opencode-wiki-historian 0.2.0

package/dist/migrate-store.d.ts

@@ -0,0 +1,52 @@
+/**
+ * Migrate persistence (plan todo 14): pre-image backup files + the
+ * path-level checkpoint. Both are plain JSON under the plan's 工件根
+ * (results/ in the plugin repo; ~/.config/opencode in the home dir), with
+ * tolerant reads — a corrupt/missing file degrades to an empty state,
+ * never a crash (the pre-image backup is the restore program of record).
+ */
+import type { Locale } from './wiki/pages.read.js';
+/** Write-side field set of one pre-migration page; null = locale absent.
+ *  publishStartDate/publishEndDate ride along so a replay restore via
+ *  updatePage (full RMW) reproduces the exact write-side contract. */
+export interface PreImageEntry {
+    readonly content: string;
+    readonly title: string;
+    readonly description: string;
+    readonly tags: readonly string[];
+    readonly isPublished: boolean;
+    readonly publishStartDate: string;
+    readonly publishEndDate: string;
+}
+export type PreImagePair = Readonly<Record<Locale, PreImageEntry | null>>;
+export interface BackupFile {
+    readonly section: string;
+    readonly createdAt: string;
+    readonly paths: Readonly<Record<string, PreImagePair>>;
+}
+export declare function sectionOf(path: string): string;
+export declare function dateKey(d: Date): string;
+export declare function defaultResultsDir(): string;
+/** Plan verbatim filename: results/pilot-backup-<section>-<date>.json */
+export declare function backupFileFor(section: string, date: string, resultsDir: string): string;
+export declare function readBackup(file: string): BackupFile;
+/** Atomic write: tmp file + rename, so a crash never leaves a truncated
+ *  backup (the restore program of record must itself be restorable). */
+export declare function writeBackup(file: string, backup: BackupFile): void;
+export interface CheckpointEntry {
+    readonly contentHash: string;
+    readonly zhHash: string | null;
+    readonly appliedAt: string;
+    readonly genre: string;
+}
+export interface CheckpointFile {
+    readonly version: 1;
+    readonly paths: Readonly<Record<string, CheckpointEntry>>;
+}
+export declare function checkpointPath(homeDir: string): string;
+export declare function readCheckpoint(homeDir: string): CheckpointFile;
+export declare function writeCheckpoint(homeDir: string, cp: CheckpointFile): void;
+/** Content hash = sha256 over whitespace-normalized text: byte-identical
+ *  content with cosmetic whitespace drift still latches the checkpoint. */
+export declare function hashText(text: string): string;
+export declare function normalizeForHash(s: string): string;

package/dist/migrate-store.js

@@ -0,0 +1,77 @@
+/**
+ * Migrate persistence (plan todo 14): pre-image backup files + the
+ * path-level checkpoint. Both are plain JSON under the plan's 工件根
+ * (results/ in the plugin repo; ~/.config/opencode in the home dir), with
+ * tolerant reads — a corrupt/missing file degrades to an empty state,
+ * never a crash (the pre-image backup is the restore program of record).
