lalachan-xyq-browser-video

name	lalachan-xyq-browser-video
description	Use when generating, preparing, monitoring, downloading, or publishing LALACHAN Xiaoyunque videos through a logged-in Chrome/CDP web UI, especially requests mentioning Lala Xia, Aya Chan, Sasa Kun, Xiaoyunque, XYQ, Seedance, 沉浸式短片, Agent 模式, reference images, reference video, Nutstore copying, low-credit generation, or avoiding the Xiaoyunque API.

LALACHAN Xiaoyunque Browser Video

Core Rule

Prefer the logged-in Xiaoyunque browser UI over the Xiaoyunque API unless the user explicitly asks for API usage. Validate visible page state before submitting: mode, model, duration, ratio, prompt, attachments, and credit estimate.

On Lachlan's workstation, $HOME/.cache/xyq-chrome is the canonical shared Chrome profile for Xiaoyunque, JLC/JLCEDA web workflows, and downloads. When a visible virtual desktop is requested, preserve that profile on display :98, CDP 9344, and noVNC 6099. Do not launch a fresh profile or substitute the book-only embedded-browser profile.

For portable docs/examples, use an ignored local config instead of personal absolute paths:

cp .config/lazyskills.env.example .config/lazyskills.local.env
$EDITOR .config/lazyskills.local.env
set -a
. .config/lazyskills.local.env
set +a

Use the bundled scripts first:

scripts/xyq_cdp_browser.py
scripts/xyq_chrome/watch_thread_dom_download.py
scripts/xyq_chrome/launch_chrome.sh

Load detailed references only when needed:

references/xyq-browser-video-generation-skill.md
references/uploaded-images-no-path-workflow.md
references/xyq-browser-automation-workflow.md
references/smooth-video-generation-runbook.md
references/continue-confirm-download-runbook.md
references/30s-agent-workflow.md
references/words-card-30s-default.md
references/lala-studio-end-to-end-runbook.md

For a complete WeChat/operator handoff covering story writing, prompt saving, browser generation, download, and LazyEdit publishing, see:

$LALACHAN_ROOT/references/lalachan-story-video-handoff-for-wechat.md
$LAZYSKILLS_ROOT/docs/lalachan-story-video-generation-handoff.md

LALACHAN Defaults

Default uploaded image order in the LALACHAN project:

words-card.jpg
LazyingArtRobot.png
display.png
patchwork-leather-notebook-luxury-clean-v2.png
raraxia.jpeg
ayachan.png
sasakun.jpeg
Trio.png

Prompt labels after upload:

• 图1: words card / 小白屏学习卡.
• 图2: LazyingArtRobot.png, robot 庄子; keep the LazyingArt logo on chest.
• 图3: LightMind AI glasses.
• 图4: handmade patchwork notebook.
• 图5: raraxia.jpeg, individual 啦啦侠 / Rara Xia reference.
• 图6: ayachan.png, individual 阿芽酱 / Aya Chan reference.
• 图7: sasakun.jpeg, individual 飒飒君 / Sasa Kun reference.
• 图8: Trio.png, three-character group identity reference.

No-Trio variant:

• If the user says not to upload Trio.png, skip only Trio.png.
• Still upload supporting assets unless separately excluded: words-card.jpg, LazyingArtRobot.png, display.png, and patchwork-leather-notebook-luxury-clean-v2.png.
• Still upload the three individual character references: raraxia.jpeg, ayachan.png, and sasakun.jpeg.
• In that case labels stop at 图7 and the prompt must not mention 图8 or a group-reference image. "No Trio" does not mean "only character photos" unless the user explicitly says only to upload character photos.

Characters:

• 啦啦侠 / Lala Xia: use raraxia.jpeg as the primary individual reference.
• 阿芽酱 / Aya Chan: use ayachan.png as the primary individual reference.
• 飒飒君 / Sasa Kun: use sasakun.jpeg as the primary individual reference.
• Trio.png: optional group identity/reference image when uploaded; omit it when the user requests no Trio.
• words-card.jpg: words card / 小白屏学习卡 style reference.
• LazyingArtRobot.png: robot 庄子; preserve the LazyingArt chest logo.
• display.png: LightMind AI glasses.
• patchwork-leather-notebook-luxury-clean-v2.png: handmade patchwork notebook/tool prop.
• raraxia.jpeg: individual 啦啦侠 / Rara Xia reference.
• ayachan.png: individual 阿芽酱 / Aya Chan reference.
• sasakun.jpeg: individual 飒飒君 / Sasa Kun reference.

