chaffjs 0.2.0 → 0.4.0

package/src/document.ts

@@ -4,7 +4,7 @@ import { gfmTableFromMarkdown } from "mdast-util-gfm-table";
 import { frontmatter } from "micromark-extension-frontmatter";
 import { frontmatterFromMarkdown } from "mdast-util-frontmatter";
 import { maskSpans } from "./mask.ts";
-import type { LanguageAdapter, ProseDocument, Section, Sentence, Span } from "./plugin.ts";
+import type { BulletList, LanguageAdapter, Paragraph, ProseDocument, Section, Sentence, Span } from "./plugin.ts";
 
 type Place = { readonly offset?: number | undefined };
 type Node = {
@@ -49,8 +49,37 @@ const linkChrome = (node: Node): Span[] => {
   ];
 };
 
-const collectMasks = (root: Node): Span[] => {
-  const spans: Span[] = [];
+/**
+ * 強調の記号（`**` `*` `_` `~~`）だけを覆う。囲まれた文字は本文なので残す。
+ *
+ * 残すと解析器が `**。` を 1 語の名詞として拾い、文の終わりが消える。
+ * 実文書で「文末が名詞」の誤検知を追って見つけた。文長にも同じだけ効いている。
+ */
+const MARKER = /^[*_~]+/u;
+
+const emphasisChrome = (node: Node, source: string): Span[] => {
+  const whole = spanOf(node);
+  if (whole === undefined) return [];
+  const width = MARKER.exec(source.slice(whole.start, whole.end))?.[0].length ?? 0;
+  if (width === 0) return [];
+  return [
+    { start: whole.start, end: whole.start + width },
+    { start: whole.end - width, end: whole.end },
+  ];
+};
+
+const EMPHASIS = new Set(["strong", "emphasis", "delete"]);
+
+/**
+ * `:::note` `:::` のようなディレクティブ。Zenn・Docusaurus・VitePress が使う記法で、
+ * 標準の Markdown には無いため段落として解析される。囲みの指定であって文章ではない。
+ */
+const DIRECTIVE = /^:::[^\n]*/gmu;
+
+const directiveSpans = (source: string): Span[] => [...source.matchAll(DIRECTIVE)].map((match) => ({ start: match.index, end: match.index + match[0].length }));
+
+const collectMasks = (root: Node, source: string): Span[] => {
+  const spans: Span[] = [...directiveSpans(source)];
   walk(root, (node) => {
     if (node.type === "link" || node.type === "linkReference") {
       spans.push(...linkChrome(node));
@@ -63,6 +92,14 @@ const collectMasks = (root: Node): Span[] => {
   return spans;
 };
 
+const emphasisSpans = (root: Node, source: string): Span[] => {
+  const spans: Span[] = [];
+  walk(root, (node) => {
+    if (EMPHASIS.has(node.type)) spans.push(...emphasisChrome(node, source));
+  });
+  return spans;
+};
+
 const textOf = (node: Node, source: string): string => {
   const parts: string[] = [];
   walk(node, (child) => {
@@ -107,10 +144,10 @@ const strongSpans = (root: Node, masked: readonly Span[]): Span[] => {
 
 const within = (span: Span, from: number, to: number): boolean => span.start >= from && span.start < to;
 
-const paragraphSpans = (root: Node): Span[] => {
+const spansOfType = (root: Node, type: string): Span[] => {
   const found: Span[] = [];
   walk(root, (node) => {
-    if (node.type !== "paragraph") return;
+    if (node.type !== type) return;
     const span = spanOf(node);
     if (span !== undefined) found.push(span);
   });
@@ -124,14 +161,17 @@ const paragraphSpans = (root: Node): Span[] => {
  * 覆った表やコードブロックを越えて次の句点まで飲み込む。段落は文が跨がない境界なので、
  * ここで切れば構造的に起きない。
  */
+const shift = (span: Span, by: number): Span => ({ start: by + span.start, end: by + span.end });
+
 const sentencesOf = (prose: string, paragraphs: readonly Span[], adapter: LanguageAdapter): Sentence[] =>
   paragraphs.flatMap((paragraph) =>
     adapter
       .segment(prose.slice(paragraph.start, paragraph.end))
       .sentences.filter((sentence) => sentence.text.trim().length > 0)
       .map((sentence) => ({
-        span: { start: paragraph.start + sentence.span.start, end: paragraph.start + sentence.span.end },
+        span: shift(sentence.span, paragraph.start),
         text: sentence.text,
+        ...(sentence.tokens === undefined ? {} : { tokens: sentence.tokens.map((token) => ({ ...token, span: shift(token.span, paragraph.start) })) }),
       })),
   );
 
