X Article Publisher

X Article Publisher

Publish Markdown content to X (Twitter) Articles editor, preserving formatting with rich text conversion.

Prerequisites

• Playwright MCP for browser automation
• User logged into X with Premium Plus subscription
• Dedicated persistent browser profile (recommended) to avoid repeated login
• Python 3.9+ with dependencies:
  • macOS: pip install Pillow pyobjc-framework-Cocoa
  • Windows: pip install Pillow pywin32 clip-util
• For Mermaid diagrams: npm install -g @mermaid-js/mermaid-cli

Scripts

Located in ~/.codex/skills/x-article-publisher/scripts/:

parse_markdown.py

Parse Markdown and extract structured data:

python parse_markdown.py <markdown_file> [--output json|html] [--html-only]

Returns JSON with: title, cover_image, content_media, content_images, content_videos, dividers (with block_index for positioning), html, total_blocks

copy_to_clipboard.py

Copy image or HTML to system clipboard (cross-platform):

# Copy image (with optional compression)
python copy_to_clipboard.py image /path/to/image.jpg [--quality 80]

# Copy HTML for rich text paste
python copy_to_clipboard.py html --file /path/to/content.html

table_to_image.py

Convert Markdown table to PNG image:

python table_to_image.py <input.md> <output.png> [--scale 2]

Use when X Articles doesn't support native table rendering or for consistent styling.

open_x_articles_browser.sh

Open X Articles with a dedicated persistent profile:

bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

Defaults:

• Profile path: ~/.codex/browser-profiles/x-articles
• Can override with env var: X_ARTICLES_PROFILE=/custom/path

prepare_article_source.py

Auto-route source input:

• Feishu/Lark URL -> download to local markdown (with video fetch)
• Local markdown path -> pass through directly

python ~/.codex/skills/x-article-publisher/scripts/prepare_article_source.py "<source>"

Output JSON includes:

• mode: feishu_url or local_markdown
• markdown_path: local markdown path to publish
• videos_downloaded / videos_appended: video handling summary
• callouts_normalized: number of Feishu callout labels such as Tip or [!TIP] removed while preserving quoted content

optimize_media_blocks.py

Reduce body media block count before upload by merging adjacent image runs into adaptive no-crop grid images:

python ~/.codex/skills/x-article-publisher/scripts/optimize_media_blocks.py article.md \
  --max-body-media 24 \
  --output article.optimized.md

Use this when parsed content_media is near or above the practical X Articles body-media limit. It preserves videos as separate upload blocks and merges only adjacent body images, leaving the first image/cover alone. Collages use a balanced grid with contain scaling: every source screenshot remains complete, without crop or cover behavior. The default final width is 1920px so text-heavy screenshots remain readable.

Persistent Profile (Recommended)

Always launch X Articles with the dedicated profile:

bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

Why:

• Avoids repeated login and lowers account risk from frequent sign-ins
• Isolated from your main Google Chrome profile
• Does not overwrite or log out your existing Chrome account session

Note:

• First run may still trigger one-time X security verification (new device/profile)
• After that, reuse the same profile path and login stays stable

Pre-Processing (Optional)

Before publishing, scan the Markdown for elements that need conversion:

Body Media Budget

X Articles can silently drop new body-media uploads after roughly 25 body media blocks. The file picker may accept a file while the editor never enters Uploading media... and the media count never increases.

Before browser upload, parse media count:

python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py article.md > article.json
jq '.content_media | length' article.json

If body media count is >= 24, run optimize_media_blocks.py and re-parse the optimized Markdown:

python ~/.codex/skills/x-article-publisher/scripts/optimize_media_blocks.py article.md \
  --max-body-media 24 \
  --output article.optimized.md > optimize.json
python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py article.optimized.md > article.json

Prefer merging adjacent image runs before upload. Do not merge videos; videos should keep their own block and original anchor. Reject or regenerate any collage that becomes an extreme vertical strip: saving media blocks never justifies clipped or unreadable screenshots in X Preview.