Main-cast identity is a paid-submit blocker:

• The generated prompt must clearly anchor 啦啦侠 / Lala Xia, 阿芽酱 / Aya Chan, 飒飒君 / Sasa Kun, and 庄子机器人 / Zhuangzi Robot to the uploaded reference images.
• Extra villains, crowds, workers, animals, or fantasy characters are allowed, but they must not replace or redesign the four main characters as unrelated humans, unrelated figurines, or new faces.
• If the visible preview/prompt setup suggests the main cast may be redesigned instead of referenced, stop before paid submission and fix the prompt or attachments.

Words-card rule:

• Treat 图1 as the visual style reference for the physical learning card.
• Use $LALACHAN_ROOT/words-card.jpg as the design reference for Codex image generation. Before Xiaoyunque upload, generate a fresh episode-specific PNG with the imagegen skill and pass the reference image to that generation call.
• For every new video, create a fresh story-relevant word card; do not reuse the previous word unless the user asks.
• Default card metadata includes English, Japanese, Japanese furigana, and one short target-language meaning. Use another target language when the episode requests it.
• Before image generation, verify that every value expresses the same intended word and is correctly written in its own language or script. Apply the same accuracy standard to every line.
• The physical card face contains only the values, one line each. Do not render language names, field labels such as English: or Japanese:, colons, bullets, numbering, or explanatory text.
• Use the in-scene prompt pattern: 图1 是已经制作好的实体学习卡，只显示上传图片中的四行正文，不加语言名称、字段标签、冒号、项目符号或编号。保持每一行文字准确清楚。它只是场景里的真实道具，不是字幕。
• Choose a word that matches the episode theme, for example battle/courage scenes can use courage / 勇気 / ゆうき / 勇气.
• Default method: use Codex imagegen with words-card.jpg as the referenced image, render the exact unlabeled value block, save the generated PNG under the current run, compare every line character-by-character, and upload that PNG as 图1.
• After generating any words card, key-scene image, character reference, or other still asset, preserve the selected result at a stable project path and open it in a detached local viewer with the persistent-image-preview skill. A preview rendered only in chat does not count as opened. Keep the viewer open while the asset is reviewed or uploaded.
• Treat an unreadable, mislabeled, misspelled, semantically inconsistent, missing, duplicated, or unverified generated card as a paid-submit blocker. Regenerate the card before Xiaoyunque; do not spend video credits hoping Xiaoyunque will repair it.
• Asking Xiaoyunque to render a new card from the base style image is only a fallback when the user explicitly accepts it or Codex image generation is unavailable and no paid submission will occur yet.
• The card is a real prop in the scene, not a subtitle overlay.

Story writing rule:

• Write stories in natural, understandable Chinese first. The viewer should understand the action, joke, and emotion without reading production notes.
• Keep one clear cause-and-effect chain per short video: problem, response, twist, payoff.
• Dialogue should sound like friends talking. Avoid pseudo-code, abstract slogans, machine-like phrases, over-explained lore, and strange translated wording.
• Use simple concrete nouns and actions. Prefer "培养皿在显微镜下慢慢亮起来" over vague phrases like "生命数据完成同步".
• If a concept is educational or technical, show it through a visible action and one plain sentence, not a lecture.
• For lab scenes, keep the lab clean and believable: gloves, sterile bench, incubator, microscope, pipette, petri dish, and careful movement. Dance can happen as small safe steps, shoulder moves, or rhythm while working, not reckless lab behavior.
• Before turning a new story into a Xiaoyunque prompt, run a critic pass using lalachan-story-critic when available. Fix exact awkward lines first, then write the final prompt.

Never paste local filesystem paths into the Xiaoyunque prompt. Paths are only for the browser file upload command. The prompt should refer only to actually uploaded labels: normally 图1 through 图8, or 图1 through 图7 when Trio.png is intentionally omitted. Explicitly ask not to draw file names or paths into the video.

Always upload the actual reference image files before generation. Do not treat typed local paths, pasted filenames, or prompt-only image descriptions as a substitute for upload. Before any paid submit, verify visible attachment evidence for every required reference image. If image upload fails or cannot be proved, stop and fix/report the blocker instead of submitting.

