Breaking News Video Pipeline

Architecture

User Trigger (微信/webchat: "突发：xxx")
  │
Phase 1: Spawn sub-agent (DeepSeek V4 Flash, thinking=high)
  │  deep search (24h freshness) → data cleaning → output ~100 char script
  ▼
Phase 2: Main agent review (char count / date / relevance)
  │  Pass → launch TTS (edge-tts, background)
  ▼
Phase 3: Exaggerated title + topic.json + info card PNG
  │
Phase 4: Assembly + subtitle line-breaking (>10 chars wrap)
  │  Output: final.mp4
  ▼
Phase 5: Browser Restart (强制，防 CDP 断连)
  │
Phase 6: Douyin Upload (upload_douyin.py)
  │
Phase 7: WeChat Delivery → user review → publish
     wechat_send.py + wechat_screenshot.py

Differences from bi-daily

Phase 1: Deep Search + Script (Sub-Agent)

Spawn a V4 Flash sub-agent to handle search and first-draft script writing.

sessions_spawn with:
  task: \x3Csee prompt template below>
  model: "volcengine-plan/deepseek-v4-flash"
  mode: "run"
  thinking: "high"

Prompt Template

Replace [keyword] with the breaking event keyword from the user's message.

你是一个突发新闻编辑。请按以下步骤执行：

1. 使用 web_search 对关键词"[keyword]"进行深度搜索
   - 强制要求：只浏览 24 小时内的报道（使用 date_after 参数）
   - 至少搜索 3 次不同角度（如：事件本身、各方反应、影响分析）
   - 阅读搜索结果获取详细内容

2. 清洗数据：
   - 排除重复信息
   - 排除未经验证的传言
   - 保留关键事实：时间、地点、人物、数字、后果

3. 撰写一条 100 字左右的短视频口播文案（中文）：
   - 开头有冲击力，直击核心
   - 不换行，正确使用中文标点（。，、）
   - 涉及"台湾"一律使用"台湾省"
   - 涉及中国领导人加职位称呼
   - 仅输出文案本身，不要任何解释

只输出文案，不要输出其他任何内容。

Output: The sub-agent returns the script text directly. Save to 突发新闻/突发\x3Cdate>/文案/script.txt.

Phase 2: Main Agent Review + TTS

2a. Review Gate (MANDATORY)

Before TTS, the main agent MUST verify:

• 字数审查: 中文字符 80-120 字。超出 → 手动裁剪。不足 → 补充关键事实。
• 日期核查: web_search 交叉验证事件日期，确认 ≤24h
• 关联度核查: 文案内容与用户触发关键词紧密相关，无偏题
• 合规审查: 无禁止类别内容（娱乐/影视/短剧/游戏/电商/采访/访谈/评论/宗教/迷信/超自然/新车上市/地方政务）

If fails: fix and re-save script.txt before proceeding.

2b. TTS + Subtitles

After review passes, launch edge-tts (can run in background, no dependency on Phase 3):

mkdir -p 突发新闻/突发\x3Cdate>/音频 && \
cd ~/.openclaw/workspace && \
edge-tts --voice zh-CN-YunjianNeural --rate=+30% \
  -f 脚本/\x3Cdate>_breaking/script.txt \
  --write-media 音频/\x3Cdate>_breaking/audio.mp3 \
  --write-subtitles 突发新闻/突发\x3Cdate>/音频/out.vtt

VTT post-processing (de-punctuation + de-overlap):

python3 skills/breaking-news/scripts/check_vtt.py 突发新闻/突发\x3Cdate>/音频/out.vtt

Phase 3: Title + Info Card

3a. Exaggerated Title（夸张标题）

From the final script, create a clickbait-style title (≤30 chars for Douyin):

• 提取最具冲击力的关键词
• 使用感叹号、问号、省略号等吸引点击
• 示例: "突发！俄无人机撞入北约大楼" / "刚刚！美伊达成停火协议"

Save to 突发新闻/标题/标题\x3Cdate>.txt:

[夸张标题]
#新闻 #突发 #时事

3b. Info Card (topic.json + gen_cards.py)

Extract from the final script + search results:

{
  "category": "国际",
  "title": "俄罗斯无人机袭击罗马尼亚公寓楼",
  "big_value": "2人受伤",
  "unit": "俄无人机坠入北约领土",
  "note": "罗马尼亚驱逐俄外交官并关闭领事馆，北约启动第四条款紧急磋商",
  "source": "NYT / NPR / euronews"
}

字段约束：

• title ≤17 中文字符，超长自动截断加省略号
• big_value ≤6 中文字符
• note ≤28 中文字符，超过 14 字自动换行
• source 列出主要信源，用 / 分隔

python3 skills/breaking-news/scripts/gen_cards.py \
  突发新闻/突发\x3Cdate>/信息卡/topic.json \
  突发新闻/突发\x3Cdate>/信息卡/

Output: s00.png (1080x1920).

Phase 4: Assembly + Subtitle Wrapping

4a. Subtitle Line-Breaking

After VTT is ready, wrap any subtitle line >10 CJK characters:

python3 skills/breaking-news/scripts/wrap_vtt.py 突发新闻/突发\x3Cdate>/音频/out.vtt 13

This inserts \ break points every 10 CJK characters. Only counts CJK codepoints — punctuation and numbers are ignored. Original file backed up as .bak.

4b. Video Assembly

Single card PNG looped for audio duration, with subtitles burned:

cd ~/.openclaw/workspace && bash 视频制作/scripts/build_breaking.sh \x3Cdate>

build_breaking.sh:

1. Detect audio duration via ffprobe
2. Convert s00.png → s00.mp4 (loop for audio duration)
3. ffmpeg single-pass: card + audio + subtitles at 1080x1920
4. Output: 导出/final_\x3Cdate>_breaking.mp4

Phase 5: Browser Restart

强制 — 每次上传前必须重启浏览器，防 CDP 端口堆积 Worker 导致断连。

browser action=stop
sleep 3
browser action=start

Phase 6: Douyin Upload

cd skills/breaking-news/scripts/upload
PYTHONUNBUFFERED=1 .venv/bin/python upload_douyin.py \x3Cdate>

Phase 7: WeChat Delivery → Publish

Step 1 — Send:

python3 skills/breaking-news/scripts/wechat_send.py \x3Cdate> --breaking

Step 2 — Screenshot verify:

python3 skills/breaking-news/scripts/wechat_screenshot.py

Step 3 — Wait for user confirmation: "可以发布"

Step 4 — Publish: Browser tool (same as bi-daily Phase 9).

Scripts

External Scripts

Directory Structure

突发新闻/
├── 标题/
│   └── 标题\x3Cdate>.txt
├── 封面/
│   └── cover.png
├── 突发\x3Cdate>/                  ← one folder per event
│   ├── 文案/script.txt
│   ├── 信息卡/
│   │   ├── topic.json
│   │   └── s00.png
│   ├── 音频/
│   │   ├── audio.mp3
│   │   └── out.vtt
│   └── 导出/final.mp4
└── 突发\x3Cdate2>/
    └── ...