Feishu Callouts / Highlight Blocks

Feishu highlighted blocks may export to Markdown as blockquotes prefixed with Tip, Note, or [!TIP]. These labels are not useful in X Articles and can break reading flow. prepare_article_source.py removes those marker lines but keeps the quoted multiline content intact. If callouts_normalized > 0, this cleanup happened.

Video Upload Preflight

Before uploading video-heavy articles, inspect video size and bitrate. Large 40-90MB/high-bitrate videos can close or destabilize the browser session even when X eventually accepts smaller files.

When a video is large or upload is unstable, transcode to a 1280px-wide H.264/AAC upload copy:

ffmpeg -y -i input.mov \
  -vf "scale='min(1280,iw)':-2" \
  -c:v libx264 -preset medium -crf 25 -pix_fmt yuv420p \
  -c:a aac -b:a 64k -movflags +faststart \
  output.x1280.mp4

Use the transcoded file for X upload, but keep the original Markdown anchor and block order. The -2 height is required: it rounds the calculated height to an even number so libx264 does not fail on outputs such as 1280x713. Run ffprobe on every output before replacing the Markdown path; a zero-byte or unreadable copy must never enter the upload queue.

GIF Upload Preflight

Treat large GIF files as video-like media, not static screenshots. Clipboard image paste can turn a GIF into a still image, and X can silently accept a large GIF file input while never adding a media block.

If a GIF is large, or if uploading it does not increase the editor media count, convert it to MP4 and insert the MP4 at the original GIF anchor:

ffmpeg -y -i input.gif \
  -vf "scale='if(gt(iw,ih),min(1280,iw),-2)':'if(gt(iw,ih),-2,min(1280,ih))',format=yuv420p" \
  -movflags +faststart -c:v libx264 -preset medium -crf 25 -an \
  input.gif.mp4

In final audit, count this converted GIF as one video-like media block at the original GIF position.

Tables → PNG

# Extract table to temp file, then convert
python ~/.codex/skills/x-article-publisher/scripts/table_to_image.py /tmp/table.md /tmp/table.png
# Replace table in markdown with: ![Table](/tmp/table.png)

Mermaid Diagrams → PNG

# Extract mermaid block to .mmd file, then convert
mmdc -i /tmp/diagram.mmd -o /tmp/diagram.png -b white -s 2
# Replace mermaid block with: ![Diagram](/tmp/diagram.png)

Dividers (---)

Dividers are automatically detected by parse_markdown.py and output in the dividers array. They must be inserted via X Articles' Insert > Divider menu (HTML <hr> tags are ignored by X).

Workflow

Strategy: "先文后媒体后分割线" (Text First, Media Second, Dividers Last)

For articles with media (images/videos) and dividers, paste ALL text content first, then insert media and dividers at correct positions using block index.

1. Route input source (prepare_article_source.py)
2. Parse Markdown once and check body-media budget
3. (Optional) Pre-process: Convert tables/mermaid to images; if body media is near/above 24, run optimize_media_blocks.py and re-parse
4. Parse final Markdown with Python script → get title, media, dividers with block_index, HTML
5. Navigate to X Articles editor
6. Upload cover image (first image)
7. Fill title
8. Copy HTML to clipboard (Python) → Paste with Cmd+V
9. Insert content images at positions specified by block_index
10. Insert content videos at positions specified by block_index
11. Insert dividers at positions specified by block_index (via Insert > Divider menu)
12. Open Preview and audit media count plus anchor order (anchor text -> following media type)
13. Save as draft (NEVER auto-publish)

Input Routing

Two trigger modes are supported:

1. Feishu URL mode Input contains a Feishu/Lark doc link (feishu.cn / larksuite.com / feishu.sg)

   • Run prepare_article_source.py "<url>"
   • It calls feishu2md dl --dump to download markdown
   • It fetches video file blocks (if any) into local static/
   • It appends <video src="..."> entries into markdown for downstream upload