Default video setup:

Mode: 沉浸式短片 by default for normal LALACHAN video generation
Model: Seedance 2.0 Mini 体验版 / cheapest visible suitable model by default
Duration target: 15s by default
Ratio: 4:3 unless the user requests otherwise
Prompt language: mainly Chinese
Always include: 不要字幕，不要生成任何字幕、说明文字、下三分之一文字或画面文字。
Post-generation: always auto-download the finished MP4, verify it, copy it to Videos/, and send it back to the requesting chat. Submit to LazyEdit only when the current request explicitly asks for LazyEdit/import/process or public publishing.

For default work, target 15秒 and prefer 沉浸式短片 with Seedance 2.0 Mini 体验版 / the cheapest visible suitable model. Use 创作 Agent / integrated-agent only when the user explicitly asks for a longer/full-song video, when the current active thread is already an Agent thread, or when short-film controls cannot satisfy the task. Before any Agent/long-video paid action, inspect the visible credit estimate. If it is not clearly the cheapest available path, or if it is a high-cost render, pause and ask for approval instead of continuing. If a generation fails, continue in the same current thread by sending a short corrective message; do not start a new session or new thread just to retry unless the current thread is unusable and the user accepts that cost/risk.

Use 30秒 or longer only when the user explicitly asks for it.

For a requested video longer than 15 seconds, select 创作 Agent (or another explicit long-video workflow) and make sure the 沉浸式短片 chip is not active. Do not split the story into 15-second clips or compress it merely because the short-film workflow is capped. Seedance 2.0 Mini 体验版 can still be selected inside Agent mode when the visible UI offers it; validate the actual selected model and credit preview before submitting.

Credit-Budget Rule

• Default to Seedance 2.0 Mini 体验版 / vipnew when the UI shows it, especially when it shows a cheap rate such as 单秒限时低至4积分.
• If Mini体验版/vipnew is unavailable, choose the cheapest visible suitable Seedance row. Do not silently upgrade to a high-credit long-video/Agent render.
• Treat visible high credit estimates, non-Mini paid long renders, recharge/payment approval, insufficient credits, disabled submit, login, CAPTCHA, or explicit user budget limits as blockers that require reporting before paid submission.
• Use normal Seedance 2.0, Fast VIP, Agent long-video, or other more expensive routes only when the user explicitly asks for that quality/duration/capability or confirms the visible cost.
• If the user asks for cheapest, Mini体验版 is the target. If the requested duration cannot be made with Mini, offer a shorter Mini version or ask before using Agent.
• If the UI shows Mini 体验版 / cheapest trial but another model is selected, do not submit. Select the cheap model first or stop and report the blocker.
• Do not continue an 智能长视频 / Agent render when it requests a large point spend unless the user has explicitly approved that exact high-cost render.
• For 4:3, verify the opened ratio menu checkmark or screenshot because the compact toolbar may still show only 比例.
• Never waste credits on avoidable retries: do not click 生成视频, 提交, or any paid action twice unless the page proves the first attempt failed without charging. If points drop or the task shows queued/running, monitor only.

Browser Workflow

Before manipulating the page, prove that CDP and noVNC expose the same logged-in Chrome profile and X display. On the canonical workstation this is normally CDP 9344, display :98, noVNC 6099, and profile ~/.cache/xyq-chrome. A healthy legacy port such as 9222 may belong to an unrelated browser. Check the Chrome process arguments, page URL, and visible noVNC tab together; do not trust a responsive port alone. Explicit LALA_STUDIO_XYQ_* values take precedence over stale generic XYQ_* settings.

1. Attach to an existing logged-in Chrome CDP endpoint, usually $XYQ_CDP_URL or another active port:

scripts/xyq_cdp_browser.py --cdp-url "$XYQ_CDP_URL" list-pages

2. Bring the Xiaoyunque page forward and inspect visible controls:

scripts/xyq_cdp_browser.py --cdp-url "$XYQ_CDP_URL" bring-to-front PAGE_ID
scripts/xyq_cdp_browser.py --cdp-url "$XYQ_CDP_URL" visible PAGE_ID

3. Select mode/model/duration/ratio by browser UI. For long work, click the exact top-level Agent mode and verify the short-film chip is inactive. Verify the requested mode, selected model row, duration, ratio, upload success, and any visible point cost/VIP/vipnew state as far as the UI allows. Do not block only because the exact preferred model or exact cost text is unavailable.