@@ -153,18 +193,62 @@ const sectionsOf = (headings: readonly Heading[], sentences: readonly Sentence[]
     });
 };
 
-export const buildDocument = (path: string, source: string, adapter: LanguageAdapter): ProseDocument => {
+const paragraphsOf = (spans: readonly Span[], sentences: readonly Sentence[], listSpans: readonly Span[]): Paragraph[] =>
+  spans
+    // 箇条書きの中の段落は「段落」として数えない。項目 1 つを 1 段落と読むと、
+    // 段落あたりの文数も長さのばらつきも、箇条書きの多い文書で壊れる。
+    .filter((span) => !listSpans.some((list) => span.start >= list.start && span.start < list.end))
+    .map((span) => ({ span, sentences: sentences.filter((sentence) => within(sentence.span, span.start, span.end)) }));
+
+/** 箇条書きは list ノードの直下の項目を数える。入れ子の項目は内側の list のものとして数える。 */
+const listsOf = (root: Node, source: string): BulletList[] => {
+  const found: BulletList[] = [];
+  walk(root, (node) => {
+    if (node.type !== "list") return;
+    const span = spanOf(node);
+    if (span === undefined) return;
+    const items = (node.children ?? []).flatMap((child) => {
+      const item = spanOf(child);
+      return child.type === "listItem" && item !== undefined ? [source.slice(item.start, item.end).replace(/\s+/gu, "").length] : [];
+    });
+    if (items.length > 0) found.push({ span, items });
+  });
+  return found;
+};
+
+/** チームが chaff.yaml に書いたもの。語彙表と同じ器に入れて、detector には出所を見せない。 */
+export type TeamRules = { readonly jargon: readonly string[]; readonly requiredSections: readonly string[] };
+
+const EMPTY_TEAM: TeamRules = { jargon: [], requiredSections: [] };
+
+/** Config から取り出す。document は Config の形を知らない。 */
+export const teamRules = (config: { readonly jargon: readonly string[]; readonly requiredSections: readonly string[] }): TeamRules => ({
+  jargon: config.jargon,
+  requiredSections: config.requiredSections,
+});
+
+export const buildDocument = (path: string, source: string, adapter: LanguageAdapter, team: TeamRules = EMPTY_TEAM): ProseDocument => {
   const root = parse(source);
-  const masked = collectMasks(root);
+  // 強調の記号は「本文でないもの」だが、太字の数を数えるときの「覆われた場所」ではない。
+  // 同じ集合にすると、太字が自分の記号のせいで覆われた場所にあることになり、1 つも数えられなくなる。
+  const blocks = collectMasks(root, source);
+  const masked = [...blocks, ...emphasisSpans(root, source)];
   const prose = maskSpans(source, masked);
-  const sentences = sentencesOf(prose, paragraphSpans(root), adapter);
+  const paragraphSpans = spansOfType(root, "paragraph");
+  const listItems = spansOfType(root, "listItem");
+  const sentences = sentencesOf(prose, paragraphSpans, adapter);
   return {
     path,
     source,
     language: adapter.id,
     lengthUnit: adapter.capabilities.lengthUnit,
-    sections: sectionsOf(headingsOf(root, source), sentences, strongSpans(root, masked), source.length),
+    capabilities: adapter.capabilities,
+    sections: sectionsOf(headingsOf(root, source), sentences, strongSpans(root, blocks), source.length),
     sentences,
-    lexicons: adapter.lexicons,
+    listSpans: listItems,
+    paragraphs: paragraphsOf(paragraphSpans, sentences, listItems),
+    lists: listsOf(root, source),
+    lexicons: { ...adapter.lexicons, "internal-jargon": team.jargon.map((pattern) => ({ pattern })) },
+    requiredSections: team.requiredSections,
   };
 };