2. Local Markdown mode Input is a local .md / .markdown file path

   • Run prepare_article_source.py "<path/to/file.md>"
   • It returns the original file directly (no download step)

Why feishu2md Misses Videos by Default

feishu2md can download media by token, but its default parser flow is image-only:

• In cmd/download.go, it only iterates parser.ImgTokens and calls DownloadImage(...)
• In core/parser.go, it only handles DocxBlockTypeImage and collects ImgTokens
• DocxBlockTypeFile (where many embedded videos are represented) is not rendered to markdown nor downloaded

So videos often disappear in generated markdown unless an extra step is added. prepare_article_source.py is that extra step for this skill.

高效执行原则 (Efficiency Guidelines)

目标: 最小化操作之间的等待时间，实现流畅的自动化体验。

1. 避免不必要的 browser_snapshot

大多数浏览器操作（click, type, press_key 等）都会在返回结果中包含页面状态。不要在每次操作后单独调用 browser_snapshot，直接使用操作返回的页面状态即可。

❌ 错误做法：
browser_click → browser_snapshot → 分析 → browser_click → browser_snapshot → ...

✅ 正确做法：
browser_click → 从返回结果中获取页面状态 → browser_click → ...

2. 避免不必要的 browser_wait_for

只在以下情况使用 browser_wait_for：

• 等待图片上传完成（textGone="正在上传媒体"）
• 等待页面初始加载（极少数情况）

不要使用 browser_wait_for 来等待按钮或输入框出现 - 它们在页面加载完成后立即可用。

3. 并行执行独立操作

当两个操作没有依赖关系时，可以在同一个消息中并行调用多个工具：

✅ 可以并行：
- 填写标题 (browser_type) + 复制HTML到剪贴板 (Bash)
- 解析Markdown生成JSON + 生成HTML文件

❌ 不能并行（有依赖）：
- 必须先点击create才能上传封面图
- 必须先粘贴内容才能插入图片

4. 连续执行浏览器操作

每个浏览器操作返回的页面状态包含所有需要的元素引用。直接使用这些引用进行下一步操作：

# 理想流程（每步直接执行，不额外等待）：
browser_navigate → 从返回状态找create按钮 → browser_click(create)
→ 从返回状态找上传按钮 → browser_click(上传) → browser_file_upload
→ 从返回状态找应用按钮 → browser_click(应用)
→ 从返回状态找标题框 → browser_type(标题)
→ 点击编辑器 → browser_press_key(Meta+v)
→ ...

5. 准备工作前置

在开始浏览器操作之前，先完成所有准备工作：

1. 解析 Markdown 获取 JSON 数据
2. 生成 HTML 文件到 /tmp/
3. 记录 title、cover_image、content_media 等信息

这样浏览器操作阶段可以连续执行，不需要中途停下来处理数据。

Step 0: Resolve Input Source

Run source resolver first:

python ~/.codex/skills/x-article-publisher/scripts/prepare_article_source.py "<source>" > /tmp/x_source.json

Then read normalized markdown path:

jq -r '.markdown_path' /tmp/x_source.json

Step 1: Parse Markdown (Python)

Use parse_markdown.py to extract all structured data:

python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py /path/to/article.md

Output JSON:

