[BUG] Skill list truncation threshold too aggressive - hides 60% of skills for ~5K token savings

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 Skill tool's available_skills section is truncated from 92 skills to 36 skills with the message:
-- Showing 36 of 92 skills due to token limits --
This hides 60% of installed skills (56 skills) to save approximately 5,000-8,000 tokens - roughly 2.5-4% of a 200K context window. The truncation threshold appears calibrated for older, smaller context windows (8K/16K era) and hasn't been updated for current model capabilities. Users cannot see their full skill inventory, making it appear skills have "vanished" when they're actually loaded but hidden from the display.

What Should Happen?

Either:
The truncation threshold should be increased to accommodate 200K context windows (showing all 92 skills costs ~10K tokens = 5% of context)
Or truncation should be configurable via settings.json
Or skills should use deferred loading per #12836

Error Messages/Logs

<!-- Showing 36 of 92 skills due to token limits -->
No error - just silent truncation that causes user confusion.

Steps to Reproduce

1. Install multiple skill sources:

• User skills in ~/.claude/skills/ (8 skills)
• Superpowers plugin with 21 skills
• Additional plugin with 61+ skills (e.g., wshobson-opus agents)

1. Start a new Claude Code session
2. Observe the Skill tool's available_skills section in the system prompt
3. Note: Only 36 of 92 skills are displayed

Metrics from reproduction:

• Total SKILL.md files: 124
• Total description text: ~33KB (~8,000-10,000 tokens)
• Skills shown: 36
• Skills hidden: 56 (60%)
• Token savings: ~5,000 tokens (2.5% of 200K context)

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.0.58

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

VS Code integrated terminal

Additional Information

Related issues:

• #12836 - Request for deferred tool loading (would solve this properly)
• #12333 - New session takes 80K tokens
• #3406 - Tool descriptions cause 10-20K overhead

Cost-benefit analysis: The current truncation saves ~5K tokens but hides 60% of functionality from the user. For a 200K context window, this is an overly aggressive optimization that harms UX more than it helps token budget. Suggested fix priority:

1. Quick fix: Raise truncation threshold to show all skills under ~150
2. Better fix: Make threshold configurable in settings.json
3. Best fix: Implement deferred skill loading per #12836