Improvement Learner
v1.1.2当需要检查 skill 质量评分、自动优化 SKILL.md 结构、追踪评估分数变化趋势、或「评分低了想知道哪里扣分」时使用。6维结构评估 + HOT/WARM/COLD 三层记忆 + Pareto front。不用于候选语义打分（用 improvement-discriminator）或全流程编排（用 impr...
⭐ 0· 64·1 current·1 all-time
by_silhouette@lanyasheng
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
OpenClaw
Benign
high confidencePurpose & Capability
Name/description match the code and SKILL.md: scripts implement an evaluation + self-improvement loop, tracking, and reporting. Required capabilities (LLM judge fallback, file I/O, local memory storage) are consistent with the described functionality.
Instruction Scope
Instructions and CLI flags are focused on reading a skill directory, evaluating SKILL.md, producing JSON reports, and writing local memory. Two points to note: (1) the code will call a local 'claude' CLI (if available) and pass SKILL.md content to it (LLM judge), which means your SKILL.md may be sent to whatever backend the 'claude' CLI is configured to use; (2) the script inserts a computed repository root into sys.path (parents[3]) to import lib.* modules — if you run the script from a different repository layout, this may cause imports from unexpected locations.
Install Mechanism
Instruction-only with included Python scripts; there is no installer that downloads or executes remote archives. No brew/npm/external download steps. Risk surface is limited to the included code being executed locally.
Credentials
The skill declares no required env vars or credentials. However, optional behavior depends on an external LLM CLI ('claude') if present — that CLI typically uses its own credentials/config (not declared by the skill). If you want to avoid sending content externally, run with --mock or ensure no LLM CLI is on PATH.
Persistence & Privilege
always:false and user-invocable:true. The scripts create local memory and reports under directories you pass (memory-dir/state-root/reports). They do not request system-wide or other-skills' config changes. Backup/revert helpers in tests are expected for this functionality.
Assessment
This skill appears to do what it claims (evaluate and iteratively improve SKILL.md). Before running it: 1) Inspect scripts/self_improve.py yourself (you already have the file) and confirm you are comfortable with the file I/O paths it will access. 2) If you do not want any SKILL.md content sent to a remote LLM, either run with --mock or ensure the 'claude' CLI is not installed/available in PATH (the script falls back to a regex mode when 'claude' is missing or --mock is used). 3) Note the script adds a repository-root path (parents[3]) to sys.path to import lib.* — run it from the intended repository layout or review the imported lib modules to ensure they are trusted. 4) Run in an isolated environment (container/VM) or point --skill-path to a copy of the skill to avoid accidental modification of production files. 5) If you need stricter guarantees about no network I/O, run the tool with network disabled or inspect for CLI usage first; otherwise the behavior is coherent with its stated purpose.Like a lobster shell, security has layers — review code before you run it.
latestvk97a74dspr9ay17mtwg6mm09ph84cj7g
License
MIT-0
Free to use, modify, and redistribute. No attribution required.