{
  "title": "Article Title",
  "cover_image": "/path/to/first-image.jpg",
  "cover_exists": true,
  "content_media": [
    {"type": "image", "path": "/path/to/img2.jpg", "original_path": "/md/dir/assets/img2.jpg", "exists": true, "block_index": 5, "after_text": "context..."},
    {"type": "video", "path": "/path/to/video.mp4", "original_path": "/md/dir/assets/video.mp4", "exists": true, "block_index": 12, "after_text": "another..."}
  ],
  "content_images": [
    {"path": "/path/to/img2.jpg", "original_path": "/md/dir/assets/img2.jpg", "exists": true, "block_index": 5, "after_text": "context..."},
    {"path": "/path/to/img3.jpg", "original_path": "/md/dir/assets/img3.jpg", "exists": true, "block_index": 18, "after_text": "another..."}
  ],
  "content_videos": [
    {"path": "/path/to/video.mp4", "original_path": "/md/dir/assets/video.mp4", "exists": true, "block_index": 12, "after_text": "another..."}
  ],
  "html": "<p>Content...</p><h2>Section</h2>...",
  "total_blocks": 45,
  "missing_media": 0,
  "missing_images": 0,
  "missing_videos": 0
}

Key fields:

• block_index: The media should be inserted AFTER block element at this index (0-indexed)
• total_blocks: Total number of block elements in the HTML
• after_text: Kept for reference/debugging only, NOT for positioning
• exists: Whether the media file was found (if false, upload will fail)
• original_path: The path resolved from Markdown (before auto-search)
• path: The actual path to use (may differ from original_path if auto-searched)
• content_media: Combined list (images + videos), preferred for unified workflows
• missing_media: Count of media files not found anywhere

Save HTML to temp file for clipboard:

python parse_markdown.py article.md --html-only > /tmp/article_html.html

Step 2: Open X Articles Editor

浏览器错误处理

如果遇到 Error: Browser is already in use 错误：

# 方案1：关闭旧会话后，用持久化 profile 重新打开（推荐）
playwright-cli close-all
bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

# 方案2：如果 browser_close 无效（锁定），提示用户手动关闭 Chrome

# 方案3：使用已有标签页，直接导航到文章页
browser_tabs action=list  # 查看现有标签
browser_navigate: https://x.com/compose/articles  # 在当前标签导航

最佳实践：每次开始前先用 browser_tabs action=list 检查状态，避免创建多余空白标签。

导航到编辑器

# 先确保使用持久化 profile 打开（不要用临时会话）
bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

重要: 页面加载后会显示草稿列表，不是编辑器。需要：

1. 等待页面加载完成: 使用 browser_snapshot 检查页面状态
2. 立即点击 "create" 按钮: 不要等待 "添加标题" 等编辑器元素，它们只有点击 create 后才出现
3. 等待编辑器加载: 点击 create 后，等待编辑器元素出现

# 1. 使用持久化 profile 打开页面
bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

# 2. 获取页面快照，找到 create 按钮
browser_snapshot

# 3. 点击 create 按钮（通常 ref 类似 "create" 或带有 create 标签）
browser_click: element="create button", ref=<create_button_ref>

# 4. 现在编辑器应该打开了，可以继续上传封面图等操作

注意: 不要使用 browser_wait_for text="添加标题" 来等待页面加载，因为这个文本只有在点击 create 后才出现，会导致超时。

If login is needed on first run, prompt user to log in manually once in the persistent profile.

Step 3: Upload Cover Image

1. Click "添加照片或视频" button
2. Use browser_file_upload with the cover image path (from JSON output)
3. Verify image uploaded

Step 4: Fill Title

• Find textbox with "添加标题" placeholder
• Use browser_type to input title (from JSON output)

Step 5: Paste Text Content (Python Clipboard)

Copy HTML to system clipboard using Python, then paste:

# Copy HTML to clipboard
python ~/.codex/skills/x-article-publisher/scripts/copy_to_clipboard.py html --file /tmp/article_html.html

Then in browser:

browser_click on editor textbox
browser_press_key: Meta+v

This preserves all rich text formatting (H2, bold, links, lists).

