← 返回 Skills 市场
Self Improvement Loop
作者
jichangtuijian
· GitHub ↗
· v4.6.13
· MIT-0
325
总下载
0
收藏
1
当前安装
44
版本数
在 OpenClaw 中安装
/install self-improvement-loop
功能描述
Per-agent feedback loop for OpenClaw — captures corrections/errors/features, detects patterns per agent workspace, notifies via per-agent channel bot, and ex...
安全使用建议
What to consider before installing:
- The package will modify your OpenClaw installation: it copies a global hook into ~/.openclaw/hooks, creates per-agent cron jobs, and edits agent AGENTS.md and memory.md files. Back up ~/.openclaw/openclaw.json and agent workspaces first.
- The installer requires python3 and the 'openclaw' CLI (not declared in the registry). Confirm those binaries exist and that you trust their network behavior.
- install.sh will invoke openclaw skill install to fetch two third-party skills (skill-creator and skill-improvement) from https://clawhub.ai — inspect those packages separately before allowing the installs.
- The cron jobs and hook run autonomously and can: scan multiple agent workspaces, write pending notification JSON files, mark learnings entries as notified/resolved/promoted, and (when users reply A/B) invoke automated flows that can create/modify skills. If you need a conservative deployment, run this in a staging environment first or on a restricted agent account.
- If you proceed: review install.sh, scripts/distill.sh, scripts/archive.sh, and scripts/agents-append.md to confirm no unexpected external endpoints or data exfiltration flows. Consider running the installer interactively and verifying each step rather than allowing unattended installs. Disable or manually review the created cron jobs and hook file if you want to limit autonomous effects.
Additional checks that would raise confidence: a signed skill package or a known, audited source for the referenced clawhub.ai slugs; an explicit manifest declaring required binaries/permissions; and a dry-run mode that shows exactly which files will be written/changed.
功能分析
Type: OpenClaw Skill
Name: self-improvement-loop
Version: 4.6.13
The skill implements a complex 'self-improvement' feedback loop that requires high-privilege access, including the ability to manage cron jobs, register system hooks, and modify the global 'openclaw.json' configuration. It intentionally injects instructions into the 'AGENTS.md' and 'memory.md' files of all agents in the system (via install.sh), which creates a significant cross-agent influence surface. While these behaviors are aligned with the stated purpose of tracking errors and suggesting skill upgrades, the broad execution of shell scripts (distill.sh, archive.sh) on user-influenced markdown content and the automated creation of crons that execute AI prompts represent a high-risk architecture. No clear evidence of intentional malice or data exfiltration was found, but the extensive system-level modifications and potential for indirect prompt injection justify a suspicious classification.
能力评估
Purpose & Capability
The skill's name/description (per-agent feedback loop) aligns with what the files implement: a hook that detects corrections, per-agent .learnings directories, per-agent cron jobs, and A/B/C/D flows to act on patterns. However the registry metadata declared no required binaries/env/config, while install.sh depends on python3 and the 'openclaw' CLI and will call gateway APIs and modify ~/.openclaw/openclaw.json and per-agent AGENTS.md/memory.md. That mismatch (undeclared dependencies and config access) is a compliance note — functionally expected for this capability but not proportionally declared.
Instruction Scope
SKILL.md + handlers instruct runtime actions that go beyond passive notification: install.sh will copy a global hook into ~/.openclaw/hooks, create per-agent cron jobs, inject lines into each agent's AGENTS.md and memory.md, and archive/update learnings files. The cron payload and scripts include scanning other skill SKILL.md files (to detect existing skills) and invoking skill-creator / skill-improvement to create or modify skills. While these actions are coherent with an automated feedback loop, they give the skill broad discretion to read and modify agent workspaces and to spawn/install new skills — more privilege than a simple notifier. The instructions also expect network installs (clawhub.ai slugs) and gateway API access.