+ */
+import * as fs from 'node:fs';
+import { createHash } from 'node:crypto';
+import { dirname, join } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { isRecord } from './jsonc.js';
+export function sectionOf(path) {
+    return path.split('/')[0];
+}
+export function dateKey(d) {
+    const y = d.getFullYear();
+    const m = String(d.getMonth() + 1).padStart(2, '0');
+    const day = String(d.getDate()).padStart(2, '0');
+    return `${y}-${m}-${day}`;
+}
+export function defaultResultsDir() {
+    return fileURLToPath(new URL('../results/', import.meta.url));
+}
+/** Plan verbatim filename: results/pilot-backup-<section>-<date>.json */
+export function backupFileFor(section, date, resultsDir) {
+    return join(resultsDir, `pilot-backup-${section}-${date}.json`);
+}
+export function readBackup(file) {
+    const parsed = readJsonTolerant(file, null);
+    if (!isRecord(parsed) || !isRecord(parsed.paths)) {
+        return { section: '', createdAt: '', paths: {} };
+    }
+    return { section: String(parsed.section ?? ''), createdAt: String(parsed.createdAt ?? ''), paths: parsed.paths };
+}
+/** Atomic write: tmp file + rename, so a crash never leaves a truncated
+ *  backup (the restore program of record must itself be restorable). */
+export function writeBackup(file, backup) {
+    writeJsonAtomic(file, backup);
+}
+export function checkpointPath(homeDir) {
+    return join(homeDir, '.config', 'opencode', 'historian-migrate.json');
+}
+export function readCheckpoint(homeDir) {
+    const parsed = readJsonTolerant(checkpointPath(homeDir), null);
+    if (!isRecord(parsed) || !isRecord(parsed.paths)) {
+        return { version: 1, paths: {} };
+    }
+    return { version: 1, paths: parsed.paths };
+}
+export function writeCheckpoint(homeDir, cp) {
+    writeJsonAtomic(checkpointPath(homeDir), cp);
+}
+// --- Hashing -----------------------------------------------------------------
+/** Content hash = sha256 over whitespace-normalized text: byte-identical
+ *  content with cosmetic whitespace drift still latches the checkpoint. */
+export function hashText(text) {
+    return createHash('sha256').update(normalizeForHash(text)).digest('hex');
+}
+export function normalizeForHash(s) {
+    return s.trim().replace(/\s+/g, ' ');
+}
+// --- fs helpers --------------------------------------------------------------
+function readJsonTolerant(file, fallback) {
+    try {
+        return JSON.parse(fs.readFileSync(file, 'utf8'));
+    }
+    catch {
+        return fallback;
+    }
+}
+function writeJsonAtomic(file, data) {
+    fs.mkdirSync(dirname(file), { recursive: true });
+    const tmp = `${file}.tmp`;
+    fs.writeFileSync(tmp, `${JSON.stringify(data, null, 2)}\n`, 'utf8');
+    fs.renameSync(tmp, file);
+}

package/dist/migrate.d.ts

@@ -0,0 +1,65 @@
+/**
+ * Migration engine — reformat side (plan todo 14): LLM-driven genre restyle.
+ * The dry-run reads the page (en, else zh), classifies the genre (explicit
+ * arg wins), reformats via the translate engine's Anthropic-compatible call
+ * path (SAME-LANGUAGE restyle — the twin is translated at apply time, never
+ * here), and scores the 10-item checklist deterministically on the DRAFT
+ * (items 9-10 are 'deferred' pre-apply; N/A per kind contract).
+ *
+ * No LLM access never crashes: every failure surfaces as a structured
+ * {ok:false, error} with the TranslateError/PageNotFoundError taxonomy.
+ * Apply lives in migrate-apply.ts (this module stays under the LOC ceiling).
+ */
+import type { GqlClient } from './wiki/client.js';
+import type { HistorianOptions } from './config.js';
+import { type Confidence, type Genre, type GenreLang } from './templates/genres.js';
+import { type ChecklistVerdict } from './migrate-score.js';
+import { type Locale, type TranslateFn } from './wiki/pages.read.js';
+export interface MigrateDeps {
+    readonly client: GqlClient;
+    readonly options: HistorianOptions;
+    readonly translate?: TranslateFn;
+    readonly fetchImpl?: typeof fetch;
+    readonly homeDir: string;
+    readonly resultsDir?: string;
+    readonly now?: () => Date;
+}
+export interface ReformatArgs {
+    readonly path: string;
+    readonly genre?: Genre;
+    /** Pilot revise-loop seam (todo 15): corrective hints appended into the
+     *  restyle user prompt when a previous draft failed checklist items. */
+    readonly reviseHints?: readonly string[];
+}
+export type ReformatOutcome = {
+    ok: true;
+    draft: string;
+    genre: Genre;
+    confidence: Confidence | 'explicit';
+    signals: readonly string[];
+    checklistResults: readonly ChecklistVerdict[];
+    urls: {
+        en: string;
+        zh: string;
+    };
+    alreadyConforms: boolean;
+    missingTwin: boolean;
+    sourceLocale: Locale;
+    sourceContent: string;
+    sourceTitle: string;
+} | {
+    ok: false;
+    error: Error;
+};
+/** Locale-aware URL pair for a server-reported path (mirrors tools/shared
+ *  reportUrls; the engine must not import the tools layer). */
+export declare function urlsOf(baseUrl: string, path: string, locale: Locale): {
+    en: string;
+    zh: string;
+};
+export declare function reformatSystemFor(lang: GenreLang): string;
+export declare function reformatPromptFor(genre: Genre, lang: GenreLang, original: string): string;
+/** Dry-run engine: read → classify (explicit genre wins) → LLM restyle →
+ *  deterministic checklist on the draft → conformance signal vs stored
+ *  content. Never writes (the tool, not this module, owns the envelope). */
+export declare function reformatPageDraft(deps: MigrateDeps, args: ReformatArgs): Promise<ReformatOutcome>;

package/dist/migrate.js

@@ -0,0 +1,111 @@
+/**
+ * Migration engine — reformat side (plan todo 14): LLM-driven genre restyle.
