Self-modification guard prompts for .claude/skills/ files despite documented exemption

Bug Description

When editing files inside .claude/skills/ directories (e.g., .claude/skills/review-all/review-state.md), Claude Code shows a self-modification permission prompt:

Do you want to make this edit to review-state.md?
 ❯ 1. Yes
   2. Yes, and allow Claude to edit its own settings for this session
   3. No

According to the permissions documentation, writes to .claude/skills/ should be exempt from the self-modification guard:

"Writes to .claude/commands, .claude/agents, and .claude/skills are exempt and do not prompt, because Claude routinely writes there when creating skills, subagents, and commands."

Steps to Reproduce

1. Create a skill that writes to a state file inside .claude/skills/<skill-name>/ (e.g., review-state.md)
2. Run the skill (or manually Edit/Write to the file)
3. The self-modification prompt appears despite the documented exemption

What was tried (none of these work)

• defaultMode: "bypassPermissions" in project settings.local.json
• skipDangerousModePermissionPrompt: true in global settings.json
• Edit(.claude/**) and Write(.claude/**) in both global and project-level allow lists
• Unrestricted Edit and Write (no path filter) in global allow list

Expected Behavior

Writes to .claude/skills/** should not trigger the self-modification permission prompt, as documented.

Actual Behavior

The self-modification prompt appears every time, regardless of permission settings or bypass mode.

Additional Context

• This is especially disruptive when skills use subagents (e.g., multi-agent review pipelines), as each subagent triggers its own prompt for every state file write
• Selecting option 2 ("allow for this session") does not persist to subagents

Environment

• Claude Code version: 2.1.81
• OS: Windows 11 (10.0.26200)
• Shell: Git Bash