← 返回 Skills 市场
Reddi Self Improving Agent
作者
Nissan Dookeran
· GitHub ↗
· v1.0.13
· MIT-0
272
总下载
0
收藏
0
当前安装
2
版本数
在 OpenClaw 中安装
/install reddi-self-improving-agent
功能描述
reddi.tech fork of self-improving-agent. Captures learnings, errors, and corrections to enable continuous improvement. Use when: a command fails, user correc...
安全使用建议
This package appears to do what it says: capture learnings/errors locally and optionally inject reminders via OpenClaw hooks. Before installing, review and be comfortable with the following: (1) it will write files into your OpenClaw workspace and may copy hook scripts into ~/.openclaw/hooks if you follow the instructions—only enable hooks you trust; (2) the scripts are readable and do not request secrets, but they will run with the agent's file-system permissions once enabled; (3) the metadata lists docker as required though bundled scripts don't invoke it — install docker only if you plan to use containerized learning pipelines referenced in the docs; (4) the SKILL.md recommends cloning from a GitHub repo if you use the upstream source — verify that repo first; and (5) if you want to limit impact, use the minimal setup option (only the activator reminder) or run extract-skill.sh with --dry-run to inspect what would be created. If you want a deeper check, provide the specific OpenClaw/OpenAI runtime environment (what env vars the host will expose, and whether hooks will run automatically) and I can point out exact lines to audit.
功能分析
Type: OpenClaw Skill
Name: reddi-self-improving-agent
Version: 1.0.13
The skill implements a 'self-improvement' mechanism that instructs the agent to modify its own core instruction files (e.g., SOUL.md, AGENTS.md, TOOLS.md) based on external inputs like command errors and user feedback. This creates a significant surface for indirect prompt injection, where malicious command output could be 'promoted' into the agent's permanent system context to alter future behavior. While the provided scripts (extract-skill.sh, error-detector.sh) and hooks are functionally benign and include basic path-traversal protections, the architectural design of allowing an AI to rewrite its own 'system prompts' is a high-risk vulnerability.
能力评估
Purpose & Capability
The skill's stated purpose is to capture learnings/errors to local markdown files and promote important items into workspace docs; the SKILL.md, hooks, and scripts all align with that. One small mismatch: metadata declares docker as a required binary, but none of the included hook scripts (activator.sh, error-detector.sh, extract-skill.sh) or handlers actually invoke docker. The README/content references docker use cases (containerized learning pipelines and examples), so the docker requirement is plausible but not strictly necessary for the shipped scripts.
Instruction Scope
Runtime instructions focus on creating and updating local .learnings/ and workspace files, copying/enabling OpenClaw hooks, and optionally running helper scripts. The hook handlers inject a virtual bootstrap file (SELF_IMPROVEMENT_REMINDER.md) into the agent context — this is consistent with the stated goal. The error-detector reads CLAUDE_TOOL_OUTPUT (an environment variable supplied by the host agent) to detect failing commands; this is documented in the SKILL.md. The instructions do require filesystem writes (creating ~/.openclaw/workspace/.learnings, copying hook files, creating skills via extract-skill.sh) but do not instruct reading unrelated secrets or contacting external endpoints.
Install Mechanism
There is no automated install spec in the registry (instruction-only), which minimizes automatic code execution. The SKILL.md recommends manual installation via ClawdHub or a git clone from a public GitHub repository — a normal, traceable pattern. All included scripts are present in the skill bundle (no download-from-URL or opaque extract steps).
Credentials
The skill declares no required environment variables or credentials. The only environment variable referenced at runtime is CLAUDE_TOOL_OUTPUT (used by the error-detector hook) and that is an expected host-provided value for tool output. The included scripts explicitly avoid requesting secrets and extract-skill.sh contains checks preventing absolute or parent (..) output paths. No tokens/keys/passwords are requested.
Persistence & Privilege
The skill does not request always:true and is user-invocable only. However, enabling the optional hooks will cause code (the hook scripts/handlers) to run as part of your OpenClaw agent lifecycle and will write files into your home/workspace (e.g., ~/.openclaw/hooks, ~/.openclaw/workspace/.learnings or ./skills/). This is expected for an integration of this kind but is a privilege worth noting: installed hooks run with the same permissions as the agent user.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install reddi-self-improving-agent - 安装完成后，直接呼叫该 Skill 的名称或使用
/reddi-self-improving-agent触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v1.0.13
- Updated SKILL.md to include explicit OpenClaw-specific metadata such as required bins (docker), network settings, and security notes.
- Streamlined the skill's description and moved technical metadata into a structured YAML block.
- Minor edits for clarity and brevity were made in the skill description.
- No changes to functionality or core usage; documentation and metadata improvement only.
v1.0.12
- Expanded and clarified documentation in SKILL.md, detailing how to log learnings, errors, and feature requests for continuous self-improvement.
- Added comprehensive quick reference tables for when and how to log each learning type, including promotion workflow.
- Provided explicit instructions and file structures for OpenClaw and generic agent setup.
- Included detailed templates and logging formats for learnings, errors, and feature request entries.
- Explained promotion targets and criteria for moving important learnings into project-wide memory files.
- Offered optional session hooks and inter-session communication support for broader adoption and utility.
元数据
常见问题
Reddi Self Improving Agent 是什么？
reddi.tech fork of self-improving-agent. Captures learnings, errors, and corrections to enable continuous improvement. Use when: a command fails, user correc... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 272 次。
如何安装 Reddi Self Improving Agent？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install reddi-self-improving-agent」即可一键安装，无需额外配置。
Reddi Self Improving Agent 是免费的吗？
是的，Reddi Self Improving Agent 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Reddi Self Improving Agent 支持哪些平台？
Reddi Self Improving Agent 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Reddi Self Improving Agent？
由 Nissan Dookeran（@nissan）开发并维护，当前版本 v1.0.13。
推荐 Skills