DiePre Vision Cognition

DiePre 视觉认知 Skill —— 将包装/模切机器视觉感知与 SOUL 推理融合的认知框架

Security Scan

The skill's README and SKILL.md describe a complex local vision/VLM system and a Python API, but there is no implementation, no declared dependencies or credentials, and several runtime behaviors are underspecified—this mismatch is suspicious and needs clarification before use.

Purpose & Capability

The skill claims a full-featured DiePre vision+SOUL cognition framework (VLM→CAD pipelines, model fine-tuning, four-way analysis) and documents a Python class API, yet the package in the registry is instruction-only with no code files, no required binaries, and no declared model weights or external service credentials. It is unclear how the described capabilities would be implemented or where required models/tools come from.

Instruction Scope

Runtime instructions require writing persistent logs (vision_log/YYYY-MM-DD.jsonl), performing four-view image analyses, supporting add_human_label, and performing local model updates. The SKILL.md and VERIFICATION_PROTOCOL also assert interception of outgoing HTTP requests containing images and forbidding external image exfiltration—but there is no implementation code to enforce these behaviors. The instructions are therefore underspecified and grant broad discretion (file I/O, local model modification, potential network activity) without clear boundaries.

Install Mechanism

No install spec is included (instruction-only), so nothing will be downloaded or written by an installer. That lowers supply-chain risk, but also means the documented Python API is not actually provided by the package as published.

Credentials

The skill requests no environment variables or credentials, yet claims to integrate VLMs, CAD pipelines and to perform model fine-tuning. If external models or services are required, credentials or dependency declarations are missing. Conversely, the skill expects filesystem access to create logs and model weight files—this local persistence is not reflected in declared requirements.

Persistence & Privilege

The skill is not flagged always:true and requests no special privileges, but its documented behavior includes persistent logging (vision_log/...), model weight checks/updates, and a heartbeat check that inspects model file integrity. These are normal for a vision skill, but because there is no code provided the actual persistence/privilege footprint is unknown and should be confirmed.

What to consider before installing

Do not install or enable this skill until the author provides the missing implementation and dependency details. Specific questions and checks to request before use: (1) Provide the Python module/package that implements DiePrevisionCognition (the registry package currently has only docs). (2) List required binaries, Python packages, model weights, and whether any external APIs or VLM/CAD services are called—if so, what credentials are required. (3) Show the code paths that write vision_log/ and perform model updates and confirm how image data is protected (is any image ever sent externally?). (4) Explain how the claimed "interception of outgoing HTTP requests" is implemented locally and provide tests proving no images are exfiltrated. Until you get concrete code and a security review, treat this skill as untrusted: run it in a sandbox, avoid using production or sensitive images, and require a code review for any component that writes or uploads images or modifies model weights.

DiePre Vision Cognition Skill

元数据

字段	值
名称	diepre-vision-cognition
版本	1.0.0
作者	KingOfZhao
发布日期	2026-03-31
置信度	96%

学术参考文献

本视觉框架的技术路线受以下前沿研究启发：

1. Generating CAD Code with Vision-Language Models — VLM生成CAD代码+迭代验证（CADCodeVerify），直接升级照片→DXF管道
2. From 2D CAD to 3D Parametric via VLM — 2D图纸→参数化3D，解决透视矫正和参数化问题
3. Tool-Augmented VLLMs as Generic CAD Task Solvers (ICCV 2025) — VLLM+工具调用做通用CAD，封装OpenCV管道为可调用Skill
4. Efficient Vision-Language-Action Models — VLA高效优化（低延迟+内存优化），适合本地部署
5. Vlaser: Synergistic Embodied Reasoning — 具身推理VLA，未来"照片→动作决策"的理论基础

核心能力

将 DiePre（模切压痕）机器视觉感知与 SOUL 认知框架融合：

1. 视觉已知/未知分离：从图像中提取确定特征（已知）与模糊区域（未知）
2. 文件记忆：每次检测结果写入 vision_log/YYYY-MM-DD.jsonl
3. 四向视觉碰撞：正视角、反转、侧光、整体布局四个维度同时分析
4. 人机闭环质检：AI 初判 → 人类复核 → 标注反馈 → 模型持续进化
5. 置信度质检输出：低于 90% 置信度的缺陷自动升级为人工复核

安装命令

clawhub install diepre-vision-cognition
# 或手动安装
cp -r skills/diepre-vision-cognition ~/.openclaw/skills/

调用方式

from skills.diepre_vision_cognition import DiePrevisionCognition

vision = DiePrevisionCognition(workspace=".")
result = vision.analyze(
    image_path="path/to/dieline.png",
    context={"material": "corrugated", "thickness_mm": 3.0}
)

print(result.confidence)     # 置信度
print(result.defects)        # 检测到的缺陷列表
print(result.collision_log)  # 四向分析详情