skillscope-optimizer

SkillScope Optimizer

Purpose

Optimize one selected SKILL.md from SkillScope's program-analysis artifacts. Use the artifact chain below as the source of truth:

skill source -> constraints + skill graph
trace events -> trace facts
graph x facts -> findings
findings -> minimal skill patch

Required Inputs

Read every path named in the launch prompt. The optimizer must have:

• Original selected skill markdown.
• constraints.json.
• skill-graph.json.
• trace-facts.json.
• findings.json.
• Task/context and result/verifier artifacts when provided.
• Native verifier artifacts when provided.

If any IR artifact is missing, stop and report which artifact is missing. Missing IR routes the workflow back to skillscope-analyzer.

Optimization Direction

Optimize only when there is a real optimization target. When the native verifier already passed, non-compliance is low, and remaining failures are only process/reporting observability gaps, use the current skill as the rerun candidate, copy it to optimized-skill.md, record the pass-through reason in the report, and set optimization-packet.json edits to an empty array.

Use this priority order:

1. Native verifier failed assertions that map to final-output or artifact contracts.
2. SkillScope violated findings on final_output or artifact targets.
3. Repeated missed findings on reachable final-output or artifact constraints.
4. High non-compliance or repeated failures across traces.
5. Process/tool/reporting failures only when they explain a native or artifact failure, repeat across traces, or the user explicitly wants observability hardening.

A single successful trace with a low-severity process-only missed finding routes to analyzer confidence, branch reachability, or UI review unless the user explicitly requests observability hardening.

Treat satisfied constraints as preservation anchors. A covered constraint, a native-verifier-backed output invariant, or a nearby branch that worked should stay textually stable except when the replacement is a clearer equivalent required by a failed neighboring constraint. Every edit that touches a covered span should state the preserved invariant in optimization-packet.json.