simt-dev 0.3.1 → 0.5.0

package/src/graph/compile.mjs

@@ -0,0 +1,475 @@
+// Skill compiler, progressive-disclosure splitter, and trigger report (plan 0003 R30).
+//
+// 1. Compiles a skill into the Agent Skills layout (`<name>/SKILL.md` + `references/*.md`)
+//    with frontmatter that passes `checkSkill` (review 0002 S44, ADR-069).
+// 2. Flags every skill body over 5,000 tokens (20,000 bytes at 4 B/token) with a
+//    deterministic split plan, and splits markdown bodies by section into `references/`
+//    so `SKILL.md` stays within the 5,000-token and 500-line budget.
+// 3. Reads R02 transcripts and run records to determine whether the agent opened a skill
+//    when the task's gold patch touched its declared `paths`, reporting trigger recall
+//    and precision per skill.
+// 4. Builds offline, proposal-only description improvements with provenance (`model`,
+//    `promptHash`, `digest`) so model output never lands directly in committed text (P2).
+
+import { createHash } from 'node:crypto';
+import { checkSkill, nameProblems, readFrontmatter } from './spec.mjs';
+import { matchesGlob } from '../resolve/resolve.mjs';
+
+export const BODY_TOKEN_BUDGET = 5000;
+export const BYTES_PER_TOKEN = 4;
+export const BODY_BYTE_BUDGET = BODY_TOKEN_BUDGET * BYTES_PER_TOKEN;
+export const BODY_LINE_BUDGET = 500;
+
+const TARGET_MAIN_BYTES = 12_000;
+const TARGET_MAIN_LINES = 350;
+const TARGET_REF_BYTES = 16_000;
+const SPEC_KEYS = new Set(['name', 'description', 'license', 'compatibility', 'metadata', 'allowed-tools']);
+
+function sha256Tag(text) {
+  return `sha256-${createHash('sha256').update(String(text ?? ''), 'utf8').digest('hex').slice(0, 16)}`;
+}
+
+function slugify(title, index) {
+  const base = String(title ?? '')
+    .toLowerCase()
+    .replace(/`+/g, '')
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '')
+    .replace(/-{2,}/g, '-')
+    .slice(0, 48);
+  const num = String(index + 1).padStart(2, '0');
+  return base ? `${num}-${base}` : `${num}-section`;
+}
+
+/** Build a conforming Agent Skills description (30..1024 chars, stating what + when). */
+export function buildConformingDescription(skill = {}, rawDescription = '') {
+  const clean = String(rawDescription ?? '').replace(/\s+/g, ' ').trim();
+  const summary = String(skill.summary ?? '').replace(/\s+/g, ' ').trim();
+  const paths = Array.isArray(skill.paths) && skill.paths.length > 0 ? skill.paths.slice(0, 4).join(', ') : '';
+  const keywords = Array.isArray(skill.keywords) && skill.keywords.length > 0 ? skill.keywords.slice(0, 5).join(', ') : '';
+
+  const whenClause = paths
+    ? `Use when touching ${paths}${keywords ? ` or working with ${keywords}` : ''}.`
+    : keywords
+      ? `Use when working with ${keywords}.`
+      : 'Use when a task matches this capability.';
+
+  let desc = clean || summary || `Instructions for ${skill.name ?? 'this skill'}.`;
+  if (!desc.endsWith('.')) desc += '.';
+  if (desc.length < 30 || !/\buse when\b/i.test(desc)) {
+    const combined = `${desc} ${whenClause}`.replace(/\s+/g, ' ').trim();
+    if (combined.length <= 1024) desc = combined;
+  }
+  if (desc.length < 30) desc = `${desc} ${whenClause}`.trim();
+  if (desc.length > 1024) desc = `${desc.slice(0, 1021).trimEnd()}...`;
+  return desc;
+}
+
+function splitMarkdownSections(bodyText) {
+  const lines = String(bodyText ?? '').split(/\r?\n/);
+  const preamble = [];
+  const sections = [];
+  let current = null;
+  let inFence = false;
+
+  for (const line of lines) {
+    if (line.trim().startsWith('```')) inFence = !inFence;
+    const h2 = !inFence ? /^##\s+(.+?)\s*$/.exec(line) : null;
+    if (h2) {
+      if (current) sections.push(current);
+      current = { title: h2[1].trim(), lines: [line] };
+      continue;
+    }
+    if (current) current.lines.push(line);
+    else preamble.push(line);