+ * The dry-run reads the page (en, else zh), classifies the genre (explicit
+ * arg wins), reformats via the translate engine's Anthropic-compatible call
+ * path (SAME-LANGUAGE restyle — the twin is translated at apply time, never
+ * here), and scores the 10-item checklist deterministically on the DRAFT
+ * (items 9-10 are 'deferred' pre-apply; N/A per kind contract).
+ *
+ * No LLM access never crashes: every failure surfaces as a structured
+ * {ok:false, error} with the TranslateError/PageNotFoundError taxonomy.
+ * Apply lives in migrate-apply.ts (this module stays under the LOC ceiling).
+ */
+import { callMessages } from './translate.js';
+import { classifyGenre, genreSkeleton } from './templates/genres.js';
+import { scoreChecklist, contentSimilar } from './migrate-score.js';
+import { PageNotFoundError } from './wiki/pages.js';
+import { readPage } from './wiki/pages.read.js';
+import { assertLocalePair, twinOf, PathValidationError } from './wiki/locale.js';
+/** Locale-aware URL pair for a server-reported path (mirrors tools/shared
+ *  reportUrls; the engine must not import the tools layer). */
+export function urlsOf(baseUrl, path, locale) {
+    const raw = (l) => `${baseUrl.replace(/\/+$/, '')}/${l}/${path}`;
+    try {
+        const pair = assertLocalePair(path, locale, baseUrl);
+        return locale === 'en' ? { en: pair.url, zh: pair.twinUrl } : { en: pair.twinUrl, zh: pair.url };
+    }
+    catch (err) {
+        if (!(err instanceof PathValidationError))
+            throw err;
+        return { en: raw('en'), zh: raw('zh') };
+    }
+}
+// --- Reformat prompt (rules.md items 1-8, embedded with machine-readable
+//      markers; the LLM restyles, scoring stays deterministic code) -----------
+const RULES = [
+    'R1 结论先行: the lead states the conclusion/action first.',
+    'R2 一页一问: the page answers exactly one question.',
+    'R3 导言占比: lead ≈10-15% of the body, one sentence per major section.',
+    'R4 句长约束: zh sentences ≤20 chars; en sentences ≤25 words.',
+    'R5 表格判据: ≥3-field structured data → table; pairs → description list.',
+    'R6 来源列: comparison/timeline tables carry a source column.',
+    'R7 时间线三列: event timelines are exactly 时间|事件|来源.',
+    'R8 行动项五要素: action items use the five columns 类型|负责人|期限|验证|状态.',
+];
+export function reformatSystemFor(lang) {
+    return [
+        `RESTYLE: ${lang}->${lang}`,
+        'You restructure an existing wiki page into the target genre skeleton.',
+        'RULES:',
+        ...RULES,
+        'PRESERVE every fact, link, and code block verbatim; restructure only.',
+        'Return the full page as Markdown. Do NOT wrap the output in code fences and do not add commentary outside the Markdown.',
+    ].join('\n');
+}
+export function reformatPromptFor(genre, lang, original) {
+    return [
+        `TARGET GENRE SKELETON (${genre}, ${lang}):`,
+        genreSkeleton(genre, lang),
+        '',
+        'ORIGINAL CONTENT TO RESTRUCTURE:',
+        original,
+    ].join('\n');
+}
+/** Append pilot revise-loop hints to the restyle user prompt (todo 15 seam);
+ *  absent hints leave the prompt byte-identical to the pre-seam form. */
+function appendReviseHints(prompt, hints) {
+    if (hints === undefined || hints.length === 0)
+        return prompt;
+    return [
+        prompt,
+        '',
+        'REVISE HINTS (the restructure MUST satisfy every hint):',
+        ...hints.map((h) => `- ${h}`),