Step 6: Insert Content Images (Text Search Positioning)

推荐方法: 使用 after_text 文字搜索定位，比 block_index 更直观可靠。

定位原理

每张图片的 after_text 字段记录了它前一个段落的末尾文字（最多80字符）。在编辑器中搜索包含该文字的段落，点击后插入图片。

操作步骤

For each content image (from content_images array), 按 block_index 从大到小的顺序：

# 1. Copy image to clipboard (with compression)
python ~/.codex/skills/x-article-publisher/scripts/copy_to_clipboard.py image /path/to/img.jpg --quality 85

# 2. 在 browser_snapshot 中搜索包含 after_text 的段落
#    找到该段落的 ref

# 3. Click the paragraph containing after_text
browser_click: element="paragraph with target text", ref=<paragraph_ref>

# 4. **关键步骤**: 按 End 键移动光标到行尾
#    这一步非常重要！避免点击到段落中的链接导致位置偏移
browser_press_key: End

# 5. Paste image
browser_press_key: Meta+v

# 6. Wait for upload (only use textGone, no time parameter)
browser_wait_for textGone="正在上传媒体"

为什么需要按 End 键？

问题: 当段落包含链接时（如 [链接文字](url)），点击段落可能会：

• 触发链接编辑弹窗
• 将光标定位在链接内部而非段落末尾

解决方案: 点击段落后立即按 End 键：

• 确保光标移动到段落末尾
• 避免链接干扰
• 图片将正确插入在该段落之后

定位策略

在 browser_snapshot 返回的结构中，搜索 after_text 的关键词：

textbox [ref=editor]:
  generic [ref=p1]:
    - StaticText: "元旦假期我在家里翻手机相册..."  # 如果 after_text 包含这段文字，点击 p1
  heading [ref=h1]:
    - StaticText: "演示"
  generic [ref=p2]:
    - StaticText: "这东西到底有多省事儿？"
    - link [ref=link1]: "Claude Code"  # 注意：段落可能包含链接
  ...

反向插入示例

如果有3张图片，block_index 分别为 5, 12, 27：

1. 先插入 block_index=27 的图片（after_text 搜索 + End + 粘贴）
2. 再插入 block_index=12 的图片
3. 最后插入 block_index=5 的图片

从大到小插入可以避免位置偏移问题。

Rich Text Anchor Matching

X can split a Markdown sentence across multiple editor text nodes when the source contains bold spans, links, or mixed Chinese/English emphasis. A full after_text string may not exist as one DOM text node.

Fallback matching order:

1. Try the normalized full anchor with Markdown markers removed (**, heading marks, blockquote marks).
2. Strip parser-only code markers (___CODE_BLOCK_START___ and ___CODE_BLOCK_END___); they are never visible in X.
3. Strip a leading Markdown/DOM list marker (-, *, +, or •) before matching because X does not include it in the rendered list-item text.
4. Try progressively shorter prefixes.
5. For split rich text, use a distinctive short prefix such as the opening clause, but only if it is unique enough in the article.
6. In Preview, re-check with a longer exact nearby phrase whenever a short anchor such as 比方说 could match an earlier paragraph.

Step 6.2: Insert Content Videos (File Upload)

视频不走剪贴板，使用文件上传（与封面相同的 file chooser 机制）。

For each content video (from content_videos array), 按 block_index 从大到小的顺序。

Use file upload, never clipboard paste. Prefer writing the hidden input[type=file] after opening Insert > Media; filechooser events can be flaky on X Articles. For video-heavy drafts, upload one video per browser pass if the editor becomes unstable, then reopen the same draft URL and continue from the next missing anchor.

