tokenQrusher:优化 AI 智能体上下文并降低成本 - Openclaw Skills
作者:互联网
2026-04-17
什么是 tokenQrusher?
tokenQrusher 是专为 Openclaw Skills 生态系统设计的专用工具,旨在解决与大语言模型 API 使用相关的高昂成本问题。它作为中间层运行,拦截智能体引导事件,确保仅将最相关的上下文发送到模型。通过将用户提示分为不同的复杂度级别,对于简单交互,它可以减少高达 99% 的 Token 开销,同时保留复杂工程任务的完整能力。
除上下文管理外,该技能还优化了 AI 智能体的后台心跳频率。这可以防止在进行电子邮件或日历坚控等周期性检查时产生不必要的 API 调用,从而进一步降低成本并提高系统效率。tokenQrusher 在构建时充分考虑了性能,以亚毫秒级的延迟运行,并使用确定性逻辑确保本地环境内可靠、线程安全的执行。
下载入口:https://github.com/openclaw/skills/tree/main/skills/qsmtco/tokenqrusher
安装与下载
1. ClawHub CLI
从源直接安装技能的最快方式。
npx clawhub@latest install tokenqrusher
2. 手动安装
将技能文件夹复制到以下位置之一
全局模式~/.openclaw/skills/
工作区
/skills/
优先级:工作区 > 本地 > 内置
3. 提示词安装
将此提示词复制到 OpenClaw 即可自动安装。
请帮我使用 Clawhub 安装 tokenqrusher。如果尚未安装 Clawhub,请先安装(npm i -g clawhub)。
tokenQrusher 应用场景
- 在 GPT-4 或 Claude 3 Opus 等昂贵模型上运行 AI 智能体时降低运营成本。
- 通过减少大模型处理简单查询时所需的上下文量来缩短响应时间。
- 通过智能间隔心跳和坚控调用来管理 API 速率限制。
- 根据任务选择性地加载身份或工具文件,自定义智能体的人设可见性。
- 为自动化智能体实施静默时间,防止非工作时段的后台活动。
- 系统在每条用户消息的 agent:bootstrap 事件期间触发 token-context 钩子。
- 分类引擎分析用户提示,以确定其是简单、标准还是复杂级别。
- 根据提示复杂度,该技能会过滤工作区文件列表,仅包含本地配置中定义的文件。
- 同时,token-heartbeat 钩子根据自定义间隔和静默时间设置,评估计划的后台任务是否到期。
- 如果检查尚未到期或处于静默时间内,系统将返回跳过 API 调用的状态。
- 最终优化的上下文和心跳计划将传回网关供智能体执行。
tokenQrusher 配置指南
要将此优化层集成到您的 Openclaw Skills 设置中,请遵循以下安装步骤:
# 从枢纽安装技能
clawhub install tokenQrusher
# 启用上下文和心跳钩子
tokenqrusher install --hooks
# 验证安装状态
tokenqrusher status
# 重启 OpenClaw 网关以应用优化
openclaw gateway restart
确保您已安装 python3 和 node 作为核心模块的先决条件。
tokenQrusher 数据架构与分类体系
该技能利用本地 JSON 配置文件来定义优化规则,确保数据绝不会离开您的系统。这符合 Openclaw Skills 的安全标准。
| 配置类型 | 路径 | 关键设置 |
|---|---|---|
| 上下文过滤器 | ~/.openclaw/hooks/token-context/config.json |
简单、标准和复杂模式的文件数组。 |
| 心跳优化器 | ~/.openclaw/hooks/token-heartbeat/config.json |
电子邮件、日历和坚控的间隔设置(以秒为单位)。 |
| 共享逻辑 | token-shared 模块 |
分类逻辑和路径穿越保护模式。 |
name: tokenQrusher
description: Token optimization system for OpenClaw reducing costs 50-80%
version: 2.1.1
author: qsmtco
license: MIT
homepage: https://github.com/qsmtco/tokenQrusher
metadata:
openclaw:
requires:
bins:
- python3
- node
emoji: "??"
tokenQrusher Skill
Overview
tokenQrusher reduces OpenClaw API costs by 50-80% through:
- Context Filtering – Loads only necessary workspace files (up to 99% reduction for simple messages)
- Heartbeat Optimization – Reduces heartbeat API calls by 75%
This is the simplified, production-ready core. Non-functional advisory components have been removed.
Components
1. Context Hook (token-context)
Event: agent:bootstrap
Filters which workspace files are loaded based on message complexity.
Config: ~/.openclaw/hooks/token-context/config.json
{
"enabled": true,
"logLevel": "info",
"dryRun": false,
"files": {
"simple": ["SOUL.md", "IDENTITY.md"],
"standard": ["SOUL.md", "IDENTITY.md", "USER.md"],
"complex": ["SOUL.md", "IDENTITY.md", "USER.md", "TOOLS.md", "AGENTS.md", "MEMORY.md", "HEARTBEAT.md"]
}
}
How it works:
- Extracts the user's latest message
- Classifies complexity (simple/standard/complex)
- Keeps only allowed files, discards the rest
- Sets
context.bootstrapFilesto the filtered list
Savings:
- Simple greetings → 2 files (99% token reduction)
- Standard tasks → 3 files (90%+ reduction)
- Complex tasks → all files (full context)
Rationale: Simple messages don’t need documentation, memory logs, or tool references. Only identity and personality are required.
2. Heartbeat Optimizer (token-heartbeat)
Event: agent:bootstrap (for heartbeat polls)
Optimizes heartbeat check schedule to dramatically reduce API calls.
