vision ocr

用于识别图片和 PDF 文档，调用你已配置的 OCR 与多模态服务输出 Markdown 结果，并可按需发送到飞书。适合截图、扫描件、表格、票据和技术文档。

This package appears coherent for OCR/PDF tasks, but check these before enabling: - Do not run node index.js --update-config (or update-config.js) unless you intend to persist VISION_* tokens into the skill's config.json; stored tokens will reside on disk and should be protected by file-system permissions. - Keep VISION_RESOLVE_OPENCLAW_SESSION and VISION_ALLOW_REMOTE_INPUT disabled (default false) unless you explicitly need CLI-session recovery or remote URL downloads; enabling them grants the skill access to OPENCLAW_* env and ~/.openclaw/runtime.json or allows downloading remote attachments. - If you enable automatic Feishu sending, ensure the optional feishu-send-files integration is trustworthy and that you really want results sent to chat targets discovered from context; otherwise keep auto-send off and use --no-send-to-feishu. - Review the local config.json and remove any secrets from the repository before sharing. If you need higher assurance, audit the remaining parts of index.js (network calls to multimodal baseUrl) to confirm it only talks to configured endpoints and does not leak content elsewhere.

Type: OpenClaw Skill Name: vision-ocr Version: 1.1.2 The vision-ocr skill provides document and PDF OCR capabilities with Feishu integration, requiring high-risk operations such as executing external processes (index.js uses execFileSync to run pdf-helper.py and feishu-send-files), downloading remote files, and reading sensitive session data from ~/.openclaw/runtime.json. While the code includes security measures like SSRF protection (isPrivateIpAddress) and file size limits, the ability to fetch remote content and access session tokens represents a significant attack surface. No evidence of intentional malice was found, but the high-privilege nature of its operations and the handling of sensitive documents justify a suspicious classification.