Xms export automation

XMS 异常件导出自动化

Overview

Automates the full workflow of exporting abnormal items from the XMS customer service system (cs-packet.i4px.com). The workflow includes browser health checks, SSO login, data export, polling for completion, file download, and optional DingTalk delivery.

Prerequisites

Create a configuration file webhook_config.json (managed by xms_config_manager.py):

{
  "xms": {
    "url": "http://cs.packet.i4px.com/",
    "username": "YOUR_USERNAME",
    "password": "YOUR_PASSWORD"
  },
  "webhooks": [
    {
      "name": "Group Name",
      "token": "dingtalk_robot_token",
      "keyword": "required_keyword",
      "folderId": "https://alidocs.dingtalk.com/i/nodes/YOUR_FOLDER_ID",
      "enabled": true,
      "schedule": {
        "timeSlots": [{"hour": 9, "minute": 0}],
        "days": [1, 2, 3, 4, 5, 6, 0]
      }
    }
  ],
  "export": {
    "poll_interval_seconds": 120,
    "max_wait_minutes": 70,
    "file_name_template": "xms异常件导出_{timestamp}.xlsx",
    "max_retries": 3,
    "retry_interval_seconds": 60,
    "browser_max_recovery_attempts": 4,
    "send_notification_on_failure": true
  }
}

• days: 0=Sunday, 1=Monday, ..., 6=Saturday
• poll_interval_seconds: Seconds between progress checks
• max_wait_minutes: Max total wait time for export completion
• max_retries: Max full-workflow retry attempts when any phase fails
• retry_interval_seconds: Seconds to wait between retry attempts
• browser_max_recovery_attempts: Max browser recovery strategies to try per attempt
• send_notification_on_failure: Whether to send a DingTalk alert when all retries exhausted

Retry & Error Handling

The automation runs inside an outer retry loop:

1. Browser recovery first: Each attempt begins with a 4-step browser health check (check size → JS resize → new tab → recreate window). If the browser MCP returns "Extension not connected" or 0x0, the agent progressively escalates recovery strategies.
2. Phase-level retries: If login, export trigger, polling, download, or DingTalk upload fails, the agent records the error, waits retry_interval_seconds, and restarts the entire workflow from browser recovery.
3. Failure notification: If all max_retries are exhausted and send_notification_on_failure is true, the agent sends a DingTalk message to every matched webhook reporting the failure reason, so you are never left in the dark.

Workflow

Phase 0: Read Config & Filter Webhooks

1. Read webhook_config.json
2. Get current Beijing time (Asia/Shanghai) and day of week
3. For each webhook where enabled=true, check if current time matches any timeSlots and current day is in days
4. Collect matching webhooks. If none match and this is a scheduled run, exit silently

Phase 0.5: Browser Window Health Check

Before any interaction, verify the browser window:

1. Check size: window.innerWidth + ',' + window.innerHeight
2. JS resize: If broken, window.moveTo(0,0); window.resizeTo(1440,900) then wait 2s and recheck
3. New tab: Call tabs_create_mcp, check new tab size
4. Recreate window: Close all tabs one by one (tabs_close_mcp only accepts a single tabId per call), then tabs_context_mcp with createIfEmpty:true
5. Small window fallback: If persistently small (e.g., 256x116) but not 0,0, DOM operations still work. Only terminate if literally 0,0 after all steps

Phase 1: Login

1. Navigate to http://cs.packet.i4px.com/
2. If redirected to sso.i4px.com, inject credentials via JavaScript:

   document.getElementById('username').value = USERNAME;
   document.getElementById('passwordOrg').value = PASSWORD;
   ['input', 'change', 'keyup'].forEach(evt => {
     document.getElementById('username').dispatchEvent(new Event(evt, {bubbles:true}));
     document.getElementById('passwordOrg').dispatchEvent(new Event(evt, {bubbles:true}));
   });
   document.getElementById('signbtn').click();
   

