promptgraph-mcp 2.8.1 → 2.8.3

package/search.js

@@ -1,216 +1,216 @@
-import { embed, cosineSimilarity } from './embedder.js';
-import { getDb, blobToVec } from './db.js';
-import { annSearch } from './ann.js';
-import { TRUST_LEVEL_BOOST } from './marketplace.js';
-
-function getHybridWeights(query) {
-  const hasTechTerms = /[A-Z]|\d/.test(query);
-  if (hasTechTerms) return { embedWeight: 0.5, bm25Weight: 0.5 }
-  return { embedWeight: 0.7, bm25Weight: 0.3 }
-}
-
-function applyRatingBoost(db, id, score) {
-  const r = db.prepare('SELECT success, fail FROM ratings WHERE skill_id = ?').get(id);
-  let boost = 1.0;
-  if (r && (r.success + r.fail) > 3) {
-    boost *= (0.85 + 0.15 * (r.success / (r.success + r.fail)));
-  }
-
-  // Trust level boost (verified/official/trusted rank higher)
-  const re = db.prepare('SELECT trust_level, popularity FROM registry_entries WHERE id = ?').get(id);
-  if (re) {
-    boost *= (TRUST_LEVEL_BOOST[re.trust_level] || 1.0);
-    // Popularity bump: top 20% of skills get +5%, rest unchanged
-    if (re.popularity > 0) {
-      const top20 = db.prepare('SELECT popularity FROM registry_entries ORDER BY popularity DESC LIMIT 1').get();
-      if (top20 && re.popularity >= top20.popularity * 0.2) {
-        boost *= 1.05;
-      }
-    }
-  }
-
-  return score * boost;
-}
-
-function normalizeBM25(raw) {
-  return Math.max(0, 1 + raw / 10);
-}
-
-async function runEmbeddingSearch(db, queryVec, topK) {
-  // Try ANN index first (fast)
-  const annResults = await annSearch(queryVec, topK);
-  if (annResults && annResults.length > 0) {
-    const bestBySkill = new Map();
-    for (const r of annResults) {
-      const prev = bestBySkill.get(r.skill_id);
-      if (!prev || r.score > prev) bestBySkill.set(r.skill_id, r.score);
-    }
-    return [...bestBySkill.entries()]
-      .map(([id, score]) => ({ id, score: Math.max(0, score) }))
-      .sort((a, b) => b.score - a.score);
-  }
-
-  // Fallback: brute force cosine (used before first reindex)
-  const chunks = db.prepare('SELECT skill_id, embedding FROM chunks').all();
-  if (chunks.length > 0) {
-    const bestBySkill = new Map();
-    for (const chunk of chunks) {
-      const score = cosineSimilarity(queryVec, blobToVec(chunk.embedding));
-      const prev = bestBySkill.get(chunk.skill_id);
-      if (!prev || score > prev) bestBySkill.set(chunk.skill_id, score);
-    }
-    return [...bestBySkill.entries()]
-      .map(([id, score]) => ({ id, score: Math.max(0, score) }))
-      .sort((a, b) => b.score - a.score);
-  }
-
-  return [];
-}
-
-function runBM25Search(db, query, topK) {
-  try {
-    const terms = query.replace(/[^\w\s]/g, ' ').trim().split(/\s+/).filter(Boolean).join(' OR ');
-    const rows = db.prepare(
-      `SELECT s.id, bm25(skills_fts) AS score FROM skills_fts
-       JOIN skills s ON skills_fts.id = s.id
-       WHERE skills_fts MATCH ? ORDER BY score LIMIT ?`
-    ).all(terms, topK);
-    return rows.map(r => ({ id: r.id, score: normalizeBM25(r.score) }));
-  } catch {
-    return [];
-  }
-}
-
-export async function search(query, topK = 5) {
-  const db = getDb();
-  const { embedWeight, bm25Weight } = getHybridWeights(query);
-  const queryVec = await embed(query);
-
-  const embedResults = await runEmbeddingSearch(db, queryVec, topK * 4);
-  const bm25Results = runBM25Search(db, query, topK * 4);
-
-  // Fallback: pure BM25 if embeddings unavailable
-  if (!embedResults.length) {
-    return bm25Results.slice(0, topK)