# 1. 在 browser_snapshot 中搜索包含 after_text 的段落并点击
browser_click: element="paragraph with target text", ref=<paragraph_ref>

# 2. 将光标移动到段落末尾
browser_press_key: End

# 3. 点击工具栏 Insert / Add Media
browser_click: element="Add Media button", ref=<add_media_ref>

# 4. 上传视频文件
browser_file_upload: paths=["/path/to/video.mp4"]

# 5. 等待上传完成
browser_wait_for textGone="正在上传媒体"

Observed X behavior:

• Video upload can create a full-page overlay that intercepts later clicks.
• Starting the next upload too early can skip media or place it under a later anchor.
• Large videos should be transcoded before upload.
• Editor-side remove-media counts can drift during long runs; final preview DOM is the source of truth.
• For adjacent mixed-media clusters (for example: video, then the next paragraph has another video plus an image), a correct count can still hide swapped order. Audit the local sequence, not only the first media after each anchor.
• If a local cluster is wrong, delete only that cluster in the editor, then reinsert it from bottom to top. Example: for anchor A -> video1 followed by anchor B -> video2 + image, insert image at B, then video2 at B, then video1 at A.
• X failed uploads can leave visible blocks saying Something went wrong. Please try again later. with a Delete block button. Delete those blocks before final Preview audit.

Step 6.5: Insert Dividers (Via Menu)

重要: Markdown 中的 --- 分割线不能通过 HTML <hr> 标签粘贴（X Articles 会忽略它）。必须通过 X Articles 的 Insert 菜单插入。

操作步骤

For each divider (from dividers array), in reverse order of block_index:

# 1. Click the block element at block_index position
browser_click on the element at position block_index in the editor

# 2. Open Insert menu (Add Media button)
browser_click on "Insert" or "添加媒体" button

# 3. Click Divider menu item
browser_click on "Divider" or "分割线" menuitem

# Divider is inserted at cursor position

与图片的插入顺序

建议先插入所有图片，再插入所有视频，最后插入分割线。三者都按 block_index 从大到小的顺序：

1. 插入所有图片（从最大 block_index 开始）
2. 插入所有视频（从最大 block_index 开始）
3. 插入所有分割线（从最大 block_index 开始）

Step 6.8: Final Preview Audit

Before reporting success, open Preview and verify both counts and order:

1. img[alt="Image"] count matches body images.
2. video[aria-label="Embedded video"] or Play Video count matches body videos.
3. Cover image exists separately from body images.
4. For each content_media item, find its after_text anchor in Preview and confirm the next visible media item has the expected type.
5. For same-anchor media runs, confirm the next N visible media items preserve the Markdown order.
6. For adjacent video/image clusters, confirm the local sequence around all involved anchors, not only the nearest media.
7. Confirm Preview does not contain Something went wrong. Please try again later..

If tag-based anchor lookup reports missing anchors even though counts are correct, use a geometry fallback. X Preview often renders ordinary paragraphs as div or span instead of p:

1. Search visible div, span, p, li, blockquote, h1, h2, h3 elements using the same normalized anchor prefixes.
2. Choose the matching visible node with the shortest normalized text; this avoids selecting an outer article container.
3. Read its bottom edge with getBoundingClientRect().bottom.
4. Sort body media by getBoundingClientRect().y, counting only img[alt="Image"] and visible video elements.
5. Confirm the first media below the anchor bottom has the expected type. For a same-anchor run, compare the next N media types with the Markdown order.

If one media item is under the wrong anchor, do not rebuild the entire draft:

1. Delete only the misplaced media block in the editor.
2. Insert the missing media at its original anchor.
3. Reinsert the misplaced video/image at its own anchor.
4. Re-run the Preview audit.

If a cluster is wrong:

1. Delete the smallest contiguous cluster that contains the misplaced items.
2. Reinsert from the latest anchor back to the earliest anchor.
3. For multiple media under the same anchor, insert in reverse original order so the final visual order matches the Markdown.
4. Re-run Preview and inspect nextThree/local media sequence around the repaired anchors.

Step 7: Save Draft

1. Verify content pasted (check word count indicator)
2. Draft auto-saves, or click Save button if needed
3. Click "预览" to verify formatting
4. Report: "Draft saved. Review and publish manually."