package/src/env.ts

@@ -0,0 +1,20 @@
+import { existsSync } from "node:fs";
+import { join } from "node:path";
+
+export const ENV_FILE = ".env";
+
+/**
+ * 作業ディレクトリの .env を読む。依存は足さない（Node が持っている）。
+ *
+ * **すでに設定されている環境変数は上書きしない。** シェルで渡したほうが勝つ。
+ * 逆にすると、一時的に別の key で試したいときに .env が黙って勝ってしまう。
+ *
+ * 読んだかどうかを返すのは、鍵がどこから来たのかを人に見せるため。
+ * 「設定したのに効かない」を、鍵の置き場所で起こさせない。
+ */
+export const loadEnvFile = (dir: string): string | undefined => {
+  const path = join(dir, ENV_FILE);
+  if (!existsSync(path)) return undefined;
+  process.loadEnvFile(path);
+  return path;
+};

package/src/init.ts

@@ -26,7 +26,8 @@ genre: ${genre}
 rules:
 `;
 
-const GITIGNORE_LINE = ".chaff-cache/";
+/** 判定のキャッシュと、鍵を書くファイル。どちらも commit しない。 */
+const GITIGNORE_LINES = [".chaff-cache/", ".env*"];
 
 type Written = { readonly path: string; readonly note: string };
 
@@ -36,13 +37,15 @@ const writeIfAbsent = (path: string, body: string, note: string): Written | unde
   return { path, note };
 };
 
-/** judge のキャッシュは commit しない。init が .gitignore に足す。 */
+/** 判定のキャッシュと .env は commit しない。init が .gitignore に足す。 */
 const ensureGitignore = (dir: string): Written | undefined => {
   const path = join(dir, ".gitignore");
-  if (!existsSync(path)) return writeIfAbsent(path, `${GITIGNORE_LINE}\n`, "作成しました");
-  if (readFileSync(path, "utf8").includes(GITIGNORE_LINE)) return undefined;
-  appendFileSync(path, `${GITIGNORE_LINE}\n`, "utf8");
-  return { path, note: `${GITIGNORE_LINE} を追記しました` };
+  if (!existsSync(path)) return writeIfAbsent(path, `${GITIGNORE_LINES.join("\n")}\n`, "作成しました");
+  const body = readFileSync(path, "utf8");
+  const missing = GITIGNORE_LINES.filter((line) => !body.split("\n").includes(line));
+  if (missing.length === 0) return undefined;
+  appendFileSync(path, `${missing.join("\n")}\n`, "utf8");
+  return { path, note: `${missing.join(" と ")} を追記しました` };
 };
 
 export const runInit = (dir: string, genre: string): string[] => {

package/src/levels.ts

@@ -1,4 +1,4 @@
-import type { Level, RuleDefinition } from "./plugin.ts";
+import type { Level, LevelTable, RuleDefinition } from "./plugin.ts";
 
 export const LEVELS: readonly Level[] = ["strict", "normal", "relaxed", "off"];
 
@@ -14,11 +14,23 @@ export type Resolved = { readonly level: Level; readonly limit: number; readonly
  * 落ちたことを呼ぶ側に伝えるのは、`chaff strict` が「変えたつもりで変わっていない」
  * 状態を作らないようにするため。
  */
-export const resolve = (rule: RuleDefinition, level: Level): Resolved => {
+/**
+ * ジャンル別の閾値。"business/email" は "business" より細かいので先に当たる。
+ * 仕様書の一文と社内メールに同じ文長を課すと、どちらかが必ず間違う。spec §9。
+ */
+const tableFor = (rule: RuleDefinition, genre: string | undefined): LevelTable => {
+  if (genre === undefined) return rule.levels;
+  const parts = genre.split("/");
+  const keys = parts.map((_, index) => parts.slice(0, parts.length - index).join("/"));
+  return keys.reduce<LevelTable | undefined>((found, key) => found ?? rule.by_genre[key], undefined) ?? rule.levels;
+};
+
+export const resolve = (rule: RuleDefinition, level: Level, genre?: string): Resolved => {
   if (level === "off") return { level, limit: 0, fellBackToNormal: false };
-  const direct = rule.levels[level];
+  const table = tableFor(rule, genre);
+  const direct = table[level];
   if (direct !== undefined) return { level, limit: direct, fellBackToNormal: false };
-  const normal = rule.levels.normal;
+  const normal = table.normal ?? rule.levels.normal;
   if (normal === undefined) throw new Error(`rule "${rule.id}" has no normal level`);
   return { level: "normal", limit: normal, fellBackToNormal: true };
 };

