痛点 1:网页端 AI 的"对话—复制—粘贴"死循环。 在 ChatGPT / Claude 网页端和 AI 对话改论文,看起来省了思考,实则陷入更耗精力的体力活:每改一段就要复制粘贴一轮——"把这一段改得更学术"→ 等回复 → 选中 → Ctrl+C → 切回 Word → Ctrl+V → 格式崩了 → 调格式 → 下一段……一个下午过去,真正动脑的时间不到半小时。
✅ md-paper:意见批量扔进去,AI 自动逐条改到 Markdown 真源里,你只在人工闸确认时动一次脑,其余体力活全免。
痛点 2:一次只能改一条意见,无法批量多 Agent 修改。 你和网页端 AI 对话时,一次只能说一件事:"帮我把引言改得更简洁"。说完等回复,再粘贴,再说下一件。就算你自己完全理解了 30 条审稿意见,把它们逐一打字成 prompt 再逐一粘贴结果,就已经耗尽了一整天——根本没有"批量多 Agent 并行修改"的可能。
✅ md-paper:蜂群模式(Swarm)——30 条审稿意见扔进去,AI 自动分类、去重、合并(
md-triage)→ 你一次性确认 → 并行派多个 agent 起草、确定性脚本串行落盘到真源(md-swarm)。从"一次改一条"升级到"一次改一整轮审稿"。
痛点 3:Word 里的图、表、题注、注释、Zotero 域——AI 一改全乱。 Word 文档里嵌着丰富的学术基础设施:图的题注("Figure 1: …")、表的题注和注释("Note: …")、Zotero 引用域、交叉引用书签……用网页端 AI 改完粘贴回 Word 时,这些全部丢失或错乱。更致命的是 Zotero Refresh 一下,之前精心调整的引用格式全部打回原形。
✅ md-paper:默认保留原 Zotero 活域和捕获的 Word 格式，图表题注与已绑定的单个引用联动。格式和对象由代码验证；具体支持范围以每篇稿件的读取、构建报告为准。
痛点 4:审稿意见太笼统,自己生成初稿耗时耗力。 "请把文章中的第一人称改为客观第三人称"、"减少语法纰漏"、"提高文献综述的覆盖面"……这类审稿意见理解起来不难,但执行起来极其耗时:你要通读全文找每一处 I/we/our,逐一替换;要逐句检查语法;要翻遍文献补漏。审稿人一句话,你要干一整天。
✅ md-paper:去 AI 味 + 语言铁律内嵌在子 Agent 契约里。笼统意见交给 AI 逐条落地——"全文去第一人称"就是一个 swarm 任务,几分钟改完,你只做最终确认。
痛点 5:改稿过程中引用静默丢失,发现时已晚。
这是开发过程中实际测量到的真实事故类型:AI 改稿时会"顺手"删掉它觉得不重要的引用、拆散成组引用 [@a; @b; @c]、把 citekey 改错一个字母——等你发现时已经过了好几轮修改,根本追溯不到是哪一步丢的。
✅ md-paper:引用默认不删硬闸——补丁模式下少一条引用直接拒写;引用体检——改完后自动列出每一条被丢的 citekey、被拆的引用组、无定义的交叉引用。图、表、公式(带编号的)也一样,每个改稿批次都被数着。
痛点 6:改完的稿子读起来像 AI 写的。 AI 重度润色后的论文有一股"AI 味"——句子工整对称、修辞堆砌、破折号泛滥、读起来不像人写的。期刊审稿人越来越容易识别 AI 生成文本,这对投稿是致命伤。
✅ md-paper:七条去 AI 味铁律内嵌在改稿流程里——少修辞、短句、打破工整、降正式度、保信息、少破折号、全覆盖。只改"怎么说",不改"说什么"。
痛点 7:AI 改稿无人复核,错了不知道错在哪。 AI 改稿最大的风险不是改得不好,而是改错了你不知道——一条引用被悄悄删掉、一段话被改窜了意思、前后术语对不上,等终审才发现已经过了好几轮。网页端 AI 给你一段结果,你只能接受或粘贴回去重问,中间没有任何“复核”环节,更没人替你盯着每一条改动。
✅ md-paper:三省制 + 多道校验——借鉴唐代三省“中书起草、门下审核、尚书执行”之意,把改稿拆成提(整理意图)/ 行(起草)/ 审(只读审计)三权分立、互不兼任。改稿全程跑多道校验:每个批次自动体检引用有没有丢、有没有凭空编造、有没有真正落地;全部改完后再跑两道只读终审——D 审计逐条核“做没做、做对没”,E 复查通读全文查前后矛盾。AI 只起草、只出报告,最终裁决权始终在你手里。
| 传统改稿（手搓 Word） | md-paper（Markdown 真源 + AI + pandoc） |
|---|---|
| 😫 手动逐条改审稿意见 | ✅ 蜂群模式:一堆意见扔进去,AI 批量改 |
| 😫 Zotero Refresh 一下引用全毁 | ✅ 不毁 Zotero 域——出稿自带活引用域,Refresh 即用 |
| 😫 插图表后题注编号全乱 | ✅ 题注与已绑定引用联动，缺失目标会报错 |
| 😫 Word 格式崩了只能 Ctrl+Z | ✅ Markdown 纯文本真源,git 版本控制 |
| 😫 改到一半发现引用对不上 | ✅ 引用默认不删硬闸 |
| 😫 网页端复制粘贴到手软 | ✅ 全流程在编辑器内,零粘贴 |
| 😫 改完一股 AI 味被审稿人识破 | ✅ 七条去 AI 味铁律 |
