AIOS Jimeng Image Ops

Overview

Use this runbook for stable image generation on Jimeng. It includes selectors, completion signals, policy-failure handling, and evidence requirements verified in a live run.

Preconditions

• Browser profile has valid Jimeng login session.
• default profile should connect to fingerprint browser via CDP (port 9222 by default).
• MCP tools available: chrome.launch_cdp, browser.connect_cdp, page.goto, page.click, page.type, page.extract_text, page.get_html, page.screenshot, browser.close.
• For page understanding, prefer page.extract_text first, then page.get_html for structure details.
• Use page.screenshot only when text/HTML evidence is insufficient, and prefer target-area capture over full-page screenshots when supported.
• Prompt is policy-safe (avoid risky terms, political/person-identifiable/sensitive wording).

Execution Flow (Updated 2026-03-13)

1. Open generation page:
   • Navigate to home: https://jimeng.jianying.com/ai-tool/home/
   • Click "图片生成" button (selector: button.button-RNHVcx:has-text("图片生成"))
   • Or use direct URL: https://jimeng.jianying.com/ai-tool/generate?enter_from=ai_feature&from_page=explore&ai_feature_name=image
2. Use page.extract_text / page.get_html to detect login/challenge markers (captcha, risk check, 2FA, login wall); if detected, ask user to complete human action and resume.
3. Confirm prompt box exists:
   • Selector: div.tiptap.ProseMirror (NOT textarea, this is a rich text editor)
4. Click prompt box to activate, then fill prompt.
5. Model & Ratio Selection (Bottom Toolbar):
   • Model selector: div.lv-select:has-text("图片") (first one, position x:344)
   • Ratio selector: Second div.lv-select (position x:458, shows "图片 4.1" or "图片 5.0")
   • Note: These may show clickable: false in snapshot but are still interactive
6. Generate Button:
   • Primary: button.lv-btn.lv-btn-primary (position x:976, y:647)
   • Note: Button may show as collapsed-submit-button class, enabled only after prompt is entered
7. Poll page.extract_text (and page.get_html when needed) until completion:
   • Success markers: latest task has image tiles + 重新编辑 and 再次生成
   • Policy failure: 你输入的文字不符合平台规则，请修改后重试
   • Timeout: still 生成中 after budget

Error Handling

• Policy failure: rewrite prompt to neutral scene/style terms and retry once.
• Timeout: refresh generation page and retry with same prompt once.
• Selector failure: recapture snapshot and update selectors from visible controls.

Evidence Standard

• Keep at least one snapshot proving success or failure.
• Save one screenshot for run summary.
• Record prompt, outcome, and retry path in final doc.

Resources

• references/run-report-2026-03-02.md: verified live-run record.
• references/selectors-and-errors.md: selector cookbook and failure taxonomy.