package/src/look-at.ts

@@ -0,0 +1,53 @@
+import type { ProseDocument } from "./plugin.ts";
+import type { Candidate } from "./semantic.ts";
+
+/**
+ * 自然文で書かれた `look_at` を、決定的な絞り込みに変える。spec §14、§26-7。
+ *
+ * 変換に LLM も対話も使わない。書き手は「どこを見るか」を説明するとき、
+ * 見るべき語を**自然に括弧でくくる**。
+ *
+ *   look_at: 「お願いします」「ご確認ください」を含む文
+ *
+ * その括弧の中を取り出すだけで、文単位の絞り込みになる。
+ * 推測を足さないので、書いたものと動くものが一致する。
+ */
+const QUOTED = /[「『"'`]([^「『』」"'`]{1,40})[」』"'`]/gu;
+
+export const wordsFrom = (lookAt: string | undefined): string[] => {
+  if (lookAt === undefined) return [];
+  const found = [...lookAt.matchAll(QUOTED)].map((match) => match[1]?.trim() ?? "").filter((word) => word.length > 0);
+  return [...new Set(found)];
+};
+
+/** 数値を見ると書いてあれば、数字のある文に限る。 */
+const WANTS_NUMBER = /数値|数字|number|figure/iu;
+
+export type Narrowing = { readonly words: readonly string[]; readonly needsNumber: boolean; readonly kept: number; readonly total: number };
+
+const NUMBER = /\d/u;
+
+const matches = (text: string, words: readonly string[], needsNumber: boolean): boolean => {
+  if (needsNumber && !NUMBER.test(text)) return false;
+  return words.length === 0 || words.some((word) => text.includes(word));
+};
+
+/**
+ * 絞り込めたら文単位で、絞り込めなければ文書全体を渡す。
+ * 全体を渡すときは「絞り込めなかった」ことを呼ぶ側に返す。黙って高いほうを選ばない。
+ */
+export const narrow = (
+  doc: ProseDocument,
+  lookAt: string | undefined,
+  whole: (doc: ProseDocument) => Candidate[],
+): { candidates: Candidate[]; narrowing: Narrowing } => {
+  const words = wordsFrom(lookAt);
+  const needsNumber = lookAt !== undefined && WANTS_NUMBER.test(lookAt);
+  const total = doc.sentences.length;
+  if (words.length === 0 && !needsNumber) return { candidates: whole(doc), narrowing: { words, needsNumber, kept: total, total } };
+  const hits = doc.sentences.filter((sentence) => matches(sentence.text, words, needsNumber));
+  return {
+    candidates: hits.map((sentence) => ({ text: sentence.text.trim(), offset: sentence.span.start })),
+    narrowing: { words, needsNumber, kept: hits.length, total },
+  };
+};

package/src/plugin.ts

@@ -3,10 +3,25 @@
 
 export type Span = { readonly start: number; readonly end: number };
 
-export type Sentence = { readonly span: Span; readonly text: string };
+/**
+ * tokens は adapter が prepare 済みのときだけ入る。無いことと「品詞が無い文」を混ぜない。
+ * token の span は文の span と同じ座標系（segment に渡した文字列の先頭が 0）。
+ * 2 つの基準を混ぜると、ずれたときに どちらが悪いか分からなくなる。
+ */
+export type Sentence = { readonly span: Span; readonly text: string; readonly tokens?: readonly Token[] };
 
