Agent Upgrade Pipeline

Operator Context

This skill upgrades a single target agent or skill through a scored, gated pipeline. It is a bottom-up quality mechanism — triggered when a specific component needs improvement — complementing the top-down system-upgrade pipeline that handles multi-component changes driven by external events.

Hardcoded Behaviors (Always Apply)

• Score before AND after: Phase 1 (AUDIT) must produce a baseline score using agent-evaluation. Phase 5 (RE-EVALUATE) must produce an after score. Never claim improvement without a measured delta.
• Plan approval gate: Phase 3 output MUST be presented to the user and explicitly approved before Phase 4 begins. Do not implement any changes without approval.
• Domain logic is off-limits: Phase 4 MUST NOT alter an agent's routing triggers, domain coverage, or core methodology. Only structural improvements, template alignment, Operator Context additions, and retro graduations are in scope unless the user explicitly directs otherwise.
• Use agent-evaluation for scoring: Do not self-assess quality. Invoke the agent-evaluation skill for objective scores.

Default Behaviors (ON unless disabled)

• Check retro graduates: Phase 1 always searches learning.db for entries targeting the agent under upgrade. Graduation candidates are surfaced in Phase 2.
• Peer comparison: Phase 2 compares the target against 2–3 agents in the same category for consistency gaps (e.g., comparing a Go agent against other Go agents).
• Regression protection: If Phase 5 delta is negative, report to user and do NOT auto-revert. User decides.