paper-video-generate-publish

name	paper-video-generate-publish
description	Use when working in the JushenRenji repo and the user asks to update the latest paper video, generate a paper explainer by title, publish or republish to Bilibili/Xiaohongshu, update Bilibili cookies, or verify upload results.

Paper Video Generate And Publish

Overview

This skill is for the JushenRenji repo's daily paper workflow: generate a single-paper video, publish it to bilibili and xiaohongshu, and handle common retry paths.

Follow the repo defaults unless the user explicitly overrides them:

• whole conversation in Chinese
• do not keep asking for confirmation
• default platforms are bilibili,xiaohongshu
• use the auto-generated title, description, tags, and cover

When to Use

Use this skill when the user asks for any of the following inside this repo:

• “更新最新视频” or “更新这篇论文”
• “发到 B站 和 小红书” or “补发 B站”
• “更新 cookie” and retry Bilibili upload
• “看下效果 / 核验结果 / 有没有被截断”

Do not use this skill for generic code changes unrelated to the paper video pipeline.

Project Facts You Must Remember

• Main entry point: src/main.py
• Default upload platforms are bilibili,xiaohongshu
• Monday rolls back to last Friday for arXiv date selection; Saturday/Sunday also roll back to Friday logic in src/main.py
• Compact platform descriptions are built in src/distribution/orchestrator.py
• Description rules already implemented:
  • include paper title
  • include paper link
  • include project link only when it actually exists
  • do not invent a project link
• Single-paper outputs are renamed to output/YYYY-MM-DD_<中文标题>.mp4 and matching .png

Standard Workflow

1. Generate and publish a paper by title

Use this command template:

cd /home/jdh/Projects/VlogCutter/JushenRenji
source ~/miniconda3/etc/profile.d/conda.sh
conda activate paperagent
mkdir -p tmp
PYTHONUNBUFFERED=1 PYTHONPATH=src TTS_MAX_WORKERS=1 TTS_SERIAL_RETRY_ATTEMPTS=2 \
python src/main.py --filename '<论文标题>' 2>&1 | tee tmp/<slug>_publish_<YYYYMMDD>.log

Notes:

• Keep --filename equal to the user-given paper title unless they ask for another query.
• Do not ask whether to publish unless the user explicitly asks to stop publishing.
• If the user asks for a newer or longer version, let the render finish; do not switch to a fast path unless they ask.

2. Verify after generation

Always verify with commands instead of trusting logs casually:

ls -lt output | head -20
ffprobe -v error -show_entries format=duration,size -of json '<video_path>'
tail -n 120 tmp/<slug>_publish_<YYYYMMDD>.log

Report at least:

• output video path
• cover path
• duration
• paper title
• paper link
• Xiaohongshu result
• Bilibili result

If Bilibili failed, quote the concrete error such as code=-101, 账号未登录.

Bilibili Retry SOP

1. Update cookies in config.yaml

The current code reads only these fields from bilibili_cookies:

• sessdata
• bili_jct
• dedeuserid
• dedeuserid_ckmd5

Ignore response-header noise such as set-cookie, Path, Domain, Expires, and SameSite.

2. Retry only Bilibili without regenerating video

Use src.distribution.orchestrator.upload_generated_content(...) against the existing output video.

Template:

cd /home/jdh/Projects/VlogCutter/JushenRenji
source ~/miniconda3/etc/profile.d/conda.sh
conda activate paperagent
PYTHONUNBUFFERED=1 PYTHONPATH=src python3 - <<'PY' 2>&1 | tee tmp/<slug>_bili_retry_<YYYYMMDD>.log
from src.distribution.orchestrator import upload_generated_content

result = upload_generated_content(
    platforms=['bilibili'],
    video_path='output/<video>.mp4',
    cover_path='output/<video>.png',
    video_title='<中文标题>',
    video_tags='人工智能,具身智能,机器人,模仿学习,强化学习,自动驾驶,具身人机',
    video_desc='<英文标题>',
    cn_titles=['<中文标题>'],
    origin_titles=['<英文标题>'],
    summaries=None,
    paper_links=['<论文链接>'],
    project_links=[''],
    bilibili_tid=188,
)
print(result)
PY

Then verify:

tail -n 80 tmp/<slug>_bili_retry_<YYYYMMDD>.log

Success means the submit result contains code: 0 and a bvid.

Description Rules

When retrying or checking content, the effective Bilibili/Xiaohongshu compact description should stay short and follow this order:

• 论文标题：...
• 论文链接：...
• 项目链接：... only if non-empty
• optional summary only if available and still within limit

For Bilibili the total limit is 250 characters. For Xiaohongshu the total limit is 300 characters.

Xiaohongshu Notes

• The repo uses the MCP-based uploader in src/distribution/xiaohongshu.py
• Docker service is auto-started when possible
• First-time login can require:

python -m src.distribution.xiaohongshu --login

• Current MCP video publish path may ignore cover_path; tell the user to adjust cover in the app if needed

Common Pitfalls

• Do not claim success before checking the final upload result in the log
• Do not say Bilibili succeeded if the returned result is empty
• Do not invent project links from organization, dataset, or demo homepages
• Do not regenerate the full video when the task is only “补发 B站”
• Do not forget that Monday's query date is the previous Friday

Fast Checklist

1. Run the generation command or the single-platform retry command.
2. Wait for the process to finish fully.
3. Verify output file and duration with ffprobe.
4. Verify upload result from the log tail.
5. Reply with concrete paths, duration, paper link, and per-platform status.