-/** 品詞は Universal Dependencies の UPOS に統一する。アダプタ独自の体系を露出させない。 */
-export type Token = { readonly span: Span; readonly surface: string; readonly pos: string; readonly lemma?: string };
+/**
+ * 品詞は Universal Dependencies の UPOS に統一する。アダプタ独自の体系を露出させない。
+ * features は UD の FEATS。「受動か」のように品詞だけでは足りず、かつ言語ごとに
+ * 見え方が違うものを、アダプタが自分の言語の知識で畳んでここに置く。例: Voice=Pass。
+ */
+export type Token = {
+  readonly span: Span;
+  readonly surface: string;
+  readonly pos: string;
+  readonly lemma?: string;
+  readonly features?: Readonly<Record<string, string>>;
+};
 
 export type Segmentation = {
   readonly sentences: readonly Sentence[];
@@ -33,6 +48,9 @@ export type AdapterCapabilities = {
   readonly lengthUnit: LengthUnit;
 };
 
+/** prepare に渡す要求。動く rule が要らないものの代金を払わせない。 */
+export type AdapterNeeds = { readonly pos: boolean };
+
 export type LanguageAdapter = {
   readonly kind: "language";
   /** BCP 47 の primary subtag。"ja" / "en"。 */
@@ -41,6 +59,11 @@ export type LanguageAdapter = {
   readonly capabilities: AdapterCapabilities;
   /** この言語である確からしさ。0..1。 */
   readonly detect: (source: string) => number;
+  /**
+   * 解析器の読み込み。capabilities は「払えばできる」の宣言で、prepare が「払う」。
+   * 日本語の辞書は初期化に 1.5 秒かかるので、要求されたときだけ呼ぶ。
+   */
+  readonly prepare?: (need: AdapterNeeds) => Promise<void>;
   readonly segment: (text: string) => Segmentation;
   /** L2 rule が word_list で引く。アダプタが自分の言語のぶんだけを持つ。 */
   readonly lexicons: Readonly<Record<string, Lexicon>>;
@@ -60,6 +83,11 @@ export type Section = {
   readonly firstSentence: Sentence | undefined;
 };
 
+export type Paragraph = { readonly span: Span; readonly sentences: readonly Sentence[] };
+
+/** 箇条書き 1 つ。項目は項目の文字数で持つ。長さのばらつきしか見ないため。 */
+export type BulletList = { readonly span: Span; readonly items: readonly number[] };
+
 /**
  * detector に渡る唯一の入り口。core が I/O を済ませてから呼ぶ。
  * detector は純関数で、fs / network / clock に触れない。spec §6。
@@ -69,10 +97,20 @@ export type ProseDocument = {
   readonly source: string;
   readonly language: string;
   readonly lengthUnit: LengthUnit;
+  /** rule の requires を突き合わせる先。満たさない rule は理由付きで skip する。 */
+  readonly capabilities: AdapterCapabilities;
   readonly sections: readonly Section[];
   readonly sentences: readonly Sentence[];
-  /** アダプタが持つ語彙表。detector は言語を知らずにこれを引く。 */
+  /** 箇条書きの範囲。体言止めのように、箇条書きでは普通で本文では困る形が見る。 */
+  readonly listSpans: readonly Span[];
+  /** 段落。文をいくつ載せているかと、長さのばらつきを見る rule が使う。 */
+  readonly paragraphs: readonly Paragraph[];
+  /** 箇条書き 1 つ。項目の数と長さのばらつきを見る rule が使う。 */
+  readonly lists: readonly BulletList[];
+  /** アダプタが持つ語彙表と、チームが chaff.yaml に足した語彙表。detector は出所を知らない。 */
   readonly lexicons: Readonly<Record<string, Lexicon>>;
+  /** この種類の文書に無いと困る見出し。チームが chaff.yaml で決める。 */
+  readonly requiredSections: readonly string[];
 };
 
 export type Severity = "error" | "warning" | "info";
@@ -116,11 +154,17 @@ export type RuleDefinition = {
   /** 4 語と数値の対応。2 つ以上。未定義の段は normal に落ちる。spec §18.1。
    *  言語別の閾値を持つ rule（max-sentence-length）は、読み込み時に言語で平坦化済み。 */
   readonly levels: LevelTable;
+  /** ジャンル別の上書き。"business" は business/* 全部に効き、"business/email" が勝つ。spec §9。 */
+  readonly by_genre: Readonly<Record<string, LevelTable>>;
   readonly how_to_find: string;
   readonly word_list: string | undefined;
   /** L4 のみ。LLM に渡す決まり。言語別。 */
   readonly what_to_check: Localized | undefined;
   readonly where: string | undefined;
+  /** adapter に要る capability。"pos" / "lemma"。満たさなければ動かさない。spec §16。 */
+  readonly requires: readonly string[];
+  /** 動かす言語。未指定は全言語。「ですます調」のように言語に固有の rule が使う。 */
+  readonly languages: readonly string[] | undefined;
   readonly use_for: readonly string[];
   readonly severity: Severity;
 };

package/src/render/explain.ts

@@ -2,14 +2,21 @@ import { definedLevels, resolve } from "../levels.ts";
 import { localized } from "./text.ts";
 import type { Level, RuleDefinition } from "../plugin.ts";
 
