openclaw - 💡(How to fix) Fix [Feature]: Skills prompt injection consumes ~10K chars with 80+ skills, causing model selection noise and context window pressure [1 comments, 2 participants]

Summary

Problem Description

With 80+ skills installed across multiple skill directories (~/.openclaw/skills, ~/.agents/skills, <workspace>/.agents/skills, etc.), the XML-formatted skills list injected into the system prompt consumes approximately 7,000–10,000 characters.

According to the documented formula: total = 195 + Σ (97 + len(name_escaped) + len(description_escaped) + len(location_escaped)) With 80 skills, this results in:

• ~2,000–2,500 tokens overhead (at ~4 chars/token)
• 15–30% of context window consumed on 8K/16K models before any user message

Problem to solve

Problem Description

With 80+ skills installed across multiple skill directories (~/.openclaw/skills, ~/.agents/skills, <workspace>/.agents/skills, etc.), the XML-formatted skills list injected into the system prompt consumes approximately 7,000–10,000 characters.

According to the documented formula: total = 195 + Σ (97 + len(name_escaped) + len(description_escaped) + len(location_escaped)) With 80 skills, this results in:

• ~2,000–2,500 tokens overhead (at ~4 chars/token)
• 15–30% of context window consumed on 8K/16K models before any user message

Impact

1. Context window pressure: Reduces effective reasoning space for complex multi-step tasks
2. Model selection noise: Forces the model to choose from 80+ unstructured skill options, increasing mismatch rate
3. No hierarchical discovery: The flat XML list provides no categorization, tagging, or priority mechanism

Current Workaround

Users are forced to maintain manual external indexes (e.g., a TOOLS.md file) to compensate for the lack of structured skill discovery within the system.

Proposed solution

为什么这会影响智能代理设计

当前的扁平化技能注入违反了智能代理架构的一个基本原则：分层抽象。 人类专家不会同时将80个工具放在工作记忆中。他们会将知识组织成领域： "这是一个金融问题 → 打开金融工具箱 → 选择特定工具" OpenClaw 目前的机制迫使模型做相反的事情： "这里有80个不相关的工具，弄清楚哪一个适合" 这会产生三种失败模式： 选择瘫痪 — 面对80多个选项，模型的概率质量分散到不相关的技能上，增加了幻觉和错误匹配率 上下文污染 — 10K字符的技能描述挤占了实际用户意图和对话历史，降低了真实任务的推理质量 技能利用率不足 — 频率较低但高度相关的技能被常见技能“淹没”，因为模型的注意力机制自然偏向熟悉的模式 真正智能的代理应该做的事情： 第一层：维护一个轻量级的“技能目录”（仅包含类别 → 技能名称） 第二层：在意图分类时，动态加载仅相关类别的完整技能定义 第三层：通过精确的工具调用执行 这反映了人类组织的运作方式——你不会给每个员工整个公司的程序手册。你给他们索引，并让他们在需要时提取相关部分。 当前这种“将所有内容都转储到上下文中”的方法在超过约20个技能后扩展性很差，并且从根本上限制了OpenClaw作为复杂多领域代理编排器的能力。

Alternatives considered

Option A: Hierarchical/Categorized Prompt Format

Instead of flat XML injection, inject a category index first:

<skills>
  <category name="Finance">
    <skill name="stock-analysis" description="..."/>
    <skill name="market-analysis" description="..."/>
  </category>
  <category name="Productivity">
    <skill name="lark-doc" description="..."/>
  </category>
</skills>
Model selects category first, then specific skill. Reduces cognitive load from O(n) to O(√n).
Option B: Lazy-Load Mechanism
Inject only a lightweight skill catalog at session start. When model expresses intent (e.g., "analyze stock"), dynamically retrieve and inject relevant skill definitions via RAG-style retrieval.
Option C: Skill Tagging + Grouped Allowlists
Extend the existing agents.defaults.skills allowlist with tag-based grouping:
{
  agents: {
    defaults: {
      skillTags: ["finance", "lark-ecosystem"],
    }
  }
}

### Impact

## Impact
1. **Context window pressure**: Reduces effective reasoning space for complex multi-step tasks
2. **Model selection noise**: Forces the model to choose from 80+ unstructured skill options, increasing mismatch rate
3. **No hierarchical discovery**: The flat XML list provides no categorization, tagging, or priority mechanism


### Evidence/examples

_No response_

### Additional information

_No response_