swun-thesis-docx-banshi1

name	swun-thesis-docx-banshi1
description	Build SWUN thesis DOCX (Format 1 / 版式1) from LaTeX using the official SWUN reference template, with post-processing fixes (TOC, chapter page breaks, indents, isLgl numbering fix, and three-line table layout normalization).
version	1.6.0

SWUN Thesis DOCX (版式1)

Generate a .docx from a SWUN thesis LaTeX project while treating the official SWUN Format 1 template as the source of truth.

Output

• main_版式1.docx in the thesis directory.

Pipeline

1. latexpand flattens main.tex
2. LaTeX preprocessing (_preprocess_latex):
   • flattens subfigure environments: replaces subfigure \ref{} with parent figure \ref{}, strips subfigure \caption/\label to prevent pandoc from inflating the figure counter, deduplicates adjacent identical refs
   • rewrites experiment figure references from .pdf to .png when PNG assets exist
   • hard-fails if any experiment figure still points to .pdf after rewrite
3. pandoc converts LaTeX -> DOCX with:
   • --reference-doc pointing at the official template
   • --citeproc + GB/T 7714-2015 numeric CSL
   • --resource-path so \\includegraphics assets are embedded
4. OOXML post-processing:
   • inserts Word TOC field
   • adds page breaks before each chapter (Heading 1)
   • adds first-line indents for body paragraphs
   • adds hanging indents for bibliography entries
   • rebuilds figure/table captions from LaTeX metadata (\label + \bilingualcaption/\caption) and DOCX anchors (fig:/tab:/tbl:):
     • hard-fails if a figure/table environment in LaTeX is missing caption or label
     • hard-fails if \bilingualcaption has an empty English title
     • figure captions are placed below figure blocks
     • table captions are placed above table blocks
     • chapter-based numbering: 图{章}-{序号} / Figure {章}-{序号} / 表{章}-{序号} / Table {章}-{序号}
     • bilingual captions emit Chinese + English lines; plain \caption{} emits Chinese line only
     • idempotent rewrite: clears nearby old captions before writing normalized ones
     • caption paragraphs are centered and keep-together hints are applied
   • fixes figure/table reference format in body text:
     • replaces pandoc's dot-format refs (图3.1, 表4.2) with hyphen format (图3-1, 表4-2)
     • normalizes standalone hyperlink ref text for fig:/tab:/tbl: anchors (3.16 -> 3-16)
     • handles both single-node refs and split-across-runs refs (pandoc often splits 图 and 3.1 into separate XML runs with whitespace runs in between)

• removes internal cross-reference hyperlinks in thesis main body (正文):
  • unwraps w:hyperlink nodes carrying internal w:anchor in chapter content and keeps plain text runs
  • preserves non-anchor external hyperlinks (e.g., DOI/URL r:id) by default
  • strips hyperlink-like run style (blue/underline) only on unwrapped anchor runs so cross-reference text renders as normal text
• normalizes \paragraph headings in thesis main body:
  • writes explicit numbering prefix as (n) 标题 on Heading5
  • resets numbering per Heading3 (subsection), fallback reset per Heading2
• fixes equation numbering (display math):
  • adds chapter-based numbering suffix: (章-序号) right-aligned
  • avoids numbering \\[ ... \\] blocks (keeps them unnumbered)
• fixes mixed Chinese/Arabic section numbering by adding w:isLgl to ilvl >= 1 (abstractNumId=0)
• enforces data-table layout and typography for three-line tables:
  • table width fills the whole text area (tblW == text width)
  • table layout is fixed (w:tblLayout w:type="fixed")
  • column widths are redistributed by table content and written back to tblGrid + tcW
  • cell text font: Chinese = 宋体 (w:eastAsia), English/digits = Times New Roman (w:ascii/w:hAnsi), size = 五号 10.5pt (w:sz="21")
  • table font normalization runs AFTER _normalize_ascii_run_fonts to prevent English runs losing eastAsia attribute
  • clears theme font attributes (asciiTheme, hAnsiTheme, eastAsiaTheme) to prevent theme override
• writes table metadata (w:tblCaption) from normalized Chinese table titles for stable downstream checks
• disables chapter numbering for backmatter headings (致谢, 参考文献, 科研成果):
  • adds paragraph-level numPr with numId=0 to override style-level numbering
  • same technique used for 目录/摘要/Abstract
• enables automatic field update on open ( in settings.xml):