-const levelLine = (rule: RuleDefinition, level: Level, current: Level): string => {
+const levelLine = (rule: RuleDefinition, level: Level, current: Level, genre: string | undefined): string => {
   const mark = level === current ? "\u2192" : " ";
   const name = level.padEnd(9);
-  return level === "off" ? `  ${mark} ${name}見ない` : `  ${mark} ${name}${resolve(rule, level).limit}`;
+  return level === "off" ? `  ${mark} ${name}見ない` : `  ${mark} ${name}${resolve(rule, level, genre).limit}`;
+};
+
+/** ジャンルで数字が変わる rule は、それを言わないと「設定したのに効かない」に見える。 */
+const genreNote = (rule: RuleDefinition, genre: string | undefined): string[] => {
+  const others = Object.keys(rule.by_genre).filter((key) => key !== genre);
+  if (others.length === 0) return [];
+  return ["", `  この数字は ${genre ?? "既定"} のものです。ほかに ${others.join(" / ")} で別の数字を持っています。`];
 };
 
 /** rule の意図と根拠を読む。指摘に納得できないときの入口。 */
-export const renderExplain = (rule: RuleDefinition, current: Level, language: string, unit: string): string => {
+export const renderExplain = (rule: RuleDefinition, current: Level, language: string, unit: string, genre?: string): string => {
   const experimental = rule.status === "experimental" ? ["  このルールはまだ試験中で、既定では動きません（--experimental で動きます）。"] : [];
   return [
     "",
@@ -20,7 +27,8 @@ export const renderExplain = (rule: RuleDefinition, current: Level, language: st
     `  直しかた: ${localized(rule.how_to_fix, language)}`,
     "",
     `  設定できる値（単位: ${unit}）:`,
-    ...definedLevels(rule).map((level) => levelLine(rule, level, current)),
+    ...definedLevels(rule).map((level) => levelLine(rule, level, current, genre)),
+    ...genreNote(rule, genre),
     "",
     `  いまは ${current} です。`,
     ...experimental,

package/src/render/rules-json.ts

@@ -2,14 +2,25 @@ import { definedLevels, resolve } from "../levels.ts";
 import type { Config } from "../config/load.ts";
 import type { Level, RuleDefinition } from "../plugin.ts";
 
-const now = (rule: RuleDefinition, config: Config): Record<string, unknown> => {
+const now = (rule: RuleDefinition, config: Config, genre: string): Record<string, unknown> => {
   const explicit = config.rules[rule.id];
   if (explicit === "off") return { level: "off", why_off: "設定で止めている" };
-  if (explicit !== undefined) return { level: explicit, limit: resolve(rule, explicit).limit };
+  if (explicit !== undefined) return { level: explicit, limit: resolve(rule, explicit, genre).limit };
   if (rule.status === "experimental" && !config.experimental) {
     return { level: "off", why_off: "experimental な rule は既定で動かさない", turn_on_with: `npx chaff lint --experimental` };
   }
-  return { level: "normal", limit: resolve(rule, "normal").limit };
+  return { level: "normal", limit: resolve(rule, "normal", genre).limit };
+};
+
+/** ジャンルで数字が変わる rule があるので、いま効いている表と既定の表の両方を出す。 */
+const levelsOf = (rule: RuleDefinition, genre: string): Record<string, unknown> => {
+  const effective = Object.fromEntries(
+    definedLevels(rule)
+      .filter((level) => level !== "off")
+      .map((level) => [level, resolve(rule, level, genre).limit]),
+  );
+  const overridden = JSON.stringify(effective) !== JSON.stringify(rule.levels);
+  return overridden ? { levels: effective, levels_default: rule.levels, levels_from: "by_genre" } : { levels: rule.levels };
 };
 
