plugin-self-eval-loop

name	plugin-self-eval-loop
description	Self-evaluation loop for the treadmill Claude plugins pack. Runs cursor-plugin-evals against bundled skills, tracks quality over time, and escalates recurring failures via PAMS. Use periodically or before publishing plugin updates.

Plugin Self-Eval Loop

When to Use

• Before publishing any plugin update
• Periodic quality check (monthly recommended)
• After adding or modifying skills in the pack
• When PAMS detects recurring correction patterns in skills from this pack

Process

1. Run evals — execute cursor-plugin-evals skill-eval against each skill in the pack
2. Parse results — identify failing skills and evaluators
3. Fix — for each failing skill, apply targeted fix (one category per iteration)
4. Re-run — re-evaluate only the failing skills
5. Iterate — max 3 iterations per skill
6. PAMS escalation — if a skill fails repeatedly (3+ sessions), escalate:
   • 2x → create a rule documenting the anti-pattern
   • 3x → create a new skill + eval to address the gap

Convergence

• All skills pass at >= 85% (OSS-equivalent threshold for plugin skills)
• No regressions in previously passing skills
• PAMS escalation triggers acted on