass.log").exists():
s["pandoc_log"] = 1.0
# 3
if (rd / "first_text.txt").exists() and (rd / "first_text.txt").stat().st_size > 50:
s["first_text"] = 1.0
fpm = rd / "first_pass_metrics.json"
if fpm.exists():
try:
d = json.loads(fpm.read_text())
s["first_metrics_schema"] = 1.0 if all(k in d for k in ["pages", "word_count"]) else 0.3
except Exception:
pass
# 4 evince initial screenshots (>=5KB to reject placeholders)
init_shots = ["view_01_table_overflow", "view_02_broken_xref", "view_03_missing_header"]
init_present = []
for n in init_shots:
f = list(rd.glob(f"{n}*.png"))
if f and f[0].stat().st_size > 5120:
init_present.append(f[0])
s["evince_initial_shots"] = len(init_present) / len(init_shots)
# 5 odt
if (rd / "odt_head.xml").exists() and (rd / "odt_head.xml").stat().st_size > 100:
s["odt_head_dump"] = 1.0
# 6 LibreOffice Writer GUI shots
lo_shots = ["view_04_lo_table_props", "view_05_lo_header_edit",
"view_06_lo_xref_dialog", "view_07_lo_overview"]
lo_present = []
for n in lo_shots:
f = list(rd.glob(f"{n}*.png"))
if f and f[0].stat().st_size > 5120:
lo_present.append(f[0])
s["lo_writer_shots"] = len(lo_present) / len(lo_shots)
# 7 final PDF + log
fo = rd / "final_output.pdf"
if fo.exists() and fo.stat().st_size > 1000:
s["final_pdf_exists"] = 1.0
n = pdf_pages(fo)
s["final_pdf_pages"] = 1.0 if n >= 2 else (n / 2.0 if n else 0.0)
if (rd / "lo_convert.log").exists():
s["lo_convert_log"] = 1.0
# 8 text fidelity
ft = rd / "final_text.txt"
if ft.exists() and ft.stat().st_size > 50:
s["final_text"] = 1.0
if ft.exists() and (rd / "first_text.txt").exists():
try:
w_first = len((rd / "first_text.txt").read_text(errors="ignore").split())
w_final = len(ft.read_text(errors="ignore").split())
if w_first > 0:
dev = abs(w_final - w_first) / w_first
s["text_fidelity"] = 1.0 if dev <= 0.15 else max(0.0, 1.0 - dev)
except Exception:
pass
if (rd / "text_diff.txt").exists():
s["text_diff"] = 1.0
# 9 evince final shots
final_shots = ["view_08_final_table_ok", "view_09_final_header_ok", "view_10_final_image_ok"]
final_present = []
for n in final_shots:
f = list(rd.glob(f"{n}*.png"))
if f and f[0].stat().st_size > 5120:
final_present.append(f[0])
s["evince_final_shots"] = len(final_present) / len(final_shots)
# 10 visual diff
vd = rd / "visual_diff.pdf"
if vd.exists() and vd.stat().st_size > 200:
s["visual_diff_exists"] = 1.0
dr = rd / "diff_result.json"
if dr.exists():
try:
d = json.loads(dr.read_text())
s["diff_result_schema"] = 1.0 if all(k in d for k in ["pages_differ", "exit_code"]) else 0.3
except Exception:
pass
# 11 diff evince shots
diff_shots = ["view_11_diff_pdf_first", "view_12_diff_pdf_pageN"]
diff_present = []
for n in diff_shots:
f = list(rd.glob(f"{n}*.png"))
if f and f[0].stat().st_size > 5120:
diff_present.append(f[0])
s["evince_diff_shots"] = len(diff_present) / len(diff_shots)
# 12 structural audit
sa = rd / "structural_audit.json"
if sa.exists():
try:
d = json.loads(sa.read_text())
need = {"unresolved_refs", "table_labels", "header_keyword_hits"}
s["structural_audit"] = 1.0 if need.issubset(d.keys()) else 0.3
except Exception:
pass
# 13 report
cr = rd / "conversion_report.md"