| 😫 AI 改稿无人复核,错了不知错在哪 | ✅ 三省制 + 多道校验,两道只读终审 |
🔗 兼容 Zotero 活引用域 · 🖼️ 图/表/题注/注释纯文本 · 🔀 交叉引用自动编号 · 🐝 并行蜂群改稿 · 🛡️ 引用默认不删 · 📝 可 git diff 的 Markdown 真源 · 🔄 去 AI 味 · 🧮 OMML→LaTeX 公式 · 🧰 全局工具链 · 🧩 开放 Agent Skills 标准——Claude Code / Codex / OpenCode / Hermes 通用 · 🏛️ 三省制(提/行/审分权)· 多道校验审计。
默认流程现在是原 Word → 当前 Markdown + 原格式档案 → Word。程序读取真实样式、局部格式、分节、图表、脚注尾注等，不根据内容猜“标题页/摘要页”。manuscript.md 会显示 custom-style="正文" / custom-style="body text" 等轻量标记，本稿的中英名称对照在 format_names.md。
| 出稿模式 | 用途 |
|---|---|
rebuild（默认） |
保留原格式和已有 Zotero 活域，可离线构建 |
live |
新建引用组；新条目须有真实 Zotero key，出稿后执行 Refresh |
static |
去掉 Zotero 域、保留显示文字，供校对；解除域可能影响排版 |
AI 改稿走受保护的格式验证入口，再由原单写者落盘。手改时也要保留格式与对象标记。完整搬迁论文项目时，带上 manifest/、media/ 等整个目录；只有 Markdown 不足以恢复原稿格式。没有格式档案的旧项目仍受支持，旧转换需明确指定 -Layout legacy。
我们的设计理念是最大程度的减少AI到论文的诸多不必要的体力活，但绝不是完全AI替代作者。 md-paper 的流程很长——摄取、整理、起草、审计、出稿,每一步都有人工闸:意见要确认才落地、改动要审阅才通过、引用增删要授权。把整个流程走完,稿子本质上还是作者本人的东西。AI 省去的是打字、粘贴、调格式、逐条发命令这些体力活,但该作者动脑的地方其实一个没少:审稿意见怎么理解、改到什么程度、引用留哪些删哪些、终稿定不定稿,全是使用者本人的决策。
我们的立场:AI 只是辅助,让改稿更便捷,但不能代替作者本人的思维与原创。 工具替你跑腿,不替你思考。
第一步:把这个仓库弄到你电脑上。 两种方式,任选一种:
- 不会命令行(推荐给小白): 点这个页面右上角绿色的
< > Code按钮 → Download ZIP → 下载后解压到一个你记得住的文件夹。 - 会用 git: 在终端运行
git clone https://github.com/pwya/md-paper.git
第二步:用 VS Code 打开这个文件夹。 打开 VS Code → 菜单 File → Open Folder(打开文件夹) → 选中你刚解压/克隆出来的 md-paper 文件夹。然后在里面开一个 Claude Code 会话(或 Codex / OpenCode / Hermes Agent 等其他 AI 编程工具,均支持)。
第三步:让 AI 替你装。 对 AI 说:
"读一下
INSTALL.md,帮我把 md-paper 装好。"
AI 会照着 INSTALL.md(一份可执行的操作手册)把五个技能接进你的 AI 工具(Claude Code / Codex / OpenCode / Hermes Agent 都行)、装 pandoc 工具链、注册保护钩子(Claude Code 原生钩子；Codex / OpenCode 用兼容适配层),你一条命令都不用自己敲。想手动装,INSTALL.md 里也逐条列了命令。
- Windows + PowerShell —— 当前支持的正式入口。默认摄取直接读 DOCX；桌面版 Microsoft Word 用于实际排版验收、章节/页码域计算和 Zotero Refresh，旧摄取流程也依赖 Word。暂不承诺完整跨平台运行。
- Python 3.10+（建议 3.12）+ lxml / Pillow，依赖安装见 INSTALL.md。PowerShell 5.1 即可。
- AI 编程工具 —— 为 Claude Code 深度打造，同时兼容 Codex / OpenCode / Hermes Agent 等 Agent Skills 工具。Claude 使用成熟原生钩子，Codex / OpenCode 可安装兼容保护层；所有环境仍由脚本内置闸门 + AGENTS.md 承担第一层安全，详见 INSTALL.md。
- Zotero + Word 插件 —— 用于后续管理引用。默认
rebuild离线保留已有活域，不要求打开 Zotero 或先对账；新引用组使用live和真实 Zotero key，查询新条目需要本机 Zotero + Better BibTeX，生成后在 Word 中 Refresh。 - 能执行本地脚本的 AI 工具。 格式读取、验证和重建由代码完成，不靠模型记住数百处排版修改。改稿和全文审计仍需要足够的上下文，模型能否遵守流程要实际验证。