+  }
+  if (current) sections.push(current);
+
+  if (sections.length <= 1 && lines.length > BODY_LINE_BUDGET) {
+    const chunkSize = 200;
+    const bodyLines = sections.length === 1 ? [...preamble, ...sections[0].lines] : preamble;
+    const intro = bodyLines.slice(0, 80);
+    const rest = bodyLines.slice(80);
+    const chunked = [];
+    for (let i = 0; i < rest.length; i += chunkSize) {
+      const partNum = Math.floor(i / chunkSize) + 1;
+      chunked.push({
+        title: `Reference Part ${partNum}`,
+        lines: [`## Reference Part ${partNum}`, '', ...rest.slice(i, i + chunkSize)],
+      });
+    }
+    return { preamble: intro.join('\n').trim(), sections: chunked };
+  }
+  return { preamble: preamble.join('\n').trim(), sections };
+}
+
+/**
+ * Compute a deterministic progressive-disclosure split plan for a skill.
+ * Every skill over 5,000 tokens (20,000 bytes) or 500 lines is flagged with a split plan.
+ */
+export function planSkillSplit(skill = {}, bodyText = null) {
+  const hasBody = typeof bodyText === 'string' && bodyText.length > 0;
+  const bodyBytes = hasBody ? Buffer.byteLength(bodyText, 'utf8') : 0;
+  const totalBytes = Math.max(Number(skill.bytes ?? 0), bodyBytes);
+  const totalTokens = Math.ceil(totalBytes / BYTES_PER_TOKEN);
+  const totalLines = hasBody ? bodyText.split(/\r?\n/).length : Math.ceil(totalBytes / 60);
+  const overTokens = totalTokens > BODY_TOKEN_BUDGET;
+  const flagged = overTokens;
+
+  if (!flagged) {
+    return {
+      flagged: false,
+      source: hasBody ? 'body' : 'registry',
+      reason: null,
+      totalBytes,
+      totalTokens,
+      totalLines,
+      mainBytes: bodyBytes || totalBytes,
+      mainTokens: Math.ceil((bodyBytes || totalBytes) / BYTES_PER_TOKEN),
+      references: [],
+    };
+  }
+
+  const reason = `${totalTokens} tokens > ${BODY_TOKEN_BUDGET} token budget`;
+
+  if (!hasBody || bodyBytes <= BODY_BYTE_BUDGET) {
+    const mainBytes = hasBody ? bodyBytes : Math.min(totalBytes, TARGET_MAIN_BYTES);
+    const mainTokens = Math.ceil(mainBytes / BYTES_PER_TOKEN);
+    const remainingBytes = Math.max(1024, totalBytes - mainBytes);
+    const topics = [...(skill.capabilities ?? []), ...(skill.keywords ?? []).slice(0, 4)];
+    const refCount = Math.max(1, Math.ceil(remainingBytes / TARGET_REF_BYTES));
+    const references = [];
+    let left = remainingBytes;
+    for (let i = 0; i < refCount; i += 1) {
+      const share = i === refCount - 1 ? left : Math.min(left, Math.ceil(remainingBytes / refCount));
+      left -= share;
+      const topic = topics[i] ?? `reference-${i + 1}`;
+      references.push({
+        path: `references/${slugify(topic, i)}.md`,
+        title: `Reference: ${topic}`,
+        bytes: share,
+        tokens: Math.ceil(share / BYTES_PER_TOKEN),
+        lines: Math.max(1, Math.ceil(share / 60)),
+      });
+    }
+    return {
+      flagged: true,
+      source: hasBody ? 'body' : 'registry',
+      reason,
+      totalBytes,
+      totalTokens,
+      totalLines,
+      mainBytes,
+      mainTokens,
+      references,
+    };
+  }
+
+  const parsed = readFrontmatter(bodyText);
+  const { preamble, sections } = splitMarkdownSections(parsed.body);
+  const keptSections = [];
+  const movedSections = [];
+  let runningBytes = Buffer.byteLength(preamble, 'utf8') + 512;
+  let runningLines = preamble.split(/\r?\n/).length + 25;
+
+  for (const section of sections) {
+    const secBytes = Buffer.byteLength(section.lines.join('\n').trim(), 'utf8');
+    const secLines = section.lines.length;
+    if (
+      movedSections.length === 0 &&
+      runningBytes + secBytes <= TARGET_MAIN_BYTES &&
+      runningLines + secLines <= TARGET_MAIN_LINES &&
+      sections.length > 1
+    ) {
+      keptSections.push(section);
+      runningBytes += secBytes;