+    ].join('\n');
+}
+// --- reformatPageDraft ------------------------------------------------------
+/** Dry-run engine: read → classify (explicit genre wins) → LLM restyle →
+ *  deterministic checklist on the draft → conformance signal vs stored
+ *  content. Never writes (the tool, not this module, owns the envelope). */
+export async function reformatPageDraft(deps, args) {
+    const en = await readPage(deps.client, args.path, 'en');
+    const source = en ?? (await readPage(deps.client, args.path, 'zh'));
+    if (source === null) {
+        return { ok: false, error: new PageNotFoundError(`page '${args.path}' does not exist (checked en and zh locales)`) };
+    }
+    const classified = classifyGenre({ title: source.title, body: source.content });
+    const genre = args.genre ?? classified.genre;
+    const user = appendReviseHints(reformatPromptFor(genre, source.locale, source.content), args.reviseHints);
+    let draft;
+    try {
+        draft = await callMessages(deps.options, { fetchImpl: deps.fetchImpl }, { system: reformatSystemFor(source.locale), user });
+    }
+    catch (err) {
+        return { ok: false, error: err instanceof Error ? err : new Error(String(err)) };
+    }
+    const twin = await readPage(deps.client, args.path, twinOf(source.locale));
+    return {
+        ok: true,
+        draft,
+        genre,
+        confidence: args.genre !== undefined ? 'explicit' : classified.confidence,
+        signals: args.genre !== undefined ? [] : classified.signals,
+        checklistResults: scoreChecklist(genre, draft),
+        urls: urlsOf(deps.options.baseUrl, source.path, source.locale),
+        alreadyConforms: contentSimilar(draft, source.content),
+        missingTwin: twin === null,
+        sourceLocale: source.locale,
+        sourceContent: source.content,
+        sourceTitle: source.title,
+    };
+}

package/dist/templates/genres.d.ts

@@ -0,0 +1,65 @@
+/**
+ * Page-genre system: bilingual G1–G5 templates, deterministic genre
+ * classification, and the 10-item self-review gate (plan todo 9).
+ *
+ * The five genres (digest "Genre templates", cross-cultural-wiki-writing):
+ *   G1 事件/复盘页  — incident postmortem (summary → metadata → background →
+ *                    timeline → impact → root cause → remediation → action
+ *                    items → lessons → appendix)
+ *   G2 对比/选型页  — comparison / selection (conclusion first → dimensions →
+ *                    object overview → comparison table → methodology →
+ *                    recommendation)
+ *   G3 清单/参考页  — inventory / reference (scope statement → structure
+ *                    mirror → entry table → maintenance note)
+ *   G4 概念/原理页  — concept / explanation (definition+rationale →
+ *                    importance-ordered aspects → how it works →
+ *                    attribution/opinion)
+ *   G5 现状卡/账本页 — current-state ledger (status block → deployed
+ *                    components table with per-row last-verified dates →
+ *                    integration → invalidation policy → agent-executable
+ *                    verification commands → append-only change log)
+ *
+ * Every skeleton embeds the harness rubric dimension-C anatomy (status block,
+ * one-line scope, Related Pages tail) and wiki.js expression pieces only —
+ * see skeletons.ts for the raw strings. Consumers (tools.ts, todo 14/15 pilot)
+ * import everything from this barrel.
