Skill runner should be handled separately

Summary

Skill runner is very annoying, no access to ~/.commandcode/skills on the agent. Too many access requests.

Expected Behavior

Seamless skill runs managed by the harness and not by LLM.

Actual Behavior

For some reason it skill calls fall back to bash cat ...,

 SHELL  [echo "===== to-spec ====="; cat ~/.commandcode/skills/to-spec/SKILL.md; echo; ls ~/.commandcode/skills/to-spec/]
 │ ===== to-spec =====
 │ ---
 │ name: to-spec

Steps to reproduce the issue

Call a skill in cmd from ~/.commandcode that calls other skills.

Command Code Version

1.38.2

Operating System

macOS

Terminal/IDE

VSCode

Shell

zsh

Session file (optional)

No response

Fix prompt (optional)

No response

Additional context

OS: macOS 27.0

首先重現來自 ~/.commandcode 中 cmd 的巢狀 skill 呼叫，並檢查 skill 呼叫在哪裡退回到 cat 等 shell 命令。完成的標準是由 harness 處理 skills（包括巢狀呼叫），且不會反覆要求存取權限，也不會執行由 LLM 產生的 shell 讀取操作。