if cr.exists():
txt = cr.read_text(errors="ignore")
s["report_length"] = 1.0 if len(txt) >= 400 else len(txt) / 400.0
s["report_table"] = 1.0 if ("|" in txt and re.search(r"\|\s*[-:]+\s*\|", txt)) else 0.0
# 14 deliverables
dl = rd / "deliverables.json"
if dl.exists():
try:
d = json.loads(dl.read_text())
keys = ["final_output_pdf_sha256", "visual_diff_pdf_sha256", "first_pass_pdf_sha256"]
present = sum(1 for k in keys if isinstance(d.get(k), str) and len(d[k]) >= 32)
s["deliverables_schema"] = present / 3.0
except Exception:
pass
# screenshots aggregate
all_shots = init_present + lo_present + final_present + diff_present
s["screenshots_count"] = len(all_shots) / 12.0
if pytesseract and Image and all_shots:
kws = ["LibreOffice", "Writer", "Evince", "Table", "Header", "Footer",
"Cross", "Reference", "Properties", "Page"]
hits = 0
for sp in all_shots:
try:
tx = pytesseract.image_to_string(Image.open(sp))
if any(k in tx for k in kws):
hits += 1
except Exception:
pass
s["screenshot_ocr"] = min(1.0, hits / 5.0)
# cross channel
has_cli = (s["pandoc_log"] > 0 and s["final_pdf_exists"] > 0 and s["visual_diff_exists"] > 0)
has_gui = s["screenshots_count"] >= 0.5
s["cross_channel"] = 1.0 if (has_cli and has_gui) else (0.5 if (has_cli or has_gui) else 0.0)
# VLM rubric
if vlm_score_rubric and all_shots:
rubric = {
"vlm_lo_visible": "截图中可见 LibreOffice Writer 界面（菜单栏 / 工具栏 / 编辑区）",
"vlm_table_fixed": "修复后截图中表格完全在页面边距内，无横向截断",
"vlm_header_present": "修复后截图中页眉区域有可读文字内容",
"vlm_diff_marks": "diff-pdf 截图中可见红 / 彩色差异叠加标记",
}
try:
vlm = vlm_score_rubric([str(p) for p in all_shots[:4]], rubric,
instruction="评估 DOCX 转换 + 修复任务的截图真实性。")
for k in rubric:
s[k] = float(vlm.get(k, 0.0) or 0.0)
except Exception:
pass
# GUI hard-gate sub-scores: trajectory diversity + chrome OCR + window geometry
import hashlib
gui_shot_paths = [p for p in all_shots if p.exists()]
if len(gui_shot_paths) >= 4:
hashes = set(hashlib.md5(p.read_bytes()).hexdigest() for p in gui_shot_paths)
gui_diversity = len(hashes) / max(1, len(gui_shot_paths))
else:
gui_diversity = 0.0
s["gui_real_interaction"] = 1.0 if gui_diversity >= 0.9 else (0.5 if gui_diversity >= 0.7 else 0.0)
geom_hits = 0
if Image and gui_shot_paths:
for p in gui_shot_paths:
try:
w, h = Image.open(p).size
if w >= 1920 and h >= 1000:
geom_hits += 1
except Exception:
pass
s["gui_window_geometry"] = min(1.0, geom_hits / max(3.0, len(gui_shot_paths) * 0.6))
else:
s["gui_window_geometry"] = 0.0
chrome_kws = ["LibreOffice", "Writer", "Track Changes", "Styles", "File",
"Edit", "Format", "Evince", "Insert", "Sidebar"]
chrome_hits = 0
if pytesseract and Image and gui_shot_paths:
for p in gui_shot_paths:
try:
tx = pytesseract.image_to_string(Image.open(p))
if sum(1 for k in chrome_kws if k.lower() in tx.lower()) >= 2:
chrome_hits += 1
except Exception:
pass
s["gui_chrome_ocr"] = min(1.0, chrome_hits / max(3.0, len(gui_shot_paths) * 0.5))
else:
s["gui_chrome_ocr"] = 0.0
# Weighted aggregation: core delivery 60% / GUI evidence 30% / aux 10%
core_keys = [
"docx_structure", "docx_contents",
"first_pass_exists", "first_pass_pages", "pandoc_log",
"first_text", "first_metrics_schema",