 /** AI に設定を書かせるときの入口。推測せずに書けるだけの情報を 1 つに入れる。spec §19.3。 */
@@ -29,10 +40,10 @@ export const rulesJson = (rules: readonly RuleDefinition[], config: Config, lang
         why: rule.why,
         how_to_fix: rule.how_to_fix,
         use_for: rule.use_for,
-        levels: rule.levels,
+        ...levelsOf(rule, genre),
         levels_you_can_set: definedLevels(rule),
         your_setting: config.rules[rule.id] === undefined ? null : { level: config.rules[rule.id], from: config.path },
-        now: now(rule, config),
+        now: now(rule, config, genre),
       })),
       how_to_change: {
         by_command: ['npx chaff relax <rule-id> --why "<理由>"', 'npx chaff strict <rule-id> --why "<理由>"', 'npx chaff off <rule-id> --why "<理由>"'],

package/src/render/semantic.ts

@@ -1,6 +1,6 @@
 import type { Finding, RuleDefinition } from "../plugin.ts";
 import type { UserCheck } from "../checks.ts";
-import type { SemanticResult } from "../run-semantic.ts";
+import type { Job as Plan, SemanticResult } from "../run-semantic.ts";
 import { MARK, localized } from "./text.ts";
 
 const RULE = 60;
@@ -9,6 +9,19 @@ const indent = (text: string, pad: string): string[] => text.split("\n").map((li
 
 export const MACHINE_BANNER = ["", `═══ 機械による判定 ${"═".repeat(RULE - 18)}`, "    同じ文章なら何度実行しても同じ結果になります", ""];
 
+/**
+ * 自分の書いた検査が、どこまで絞り込めたかを見せる。
+ * 絞り込めないまま全文を読んでいることに気づけないと、遅さの原因が分からない。
+ */
+const narrowingLines = (result: SemanticResult): string[] =>
+  result.narrowed.flatMap(({ name, narrowing }) => {
+    if (narrowing.words.length === 0 && !narrowing.needsNumber) {
+      return [`    ${name}: 見るところを絞れず、全文を読みました`, `      look_at に「」で語を書くと、その語を含む文だけになります`];
+    }
+    const what = [...narrowing.words.map((word) => `「${word}」`), ...(narrowing.needsNumber ? ["数字"] : [])].join(" ");
+    return [`    ${name}: ${what} を含む ${narrowing.kept} 文だけを読みました（全 ${narrowing.total} 文）`];
+  });
+
 export const aiBanner = (result: SemanticResult): string[] => [
   "",
   `═══ AI による判定 ${"═".repeat(RULE - 17)}`,
@@ -16,6 +29,7 @@ export const aiBanner = (result: SemanticResult): string[] => [
   "    あります。おかしいと思ったら、そのまま無視して構いません。",
   "",
   `    ${result.sentencesSeen} 文のうち ${result.asked} 箇所を読みました（残りは機械が対象外と判断）`,
+  ...narrowingLines(result),
   "",
 ];
 
@@ -58,3 +72,26 @@ export const renderSemantic = (result: SemanticResult, rules: readonly RuleDefin
   ...result.findings.flatMap((finding) => block(finding, nameOf(finding.rule, rules, checks, language))),
   ...(result.skipped.length > 0 ? [`  ${result.skipped.length} 件の検査は、見るところが無いため動いていません。`, ""] : []),
 ];
+
+/**
+ * 何が送られるのかを、API を呼ばずに見せる。`chaff test --dry-run`。
+ *
+ * 鍵が無い環境でも二段構えの 1 段目まで確かめられる。全文が送られようとしていることに
+ * 請求書で気づくのでは遅い。spec §14。
+ */
+const howNarrowed = (job: Plan): string => {
+  const narrowing = job.narrowing;
+  if (narrowing === undefined) return "機械で絞り込み済み";
+  const quoted = narrowing.words.map((word) => `「${word}」`);
+  const parts = narrowing.needsNumber ? [...quoted, "数字"] : quoted;
+  if (parts.length === 0) return "絞り込めず全文";
+  return `${parts.join(" ")} を含む文`;
+};
+
+export const renderPlan = (path: string, jobs: readonly Plan[], sentences: number): string[] => {
+  const asked = jobs.reduce((sum, job) => sum + job.candidates.length, 0);
+  const lines = jobs.map((job) => `    ${String(job.candidates.length).padStart(3)} 箇所  ${job.name}  （${howNarrowed(job)}）`);
+  const first = jobs.find((job) => job.candidates.length > 0);
+  const sample = first === undefined ? [] : ["", "  最初の 1 件に送る文章:", "", ...indent(String(first.candidates[0]?.text ?? "").slice(0, 300), "    ")];
+  return ["", `${path}   全 ${sentences} 文のうち ${asked} 箇所を送ります（API は呼んでいません）`, "", ...lines, ...sample, ""];
+};