Refactored Architecture

• Thin CLI entrypoints:
  • scripts/build_docx_banshi1.py
  • scripts/verify_extra.py
  • scripts/gate_loop.sh
  • scripts/gate_loop_runner.py
• Core build module:
  • scripts/modules/docx_builder.py
• Facade modules (phase 3/4 extraction):
  • scripts/modules/latex_parser.py
  • scripts/modules/style_processor.py
  • scripts/modules/reference_handler.py
  • scripts/modules/figure_table_handler.py
  • scripts/modules/equation_handler.py
  • scripts/modules/template_loader.py
  • scripts/modules/post_processor.py
• Verification modules:
  • scripts/verification/report_generator.py
  • scripts/verification/structure_checker.py
  • scripts/verification/style_checker.py
  • scripts/verification/reference_checker.py
  • scripts/verification/content_checker.py
• Gate-loop phase modules:
  • scripts/phase_checks/phase1_structure.py
  • scripts/phase_checks/phase2_style.py
  • scripts/phase_checks/phase3_caption.py
  • scripts/phase_checks/phase4_crossref.py
  • scripts/phase_checks/phase5_content.py
  • scripts/phase_checks/phase6_visual.py

Prerequisites

• Tools: pandoc, latexpand, python3
• Thesis workspace contains main.tex (and usually backmatter/references.bib, figures/media directories, etc.)
• Reference template (resolution order):
  1. SWUN_REFERENCE_DOCX env var
  2. {thesis_dir}/reference_thesis.docx (per-thesis override)
  3. skill-bundled slim template assets/swun_banshi1_reference.docx (default, self-contained)
• The bundled slim template is generated by scripts/make_slim_template.py from the full reference thesis: it keeps styles/numbering/settings/headers/footers, the cover pages (incl. the first inline sectPr that pandoc reads for page geometry), and one sample figure/table caption paragraph for caption-profile extraction, while stripping all other body content and unreferenced media (7.4MB -> ~90KB)

Usage

# Default: /Users/bit/LaTeX/SWUN_Thesis
bash /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/main.sh

# Or specify thesis dir explicitly
bash /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/main.sh /path/to/SWUN_Thesis

# Optionally override reference template for the build script
export SWUN_REFERENCE_DOCX="/path/to/reference_thesis_or_template.docx"
export SWUN_CSL="/path/to/china-national-standard-gb-t-7714-2015-numeric.csl"
export SWUN_BIB="/path/to/references.bib"

# Regenerate the bundled slim template from a full reference thesis
python3 /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/make_slim_template.py \
  --input /path/to/full_reference.docx \
  --output /Users/bit/.claude/skills/swun-thesis-docx-banshi1/assets/swun_banshi1_reference.docx

Direct CLI examples

python3 /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/build_docx_banshi1.py /Users/bit/LaTeX/SWUN_Thesis
python3 /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/verify_extra.py /Users/bit/LaTeX/SWUN_Thesis/main_版式1.docx

Gate-Loop Usage (6-Phase)

# Run all phases (default)
bash /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/gate_loop.sh /Users/bit/LaTeX/SWUN_Thesis --skip-build

# Run only one phase
python3 /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/gate_loop_runner.py \
  /Users/bit/LaTeX/SWUN_Thesis \
  --phase 1 --skip-build

# Tune retry count (used when codex auto-fix is enabled)
python3 /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/gate_loop_runner.py \
  /Users/bit/LaTeX/SWUN_Thesis \
  --max-retry 2

Built-in Verification (via main.sh)

main.sh runs a 5-step pipeline:

1. Build DOCX by running build_docx_banshi1.py
   • includes hard validation that experiment figures are embedded as PNG
2. Run regression samples for ref normalization via ref_hyphen_regression.py
3. Run abstract section regression samples via abstract_section_regression.py
   • guards against duplicate 摘要 heading when heading/first paragraph are separated by blank lines
   • guards against truncating multi-paragraph English abstract
4. Run generic structural checks with verify_extra.py
   • includes regression guard: no forced page break between "摘要"/"Abstract" headings and their first content paragraph
   • includes regression guard: \paragraph mapped Heading5 line MUST have first-line indent (matching PDF layout), and its first following body paragraph must also keep first-line indent
   • hard check: figure/table hyperlink refs (fig:/tab:/tbl:) must not contain dot numbering (3.16); must use hyphen (3-16)
   • hard check: main-body internal cross-reference links must be plain text (no w:anchor hyperlink nodes, no hyperlink-like blue/underline style residue)
   • hard check: references must NOT contain DOI external hyperlinks (w:hyperlink r:id with DOI text); builder strips them
   • hard check: main-body Heading5 (\paragraph) must start with explicit (n) prefix