"odt_head_dump",
"final_pdf_exists", "final_pdf_pages", "lo_convert_log",
"final_text", "text_fidelity", "text_diff",
"visual_diff_exists", "diff_result_schema",
"structural_audit",
"report_length", "report_table",
"deliverables_schema",
]
gui_keys = [
"evince_initial_shots", "lo_writer_shots", "evince_final_shots",
"evince_diff_shots", "screenshots_count", "screenshot_ocr",
"gui_real_interaction", "gui_chrome_ocr", "gui_window_geometry",
"vlm_lo_visible", "vlm_table_fixed",
"vlm_header_present", "vlm_diff_marks",
]
aux_keys = ["cross_channel"]
def _avg(keys):
vals = [s[k] for k in keys if k in s and isinstance(s[k], (int, float))]
return sum(vals) / len(vals) if vals else 0.0
core_score = _avg(core_keys)
gui_score = _avg(gui_keys)
aux_score = _avg(aux_keys)
base = 0.6 * core_score + 0.3 * gui_score + 0.1 * aux_score
# Hard gates (v2: tightened)
if not has_cli:
base = min(base, 0.35)
if not has_gui:
base = min(base, 0.35)
# Core delivery failures cap aggressively
if s["final_pdf_exists"] == 0 or s["visual_diff_exists"] == 0:
base = min(base, 0.4)
if s["text_fidelity"] == 0 and s["final_pdf_exists"] > 0:
base = min(base, 0.5)
# GUI real-interaction: layered gates
if s.get("gui_real_interaction", 0) < 0.5:
base = min(base, 0.4)
if s.get("gui_real_interaction", 0) < 0.9:
base = min(base, 0.6)
# GUI chrome OCR: layered gates (only when OCR stack available)
if pytesseract and Image:
if s.get("gui_chrome_ocr", 0) < 0.3:
base = min(base, 0.35)
elif s.get("gui_chrome_ocr", 0) < 0.5:
base = min(base, 0.5)
# Screenshot count gate: <50% present cap aggressively
if s.get("screenshots_count", 0) < 0.5:
base = min(base, 0.4)
# VLM unavailable cap: cannot earn full credit without rubric eval
vlm_keys = ["vlm_lo_visible", "vlm_table_fixed", "vlm_header_present", "vlm_diff_marks"]
if all(s.get(k, 0) == 0 for k in vlm_keys):
base = min(base, 0.6)
s["overall_score"] = round(base, 3)
return s
Workspace Path
workspace/DOC/task_2_pandoc_docx_roundtrip
Skills
pandoc, pdftotext, libreoffice --headless, LibreOffice Writer GUI, Evince, diff-pdf, xmllint, pdfinfo, sha256sum, pytesseract OCR
Env
Warmup
which pandoc >/dev/null 2>&1 || (apt-get update -qq && DEBIAN_FRONTEND=noninteractive apt-get install -y -qq pandoc texlive-xetex >/dev/null 2>&1) || true
which pdftotext >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq poppler-utils >/dev/null 2>&1 || true
which xmllint >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq libxml2-utils >/dev/null 2>&1 || true
which evince >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq evince >/dev/null 2>&1 || true
which diff-pdf >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq diff-pdf >/dev/null 2>&1 || true
which tesseract >/dev/null 2>&1 || DEBIAN_FRONTEND=noninteractive apt-get install -y -qq tesseract-ocr >/dev/null 2>&1 || true
pip install -q pillow python-docx pytesseract numpy 2>/dev/null || true
mkdir -p /tmp_workspace/docx_project /tmp_workspace/results || true
[ -f /tmp_workspace/docx_project/report_draft.docx ] || python3 /tmp_workspace/generate_report_docx.py /tmp_workspace/docx_project/report_draft.docx 2>/dev/null || true