+      runningLines += secLines;
+    } else {
+      movedSections.push(section);
+    }
+  }
+  if (movedSections.length === 0 && sections.length > 0) movedSections.push(keptSections.pop());
+
+  const references = movedSections.map((sec, idx) => {
+    const content = `# ${sec.title}\n\n${sec.lines.slice(1).join('\n').trim()}\n`;
+    const bytes = Buffer.byteLength(content, 'utf8');
+    return {
+      path: `references/${slugify(sec.title, idx)}.md`,
+      title: sec.title,
+      bytes,
+      tokens: Math.ceil(bytes / BYTES_PER_TOKEN),
+      lines: content.split('\n').length,
+      content,
+    };
+  });
+
+  const keptBody = [preamble, ...keptSections.map((s) => s.lines.join('\n').trim())].filter(Boolean).join('\n\n');
+  const refIndexLines = [
+    '## References',
+    '',
+    'Progressive-disclosure reference files (load on demand when needed):',
+    ...references.map((r) => `- [${r.title}](${r.path}) (~${r.tokens} tokens, ${r.bytes} bytes)`),
+  ];
+  const compiledBody = `${keptBody}\n\n${refIndexLines.join('\n')}\n`;
+  const mainBytes = Buffer.byteLength(compiledBody, 'utf8');
+
+  return {
+    flagged: true,
+    source: 'body',
+    reason,
+    totalBytes,
+    totalTokens,
+    totalLines,
+    mainBytes,
+    mainTokens: Math.ceil(mainBytes / BYTES_PER_TOKEN),
+    compiledBody,
+    references,
+  };
+}
+
+function serializeFrontmatter({ name, description, license, compatibility, allowedTools, metadata }) {
+  const lines = ['---', `name: ${name}`, `description: ${JSON.stringify(description)}`];
+  if (license) lines.push(`license: ${JSON.stringify(String(license))}`);
+  if (compatibility) lines.push(`compatibility: ${JSON.stringify(String(compatibility).slice(0, 500))}`);
+  if (allowedTools) lines.push(`allowed-tools: ${JSON.stringify(String(allowedTools))}`);
+  const metaEntries = Object.entries(metadata ?? {}).filter(([, v]) => v !== undefined && v !== null && v !== '');
+  if (metaEntries.length > 0) {
+    lines.push('metadata:');
+    for (const [k, v] of metaEntries.sort(([a], [b]) => a.localeCompare(b))) {
+      lines.push(`  ${k}: ${JSON.stringify(String(v))}`);
+    }
+  }
+  lines.push('---');
+  return lines.join('\n');
+}
+
+/** Compile a skill into Agent Skills conforming `SKILL.md` + optional `references/*.md`. */
+export function compileSkill(skill = {}, rawText = null) {
+  const name = String(skill.name ?? '').trim();
+  const brokenName = nameProblems(name);
+  if (brokenName) throw new Error(`skill "${name}" cannot be compiled: name ${brokenName}`);
+
+  const parsed = rawText ? readFrontmatter(rawText) : { meta: {}, present: [], body: '' };
+  const description = buildConformingDescription(skill, parsed.meta.description);
+  const metadata = {};
+  for (const [k, v] of Object.entries(parsed.meta.metadata ?? {})) metadata[k] = String(v);
+  for (const key of parsed.present) {
+    if (!SPEC_KEYS.has(key) && parsed.meta[key] !== undefined) metadata[key] = String(parsed.meta[key]);
+  }
+
+  const fmBlock = serializeFrontmatter({
+    name,
+    description,
+    license: parsed.meta.license,
+    compatibility: parsed.meta.compatibility,
+    allowedTools: parsed.meta['allowed-tools'],
+    metadata,
+  });
+
+  const defaultBody = parsed.body.trim() || `# ${name}\n\n${skill.summary ?? description}`;
+  const fullCandidate = `${fmBlock}\n\n${defaultBody}\n`;
+  const splitPlan = planSkillSplit(skill, rawText ? fullCandidate : null);
+  const finalBody = splitPlan.compiledBody ? splitPlan.compiledBody.trim() : defaultBody;
+  const mainText = `${fmBlock}\n\n${finalBody}\n`;
+  const mainBytes = Buffer.byteLength(mainText, 'utf8');
+  const mainTokens = Math.ceil(mainBytes / BYTES_PER_TOKEN);
+  const conformance = checkSkill({ directoryName: name, fileName: 'SKILL.md', text: mainText, at: name });
+
+  return {
+    name,