3. Handle "already logged in elsewhere" dialog: If a dialog appears saying "此账号已在别的地方登录" or similar, click the "是"/"确定" button to continue
4. Handle QR code / device verification: If after clicking login, a QR code scan, mobile device authentication popup, or slider captcha appears that the Agent cannot complete automatically:
   • Do NOT retry repeatedly
   • Immediately notify the user via IM (e.g., 小Q channel)
   • Wait 3 minutes, then recheck if the page has redirected to XMS homepage
   • If still not redirected after timeout, record error and exit to failure notification
5. Handle "server exception" error: If at any point the page displays "服务器异常" (or "系统异常"/"服务异常"):
   • Close the current tab using tabs_close_mcp
   • Create a new tab using tabs_create_mcp
   • Re-navigate to http://cs.packet.i4px.com/ in the new tab
   • Restart the login flow from Phase 1 Step 1
   • This counts toward the outer retry loop's max_retries

   Do NOT repeatedly interact with the error page. Always close and recreate the tab.

6. Login timeout: If the login page URL does not change within 3 minutes and no special scenario above occurs, record "XMS login timeout" and handle via retry logic
7. Wait for redirect to cs-packet.i4px.com/index

Phase 2: Export Data

1. Click 异常件管理 in left sidebar to expand submenu
2. Click 自有业务异常件管理
3. Clear the shelf organization field (critical):
   • The visible field is id=txtUpShelfOgCodeShow (default value: "客户服务部")
   • The hidden field is id=txtUpShelfOgCode (default value: "D000414")
   • Clear both via JavaScript: .value = '' then dispatch input/change events
   • Alternative: left-click inside the input field on the right side of the text; the text should disappear automatically
4. Click 查询 (id=abnormalSearch)
5. Wait for data, then click 导出原因(新) button (top-right)

Phase 3: Poll for Completion

1. Click 下载中心 to open the export records dialog
2. Read the drawer table. First data row shows: task name, file size, progress %, status, export time, download action
3. If progress is not 100%, wait poll_interval_seconds, click 刷新, and recheck
4. Repeat until 100% or max_wait_minutes exceeded

Modal reading tip: read_page and get_page_text cannot see modal overlays. Use computer screenshot or JavaScript DOM queries on .next-drawer to read the export record table.

Phase 4: Download

1. Read the 导出时间 from the first row (format: 2026-05-09 22:59:29)
2. Generate filename: xms异常件导出_2026-05-09 22_59_29.xlsx (replace colons with underscores for Windows compatibility)
3. Click the 下载 button in the first row's action cell
4. Find the downloaded file in ~/Downloads/ (filename starts with xms-cts-)
5. Copy to output directory with the generated filename

Phase 5: DingTalk Delivery (Optional)

For each matched webhook:

1. Call get_file_upload_info for upload credentials
2. Upload the xlsx via HTTP PUT to resourceUrl
3. Call commit_uploaded_file with folderId and filename to get nodeId
4. Generate permanent link: https://alidocs.dingtalk.com/i/nodes/{nodeId}
5. Send message via send_message_by_custom_robot including the webhook's keyword and the permanent link

Verified Element References

Element	ID / Text	Notes
SSO username	id=username	Not loginName
SSO password	id=passwordOrg	Not loginPwd
SSO login button	id=signbtn	Not loginBtn
Shelf org visible	id=txtUpShelfOgCodeShow	Default: "客户服务部"
Shelf org hidden	id=txtUpShelfOgCode	Default: "D000414"
Query button	id=abnormalSearch
Export reason button	text=导出原因(新)
Download center	text=下载中心	Opens drawer
Refresh button	text=刷新	In drawer
Drawer table	.next-drawer table tr	First data row = latest export

Critical Rules

• Model: When creating the QoderWork cron job, set payload.model to "qwork-auto" (standard model). Do not use flagship model for production.
• Small window: 256x116 is functional. DOM queries, .click(), and .value assignments work even in tiny windows. Only abort at 0,0.
• tabs_close_mcp: Only accepts a single integer tabId per call. Close tabs one by one.
• Filename: Use the export record's "导出时间", not system time. Replace : with _ for Windows filenames.
• DingTalk links: Always use https://alidocs.dingtalk.com/i/nodes/{nodeId}, never temporary OSS links.
• Keywords: Every robot message must contain that webhook's configured keyword.
• Robot token parameter: Use robotToken (not access_token) when calling send_message_by_custom_robot.
• Upload Content-Type: When uploading xlsx via HTTP PUT to resourceUrl, set Content-Type header to an empty string "", and include the returned Authorization and x-oss-date headers.