What: Teach the planner to autonomously generate new SkillConfig variants based on lineage winners.
Why: Closes the loop from experiment results back into skill design — the planner reads lineage, identifies what worked (e.g., "v2 won because it added step-by-step reasoning"), and proposes v3 with targeted modifications (e.g., add conciseness constraints to the winning pattern).
Pros: Fully autonomous skill improvement without human hypothesis injection.
Cons: Planner needs access to lineage data — adds a new input to plan_next().
Context: Depends on SkillConfig type and lineage.jsonl both being shipped. Planner's _next_prompt would gain a SKILL_LINEAGE section summarizing prior skill versions and their scores. _to_step would emit a SkillConfig (not just a system_prompt string) when operating in skill-synthesis mode.
Effort: M (human: ~2d / CC: ~25min) | Priority: P2
Depends on: SkillConfig (#4), skill lineage tracking (#6) from tool/skill A/B plan