Disabled plugin skills still appear in system prompt (example-skills)

Bug Description

Plugin skills from disabled plugins still appear in the system prompt's skill list, consuming context window tokens even when explicitly disabled in .claude/settings.json.

Steps to Reproduce

1. In .claude/settings.json, set a plugin to false:

``json
"enabledPlugins": {
"example-skills@anthropic-agent-skills": false
}
``

1. Start a new Claude Code session
2. Observe the system-reminder skill list

Expected Behavior

Skills from example-skills@anthropic-agent-skills should NOT appear in the system prompt when the plugin is disabled.

Actual Behavior

All ~15 example-skills:* skill descriptions (theme-factory, doc-coauthoring, claude-api, xlsx, pdf, algorithmic-art, internal-comms, skill-creator, canvas-design, pptx, slack-gif-creator, webapp-testing, frontend-design, mcp-builder, brand-guidelines, docx, web-artifacts-builder) appear in the system prompt alongside their enabled counterpart document-skills:*.

This means duplicate skill descriptions (document-skills + example-skills) consume context even though only document-skills is enabled.

Impact

Each skill description is ~50-150 tokens. With ~15 duplicated skills, that's ~750-2,250 unnecessary tokens in every conversation's system prompt. For users with many plugins, this compounds and contributes to a high baseline context usage (35%+ of window consumed before any user interaction).

Environment

• Claude Code CLI (latest)
• macOS Darwin 25.3.0
• Settings file: .claude/settings.json with enabledPlugins configuration