+    description,
+    mainPath: `${name}/SKILL.md`,
+    mainText,
+    mainBytes,
+    mainTokens,
+    references: (splitPlan.references ?? []).map((r) => ({
+      path: r.path,
+      title: r.title,
+      bytes: r.bytes,
+      tokens: r.tokens,
+      lines: r.lines,
+      ...(r.content ? { content: r.content } : {}),
+    })),
+    splitPlan: {
+      flagged: splitPlan.flagged,
+      source: splitPlan.source,
+      reason: splitPlan.reason,
+      totalBytes: splitPlan.totalBytes,
+      totalTokens: splitPlan.totalTokens,
+      mainBytes: rawText ? mainBytes : splitPlan.mainBytes,
+      mainTokens: rawText ? mainTokens : splitPlan.mainTokens,
+      references: (splitPlan.references ?? []).map(({ path, title, bytes, tokens, lines }) => ({
+        path,
+        title,
+        bytes,
+        tokens,
+        lines,
+      })),
+    },
+    conformance,
+  };
+}
+
+/** Return the sorted names of skills whose declared `paths` globs match any file in `touchedPaths`. */
+export function matchSkillsForPaths(skills = [], touchedPaths = []) {
+  const cleanPaths = (touchedPaths ?? []).map((p) => String(p ?? '').replace(/^\/+/, '').trim()).filter(Boolean);
+  if (cleanPaths.length === 0) return [];
+  const matched = [];
+  for (const skill of skills) {
+    if (!skill?.name || skill.aliasOf) continue;
+    const globs = Array.isArray(skill.paths) ? skill.paths : [];
+    if (globs.length > 0 && cleanPaths.some((file) => globs.some((glob) => matchesGlob(file, glob)))) {
+      matched.push(skill.name);
+    }
+  }
+  return [...new Set(matched)].sort();
+}
+
+/** Extract the sorted list of skill names opened by an agent from an R02 transcript. */
+export function extractOpenedSkills(transcriptInput, knownSkillNames = null) {
+  const known = knownSkillNames ? new Set(knownSkillNames) : null;
+  const opened = new Set();
+  const recordIfKnown = (candidate) => {
+    const name = String(candidate ?? '').trim().toLowerCase();
+    if (!name || nameProblems(name) !== null) return;
+    if (!known || known.has(name)) opened.add(name);
+  };
+
+  if (Array.isArray(transcriptInput)) {
+    for (const item of transcriptInput) {
+      if (typeof item === 'string') recordIfKnown(item);
+      else if (item && typeof item === 'object') {
+        for (const s of extractOpenedSkills(JSON.stringify(item), known)) opened.add(s);
+      }
+    }
+    return [...opened].sort();
+  }
+
+  const text = typeof transcriptInput === 'string' ? transcriptInput : JSON.stringify(transcriptInput ?? '');
+  for (const m of text.matchAll(/(?:^|[\s"'`:=(/])skills\/([a-z0-9-]+)(?:\/SKILL\.md|\.md|\/|[\s"'`),]|$)/gi)) {
+    recordIfKnown(m[1]);
+  }
+  for (const m of text.matchAll(/(?:^|[\s"'`:=(/])([a-z0-9-]+)\/SKILL\.md\b/g)) recordIfKnown(m[1]);
+  for (const m of text.matchAll(/\bsimt(?:-dev)?\s+skills\s+show\s+([a-z0-9-]+)\b/gi)) recordIfKnown(m[1]);
+  for (const m of text.matchAll(/"(?:skill|skill_name|skillName)"\s*:\s*"([a-z0-9-]+)"/gi)) recordIfKnown(m[1]);
+  for (const m of text.matchAll(/"name"\s*:\s*"skills_show"[\s\S]{0,200}?"name"\s*:\s*"([a-z0-9-]+)"/gi)) {
+    recordIfKnown(m[1]);
+  }
+  return [...opened].sort();
+}
+
+/** Compute per-skill trigger recall/precision across R02 runs and attach split plans. */
+export function evaluateSkillTriggers(registry, runs = [], { bodiesBySkill = new Map(), goldPathsByTask = new Map() } = {}) {
+  const skills = [...(registry?.skills ?? [])].sort((a, b) => a.name.localeCompare(b.name));
+  const knownNames = new Set(skills.map((s) => s.name));
+
+  const normalizedRuns = (runs ?? []).map((run) => {
+    const taskId = String(run?.task ?? '');
+    const expectedFromRun = Array.isArray(run?.skills?.expected)
+      ? run.skills.expected
+      : Array.isArray(run?.expectedSkills) ? run.expectedSkills : null;
