[BUG] No mechanism to un-load or not reload skills after compaction - slowly accumulates skills over long horizon tasks

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The new feature to restore skills and slashcommands is a good feature. It saves me from having to watch claude like a hawk and yell at it to reload skills across compaction events!

But there's an obvious and hilarious issue with this.

When claude works on a long horizon task, it eventually starts to accumulate skills. After an hour or so of work, I see something like:

```claude-code-session
⎿  Slash commands restored (editing-swift-code-safely, creating-meaningful-data-visualizations, elderly-user-interface-patterns, using-claude-code-tools-correctly, understanding-deedee-architecture,
validating-health-alerts-with-medical-evidence, healthkit-integration-patterns, refactoring-without-breaking)

(Strangely, this very rarely happens with my actual SlashCommands, even though it says "Slash commands restored", which is marginally frustrating since it means I need to refactor some of them into skills)

As it goes longer and longer and loads more skills, the actual free context window shrinks further and further! If I'm orchestrating a complex task that really needs to progressively build up _some context_ (because I don't have proper TDD or E2E for the in-context re-learning), it leaves less and less room for actual work to get done.

### What Should Happen?

Designing a proper solution for this would be hard. I don't think you can do it without increasing the complexity of the prompt-driven compaction and summarization to decide which skills are still relevant.

### Error Messages/Logs

```shell

Steps to Reproduce

1. Create more than a dozen skills, skills that you actually find useful enough to trigger in a long horizon task
2. Artificially inflate their contents, the more the easier to reproduce.
3. Begin some long horizon task that triggers them infrequently, but with a degree of certainty across compactions
4. Watch as they accumulate?

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.0.76

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

I mean, honestly, this is a good problem to have. I'd much prefer to have this problem with skills loading rather than not loading or not having skills.