Command Code failed to read globally installed skill

Summary

Command code failed to read and execute globally installed skill. The skill registration found it, but the file permissions didn't follow.

Expected Behavior

Command code should be able to find and read globally installed skill.

Actual Behavior

Command code able to find globally installed skill, but unable to read it because the file is located outside of the workspace.

Steps to reproduce the issue

1. Install a skill globally, in my case i install exa-search skill globally
2. Write a prompt that tell command code to find information online (to trigger it to use exa search). In my case, i ask command code to find information to compare Step 2.7 Flash vs Deepseek v4 Flash for coding related task
3. Press ctrl+o to see the model thinking or read the result

Command Code Version

0.30.1

Operating System

macOS

Terminal/IDE

ghostty

Shell

fish

Additional context

Here is the model COT in my test:

The user asked about comparing Step 3.7 Flash vs DeepSeek V4 Flash for coding tasks. I
tried to use the explore agent, but the skill file for exa-search is outside the
workspace so I couldn't read it. The explore agent returned a general comparison based
on public information.

The user is asking a knowledge question about recent AI model releases - specifically
Step 3.7 Flash vs DeepSeek V4 Flash. This is a question I should answer directly rather
than searching the codebase, per the taste rules: "For questions unrelated to the
current project, do not search the codebase; provide general knowledge answers instead."

However, I also have access to the exa-search skill which is designed for semantic
search. The issue is the SKILL.md file is outside my workspace, so I couldn't read it
via read_file. I could potentially still invoke the skill through shell commands or
other means, but the skill might need to be invoked differently.

Reproduce the issue on macOS with a globally installed exa-search skill, then trace the skill registration and file-reading path used by Command Code. Done means a skill located outside the workspace can be read and executed, while the existing reproduction flow succeeds.