• hard check: each main-body fig:/tab:/tbl: anchor must have Chinese caption line with correct position and chapter-local numbering; if English caption line exists, its format/numbering must also be valid
• hard check: figure captions must be below figure blocks, table captions must be above table blocks

5. Run table-specific checks to enforce:
   • data table full-width (tblW=dxa and equals text width)
   • fixed table layout (tblLayout=fixed)
   • tblGrid width sum equals table width
   • table captions are above tables (no caption-below)
   • table captions follow chapter numbering order (表{章}-{序号})
   • when no data table exists, defaults to SKIP (set SWUN_TABLE_VERIFY_ALLOW_EMPTY=0 to enforce hard-fail)

Guardrails

• Never recreate styles/numbering from scratch. Always use the official template as --reference-doc.
• Do not manually type chapter/section heading numbering; rely on the template multilevel list mapping.
• \paragraph (Heading5) in main body is the exception: numbering is explicitly normalized to (n) text during post-processing.
• For three-line tables, do not keep auto table width and keep captions above tables.
• Figure/table references must use caption-style hyphen numbering (图3-16, 表4-2), not dot numbering (图3.16, 表4.2).
• Main body (正文) internal cross-reference links must be plain text (no blue underline clickable links).
• Experiment figures must be embedded as PNG in DOCX (fig_3_*, fig_4_*); PDF embeds are treated as build errors.

触发条件 / Trigger

触发词

本 skill 在以下情形下被激活：

• /swun-docx — 快捷触发词
• 构建版式1 / build docx banshi1 / build banshi1 docx
• 生成 DOCX / 导出 DOCX / DOCX 构建
• swun-thesis-docx / swun thesis docx
• 用户要求构建 SWUN 论文的 Word 格式文件时

应触发（Should Trigger）

用户: 帮我构建一下版式1 DOCX
→ 触发本 skill，调用 scripts/main.sh

用户: /swun-docx
→ 触发本 skill

用户: 论文写完了，需要导出 Word 给导师
→ 触发本 skill（SWUN 毕业论文场景）

不应触发（Should NOT Trigger）

用户: 帮我修改 chapter3.tex 的内容
→ 不触发（这是 LaTeX 编辑任务，不是 DOCX 构建）

用户: 检查论文的引用格式
→ 不触发（这是审查任务）

用户: 构建普通 Word 文档
→ 不触发（必须是 SWUN 学位论文项目）

Constraints / 约束

CAN（允许）

• 可以调用 pandoc + latexpand 转换 LaTeX → DOCX
• 可以对生成的 DOCX 进行 OOXML 后处理（TOC、段落样式、标题、脚注等）
• 可以覆盖写入 main_版式1.docx（每次构建均覆盖）
• 可以读取 main.tex 及其 \include 的子文件（只读）
• 可以修改构建临时文件（/tmp/ 下的中间产物）
• 可以运行回归测试脚本（ref_hyphen_regression.py、abstract_section_regression.py）

CANNOT（禁止）

• 不可修改 chapters/*.tex、backmatter/references.bib 等源文件（构建不改源）
• 不可删除用户的 LaTeX 源文件或图片资产
• 不可重建 Word 样式/编号（必须从官方模板继承，禁止从零创建）
• 不可手动写入章节标题编号（由模板多级列表映射处理）
• 不可跳过 PNG 验证步骤（实验图必须为 PNG，PDF 嵌入视为构建错误）
• 不可将 DOI 超链接写入参考文献区（项目规则：最终输出无 DOI）
• 不可使用 ZIP 重打包方式修改 DOCX（会损坏文件内部引用）

