Multi-Skill-Eval | 集成化技能评估系统

集成化多方法技能评估系统。整合静态分析(skill-assessment)、Rubric质量打分(skill-evaluator)和自主基准测试(skill-eval)。用于全面评估、对比、审计或改进OpenClaw技能。覆盖文档完整性、代码质量、25项Rubric打分、多模型基准测试。 触发词(中文): 评估技...

This package appears coherent for its stated purpose: it runs local static checks, rubric grading, and an agent-driven benchmark. Before installing/running: (1) review the scripts (especially the truncated ones) for any network calls, subprocess calls, or file-write operations you don't expect; (2) run in an isolated environment (container or VM) if you plan to evaluate untrusted skills; (3) when using the benchmark mode, be aware it will cause the agent to read the full target skill directory and spawn subagents — do not point it at directories containing credentials or sensitive secrets. If you want a higher-confidence assessment, provide the remaining script contents (static-analyze.py and the truncated parts) so I can check for external network calls, subprocess.exec/ shell injection, or hardcoded secrets.

Type: OpenClaw Skill Name: multi-skill-eval Version: 1.0.2 The bundle is a comprehensive toolkit designed for auditing, benchmarking, and improving OpenClaw skills. It contains several Python scripts (e.g., static-analyze.py, eval-skill.py, grade-assertions.py) that perform legitimate static analysis, such as checking for hardcoded secrets, verifying Python syntax, and ensuring documentation completeness. The logic is transparent, uses standard libraries, and aligns with the stated purpose of a skill evaluation system. While the documentation mentions a 'Self-Evolution' feature for rewriting skill instructions, it is explicitly marked as 'not yet implemented,' and the existing code contains no evidence of data exfiltration, unauthorized remote execution, or malicious intent.