Critical Rules

1. NEVER publish - Only save draft
2. First image = cover - Upload first image as cover image
3. Rich text conversion - Always convert Markdown to HTML before pasting
4. Use clipboard API - Paste via clipboard for proper formatting
5. Block index positioning - Use block_index for precise image/video/divider placement
6. Reverse order insertion - Insert images/videos/dividers from highest to lowest block_index
7. H1 title handling - H1 is used as title only, not included in body
8. Dividers via menu - Markdown --- must be inserted via Insert > Divider menu (HTML <hr> is ignored)
9. Video upload method - Videos must use file upload, not clipboard paste
10. Session stability - Always open X with dedicated persistent profile, never temporary profile
11. Media budget - Keep body media at or below 24 when possible; X may silently ignore uploads after roughly 25 body media blocks
12. Verify by position, not count only - Final audit must check media count and key anchors (anchor text -> following media type/order), because a correct video count can still hide one misplaced video
13. Callout cleanup - Remove Feishu callout markers (Tip, [!TIP]) while preserving the highlighted text and line breaks
14. Video preflight - Transcode oversized/high-bitrate videos to 1280px H.264/AAC before upload when stability matters
15. Resume instead of restarting - If browser/session drift happens mid-upload, reopen the existing draft URL and continue from missing anchors; do not paste the body again
16. GIF fallback - Upload GIFs through file input when preserving animation matters; if X silently ignores a large GIF, convert it to MP4 and upload that at the same anchor
17. Adjacent cluster repair - For nearby video/image groups, verify local order in Preview; repair by deleting only that small cluster and reinserting from bottom to top
18. Failed block cleanup - Delete Something went wrong / Delete block placeholders before reporting a draft ready

Supported Formatting

Example Flow

User: "Publish /path/to/article.md to X"

# Step 0: Resolve source (URL or local markdown)
python ~/.codex/skills/x-article-publisher/scripts/prepare_article_source.py "/path/to/article.md_or_feishu_url" > /tmp/x_source.json
MD_PATH="$(jq -r '.markdown_path' /tmp/x_source.json)"

# Step 1: Parse Markdown
python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py "$MD_PATH" > /tmp/article.json
python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py "$MD_PATH" --html-only > /tmp/article_html.html

If /tmp/article.json has too many body media blocks, optimize and re-parse before opening X:

if [ "$(jq '.content_media | length' /tmp/article.json)" -ge 24 ]; then
  python ~/.codex/skills/x-article-publisher/scripts/optimize_media_blocks.py "$MD_PATH" \
    --max-body-media 24 \
    --output /tmp/article.optimized.md > /tmp/article_optimize.json
  MD_PATH=/tmp/article.optimized.md
  python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py "$MD_PATH" > /tmp/article.json
  python ~/.codex/skills/x-article-publisher/scripts/parse_markdown.py "$MD_PATH" --html-only > /tmp/article_html.html
fi

2. Navigate to https://x.com/compose/articles
3. Upload cover image (browser_file_upload for cover only)
4. Fill title (from JSON: title)
5. Copy & paste HTML:

   python ~/.codex/skills/x-article-publisher/scripts/copy_to_clipboard.py html --file /tmp/article_html.html
   

   Then: browser_press_key Meta+v
6. For each content image, in reverse order of block_index:

   python copy_to_clipboard.py image /path/to/img.jpg --quality 85
   

   • Click block element at block_index position
   • browser_press_key Meta+v
   • Wait until upload complete
7. For each content video, in reverse order of block_index:
   • Click block element at block_index position
   • Click Insert/Add Media
   • browser_file_upload with /path/to/video.mp4
   • Wait until upload complete
8. Verify in preview
9. "Draft saved. Please review and publish manually."