Error Handling / 错误处理

常见错误与处理策略

错误类型	触发条件	处理方式
Build failure	pandoc / latexpand 返回非零退出码	立即终止，输出完整错误日志，不生成部分 DOCX
PDF embed error	fig_3_* / fig_4_* 实验图以 PDF 嵌入	hard-fail，提示用户将对应图转为 PNG
Missing caption/label	LaTeX 浮动体缺少 \caption 或 \label	hard-fail，输出缺失的环境位置
Empty English caption	\bilingualcaption 英文标题为空	hard-fail，提示补全英文标题
Timeout	pandoc / 后处理脚本超过 300s	超时退出，提示检查文档大小或依赖工具
Template not found	官方模板文件不存在于默认路径	报错并提示设置 SWUN_TEMPLATE_DOCX 环境变量
Crash / 异常	Python 后处理脚本未捕获异常	输出完整 traceback，保留已生成的中间 DOCX 以供调试
Gate loop failure	6 阶段校验中任一阶段不通过	输出失败阶段和具体违规项，等待人工修复或 codex 自动修复

自动恢复策略

• 回归测试失败：打印失败用例的 input/expected/actual，不中断整体流程（继续后续阶段）
• 表格样式警告：SWUN_TABLE_VERIFY_ALLOW_EMPTY=1（默认）时无三线表则跳过（SKIP），不报错
• subagent 隔离：每次构建在独立 Python 进程中运行，崩溃不影响主会话

评估 / Evaluation

构建质量评估

gate-loop（gate_loop.sh / gate_loop_runner.py）按 6 个阶段逐步评估输出质量：

阶段	评估脚本	检查内容
Phase 1	phase1_structure.py	标题层级、目录结构、页码分区
Phase 2	phase2_style.py	正文样式、字体、段落缩进
Phase 3	phase3_caption.py	图表标题格式、位置、编号
Phase 4	phase4_crossref.py	正文内交叉引用格式（连字符编号）
Phase 5	phase5_content.py	参考文献、摘要关键词
Phase 6	phase6_visual.py	可视化抽样检查（可选）

评分标准（keep / discard）

• keep：所有阶段 PASS 或 SKIP，无 FAIL
• discard：任一阶段 FAIL 且无法自动修复
• 回归测试：ref_hyphen_regression.py + abstract_section_regression.py 全部 PASS 为基准条件

辅助评估工具

# 运行 gate-loop 全部阶段
bash scripts/gate_loop.sh /path/to/SWUN_Thesis

# 运行 normalize_template 检查模板状态
python3 scripts/normalize_template.py /path/to/template.docx

# 运行 verify_extra 结构检查
python3 scripts/verify_extra.py /path/to/main_版式1.docx

Execution Isolation / 执行隔离

• 本 skill 通过 subagent 派发执行，每次构建在独立 Python 进程中运行
• 构建脚本之间无共享全局状态（无单例、无全局缓存）
• 多个 SWUN 论文项目可并发构建（使用不同 thesis_dir 参数），互不干扰
• Gate-loop 各阶段顺序执行，不并发（保证前置阶段数据供后置阶段使用）
• 临时文件写入 /tmp/，构建完成后可安全清理

Setup / 安装与配置

1. 确保依赖工具已安装：

   brew install pandoc latexmk
   pip3 install python-docx lxml
   

2. 确认内置瘦身模板存在（skill 自包含，默认无需外部模板）：

   ls /Users/bit/.claude/skills/swun-thesis-docx-banshi1/assets/swun_banshi1_reference.docx
   

3. （可选）配置环境变量覆盖默认路径：

   export SWUN_REFERENCE_DOCX="/path/to/reference_thesis.docx"
   export SWUN_CSL="/path/to/gb-t-7714-2015-numeric.csl"
   export SWUN_BIB="/path/to/references.bib"
   

4. 验证安装：

   bash /Users/bit/.claude/skills/swun-thesis-docx-banshi1/scripts/main.sh /Users/bit/LaTeX/SWUN_Thesis
   

输出格式 / Output Format

构建成功时输出示例:
  [build] latexpand done
  [build] pandoc done
  [postprocess] TOC inserted
  [postprocess] chapter page breaks done
  [verify] Phase 1: PASS
  [verify] Phase 2: PASS
  ...
  [done] main_版式1.docx written (N pages)

构建失败时输出示例:
  [ERROR] experiment figure fig_4_2 is embedded as PDF, not PNG
  Build aborted.

输出文件：{thesis_dir}/main_版式1.docx