Install Mechanism
The skill is delivered with its code files (no remote download of the main skill), but install.sh uses the 'openclaw skill install' command to pull additional skills from https://clawhub.ai/<slug>. This is an expected pattern for extensible platforms, but it means installing this skill will trigger further remote installs. No URL shorteners or suspicious IPs are used, but the install step relies on the platform CLI/network and will write files under ~/.openclaw and create crons/hooks.
Credentials
Registry metadata lists no required env vars/credentials, yet the code reads/writes HOME/.openclaw/openclaw.json, writes to ~/.openclaw/hooks, inspects ~/.openclaw/workspace/skills/* and per-agent workspaces, and the cron setup uses a Gateway API (Gateway :18789) per the documentation. install.sh requires python3 and the 'openclaw' CLI. The package will therefore access configuration and credentials implicit to the OpenClaw environment and can modify message routing (bindings). Those capabilities are materially broader than the registry declaration and should be explicitly acknowledged by an installer.
Persistence & Privilege
The installer places a global hook (handler.js) into ~/.openclaw/hooks and creates persistent crons per agent. Those artifacts will run autonomously on schedule (cron jobs) and the hook receives agent events (bootstrap, message:preprocessed). Combined with the ability to call skill-creator/skill-improvement (potentially creating/modifying other skills) and to update openclaw.json bindings, the skill gains lasting, high-impact presence. While not 'always:true', this persistent autonomy plus ability to install/modify skills increases blast radius and warrants careful review and least-privilege controls.
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 self-improvement-loop - 安装完成后，直接呼叫该 Skill 的名称或使用
/self-improvement-loop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
v4.6.13
fix: learnings format 替换逻辑 — Tier 3 检查 Pattern-Key 字段，无则替换模板
v4.6.12
fix: install.sh hook检测跳过 openclaw hooks list（挂起问题）; fix: archive.sh pending_dir 硬编码修复（argv传递）
v4.6.7
self-improvement-loop v4.6.7
- Metadata updated for new version release (no code or documentation changes).
- Version number incremented from 4.6.6 to 4.6.7 in _meta.json.
v4.6.6
setup_crons.py: 修复with块内缩进错误（cfg/accounts/for循环应缩进12格）；删除所有Windows换行符
v4.6.5
setup_crons.py：修复moling硬编码的语法错误（缩进问题）
v4.6.4
setup_crons.py：删除moling硬编码账户名，改为动态读accounts.keys()；cron-payloads.json：删除真实cron job id字段
v4.6.3
**Improved notification reliability, permission transparency, and matching logic.**
- Added entry_id field to pending notification JSON; write_notified.py prioritizes exact MD title ID, preventing incorrect skill matches.
- AGENTS.md injection now prompts for confirmation, allowing users to skip with ‘s’ for better safety.
- Memory distillation and notification cron are now hourly instead of every 30 minutes.
- "match-existing-skill.sh" script removed; skill-matching handled directly via SKILL.md scanning.
- Permissions declaration added for clarity, outlining exact read/write/exec/cron requirements.
- Minor flow and documentation updates for clearer user experience and integration guidance.
v4.6.1
self-improvement-loop v4.6.1
- Updated skill and learnings documentation for improved clarity and structure.
- Refined archive and notification scripts for enhanced reliability and silent operation.
- Improved memory distillation and pattern detection processes.
- Minor script and config adjustments to streamline closed-loop handling and reduce potential for orphaned entries.
v4.5.1
self-improvement-loop v4.5.1
- Updated documentation in SKILL.md for improved clarity and flow; restructuring and minor clarifications.
- Expanded and cleaned up explanatory sections for notification outcomes, fallback, and flow diagrams.
- No functional or behavioral logic changes to scripts or core loop; update is documentation-focused.
- Internal references and examples remain consistent with previous version for stability.
v4.5.0
self-improvement-loop v4.5.0
- Updated to version 4.5.0.
