ClawHub

tool-skill-manager

v0.1.0
Download

Inspect real model-visible tools and skills, then enable or disable them from chat

tool-skill-manager·runtime tool-skill-manager·by @neaoo
Code Pluginsource linkedCommunity code plugin. Review compatibility and verification before install.

README

Tool & Skill Manager

Tool & Skill Manager hero banner

OpenClaw Plugin Prompt Surface Audit Context Cleanup Bilingual README

Inspect the exact tools and skills the model can really see, then remove the noise from chat.
直接从模型视角审计 tools / skills，并在聊天里清理掉那些长期污染上下文的噪音。

If OpenClaw feels less smart, audit the prompt surface before blaming the model.
如果 OpenClaw 看起来变钝了，先看它到底背了多少乱七八糟的工具和技能。

Preview • Quick Start • Why It Matters • 中文说明 • English Guide

Preview

Before and after cleanup preview

Why This Matters

OpenClaw often gets "worse" for a very ordinary reason: context clutter.

  • too many tools installed just to try them once
  • too many skills with overlapping descriptions
  • too many outdated experiments still present in the system prompt
  • too many choices competing for the model's routing attention

What looks like model weakness is often prompt pollution:

  • wrong tool calls
  • missed skill usage
  • unstable routing
  • inconsistent answers

Tool & Skill Manager gives you a model-perspective audit trail for that exact problem. It captures the real tool list and skill list that were visible at prompt-build time, then lets you disable or re-enable them directly from chat.

Quick Workflow

Capture inspect clean workflow

Quick Start

Run these commands in a terminal on the machine where OpenClaw is installed and running. Do not run them inside chat.

git clone https://github.com/NeAoo/Tool-Skill-Manager.git
cd Tool-Skill-Manager
openclaw plugins install .
openclaw plugins enable tool-skill-manager
openclaw gateway restart

Then use it from chat:

/tools
/skills
/tools off 3 5
/skills off "Agent Browser" summarize

If there is no real snapshot yet, /tools or /skills will auto-generate a minimal bootstrap snapshot using the configured default model. The next real run replaces it. After /tools off/on or /skills off/on, the plugin refreshes the snapshot automatically in the background.

What You Get

Real Prompt Surface
Captures what the model actually saw at llm_input, not what you assume was loaded from config. 		Chat-Native Cleanup
Disable or re-enable noisy tools and skills without leaving chat or hand-editing large config files. 		Snapshot History
Keeps latest.json, LATEST.md, and catalog.json for inspection and later cleanup.
Tools + Skills
Audits both tool visibility and skill visibility, because either one can pollute the prompt surface. 		Fast Triage
See the visible list, hide the worst offenders, run again, and confirm the context got cleaner. 		No Guesswork
Great for cases where the right tool exists, but the model keeps choosing the wrong one anyway.

Why This Is Better Than Reading Config

Your config is not the final prompt.

OpenClaw can assemble the final model-visible prompt from multiple places:

  • openclaw.json
  • installed tools
  • installed skills
  • plugin state
  • runtime defaults

This plugin inspects the model-visible result after prompt assembly, which is the thing that actually affects behavior.

flowchart TD
    A[openclaw.json] --> E[Prompt builder]
    B[Installed tools] --> E
    C[Installed skills] --> E
    D[Plugin and runtime state] --> E
    E --> F[Final system prompt seen by the model]
    F --> G[Tool and Skill Manager captures the visible list]

Commands

Tools

  • /tools
  • /tools detail
  • /tools off 3 5
  • /tools off feishu_sheet feishu_task_task
  • /tools on 1 2
  • /tools help

Skills

  • /skills
  • /skills detail
  • /skills off 2 4
  • /skills off "Agent Browser" summarize
  • /skills on 1 2
  • /skills help

Command Rules

  • /tools and /skills default to the latest detail view
  • off uses indexes from Model-Visible ...
  • on uses indexes from Model-Hidden ...
  • you can target by number or by name

Files Written By The Plugin

Default output directory:

~/.openclaw/logs/tool-skill-manager

Default files:

  • latest.json
  • LATEST.md
  • catalog.json

How It Works

  • uses OpenClaw's official llm_input hook
  • parses the model-visible tools from the captured system prompt
  • parses the model-visible skills from the captured system prompt
  • stores the latest snapshot and a small seen-before catalog
  • updates tools.deny and skills.entries.*.enabled

中文说明

这个插件到底是干什么的

Tool & Skill Manager 不是一个普通的配置查看器，而是一个从模型视角做排查的审计插件。

它会在 OpenClaw 构建最终 prompt 的时候，抓取模型真正能看到的 tools 和 skills，然后让你直接在聊天里把这些内容查看清楚、关闭掉、再重新启用。

它的核心价值不是“帮你看配置”，而是回答这个更关键的问题：

模型这一次真正看到了什么？到底是什么在长期影响它的工具选择？

为什么很多 OpenClaw 会越用越钝

很常见的原因不是模型本身，而是上下文开始脏了。

  • 之前为了玩、为了试新东西，装了很多 tools 和 skills
  • 有些描述高度重叠，模型容易在相似候选项里选错
  • 有些已经不再使用，但仍然长期留在 prompt 里
  • 工具和技能越堆越多，模型的决策表面越来越拥挤

