Agent Skills: narrative-tests

When to use

• Author wants stronger control: “改一处就知道后面哪里拽歪”
• After any major deviation or canon change

Test writing principles

• Tests are author intent made explicit, not bureaucracy.
• Prefer rules that are:
  • specific (who/what/when)
  • checkable (can locate failure in concrete text)
  • stable (won’t change every chapter)

Recommended categories

• Canon rules: power system, forbidden knowledge, hard constraints
• Timeline rules: ordering, causality, “X must happen before Y”
• Character rules: motivation consistency, relationship milestones
• Information revelation: “before chapter N, reader must not know X”

Output files

• narrative_tests.md: test suite definition (author editable)
• narrative_test_report.md: run result (failures + location + fix options)

v1 Test Format (deterministic)

当前实现（v1）支持在 narrative_tests.md 中嵌入一个 ```json 代码块：

{
  "suiteId": "default",
  "cases": [
    {
      "id": "no-spoiler-before-1",
      "severity": "ERROR",
      "type": "must_not_contain_before_chapter",
      "untilChapter": 1,
      "pattern": "SPOILER",
      "message": "Chapter 1 must not contain SPOILER."
    }
  ]
}

支持的 type（v1）：

• must_contain
• must_not_contain
• must_not_contain_before_chapter

Failure handling (must produce options)

For each failure, provide:

1. where it breaks (chapter/beat/excerpt)
2. why it breaks (rule violated, drift introduced by what change)
3. routes (choose 1):
   • minimal patch (small local edit)
   • propagate change (update outline/canon + future chapters)
   • branch rewrite (keep mainline stable)