- pandoc 工具链 —— 由
setup_md_tools.ps1自动安装:从官方发布页下载锁定版 pandoc 3.9.0.2 + crossref 0.3.24a(两者必须配套,脚本自动搞定,你什么都不用下载/上传)。⚠️ 国内网络往往下载不了 GitHub 上的 pandoc,可能需要开代理 / VPN;也可以加-Mirror https://<镜像>走镜像。
五阶段流水线。永远从 md-unpack 开始、md-build 结束,中间看你改多大。
flowchart LR
A["📄 原稿.docx"] --> B["📥 md-unpack<br/>摄取"]
B --> C["📝 manuscript.md<br/>（真源）"]
C --> D{"改稿规模？"}
D -->|"小修"| E["✏️ 手改"]
D -->|"AI 改一两处"| K["🎯 md-iterate"]
D -->|"一大堆乱意见"| T["🗂️ md-triage<br/>整理→你确认"]
T --> F["🐝 md-swarm<br/>AI 并行改稿"]
F -.->|"可选:继续用 md-iterate 微调"| K
E --> G["📤 md-build<br/>编译"]
K --> G
F --> G
G --> H["📗 Word .docx"]
顺序是活的——除了「md-unpack 永远第一步、md-build 永远最后一步」,中间的技能你可以按自己的需求自由组合、反复使用。三种典型顺序:
-
小修:
md-unpack→ 手改manuscript.md→md-build -
单点 AI 润色:
md-unpack→md-iterate→md-build -
大改:
md-unpack→md-triage→md-swarm→ (可选:再用md-iterate微调几处) →md-build -
md-unpack—— 摄取。 Word.docx→manuscript.md(Markdown 真源)。永远第一步。💡 Markdown 提供可读、可修改、可对比的正文；格式档案保存不适合交给模型逐项重写的 Word 属性。两者共同参与构建。
-
md-triage—— 整理。 一堆修订意图 → 你审核确认的清单。(仅大改。) -
md-swarm—— 批量改。 多 AI agent 并行、引用安全地落地清单。(仅大改。) -
md-iterate—— 润色单段。(小修。) -
md-build—— 出稿,编回 Word。永远最后一步。
一句话: md-unpack → (手改 · 或 md-iterate · 或 md-triage + md-swarm) → md-build。
📖 完整演练、可见格式标记、修改范围、模式选择和排错:见**用户完全手册**。
- 格式保留有边界。 支持原格式恢复、现有普通图片替换、表格文字修改及单个题注/章节引用；复杂对象通常保留原生部件。增删表格行列、任意合并、SmartArt 自由编辑、复杂成组引用等不在当前范围。
- 曾观察到 Codex Desktop 的 Hook 调度回归（2026-07-28） —— 部分 Codex Desktop / 非交互 CLI 版本会显示 Hook 已 Active/Trusted，却不真正执行处理器（openai/codex#21639）。在 Codex 中运行写入型
md-swarm/md-iterate前，必须亲眼看到 current-session live-probe 两个动作都 DENY；否则切换到已通过验证的宿主，当前推荐 Claude Code。 - 改字可能改变分页，字体和 Word 版本也会影响版面。static 的开发验收仍有两处已知排版差异，见用户完全手册；不承诺任意稿件逐像素一致。
📝 每个版本改了什么:见 更新日志 CHANGELOG.md。
工作流代码 Apache-2.0 © 2026 潘王雨昂 (Yuang Panwang)。第三方:pandoc 与 pandoc-crossref(GPL-2.0)安装时下载、不随仓库分发;内置 Zotero/Lua 过滤器为 MIT——见 NOTICE。
欢迎关注微信公众号:计算公共治理 · 人类有趣行为实验室
Pain 1 — The web-AI "chat → copy → paste" death loop. Revising a paper by chatting with ChatGPT / Claude in the browser looks like it saves thinking, but it traps you in worse manual labor: every paragraph is a copy-paste round — "make this more academic" → wait → select → Ctrl+C → back to Word → Ctrl+V → formatting broke → fix it → next paragraph… An afternoon gone, under 30 minutes of real thinking.