Best Practices

为什么用 block_index 而非文字匹配？

1. 精确定位: 不依赖文字内容，即使多处文字相似也能正确定位
2. 可靠性: 索引是确定性的，不会因为文字相似而混淆
3. 调试方便: after_text 仍保留用于人工核验

为什么用 Python 而非浏览器内 JavaScript？

1. 本地处理更可靠: Python 直接操作系统剪贴板，不受浏览器沙盒限制
2. 图片压缩: 上传前压缩图片 (--quality 85)，减少上传时间
3. 代码复用: 脚本固定不变，无需每次重新编写转换逻辑
4. 调试方便: 脚本可单独测试，问题易定位

等待策略

重要发现: Playwright MCP 的 browser_wait_for 实际行为是 先等待 time 秒，再检查条件，而非轮询！

// 实际执行的代码：
await new Promise(f => setTimeout(f, time * 1000));  // 先固定等待
await page.getByText("xxx").waitFor({ state: 'hidden' });  // 再检查

正确用法:

• ✅ 只用 textGone，不设 time：让 Playwright 自己轮询等待
• ✅ 只用 time：固定等待指定秒数
• ❌ 同时用 textGone + time：会先等 time 秒再检查，浪费时间

# 推荐：只用 textGone，让它自动等待条件满足
browser_wait_for textGone="正在上传媒体"

# 或者：用 browser_snapshot 轮询检查状态
# 每次操作后检查返回的页面状态，无需额外等待

图片插入效率

每张图片的浏览器操作从5步减少到2步：

• 旧: 点击 → 添加媒体 → 媒体 → 添加照片 → file_upload
• 新: 点击段落 → Meta+v

封面图 vs 内容图

• 封面图: 使用 browser_file_upload（因为有专门的上传按钮）
• 内容图: 使用 Python 剪贴板 + 粘贴（更高效）

故障排除

MCP 连接问题

如果 Playwright MCP 工具不可用（报错 No such tool available 或 Not connected）：

方案1：重新连接 MCP（推荐）

执行 /mcp 命令，选择 playwright，选择 Restart

方案2：清理残留进程后重连

# 杀掉所有残留的 playwright 进程
pkill -f "mcp-server-playwright"
pkill -f "@playwright/mcp"

# 然后执行 /mcp 重新连接

配置文件位置: ~/.codex/mcp_servers.json

浏览器错误处理

如果遇到 Error: Browser is already in use 错误：

# 方案1：关闭全部会话后，用持久化 profile 重开（推荐）
playwright-cli close-all
bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

# 方案2：杀掉 Chrome 进程
pkill -f "Chrome.*--remote-debugging"
# 然后重新打开持久化会话
bash ~/.codex/skills/x-article-publisher/scripts/open_x_articles_browser.sh

图片位置偏移

如果图片插入位置不正确（特别是点击含链接的段落时）：

原因: 点击段落时可能误触链接，导致光标位置错误

解决方案: 点击后必须按 End 键移动光标到行尾

# 正确流程
1. browser_click 点击目标段落
2. browser_press_key: End        # 关键步骤！
3. browser_press_key: Meta+v     # 粘贴图片
4. browser_wait_for textGone="正在上传媒体"

媒体路径找不到

如果 Markdown 中的相对路径媒体找不到（如 ./assets/image.png、./assets/clip.mp4 实际在其他位置）：

自动搜索: parse_markdown.py 会自动在以下目录搜索同名文件：

• ~/Downloads
• ~/Desktop
• ~/Pictures

stderr 输出示例:

[parse_markdown] Image not found at '/path/to/assets/img.png', using '/Users/xxx/Downloads/img.png' instead

JSON 字段说明:

• original_path: Markdown 中指定的路径（解析后的绝对路径）
• path: 实际使用的路径（如果自动搜索成功，会不同于 original_path）
• exists: true 表示找到文件，false 表示未找到（上传会失败）

如果仍然找不到:

1. 检查 JSON 输出中的 missing_media / missing_images / missing_videos 字段
2. 手动将图片复制到 Markdown 文件同目录的 assets/ 子目录
3. 或修改 Markdown 中的图片路径为绝对路径