Skill 自动评估和改进管线。9 维结构评分（含 LLM-as-Judge）、4 角色加权、 类别修正系数（tool/knowledge/orchestration/rule）、Pareto front 回归保护 （security 2%/efficiency 10%/其他 5%）、trace-aware 失败...

What to consider before installing/running this skill: - Missing declarations: The registry says 'no required env vars / config paths', but the code and docs expect LLM credentials (e.g., ANTHROPIC_API_KEY or a 'claude -p' CLI) and access to your local Claude session logs (~/.claude/projects). Ask the publisher to document required secrets and file access clearly before use. - Sensitive reads: The session-feedback analyzer reads ~/.claude/projects JSONL files. Those logs may contain private prompts, code, or data. Do not run this skill on a machine with sensitive Claude sessions unless you audit exactly what it reads and stores. - File modifications: The executor can apply edits to SKILL.md files and will create backups/receipts. Review the execute/rollback code and test in a sandbox or a copy of your skills directory to confirm behavior and backup integrity before running on real skill files. - LLM/network calls: The evaluator/discriminator call LLM backends (claude CLI or Anthropic API). Evaluate privacy/cost implications and consider using mock/backends or offline testing first. - Prompt-injection indicator: The SKILL.md contains unicode control characters flagged by a static check. Inspect the SKILL.md for invisible characters and remove or justify them; such characters can cause unexpected parsing when documents are used as prompts. - Least privilege: Provide only the minimum credentials and limit filesystem scope (run in a container or on a copy) until you validate the behavior. Prefer read-only dry-run modes and use '--dry-run' where available. - Review and test: Because the repository is large and powerful, run the pipeline on non-production examples, read the code paths that perform subprocess calls and file writes (improvement-executor, autoloop-controller, session-feedback-analyzer), and ask the maintainer for an explicit list of env vars, required binaries/CLIs, and config paths. If you want, I can extract the exact lines that reference ANTHROPIC_API_KEY, the paths under ~/.claude, and the code paths that perform file writes so you can inspect them before running.

Type: OpenClaw Skill Name: auto-improvement-orchestrator-skill Version: 1.0.3 The bundle is a highly sophisticated and well-engineered orchestration system designed for the automated evaluation and iterative improvement of AI agent skills. It implements a complex five-stage pipeline (Propose, Discriminate, Evaluate, Execute, and Gate) that includes advanced features such as Pareto front regression protection (lib/pareto.py), multi-role blind panel scoring (score.py), and Karpathy-style self-improvement loops (self_improve.py). While the system utilizes high-privilege capabilities—including modifying local skill files (execute.py), reading Claude session logs for implicit feedback (analyze.py), and executing AI-generated prompts via the 'claude' CLI (task_runner.py)—these behaviors are transparently documented and strictly necessary for the stated goal of skill optimization. The presence of security-conscious design patterns, such as path traversal validation in the PytestJudge (judges.py) and explicit opt-out checks for log analysis, confirms that the bundle is a legitimate developer tool rather than malicious software.