Typing custom skill names triggers unnecessary compaction

Description

When typing a custom skill name (e.g. /refresh), Claude Code triggers a compaction cycle even though no meaningful work has been done. The skill invocation itself is lightweight — it just creates and deletes a task to refresh the TaskList — but the act of typing it causes Claude Code to compact the conversation.

This creates a disruptive UX: the operator types a quick skill command, and instead of instant execution, they wait 30-60 seconds for compaction before anything happens.

Steps to Reproduce

1. Have a conversation with moderate context usage (~17% in my case)
2. Type a custom skill (e.g. /refresh — a skill that simply creates and deletes a task)
3. Observe that compaction triggers before the skill executes

Expected Behavior

Skill invocation should not trigger compaction. The skill input is a short command, not a large context addition. Compaction should only trigger based on actual context growth, not skill dispatch.

Actual Behavior

Compaction fires immediately after typing the skill, adding ~44 seconds of delay before the skill can execute.

Screenshot

!Compaction triggered by /refresh skill

The screenshot shows:

• /refresh was typed
• "Compacting..." appears for 44 seconds
• Context usage was only at 17%
• The skill itself is trivial (create + delete a task)

Environment

• Claude Code CLI
• Model: Opus (1M context)
• Context usage: ~17% at time of compaction trigger