+ */
+export { G1_EN, G1_ZH, G2_EN, G2_ZH, G3_EN, G3_ZH, G4_EN, G4_ZH, G5_EN, G5_ZH, } from './skeletons.js';
+export type Genre = 'G1' | 'G2' | 'G3' | 'G4' | 'G5';
+export type GenreLang = 'en' | 'zh';
+export type Confidence = 'high' | 'medium' | 'low';
+export declare const GENRES: readonly Genre[];
+export interface ClassifyInput {
+    readonly title: string;
+    readonly body: string;
+}
+export interface ClassifyResult {
+    readonly genre: Genre;
+    readonly confidence: Confidence;
+    /** The keyword cues that were matched (deterministic order, no duplicates). */
+    readonly signals: readonly string[];
+}
+/** One gate item of the 10-item self-review checklist (plan todo 9). */
+export interface ChecklistItem {
+    /** 1..10 — items are always returned as the full contiguous gate. */
+    readonly id: number;
+    readonly label: string;
+    /** kind='genre-specific' items are scored N/A=PASS when the page's genre is
+     *  not in `appliesTo` (pilot scoring, todo 14/15). */
+    readonly kind: 'content' | 'post-write' | 'genre-specific';
+    readonly appliesTo: readonly Genre[] | 'all';
+}
+/** Full markdown skeleton for a genre × language pair. Pure string data —
+ *  the author copies it, replaces placeholders, and fills the commented slots. */
+export declare function genreSkeleton(genre: Genre, lang: GenreLang): string;
+/** Deterministic keyword scoring. No randomness, no state:
+ *  - each genre scores Σ(title hits × 3 + body hits × 1) over its vocabulary;
+ *  - body is truncated to the first 200_000 chars (guard), latin keywords are
+ *    word-boundary matched;
+ *  - all-zero scores → G4/low (garbage fallback);
+ *  - a tie between the two top scores (both > 0) → G4/medium (generalist
+ *    genre wins the ambiguous case);
+ *  - otherwise the top genre wins; confidence = margin ≥ 3 → high, else medium.
+ *  `signals` names the matched cues (used by the pilot engine, todo 14/15). */
+export declare function classifyGenre(input: ClassifyInput): ClassifyResult;
+export declare function selfReviewChecklist(genre: Genre): readonly ChecklistItem[];

package/dist/templates/genres.js

@@ -0,0 +1,228 @@
+/**
+ * Page-genre system: bilingual G1–G5 templates, deterministic genre
+ * classification, and the 10-item self-review gate (plan todo 9).
+ *
+ * The five genres (digest "Genre templates", cross-cultural-wiki-writing):
+ *   G1 事件/复盘页  — incident postmortem (summary → metadata → background →
+ *                    timeline → impact → root cause → remediation → action
+ *                    items → lessons → appendix)
+ *   G2 对比/选型页  — comparison / selection (conclusion first → dimensions →
+ *                    object overview → comparison table → methodology →
+ *                    recommendation)
+ *   G3 清单/参考页  — inventory / reference (scope statement → structure
+ *                    mirror → entry table → maintenance note)
+ *   G4 概念/原理页  — concept / explanation (definition+rationale →
+ *                    importance-ordered aspects → how it works →
+ *                    attribution/opinion)
+ *   G5 现状卡/账本页 — current-state ledger (status block → deployed
+ *                    components table with per-row last-verified dates →
+ *                    integration → invalidation policy → agent-executable
+ *                    verification commands → append-only change log)
+ *
+ * Every skeleton embeds the harness rubric dimension-C anatomy (status block,
+ * one-line scope, Related Pages tail) and wiki.js expression pieces only —
+ * see skeletons.ts for the raw strings. Consumers (tools.ts, todo 14/15 pilot)
+ * import everything from this barrel.
+ */
+import { G1_EN, G1_ZH, G2_EN, G2_ZH, G3_EN, G3_ZH, G4_EN, G4_ZH, G5_EN, G5_ZH, } from './skeletons.js';
+// Contract re-exports: raw skeleton data stays reachable through the barrel
+// (todo 13 renders these double-checked strings into skill references).
+export { G1_EN, G1_ZH, G2_EN, G2_ZH, G3_EN, G3_ZH, G4_EN, G4_ZH, G5_EN, G5_ZH, } from './skeletons.js';
+export const GENRES = ['G1', 'G2', 'G3', 'G4', 'G5'];
+// --- genreSkeleton ----------------------------------------------------------
+const SKELETONS = {
+    G1: { en: G1_EN, zh: G1_ZH },
+    G2: { en: G2_EN, zh: G2_ZH },
+    G3: { en: G3_EN, zh: G3_ZH },
+    G4: { en: G4_EN, zh: G4_ZH },
+    G5: { en: G5_EN, zh: G5_ZH },
+};
+/** Full markdown skeleton for a genre × language pair. Pure string data —
+ *  the author copies it, replaces placeholders, and fills the commented slots. */
+export function genreSkeleton(genre, lang) {
+    return SKELETONS[genre][lang];
+}
+// --- classifyGenre ----------------------------------------------------------
+/** Title hits weigh TITLE_WEIGHT× a body hit: a title naming the genre (e.g.
