Feature Request: Config-Based Skill Disabling

Summary

Add support for disabling skills via configuration without manual filesystem changes.

Motivation

Currently, all skills in /usr/lib/node_modules/openclaw/skills/ are loaded into the system prompt context, consuming ~16k tokens per session regardless of whether they're actively used.

For users with specific workflows (e.g., primarily email/calendar management), many skills go unused but still contribute to token costs.

Proposed Solution

Add a skills.disabled array to openclaw.json:

{
  "skills": {
    "install": {
      "nodeManager": "npm"
    },
    "disabled": [
      "prose",
      "clawflow",
      "gh-issues",
      "tmux",
      "video-frames"
    ]
  }
}

When a skill is listed in disabled, OpenClaw should:

1. Skip loading its SKILL.md into the system prompt
2. Hide it from the <available_skills> context block
3. Still allow re-enabling by removing it from the array

Benefits

• Token savings: ~8-10k tokens/session for users who disable 50-60% of unused skills
• User control: No need to manually delete/rename system directories
• Update-safe: Changes persist across OpenClaw version updates
• Reversible: Easy to re-enable skills as needed

Alternatives Considered

1. Manual skill deletion — breaks updates, risky
2. Custom skill directories only — loses access to built-in skills
3. Accept current state — works, but wasteful for specialized workflows

Use Case

A user running primarily Gmail/Calendar automation doesn't need:

• Video frame extraction
• Tmux session control
• Multi-agent workflow orchestration (ClawFlow)
• GitHub issue auto-fixing

Disabling these would save ~10k tokens/session while keeping core functionality intact.

Implementation Notes

• Should validate skill names against available skills (warn if unknown skill listed)
• Consider adding openclaw config set skills.disabled --append <skill-name> CLI helper
• Maybe support glob patterns: "disabled": ["*-triage", "video-*"]

Priority: Medium
Impact: Token cost reduction for power users
Complexity: Low (filter available skills during prompt assembly)