- SKILL.md revised (contents consistent; version incremented).
- Minor maintenance: version bump, no functional or feature changes detected.
v4.4.9
- Improved documentation in SKILL.md for clarity and formatting.
- Minor maintenance updates to scripts/distill_json.py (details not shown).
- No changes to core logic or functionality.
v4.4.8
**Pattern-Key format validation and notification improvements**
- Pattern-Key format is now validated: keys not matching the `X.Y.Z` structure are aggregated by category and marked as `pk_format_invalid: true` in JSON output, prompting AI or user correction.
- Documentation clarifies that the user option set is A/B/C/D (was previously ambiguous in places).
- Batch processing/fallback mechanism: handlers now scan all pending notification files to avoid orphaned patterns if multiple are triggered before user action.
- Expanded and clarified documentation, including new English-language flowcharts and explicit feature descriptions.
- No breaking changes to core loop logic or deployment scripts.
v4.4.7
- Documentation updated: clarified that 主代理主动写入 (active detection) depends on workspace configuration (HEARTBEAT.md).
- No changes to code or core logic; this version improves process transparency in the SKILL.md.
v4.4.6
self-improvement-loop v4.4.6
- Expanded the core decision loop and notification options to include “D” (Promote), covering A/B/C/D user replies.
- Added and documented the complete 闭环流程图 (“full closed-loop flow chart”) in SKILL.md, detailing all feedback paths and status transitions.
- Updated detection flow: Hook now pushes AI reminders for writing learnings instead of directly writing them.
- Refined documentation for technical workflow, user interaction, and multi-path experience ingestion.
- Removed files related to previous self-reflection and heartbeat handling for simplification and clarity.
v4.4.1
v4.4.1 adds a new "D" (Promotion) option for user decision handling.
- Added D (Promotion) response: users can now choose to promote patterns to SOUL/AGENTS/TOOLS.
- Updated documentation to reference A/B/C/D options (previously A/B/C only).
- Response table and handling updated: "User picks D | dormant, promote".
- Minor clarifications in notification handling.
v4.4.0
self-improvement-loop v4.4.0
- Added 闭环流程图.md reference file.
- Documented the role of memory-daily-distill as secondary in SKILL.md.
v4.3.9
- Features section reorganized and merged in documentation for clarity.
- All filesystem paths now consistently use `$HOME`, `~`, or `os.path.expanduser()` (no hardcoded absolute paths).
- Changelog updated to reflect merged features documentation and path changes.
v4.3.8
v4.3.8: Hardcoded paths replaced for improved portability
- All scripts now use `$HOME` or `os.path.expanduser` instead of hardcoded absolute paths.
- File reading and writing operations now correctly resolve user home directories.
- Enhances compatibility across different environments.
v4.3.7
self-improvement-loop v4.3.7
- install.sh now appends the A/B/C handler to AGENTS.md upon installation
- Added scripts/agents-append.md for modular handler updates
- Improved documentation to reflect new A/B/C handler location in AGENTS.md
v4.3.6
self-improvement-loop v4.3.6
- The A/B/C notification handler is now documented in AGENTS.md instead of HEARTBEAT.md.
- install.sh no longer appends the A/B/C section to HEARTBEAT.md.
- Documentation updates to reflect the above changes.
元数据
常见问题
Self Improvement Loop 是什么？
Per-agent feedback loop for OpenClaw — captures corrections/errors/features, detects patterns per agent workspace, notifies via per-agent channel bot, and ex... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 325 次。
如何安装 Self Improvement Loop？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install self-improvement-loop」即可一键安装，无需额外配置。
Self Improvement Loop 是免费的吗？
是的，Self Improvement Loop 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Self Improvement Loop 支持哪些平台？
Self Improvement Loop 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 Self Improvement Loop？
由 jichangtuijian（@jichangtuijie）开发并维护，当前版本 v4.6.13。
推荐 Skills