+ *  "Postmortem") is the strongest signal; body cues still move the needle. */
+const TITLE_WEIGHT = 3;
+const BODY_WEIGHT = 1;
+/** Never scan more than this many chars of body (truncation guard). */
+const BODY_TRUNCATE_CHARS = 200_000;
+/** Keyword vocabulary, one set per genre. Latin cues are matched on word
+ *  boundaries (case-insensitive) so "vs" never fires inside "moves"/"versus".
+ *  Keep this table in sync with skills references when it changes. */
+const GENRE_KEYWORDS = {
+    G1: ['故障', '复盘', '事故', 'incident', 'postmortem', 'outage'],
+    G2: ['对比', '选型', 'vs', 'versus', 'compare', 'benchmark', 'alternatives'],
+    G3: ['清单', '列表', 'inventory', 'checklist', 'catalog', '命令速查'],
+    G4: ['原理', '为什么', 'how it works', '概念', '机制'],
+    // Deployed-state cues only — bare 部署/版本 would collide with G3 "部署清单"
+    // and general changelog talk, regressing existing corpus classifications.
+    G5: ['现状卡', '当前状态', '已部署', '部署物', '现役', '上线', '端口', '上次核实', '失效策略', '验证命令', 'current state', 'deployed', 'last verified', 'running now'],
+};
+function escapeRegex(s) {
+    return s.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+}
+/** Occurrence count of one keyword in a haystack: word-boundary regex for
+ *  latin cues, plain substring scan for CJK (non-overlapping). */
+function countMatches(haystack, keyword) {
+    if (/[A-Za-z]/.test(keyword)) {
+        return (haystack.match(new RegExp(`\\b${escapeRegex(keyword)}\\b`, 'gi')) ?? []).length;
+    }
+    let count = 0;
+    let idx = haystack.indexOf(keyword);
+    while (idx !== -1) {
+        count += 1;
+        idx = haystack.indexOf(keyword, idx + keyword.length);
+    }
+    return count;
+}
+/** Deterministic keyword scoring. No randomness, no state:
+ *  - each genre scores Σ(title hits × 3 + body hits × 1) over its vocabulary;
+ *  - body is truncated to the first 200_000 chars (guard), latin keywords are
+ *    word-boundary matched;
+ *  - all-zero scores → G4/low (garbage fallback);
+ *  - a tie between the two top scores (both > 0) → G4/medium (generalist
+ *    genre wins the ambiguous case);
+ *  - otherwise the top genre wins; confidence = margin ≥ 3 → high, else medium.
+ *  `signals` names the matched cues (used by the pilot engine, todo 14/15). */
+export function classifyGenre(input) {
+    const title = input.title.toLowerCase();
+    const rawBody = input.body;
+    const body = (rawBody.length > BODY_TRUNCATE_CHARS ? rawBody.slice(0, BODY_TRUNCATE_CHARS) : rawBody).toLowerCase();
+    const scores = new Map();
+    const signals = [];
+    for (const genre of GENRES) {
+        let score = 0;
+        for (const keyword of GENRE_KEYWORDS[genre]) {
+            const titleHits = countMatches(title, keyword);
+            const bodyHits = countMatches(body, keyword);
+            if (titleHits > 0 || bodyHits > 0)
+                signals.push(keyword);
+            score += titleHits * TITLE_WEIGHT + bodyHits * BODY_WEIGHT;
+        }
+        scores.set(genre, score);
+    }
+    const [top, second] = [...GENRES].sort((a, b) => (scores.get(b) ?? 0) - (scores.get(a) ?? 0));
+    const topScore = scores.get(top) ?? 0;
+    const secondScore = scores.get(second) ?? 0;
+    if (topScore === 0)
+        return { genre: 'G4', confidence: 'low', signals: [] };
+    const margin = topScore - secondScore;
+    if (margin === 0) {
+        // Ambiguous: two genres scored equally — the generalist G4 wins the tie.