+    const goldPaths = run?.goldFiles ?? goldPathsByTask.get(taskId) ?? [];
+    const expected = expectedFromRun
+      ? expectedFromRun.filter((n) => knownNames.has(n))
+      : matchSkillsForPaths(skills, goldPaths);
+    const openedFromRun = Array.isArray(run?.skills?.opened)
+      ? run.skills.opened
+      : Array.isArray(run?.openedSkills) ? run.openedSkills : null;
+    const opened = openedFromRun
+      ? openedFromRun.filter((n) => knownNames.has(n))
+      : extractOpenedSkills(run?.transcript ?? '', knownNames);
+    return { task: taskId, arm: String(run?.arm ?? 'default'), expected: new Set(expected), opened: new Set(opened) };
+  });
+
+  const rows = skills.map((skill) => {
+    let tp = 0;
+    let fp = 0;
+    let fn = 0;
+    for (const run of normalizedRuns) {
+      const exp = run.expected.has(skill.name);
+      const opn = run.opened.has(skill.name);
+      if (exp && opn) tp += 1;
+      else if (!exp && opn) fp += 1;
+      else if (exp && !opn) fn += 1;
+    }
+    const expected = tp + fn;
+    const opened = tp + fp;
+    const splitPlan = planSkillSplit(skill, bodiesBySkill.get(skill.name) ?? null);
+    return {
+      name: skill.name,
+      bytes: skill.bytes ?? 0,
+      tokens: splitPlan.totalTokens,
+      paths: skill.paths ?? [],
+      expected,
+      opened,
+      tp,
+      fp,
+      fn,
+      recall: expected > 0 ? tp / expected : null,
+      precision: opened > 0 ? tp / opened : null,
+      flagged: splitPlan.flagged,
+      splitPlan: {
+        flagged: splitPlan.flagged,
+        source: splitPlan.source,
+        reason: splitPlan.reason,
+        totalBytes: splitPlan.totalBytes,
+        totalTokens: splitPlan.totalTokens,
+        mainBytes: splitPlan.mainBytes,
+        mainTokens: splitPlan.mainTokens,
+        references: (splitPlan.references ?? []).map(({ path, title, bytes, tokens, lines }) => ({
+          path,
+          title,
+          bytes,
+          tokens,
+          lines,
+        })),
+      },
+    };
+  });
+
+  return {
+    runCount: normalizedRuns.length,
+    skillCount: rows.length,
+    flaggedCount: rows.filter((r) => r.flagged).length,
+    skills: rows,
+  };
+}
+
+/** Create a proposal-only description improvement with provenance (Principle P2). */
+export async function proposeSkillDescription(
+  skill,
+  { currentDescription = '', triggerStats = null, proposer = null, model = 'deterministic-template' } = {},
+) {
+  const prompt = [
+    `Skill: ${skill.name}`,
+    `Summary: ${skill.summary ?? ''}`,
+    `Paths: ${(skill.paths ?? []).join(', ') || '(none)'}`,
+    `Keywords: ${(skill.keywords ?? []).join(', ') || '(none)'}`,
+    `Capabilities: ${(skill.capabilities ?? []).join(', ') || '(none)'}`,
+    `Current description: ${currentDescription || skill.summary || ''}`,
+    triggerStats
+      ? `Trigger stats: expected=${triggerStats.expected}, opened=${triggerStats.opened}, recall=${triggerStats.recall ?? 'n/a'}, precision=${triggerStats.precision ?? 'n/a'}`
+      : 'Trigger stats: unmeasured',
+  ].join('\n');
+
+  const promptHash = sha256Tag(prompt);
+  let candidate = '';
+  let usedModel = model;
+  if (typeof proposer === 'function') {
+    const raw = await proposer({ skill, prompt, triggerStats });
+    if (typeof raw === 'string' && raw.trim()) candidate = raw.trim();
+    else if (raw && typeof raw.description === 'string') {
+      candidate = raw.description.trim();
+      if (raw.model) usedModel = String(raw.model);
+    }
+  }
+  const proposedDescription = buildConformingDescription(skill, candidate);
+  return {
+    schemaVersion: 1,
+    skill: skill.name,
+    status: 'proposal',
+    currentDescription: currentDescription || skill.summary || '',
+    proposedDescription,
+    provenance: { model: usedModel, promptHash, digest: sha256Tag(proposedDescription) },
+  };
+}