于是就会出现这些现象：

  • 明明有正确工具，却总是调用错
  • 明明有对应 skill，却没有触发
  • 回答开始绕路、犹豫、风格不稳定

这时候你需要的不是再猜配置，而是直接看看模型到底看到了什么。

你可以怎么用

安装并启用插件后，直接执行：

/tools
/skills

如果你发现里面有很多以前装着玩的、早就不用的、或者名称描述高度重叠的项目，就可以直接清理：

/tools off 3 5
/tools off feishu_sheet feishu_task_task
/tools on 1 2

/skills off 2 4
/skills off "Agent Browser" summarize
/skills on 1 2

一个非常实用的清理流程

  1. 先跑 /tools 或 /skills；如果还没有真实快照，插件会自动生成一个最小 bootstrap 快照。
  2. 跑 /tools detail，看模型当前到底背了多少工具。
  3. 跑 /skills detail，看模型当前到底背了多少技能。
  4. 关掉那些只装过一次、长期不用、含义重复、明显会干扰路由的项目。
  5. 直接再跑 /tools 和 /skills，查看自动刷新的最新结果。
  6. 对比前后列表，确认上下文真的变干净了。

为什么它比直接看配置更有用

因为配置文件不等于最终 prompt。

真正影响模型行为的是 prompt 组装完成之后的结果，而不是你以为它加载了什么。这个插件抓的是模型真正可见的最终表面，所以能更准确地帮你定位问题。

命令一览

Tools

  • /tools
  • /tools detail
  • /tools off 3 5
  • /tools off feishu_sheet feishu_task_task
  • /tools on 1 2
  • /tools help

Skills

  • /skills
  • /skills detail
  • /skills off 2 4
  • /skills off "Agent Browser" summarize
  • /skills on 1 2
  • /skills help

快照与状态文件

默认会写到：

~/.openclaw/logs/tool-skill-manager

会生成：

  • latest.json
  • LATEST.md
  • catalog.json

如果当前还没有真实快照，执行 /tools 或 /skills 时会自动生成一个最小化的 bootstrap 快照，并在下一次真实运行后替换。

适合谁

  • 装过很多插件、工具、技能，已经记不清谁在生效
  • 怀疑模型经常选错工具，但一直缺少证据
  • 想从模型视角而不是配置视角排查问题
  • 想给 OpenClaw 做一次真正的上下文瘦身

一句话总结

如果你的 OpenClaw 看起来变笨了，先别急着换模型。

先看它到底背了多少你早就不需要、却还在持续影响它判断的 tools 和 skills。

English Guide

What This Plugin Actually Does

Tool & Skill Manager is not just a config viewer. It is a model-perspective audit plugin.

It captures the exact tools and skills that were actually visible to the model during prompt assembly, then lets you inspect, disable, and re-enable them directly from chat.

It answers the question that matters most:

What did the model really see, and which old tools or skills are still affecting routing quality?

Why OpenClaw Can Feel Less Smart Over Time

Often the problem is not the model. The problem is context clutter.

  • too many tools installed for one-off experiments
  • too many skills competing with similar descriptions
  • old prompt baggage that never got cleaned up
  • an overcrowded routing surface for the model

That usually shows up as:

  • wrong tool calls
  • missed skill usage
  • hesitant or unstable behavior
  • the feeling that the agent got "dumber"

This plugin helps you inspect and clean that surface from the model's point of view.

Basic Usage

Send one normal message first so the plugin can capture a real run, then use:

/tools
/skills

When you find stale or noisy entries, clean them directly from chat:

/tools off 3 5
/tools off feishu_sheet feishu_task_task
/tools on 1 2

/skills off 2 4
/skills off "Agent Browser" summarize
/skills on 1 2

A Practical Cleanup Loop

  1. Run /tools or /skills; if no real snapshot exists yet, the plugin will generate a minimal bootstrap snapshot automatically.
  2. Run /tools detail and inspect the current visible tools.
  3. Run /skills detail and inspect the current visible skills.
  4. Disable old, redundant, or confusing entries.
  5. Run /tools and /skills again and inspect the automatically refreshed lists.
  6. Confirm the prompt surface is cleaner than before.

Snapshot Behavior

The plugin stores its latest snapshot and a small historical catalog under:

~/.openclaw/logs/tool-skill-manager

By default it writes:

  • latest.json
  • LATEST.md
  • catalog.json

If no real snapshot exists yet, /tools or /skills will auto-generate a minimal bootstrap snapshot using the configured default model. That bootstrap snapshot is clearly labeled and replaced by the next real captured run. After /tools off/on or /skills off/on, the plugin refreshes the snapshot automatically in the background.

Install

Run these commands in a terminal on the machine where OpenClaw is installed and running. Do not run them inside chat.

git clone https://github.com/NeAoo/Tool-Skill-Manager.git
cd Tool-Skill-Manager
openclaw plugins install .
openclaw plugins enable tool-skill-manager
openclaw gateway restart

License

MIT

Capabilities

Tags
executes-code
configSchema
Yes
Executes code
Yes
HTTP routes
0
Runtime ID
tool-skill-manager

Compatibility

Built With Open Claw Version
0.1.0
Plugin Api Range
2026.3.23-2

Verification

Tier
source linked
Scope
artifact only
Summary
Validated package structure and linked the release to source metadata.
Source
github.com/NeAoo/Tool-Skill-Manager
Commit
7e5115a2d50c
Tag
main
Provenance
No
Scan status
clean

Tags

latest
0.1.0