✅ md-paper: drop the comments in, AI applies them one by one to the Markdown source. You think once, at the approval gate; the grunt work is gone.
Pain 2 — One comment at a time; no batch, no multi-agent. Chatting with a browser AI, you can only say one thing at a time: "make the intro more concise." Even if you fully understand 30 reviewer comments, typing each into a prompt and pasting each result back already burns a full day — parallel multi-agent revision is simply impossible.
✅ md-paper: Swarm mode — drop in 30 comments, AI classifies/dedups/merges them (
md-triage) → you approve once → parallel agents draft, a deterministic script lands them serially into the source (md-swarm). From "one comment at a time" to "one whole review round at a time."
Pain 3 — Figures, tables, captions, notes, Zotero fields all scramble the instant AI touches Word. A Word manuscript embeds rich academic infrastructure: figure captions, table captions and notes, Zotero citation fields, cross-reference bookmarks. Paste browser-AI output back into Word and it's all lost or scrambled — and one Zotero Refresh resets every citation you carefully adjusted.
✅ md-paper preserves captured Word formatting and original live Zotero fields by default. Captions and bound individual references stay linked. Code validates format/object bindings; each manuscript gets capture and build reports describing its supported and uneditable parts.
Pain 4 — Vague comments take forever to execute. "Change first person to objective third person," "reduce grammatical slips," "broaden the literature review" — easy to understand, brutally slow to do: read the whole paper for every I/we/our and replace each; check every sentence's grammar; hunt down missing references. One reviewer sentence, one full day of work.
✅ md-paper: de-AI rules + language rules are baked into the agent contract. A vague comment becomes an executable task — "remove all first person" is a single swarm job, done in minutes; you just approve.
Pain 5 — Citations vanish silently mid-revision, discovered too late.
A real, measured failure type from development: while revising, AI quietly drops a citation it deems unimportant, splits a group [@a; @b; @c], or mistypes one letter of a citekey — and by the time you notice, several rounds have passed and it's untraceable.
✅ md-paper: a never-drop-a-citation hard gate — in patch mode, losing one citation is hard-rejected; a ref-check lists every dropped citekey, split group, and dangling cross-reference after each revision. Figures, tables, and (numbered) equations are counted every batch too.
Pain 6 — The revised paper reads like a machine wrote it. Heavily AI-polished prose has an "AI smell" — symmetric sentences, piled-up rhetoric, em-dashes everywhere. Reviewers increasingly spot machine-generated text, and that's fatal for a submission.
✅ md-paper: seven de-AI rules baked into the flow — less rhetoric, short sentences, break the symmetry, lower the formality, keep the information, fewer em-dashes, cover everything. Change only how it's said, never what.
Pain 7 - AI revision with no review; errors hide until too late. The real risk of AI revision is not bad writing - it is wrong writing you never notice: a citation silently dropped, a passage whose meaning got twisted, terminology that no longer matches. By the time you catch it at final review, several rounds have passed. A browser AI hands you a result and you either accept it or paste it back and ask again - there is no "review" step, no one watching each change.