4. Upload and verify reference images:

scripts/xyq_cdp_browser.py --cdp-url "$XYQ_CDP_URL" upload-images-verify PAGE_ID \
  words-card.jpg \
  LazyingArtRobot.png display.png patchwork-leather-notebook-luxury-clean-v2.png \
  raraxia.jpeg ayachan.png sasakun.jpeg Trio.png \
  --timeout 180 \
  --screenshot outputs/xyq-run/after-upload.png

5. Fill the saved prompt:

scripts/xyq_cdp_browser.py --cdp-url "$XYQ_CDP_URL" type-prompt PAGE_ID references/prompts/example.md --wait 2

6. If the submit button stays disabled, inspect upload item classes. Wait until every uploaded file item is success; a single uploading reference blocks submission.

7. If the current thread pauses, fails, or needs a correction, first send a short message in the same thread. Avoid new sessions because they can waste credits and confuse the user. Use 创作 / new-session only when the current thread is truly unusable or the user explicitly asks.

8. Submit only when the user asked for generation and the pre-submit contract is satisfied.

Agent Confirmation Pauses

Long-video / Agent workflows may stop after storyboard or reference-material creation and ask the user to confirm before generating video. If the page asks for continuation, reply in the same thread with a short message such as 继续生成视频。; do not start a new generation. Use type-prompt or real browser keystrokes so the chat input enables the send button.

scripts/xyq_cdp_browser.py type-prompt PAGE_ID outputs/xyq-run/continue.md --wait 0.3
scripts/xyq_cdp_browser.py click PAGE_ID SEND_BUTTON_X SEND_BUTTON_Y

Watch, Download, Copy

Monitor the submitted thread through the browser page:

scripts/xyq_chrome/watch_thread_dom_download.py \
  --cdp-url "$XYQ_CDP_URL" \
  --page-id PAGE_ID \
  --thread-url "THREAD_URL" \
  --output-dir outputs/xyq-run \
  --filename result_30s.mp4 \
  --expected-duration 30 \
  --copy-to Videos \
  --copy-to "$NUTSTORE_AUTOPUBLISH"

After every successful Xiaoyunque generation, do not stop at the browser result. Download the final MP4 automatically, verify it with ffprobe, apply the 5-second duration tolerance unless exact duration was requested, copy it to Videos/, and send it back to the requesting chat. Only submit it to LazyEdit when the current request explicitly asks for LazyEdit/import/process or public publishing. Direct LazyEdit CLI upload is preferred when available; Nutstore AutoPublish import is an acceptable fallback. For LazyEdit-only requests, use --no-publish; for public publish requests, publish exactly once to the requested platforms.

Do not scan every page-level video element and accept the first downloadable file. Xiaoyunque pages can contain promotional media and stale results. Scope discovery to the current result card or visible preview, pass --expected-duration, and reject a candidate whose probed duration falls outside the requested tolerance. Hash the accepted download and copied outputs when exact identity matters.

Direct LazyEdit handoff pattern:

cd "$LAZYEDIT_ROOT"
python scripts/lazyedit_publish.py \
  --video "$LALACHAN_ROOT/Videos/VIDEO.mp4" \
  --title VIDEO_COMPLETED \
  --use-current-settings \
  --correction-prompt-file "$LALACHAN_ROOT/references/prompts/PROMPT.md" \
  --metadata-prompt-file temp/METADATA_BRIEF.md \
  --correct-subtitles \
  --correction-source polished \
  --no-publish \
  --wait \
  --poll-seconds 10

Nutstore fallback pattern:

cp -f "$LALACHAN_ROOT/Videos/VIDEO.mp4" \
  "$NUTSTORE_AUTOPUBLISH/VIDEO_COMPLETED.mp4"

Protected everphoto URLs may fail from unauthenticated direct HTTP. First test browser-context fetch with --await-promise; if it returns 200 video/mp4, pull the active video.currentSrc from the page and download with browser-like Referer and User-Agent headers. The bundled watcher can now use this same browser-context path by triggering an in-page blob download and copying the downloaded MP4 from ~/Downloads. If protected URL download still fails but the page is complete and has a visible top-right 下载 button, click the page button and then copy the newest MP4 from ~/Downloads. Quote the file path or use null-delimited find -print0 patterns because Xiaoyunque may save names such as final_video (5).mp4. See references/continue-confirm-download-runbook.md for the full fallback sequence.

Completion Check

Before final response, report only verified facts:

• prompt/reference files saved;
• mode/model/duration/ratio seen on page;
• attachment filenames verified;
• prompt verified to contain no local image paths;
• credit charge or blocker observed;
• local MP4 path, ffprobe dimensions/duration, and copy targets.
• LazyEdit handoff status: video id for direct CLI upload, or Nutstore _COMPLETED path/import evidence for folder handoff.