t-cpm自定义技能

基于conding-plan-models.json配置的AI强化识图筛选工具，支持关键词三维深度分析、多模型图片识别、严格内容审核、批量图片筛选，自动删除不符合要求的图片。触发场景：(1) 本地图片批量匹配关键词筛选 (2) 图片内容审核/合规校验 (3) 自定义规则图片识别分类 (4) 对接其他图片源的二次筛...

This skill will upload your images (base64-embedded) to the model endpoint specified in /root/.OpenClaw/workspace/conding-plan-models.json and — by default — permanently delete images judged non-conforming. Before installing or running: 1) Inspect the config file and confirm the Base URL and API key point to a trusted service; 2) Do not run with production or sensitive images until you validate behavior; run with --delete_invalid False and test on a small sample first; 3) Backup any images or directories you run this on (deletion is permanent); 4) Consider running the script in an isolated environment or container and review network activity to ensure images are not sent to an unexpected endpoint; 5) If you cannot verify the endpoint or if policy forbids uploading images, do not use this skill. If you want help checking the config file format or testing safely, provide a redacted sample and I can advise further.

Type: OpenClaw Skill Name: t-cpm Version: 1.0.0 The skill bundle implements an automated image filtering tool that deletes local files based on AI model judgments, which poses a significant risk of unintended data loss due to potential AI hallucinations or prompt injection. The script `scripts/image_filter_pipeline.py` reads sensitive API credentials from a hardcoded local configuration file (`/root/.OpenClaw/workspace/conding-plan-models.json`) and performs file deletions (`os.remove`) based on remote API responses. While these behaviors are documented as the tool's primary purpose in `SKILL.md`, the combination of automated file system modification and credential handling is inherently high-risk.