✅ md-paper: three-province separation + multi-pass auditing. Borrowing the Tang-dynasty Three Departments (draft / review / execute), revision is split into three non-overlapping roles: propose (organize intent) / execute (draft) / audit (read-only review). The whole run goes through multiple checkpoints: every batch auto-checks whether citations were dropped, fabricated, or actually landed; once all edits are done, two read-only final audits run - D audits each comment ("done? done right?"), E reads the full text for internal contradictions. AI only drafts and reports; the final call is always yours.
| Editing Word by hand | md-paper (Markdown source + AI + pandoc) |
|---|---|
| 😫 Revise every reviewer comment one by one | ✅ Swarm: drop the whole pile in, AI revises in batch |
| 😫 One Zotero Refresh wrecks your citations | ✅ Live Zotero fields survive — Refresh anytime |
| 😫 Figure/table numbers scramble after edits | ✅ Linked captions and bound references, with missing-target checks |
| 😫 Word formatting breaks; Ctrl+Z is your only hope | ✅ Plain-text Markdown source — git-versioned |
| 😫 You discover a lost citation three rounds too late | ✅ Never-drop-a-citation hard gate |
| 😫 Endless copy-paste between browser and Word | ✅ Everything in the editor — zero pasting |
| 😫 The result reeks of "AI writing" | ✅ Seven de-AI rules lower the machine-written signal |
| 😫 AI revises with no review; errors hide | ✅ Three-province separation + multi-pass audits, two read-only final reviews |
🔗 Zotero-native live citation fields · 🖼️ figures/tables/captions/notes as plain text · 🔀 auto-numbered cross-references · 🐝 parallel swarm revision · 🛡️ never-drop-a-citation guard · 📝 git-diffable Markdown source · 🔄 de-AI humanizer · 🧮 OMML→LaTeX equations · 🧰 one shared toolchain · 🧩 open Agent Skills standard — works in Claude Code / Codex / OpenCode / Hermes · 🏛️ three-province separation (propose/execute/audit) + multi-pass auditing.
The default workflow is now original Word → current Markdown plus a captured format profile → Word. Code reads actual styles, local formatting, sections, figures/tables, and footnotes/endnotes; it does not guess semantic “title pages” or “abstract pages.” Visible custom-style="正文" / custom-style="body text" markers share a per-manuscript bilingual map in format_names.md.
| Mode | Purpose |
|---|---|
rebuild (default) |
Retain captured formatting and existing live Zotero fields offline |
live |
Create citation groups using real Zotero keys; Refresh in Word afterward |
static |
Remove Zotero field controls while retaining display text for proofing; layout may change |
AI edits pass format validation before the original single writer applies them. Human edits must also retain format/object markers. Move the whole paper project, including manifest/ and media/; Markdown alone cannot restore source formatting. Old projects without a format profile remain supported; explicitly choose -Layout legacy for the old conversion path.
Our design philosophy is to minimize, to the greatest extent, the many unnecessary manual chores between AI and a finished paper - but never to have AI replace the author. The md-paper pipeline is long - ingest, organize, draft, audit, compile - and every stage has a human gate: comments land only after confirmation, edits pass only after review, citation changes need authorization. Walk the whole pipeline and the manuscript is still fundamentally the author's own work. What AI removes is the typing, pasting, format-fixing, and issuing commands one by one; every place that needs the author's thinking is still there: how to read the reviewer comments, how far to go, which citations to keep or cut, whether the final draft is final - all the user's own decisions.
Our stance: AI is only an aid - it makes revision more convenient, but it cannot replace the author's own thinking and originality. The tool runs errands; it does not think for the author.
Step 1 — Get this repo onto your computer. Either way works:
- No command line (easiest): click the green
< > Codebutton at the top of this page → Download ZIP → unzip it to a folder you'll remember. - If you use git:
git clone https://github.com/pwya/md-paper.git
Step 2 — Open the folder in VS Code. In VS Code: File → Open Folder → pick the md-paper folder you just unzipped/cloned. Then start a Claude Code session inside it (or Codex / OpenCode / Hermes Agent — all supported).
Step 3 — Let an AI install it. Tell your AI:
"Read
INSTALL.mdand set up md-paper for me."
The AI follows INSTALL.md — an executable runbook — to link the five skills into your AI tool, install the pandoc toolchain, and register protection hooks (native Claude hooks plus Codex/OpenCode adapters). You don't type a single command yourself. Prefer manual? INSTALL.md lists every command.
- Windows + PowerShell are the supported entrypoint environment. Default capture reads DOCX directly. Desktop Microsoft Word is needed for visual validation, heading/page field calculation, Zotero Refresh, and legacy ingest. Full cross-platform operation is not promised.
