Refactor plan

Judge first

references/rubric.md separates fragile from fast-moving from coupled, and sets confidence. The counts from system-history are leads. A plan needs a named repeated mechanism, taken from diffs you read.

Scaffold

pnpm exec tsx .agents/skills/refactor-plan/scripts/new-plan.ts --run <id> \
  --slug <kebab> --title "<imperative line>" --systems a,b \
  --area <Framework|Slides|Design|Clips|Content|Dispatch|General> [--force]

It reads the run's analysis.json. Every system must be one analyze scored. When the run has a bug.json, the plan becomes a bug plan:

• it adds a Trigger section, and quotes the report when it has no link;
• its paths are the focus files only;
• its frontmatter gets trigger: bug and the report as source.

Plans go to <plansDir>/<run>/<slug>.md (see config.json).

Fill it

Replace every TODO(agent) marker, summary, and confidence, following references/plan-guide.md. Keep slugs stable across runs. The fingerprint comes from the systems and the slug, and ticket dedup keys on it.

A plan is finished when fragility-common's findings-upsert.ts accepts it.