claude -p does not trigger skills — skill-creator eval loop always shows 0% recall

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?

Problem

When using the skill-creator plugin's run_loop.py to evaluate skill description triggering, claude -p never invokes any skills. All should-trigger queries show rate=0/3, resulting in 0% recall and 50% accuracy (only negatives pass).

Steps to reproduce

1. Install skill-creator plugin: claude plugin install skill-creator
2. Create a skill with a description in SKILL.md frontmatter
3. Create an eval set with should-trigger queries
4. Run: python3 -m scripts.run_loop --eval-set eval.json --skill-path ~/.claude/skills/my-skill --model claude-opus-4-6 --max-iterations 5

Expected

claude -p should load and trigger skills based on the description, same as in an interactive session.

Actual

Skills are never triggered. Every should-trigger query returns rate=0/3. The skills work correctly in interactive Claude Code sessions.

Environment

• Claude Code 2.1.80
• macOS Darwin 25.3.0
• Model: claude-opus-4-6

What Should Happen?

When claude -p is called with a prompt that matches a skill's description, the skill should be triggered and consulted — the same way it works in an interactive Claude Code session. The run_loop.py script from the skill-creator plugin relies on this to
evaluate skill description quality.

Error Messages/Logs

Steps to Reproduce

1. Install the skill-creator plugin:

claude plugin install skill-creator

1. Create a simple test skill:

mkdir -p ~/.claude/skills/test-skill
cat > ~/.claude/skills/test-skill/SKILL.md << 'EOF'
---
name: test-skill
description: Use this skill whenever the user asks about Sevdesk, invoices, bookkeeping, or billing operations.
---
# Test Skill
This is a test skill.
EOF

1. Run a prompt that should trigger the skill:

claude -p "Erstelle eine Rechnung in Sevdesk für den Kunden Mueller, 1.500 EUR netto"

1. Observe that the skill is never triggered. Verify by checking the output — no skill consultation happens.
2. The same prompt in an interactive claude session correctly triggers the skill.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.80

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_