+        return { genre: 'G4', confidence: 'medium', signals };
+    }
+    return { genre: top, confidence: margin >= TITLE_WEIGHT ? 'high' : 'medium', signals };
+}
+// --- selfReviewChecklist ----------------------------------------------------
+/**
+ * The 10-item self-review gate (plan todo 9). The FULL gate is always
+ * returned (ids 1..10 contiguous); the pilot engine (todo 15) applies items
+ * via the kind/appliesTo contract:
+ *  - content items (1–3, 7, 8)     — scored on the rewritten draft (dry-run);
+ *  - genre-specific items (4–6)    — N/A=PASS when the page's genre is not in
+ *    appliesTo (G2 pages only have a source column; only G1 pages have
+ *    timeline sources and the action-item five essentials; G5 pages swap in
+ *    the ledger variants: per-row last-verified, verification commands, no
+ *    narrative prose);
+ *  - post-write items (9, 10)      — scored after apply + written into the
+ *    pilot report (never scored pre-write, where they are FAIL by
+ *    construction).
+ * `genre` is validated (unknown values throw) so callers cannot silently
+ * score against a partial gate.
+ */
+/** G5 swaps the genre-specific items 4–6 for ledger gates; everything else
+ *  (1–3, 7–10) is shared with the base gate. */
+const G5_CHECKLIST_VARIANTS = {
+    4: {
+        id: 4,
+        label: '部署物清单每行带「上次核实于/Last verified」列，缺日期即过期（ledger component rows carry a last-verified date）',
+        kind: 'genre-specific',
+        appliesTo: ['G5'],
+    },
+    5: {
+        id: 5,
+        label: '验证方法节存在：每个组件对应一条 agent 可直接执行的复核命令（verification section: an executable re-check command per component）',
+        kind: 'genre-specific',
+        appliesTo: ['G5'],
+    },
+    6: {
+        id: 6,
+        label: '无叙事正文：现状卡 = 状态块 + 表格，叙述历史链向 G1 事件页（table+status page: no narrative prose bodies; history links to G1 pages）',
+        kind: 'genre-specific',
+        appliesTo: ['G5'],
+    },
+};
+export function selfReviewChecklist(genre) {
+    if (!GENRES.includes(genre)) {
+        throw new Error(`unknown genre: ${genre}`);
+    }
+    const items = [
+        {
+            id: 1,
+            label: '导言占比 10–15%：导言 ≈ 正文的 10–15%，每个重要小节在导言至少占一句（lead ≈10–15% of body; every major section ≥1 sentence in lead）',
+            kind: 'content',
+            appliesTo: 'all',
+        },
+        {
+            id: 2,
+            label: '句长上限：中文句 ≤20 字、英文句 ≤25 词（sentence cap: zh ≤20 chars, en ≤25 words）',
+            kind: 'content',
+            appliesTo: 'all',
+        },
+        {
+            id: 3,
+            label: '表格判据：≥3 字段的结构化枚举入表，成对数据用描述列表（≥3 fields → table; pairs → description list）',
+            kind: 'content',
+            appliesTo: 'all',
+        },
+        {
+            id: 4,
+            label: '对比表/枚举表每行有来源列，行序固定、无合并单元格（comparison/enumeration table has a source column）',
+            kind: 'genre-specific',
+            appliesTo: ['G2'],
+        },
+        {
+            id: 5,
+            label: '时间线每行有来源列，仅日志可证事实（timeline rows carry sources; log-verifiable facts only）',
+            kind: 'genre-specific',
+            appliesTo: ['G1'],
+        },
+        {
+            id: 6,
+            label: '行动项五要素 = 类型|负责人|期限|验证|状态 五列，措施是行内容（action items: five essentials as columns）',
+            kind: 'genre-specific',
+            appliesTo: ['G1'],
+        },
+        {
+            id: 7,
+            label: '无杂项筐：除 参见/附录 之外没有 "其他/杂项" 类 catch-all 小节（no catch-all sections outside See-Also/Appendix）',
+            kind: 'content',
+            appliesTo: 'all',
+        },
+        {
+            id: 8,
+            label: '无溢美词：领先/强大/灵活/高效 等 bare claim 改事实或删除（no unbacked praise: 领先/强大/灵活/高效 without evidence）',
+            kind: 'content',
+            appliesTo: 'all',
+        },
+        {
+            id: 9,
+            label: '双语 URL 已回报：报告含 /en/ 与 /zh/ 两个可访问 URL（post-write: report carries both /en/ and /zh/ URLs）',
+            kind: 'post-write',
+            appliesTo: 'all',
+        },
+        {
+            id: 10,
+            label: '孪生页已建 或 zh_status:pending 已记录并在报告中声明（twin created OR zh_status pending recorded and declared in report）',
+            kind: 'post-write',
+            appliesTo: 'all',
+        },
+    ];
+    if (genre === 'G5') {
+        return items.map((item) => G5_CHECKLIST_VARIANTS[item.id] ?? item);
+    }
+    return items;
+}

package/dist/templates/skeletons.d.ts

@@ -0,0 +1,48 @@
+/**
+ * Bilingual G1–G4 page skeletons (template data).