- Python 3.10+ (3.12 recommended), lxml / Pillow, and PowerShell 5.1 or later. See INSTALL.md for dependency installation.
- AI coding tool — built first for Claude Code and compatible with Codex / OpenCode / Hermes Agent. Claude uses its native hooks; Codex and OpenCode can install compatible adapters. Script gates plus AGENTS.md remain the authoritative first layer everywhere; see INSTALL.md.
- Zotero + its Word plugin manage citations after export. Default
rebuildpreserves existing live fields offline without reconciliation. New citation groups useliveand real keys; looking up new items requires local Zotero + Better BibTeX, followed by Refresh in Word. - An AI tool that can execute local scripts. Code captures, validates, and rebuilds formatting without requiring the model to remember hundreds of layout edits. Drafting and full-text audits still require adequate context and verified workflow compliance.
- pandoc toolchain — installed automatically by
setup_md_tools.ps1, which downloads the pinned pandoc 3.9.0.2 + pandoc-crossref 0.3.24a from the official releases (the two versions must match — the installer handles it; you download nothing).⚠️ If GitHub downloads fail (common in restricted networks), you may need a proxy / VPN; or pass-Mirror https://<a-github-mirror>.
Five-stage pipeline. Always start with md-unpack, finish with md-build; the middle depends on how much you're changing.
flowchart LR
A["📄 paper.docx"] --> B["📥 md-unpack<br/>ingest"]
B --> C["📝 manuscript.md<br/>(source of truth)"]
C --> D{"how big is<br/>the revision?"}
D -->|"tiny fix"| E["✏️ hand-edit"]
D -->|"one passage, with AI"| K["🎯 md-iterate"]
D -->|"a pile of comments"| T["🗂️ md-triage<br/>organize → you approve"]
T --> F["🐝 md-swarm<br/>parallel AI revision"]
F -.->|"optional: fine-tune with md-iterate"| K
E --> G["📤 md-build<br/>compile"]
K --> G
F --> G
G --> H["📗 Word .docx"]
The order is flexible — apart from "md-unpack always first, md-build always last", you can combine and repeat the middle skills however your revision demands. Three typical orders:
-
Tiny fix:
md-unpack→ hand-editmanuscript.md→md-build -
Single-passage AI polish:
md-unpack→md-iterate→md-build -
Major revision:
md-unpack→md-triage→md-swarm→ (optional: fine-tune a few spots withmd-iterate) →md-build -
md-unpack— ingest. Word.docx→manuscript.md(Markdown source of truth). Run first.💡 Markdown keeps prose readable and diffable; the captured profile retains Word properties that should not depend on model-generated formatting instructions. Both are required to build the document.
-
md-triage— organize. A pile of revision intents → a checklist you approve. (Big revisions only.) -
md-swarm— batch-revise. Parallel AI agents apply the approved checklist, citation-safe. (Big revisions only.) -
md-iterate— polish one passage with AI. (Small edits.) -
md-build— compile back to Word. Run last.
In one line: md-unpack → (hand-edit · or md-iterate · or md-triage + md-swarm) → md-build.
📖 Full walkthrough, format markers, supported edits, build modes, and troubleshooting: User Guide (中文).
- Formatting preservation has explicit limits. Existing ordinary pictures, table text, and individual caption/heading references are supported. Complex objects may be retained as native parts; arbitrary table row/column changes, merges, SmartArt editing, and grouped/range references are outside the current scope.
- A Codex Desktop Hook-dispatch regression was observed on 2026-07-28 — some Codex Desktop / non-interactive CLI builds report Hooks as Active/Trusted without invoking the handler (openai/codex#21639). Before any write-capable
md-swarm/md-iteraterun in Codex, require both current-session live-probe actions to DENY; otherwise switch to a verified host, currently Claude Code. - Changing text can change pagination. Fonts and Word versions matter. Static mode still has two known rendering differences in development validation; see the User Guide. No pixel-perfect guarantee is made for arbitrary manuscripts.
📝 What changed in each version: see the CHANGELOG.
Workflow code Apache-2.0 © 2026 Yuang Panwang (潘王雨昂). Third-party: pandoc & pandoc-crossref (GPL-2.0) are downloaded at setup, not redistributed here; bundled Zotero/Lua filters are MIT — see NOTICE.
Yuang Panwang 潘王雨昂 · Yu Zhang 张宇
WeChat official accounts: 计算公共治理 (Computational Public Governance) · 人类有趣行为实验室 (Fun Human Behavior Lab)