package/src/rule-load.ts

@@ -31,6 +31,16 @@ const flattenLevels = (raw: unknown, language: string): LevelTable | undefined =
   return isLevelTable(forLanguage) ? normalize(forLanguage) : undefined;
 };
 
+/** ジャンル別の上書きも、言語別の levels と同じ形で書ける。読めないものは黙って落とさず捨てる。 */
+const genreTables = (raw: unknown, language: string): Readonly<Record<string, LevelTable>> => {
+  if (!isRecord(raw)) return {};
+  const entries = Object.entries(raw).flatMap(([genre, table]) => {
+    const flattened = flattenLevels(table, language);
+    return flattened === undefined ? [] : [[genre, flattened] as const];
+  });
+  return Object.fromEntries(entries);
+};
+
 const SEVERITIES: readonly Severity[] = ["error", "warning", "info"];
 const isSeverity = (value: unknown): value is Severity => SEVERITIES.some((entry) => entry === value);
 
@@ -58,6 +68,8 @@ const missingFields = (raw: Record<string, unknown>, levels: LevelTable | undefi
 
 const localizedOf = (value: unknown): Record<string, string> => (isLocalized(value) ? value : {});
 
+const stringList = (value: unknown): string[] | undefined => (Array.isArray(value) ? value.map((entry) => String(entry)) : undefined);
+
 const toRule = (raw: unknown, language: string, file: string): RuleDefinition => {
   if (!isRecord(raw)) throw new Error(`${file}: rule は object であること`);
   const levels = flattenLevels(raw["levels"], language);
@@ -73,17 +85,32 @@ const toRule = (raw: unknown, language: string, file: string): RuleDefinition =>
     how_to_fix: localizedOf(raw["how_to_fix"]),
     message: localizedOf(raw["message"]),
     levels,
+    by_genre: genreTables(raw["by_genre"], language),
     how_to_find: String(raw["how_to_find"]),
     word_list: typeof raw["word_list"] === "string" ? raw["word_list"] : undefined,
     what_to_check: isLocalized(raw["what_to_check"]) ? raw["what_to_check"] : undefined,
     where: typeof raw["where"] === "string" ? raw["where"] : undefined,
+    requires: stringList(raw["requires"]) ?? [],
+    languages: stringList(raw["languages"]),
     use_for: Array.isArray(raw["use_for"]) ? raw["use_for"].map((entry) => String(entry)) : [],
     severity: isSeverity(raw["severity"]) ? raw["severity"] : "warning",
   };
 };
 
+/**
+ * 読めない rule ファイルは、どれが何で読めないかを言う。
+ * yaml の例外をそのまま投げると、スタックトレースだけが出てファイル名が出ない。
+ */
+const parseRule = (dir: string, file: string): unknown => {
+  try {
+    return parse(readFileSync(join(dir, file), "utf8"));
+  } catch (error) {
+    throw new Error(`${file} を読めません: ${error instanceof Error ? error.message.split("\n")[0] : String(error)}`, { cause: error });
+  }
+};
+
 export const loadRules = (language: string, dir: string = RULES_DIR): RuleDefinition[] =>
   readdirSync(dir)
     .filter((file) => file.endsWith(".yaml"))
     .sort((left, right) => left.localeCompare(right, "en"))
-    .map((file) => toRule(parse(readFileSync(join(dir, file), "utf8")), language, file));
+    .map((file) => toRule(parseRule(dir, file), language, file));