+ *
+ * Provenance: `.omo/research/cross-cultural-wiki-writing-digest.md` (Genre
+ * templates + SYN-1..20) and `.omo/research/wikijs-2x-capabilities-digest.md`
+ * (expression syntax), vendored at `docs/research/*`.
+ *
+ * Contract of every skeleton constant:
+ *  - Rubric dimension C anatomy: H1 placeholder → status line
+ *    `**状态/Status**: Active <!-- or Historical/Superseded --> · **日期/Date**: YYYY-MM-DD`
+ *    → one-line scope (`This page answers:` / `本页回答：`) → tail
+ *    `Related Pages`/`相关页面` section with a real-link hint (SYN-9 fixed tail).
+ *  - wiki.js 2.x expression pieces only: blockquote admonitions
+ *    (`> …` + `{.is-info}`), `{.dense}` tables, `[^1]` footnotes.
+ *  - FORBIDDEN: `{{toc}}`, `:::` containers, YAML frontmatter (`---` at pos 0) —
+ *    wiki.js does not support them (they render as body text or v-pre escapes).
+ *  - Language-native: zh skeletons use 中文节标题 and zh author guidance,
+ *    en skeletons English; the bilingual pairs correspond section-for-section.
+ *  - Inline `<!-- … -->` comments carry author guidance per section (what to
+ *    write, length caps); placeholders are marked 占位/placeholder and must be
+ *    replaced by the author before publishing.
+ *
+ * // allow: SIZE_OK — pure template data, one constant per (genre, lang);
+ *     split across files would buy nothing (each pair is a single narrative).
+ */
+/** G1 — 事件/复盘页 (incident postmortem), zh. */
+export declare const G1_ZH: string;
+/** G1 — incident postmortem, en (section-for-section twin of G1_ZH). */
+export declare const G1_EN: string;
+/** G2 — 对比/选型页 (comparison / selection), zh. */
+export declare const G2_ZH: string;
+/** G2 — comparison / selection, en (section-for-section twin of G2_ZH). */
+export declare const G2_EN: string;
+/** G3 — 清单/参考页 (inventory / reference), zh. */
+export declare const G3_ZH: string;
+/** G3 — inventory / reference, en (section-for-section twin of G3_ZH). */
+export declare const G3_EN: string;
+/** G4 — 概念/原理解析页 (concept / explanation), zh. */
+export declare const G4_ZH: string;
+/** G4 — concept / explanation, en (section-for-section twin of G4_ZH). */
+export declare const G4_EN: string;
+/** G5 — 现状卡/部署现状账本页 (current-state ledger), zh. Status + tables only:
+ *  one authoritative snapshot of what is deployed/running NOW, per-row
+ *  last-verified dates, agent-executable re-check commands. Narrative history
+ *  belongs to G1 event pages, linked from 变更记录. */
+export declare const G5_ZH: string;
+/** G5 — current-state ledger, en (section-for-section twin of G5_ZH). */
+export declare const G5_EN: string;