package/src/index/gate.mjs

@@ -9,14 +9,9 @@ import { readFile } from 'node:fs/promises';
 import { join } from 'node:path';
 
 import { isTestPath } from '../bench/mine.mjs';
+import { langOf } from './store.mjs';
 import { importsOf } from './tests.mjs';
 
-const LANG = {
-  js: 'javascript', mjs: 'javascript', cjs: 'javascript', jsx: 'javascript',
-  ts: 'typescript', tsx: 'typescript', kt: 'kotlin', kts: 'kotlin', java: 'java',
-  py: 'python', go: 'go', rs: 'rust', swift: 'swift', dart: 'dart', rb: 'ruby', cs: 'csharp',
-};
-
 /**
  * Test files that import one of `changed`, directly or through other files.
  * A changed file that is itself a test is included: the edit is the test.
@@ -92,10 +87,3 @@ export async function readEntries(root, files) {
 export async function readImportEdges(root, files) {
   return importEdges(await readEntries(root, files));
 }
-
-function langOf(file) {
-  const base = file.slice(file.lastIndexOf('/') + 1);
-  const dot = base.lastIndexOf('.');
-  if (dot <= 0) return '';
-  return LANG[base.slice(dot + 1)] ?? '';
-}

package/src/index/govern.mjs

@@ -6,22 +6,35 @@
 // and a lesson do the same with their own path list. Nothing here is stored: the
 // answer is the records on disk.
 
+import { parseDecisionFrontmatter, splitFrontmatter } from '../memory/decide.mjs';
+
 const SIBLING = /^\*{0,2}[A-Z][\w-]*(?:[ \t]+[\w-]+){0,3}:\*{0,2}(?:[ \t]|$)/;
 const HEADING = /^#{1,6}\s/;
 const TOOL = /^(?:simt-dev|simt)\b/;
 
-/** Backticked tokens of the Affects paragraph, in order. */
+/**
+ * Affected tokens of a decision record: reads `simt-affects:` frontmatter first (R34),
+ * falling back to backticked tokens of the `Affects:` paragraph (R25).
+ */
 export function affectsTokens(text) {
-  const lines = String(text).split('\n');
+  const fm = parseDecisionFrontmatter(text);
+  const { body } = splitFrontmatter(text);
+  const lines = String(body).split('\n');
   const start = lines.findIndex((line) => /^\s*\*{0,2}Affects:\*{0,2}/.test(line));
-  if (start < 0) return [];
-  const parts = [lines[start].replace(/^\s*\*{0,2}Affects:\*{0,2}\s*/, '')];
-  for (let i = start + 1; i < lines.length; i += 1) {
-    const line = lines[i];
-    if (line.trim() === '' || SIBLING.test(line) || HEADING.test(line)) break;
-    parts.push(line);
+  const proseTokens = [];
+  if (start >= 0) {
+    const parts = [lines[start].replace(/^\s*\*{0,2}Affects:\*{0,2}\s*/, '')];
+    for (let i = start + 1; i < lines.length; i += 1) {
+      const line = lines[i];
+      if (line.trim() === '' || SIBLING.test(line) || HEADING.test(line)) break;
+      parts.push(line);
+    }
+    proseTokens.push(...[...parts.join('\n').matchAll(/`([^`]+)`/g)].map((match) => match[1].trim()).filter(Boolean));
+  }
+  if (fm.hasAffects) {
+    return [...new Set([...fm.affects, ...proseTokens])];
   }
-  return [...parts.join('\n').matchAll(/`([^`]+)`/g)].map((match) => match[1].trim()).filter(Boolean);
+  return proseTokens;
 }
 
 /**
@@ -205,8 +218,7 @@ function extension(file) {
 }
 
 function frontmatter(text) {
-  const match = /^---\n([\s\S]*?)\n---/.exec(String(text));
-  return match ? match[1] : '';
+  return splitFrontmatter(text).front ?? '';
 }
 
 function quotedList(front, key) {