Feature request: evaluate and score user-created subagents and skills with keep/remove recommendations

Summary

Add a built-in Claude Code command (e.g., /audit-subagents, /audit-skills, or a unified /audit-extensions) that analyzes the user's custom subagents (.claude/agents/*.md) and skills (.claude/skills/*/SKILL.md), scores each one on value contribution, and recommends whether to keep, revise, or remove them. Surface the same analysis as a dedicated section in the HTML report produced by /insights.

Problem

Power users accumulate dozens of custom subagents and skills over time. In practice:

• Many are created once for a narrow task and never used again.
• Some overlap heavily with each other or with built-in capabilities.
• Some have drifted out of sync with the project (referencing files/paths that no longer exist, or describing behavior the codebase has moved past).
• Descriptions are often too vague for the dispatcher to match against — so the skill/agent silently never fires.
• There's no principled way to decide which ones to prune, short of manual review.

The result is cruft: longer context, worse dispatch precision (more candidate skills/agents diluting the match), and higher cognitive load when writing new ones.

Proposed behavior

A command that, for each custom subagent and skill:

1. Usage signal — count how often it has been invoked (from transcripts / session logs) over a user-specified window.
2. Description quality score — check whether the description is specific enough to fire reliably (trigger keywords, SKIP clauses, concrete examples) vs. vague.
3. Overlap detection — cluster items with substantially similar descriptions/triggers and flag redundancy.
4. Staleness check — detect references to files, paths, tools, or commands that no longer exist in the repo.
5. Value rating — a composite score (e.g., 0–10) with a one-line justification.
6. Recommendation — KEEP / REVISE / REMOVE, with the reason and, for REVISE, a suggested edit.

Output as a ranked table so the user can quickly sweep through and approve removals in bulk.

/insights integration

Add a new section to the HTML report generated by /insights titled e.g. "Custom Extensions Health" that renders the same audit data in a browser-friendly format:

• Summary tiles at the top: total custom subagents, total custom skills, count flagged REMOVE, count flagged REVISE, count unused in the reporting window.
• Sortable / filterable table with one row per subagent and skill, columns: name, type (agent/skill), scope (user/project/plugin), invocations in window, description-quality score, overlap group, staleness flags, composite value score, recommendation, and expandable justification.
• Visual cues: red row for REMOVE, yellow for REVISE, green for KEEP; a small bar or sparkline showing invocation counts over time.
• A "suggested revisions" drawer per row showing the rewritten description/trigger text, copy-to-clipboard for quick application.
• Link at the top of the section pointing to the raw /audit-extensions command so users can re-run interactively.

This keeps the audit discoverable for users who already run /insights periodically, without requiring them to remember a separate command.

Why this belongs in Claude Code

Only Claude Code has visibility into both the definitions and the actual invocation history. A user can't easily compute "this skill has fired 0 times in 90 days" without building their own tooling on top of transcripts.

Nice-to-haves

• Dry-run vs. apply mode (with --apply actually deleting or rewriting files after confirmation).
• Per-scope filtering (--user, --project, --plugin).
• Export as JSON for scripting.
• CLI flag on /insights (e.g., --skip-extensions-audit) for users who don't want this section.