Config: ~/.openclaw/hooks/token-heartbeat/config.json
{
"enabled": true,
"intervals": {
"email": 7200,
"calendar": 14400,
"weather": 14400,
"monitoring": 7200
},
"quietHours": { "start": 23, "end": 8 }
}
How it works:
- Determines if enough time has elapsed since last check
- Skips checks that are not due
- Honors quiet hours (23:00–08:00 by default)
- Returns
HEARTBEAT_OKwhen nothing needs checking
Optimization Table:
| Check | Default Interval | Optimized | Reduction |
|---|---|---|---|
| 60 min | 120 min | 50% | |
| Calendar | 60 min | 240 min | 75% |
| Weather | 60 min | 240 min | 75% |
| Monitoring | 30 min | 120 min | 75% |
Result: 48 checks/day → 12 checks/day (75% fewer API calls)
3. Shared Module (token-shared)
Pure functions used by both hooks:
classifyComplexity(message)→ complexity levelgetAllowedFiles(level, config)→ file listisValidFileName(name)→ path traversal protectionloadConfigCached(logFn)→ 60s TTL config caching- Maybe/Either patterns (no exceptions for control flow)
External Endpoints
This skill does not call any external network endpoints. All operations are local to your machine.
Security & Privacy
- 100% local execution – No data leaves your system.
- No external telemetry – No calls to third?party servers.
- Configuration read only from local hook configs.
- File validation prevents path traversal attacks.
- No environment secrets required.
Model Invocation Note
Hooks run automatically:
token-contextruns on everyagent:bootstrap(every user message).token-heartbeatruns on heartbeat polls. No manual intervention needed after enabling the hooks.
Trust Statement
By installing this skill, you trust that the code operates locally and does not transmit your workspace data. Review the open?source implementation on GitHub before installing.
CLI Commands
After installation, the tokenqrusher command is available.
tokenqrusher context
Recommends which context files should be loaded for a given prompt.
$ tokenqrusher context "hi"
Complexity: simple (confidence: 95%)
Files: SOUL.md, IDENTITY.md
Savings: 71%
tokenqrusher status [--verbose]
Shows system status.
$ tokenqrusher status
=== tokenQrusher Status ===
Hooks:
? token-context (Filters context)
? token-heartbeat (Optimizes heartbeat)
tokenqrusher install [--hooks] [--all]
Installs/enables hooks.
$ tokenqrusher install --hooks
? Enabled: token-context
? Enabled: token-heartbeat
Configuration
All hook configs are JSON files stored in: ~/.openclaw/hooks/
You can edit these to customize behavior (e.g., adjust file lists, intervals, or quiet hours). Changes are reloaded after 60 seconds (config cache TTL).
Design Principles
- Deterministic – Same input → same output.
- Pure functions – No hidden side effects.
- Immutability – Frozen data structures.
- No exceptions for control flow – Uses Result/Either types.
- Thread?safe – RLock protected shared state.
- Exhaustive typing – Full type hints.
Performance
| Operation | Latency | Memory |
|---|---|---|
| Context classification | <1?ms | <1?MB |
| Heartbeat check | <0.5?ms | <0.5?MB |
Negligible overhead per agent message.
Troubleshooting
Hook not loading?
# Check status
openclaw hooks list
# Should show "? ready" next to token-context and token-heartbeat
# Enable if missing
openclaw hooks enable token-context
openclaw hooks enable token-heartbeat
# Restart gateway
openclaw gateway restart
Changes not taking effect?
Config cache TTL is 60?seconds. Restart the gateway for immediate effect.
Migration from v2.0.x
v2.1.0 removes non?functional components. If upgrading:
-
Disable and remove old hooks:
openclaw hooks disable token-model openclaw hooks disable token-usage openclaw hooks disable token-cron rm -rf ~/.openclaw/hooks/token-model rm -rf ~/.openclaw/hooks/token-usage rm -rf ~/.openclaw/hooks/token-cron -
Update the skill:
clawhub update tokenQrusher -
Re?install remaining hooks:
tokenqrusher install --hooks -
Restart gateway:
openclaw gateway restart
Removed commands: tokenqrusher model, budget, usage, optimize. They no longer exist.
License
MIT. See LICENSE file in repository.
Credits
- Design & Implementation: Lieutenant?Qrusher (qsmtco)
- Review: Captain?JAQ (SMTCo)
- Framework: OpenClaw Team
Built with OpenClaw: https://github.com/openclaw/openclaw
相关推荐
专题
+ 收藏
+ 收藏
+ 收藏
+ 收藏
+ 收藏
+ 收藏
最新数据
相关文章
自动更新器:自动化 AI 智能体与技能更新 - Openclaw Skills
自动更新程序:Openclaw 技能的自动化维护
自动更新器:自动化智能体与技能维护 - Openclaw Skills
自动更新器:自动执行核心与技能更新 - Openclaw Skills
Agent Browser:AI 智能体网页自动化 CLI - Openclaw Skills
Hacker News 发布器:自动进行 HN 提交和评论 - Openclaw Skills
ClawdHub CLI: 管理和发布 Openclaw 技能 - Openclaw Skills
ClawdHub CLI:管理并发布智能体技能 - Openclaw Skills
bird: 面向 AI 智能体的 X/Twitter CLI 技能 - Openclaw Skills
bird: X/Twitter CLI 与 AI 智能体集成 - Openclaw Skills
AI精选