-      .map(r => skillWithSnippet(db, r.id, applyRatingBoost(db, r.id, r.score)))
-      .filter(Boolean);
-  }
-
-  // Fallback: pure embedding if BM25 returns nothing
-  if (!bm25Results.length) {
-    return embedResults.slice(0, topK)
-      .map(({ id, score }) => skillWithSnippet(db, id, applyRatingBoost(db, id, score)))
-      .filter(Boolean);
-  }
-
-  // Hybrid: combine normalized scores from both signals
-  const combined = new Map();
-  for (const { id, score } of embedResults) {
-    combined.set(id, { embedScore: score, bm25Score: 0 });
-  }
-  for (const { id, score } of bm25Results) {
-    const entry = combined.get(id);
-    if (entry) {
-      entry.bm25Score = score;
-    } else {
-      combined.set(id, { embedScore: 0, bm25Score: score });
-    }
-  }
-
-  const ordered = [...combined.entries()]
-    .map(([id, s]) => ({
-      id,
-      score: embedWeight * s.embedScore + bm25Weight * s.bm25Score,
-    }))
-    .sort((a, b) => b.score - a.score)
-
-  const rerankerEnabled = process.env.PG_RERANKER !== '0'
-
-  if (rerankerEnabled) {
-    const { Reranker } = await import('./src/reranker/reranker.js')
-    const reranker = new Reranker()
-    const topN = ordered.slice(0, Math.max(50, topK * 6))
-      .map(({ id, score }) => {
-        const s = skillWithSnippet(db, id, score)
-        return s ? { id, text: s.snippet, score } : null
-      })
-      .filter(Boolean)
-    const reranked = await reranker.rerank(query, topN, topK)
-    return reranked.map(r => skillWithSnippet(db, r.id, applyRatingBoost(db, r.id, r.score))).filter(Boolean)
-  }
-
-  return ordered.slice(0, topK)
-    .map(({ id, score }) => skillWithSnippet(db, id, applyRatingBoost(db, id, score)))
-    .filter(Boolean);
-}
-
-function skillWithSnippet(db, id, score) {
-  const skill = db.prepare('SELECT id, name, description, path, source, content, version, author, license, updated_at, downloads, verified FROM skills WHERE id = ?').get(id);
-  if (!skill) return null;
-  const { content, ...rest } = skill;
-  const snippet = content?.replace(/^---[\s\S]*?---\n?/, '').trim().slice(0, 400) || '';
-  return { ...rest, score, snippet };
-}
-
-export function getContext(nameOrId) {
-  const db = getDb();
-  let skill = db.prepare('SELECT * FROM skills WHERE id = ?').get(nameOrId);
-  if (!skill) {
-    const res = resolveId(db, nameOrId);
-    if (res.error) return res;
-    skill = db.prepare('SELECT * FROM skills WHERE id = ?').get(res.id);
-  }
-  if (!skill) return null;
-  const callees = db.prepare('SELECT to_skill FROM edges WHERE from_skill = ?').all(skill.id).map(r => r.to_skill);
-  const callers = db.prepare('SELECT from_skill FROM edges WHERE to_skill = ?').all(skill.id).map(r => r.from_skill);
-
-  return { ...skill, callees, callers };
-}
-
-function resolveId(db, nameOrId) {
-  const byId = db.prepare('SELECT id FROM skills WHERE id = ?').get(nameOrId);
-  if (byId) return { id: byId.id };
-  const byName = db.prepare('SELECT id FROM skills WHERE name = ?').all(nameOrId);
-  if (byName.length === 1) return { id: byName[0].id };
-  if (byName.length > 1) {
-    const candidates = byName.map(r => r.id).join(', ');
-    return { error: `Ambiguous name "${nameOrId}" — multiple skills match. Use a full id: ${candidates}` };
-  }
-  return { error: `Skill not found: ${nameOrId}` };
-}
-
-export function getCallers(nameOrId) {
-  const db = getDb();
-  const res = resolveId(db, nameOrId);
-  if (res.error) return res;
-  return db.prepare('SELECT from_skill FROM edges WHERE to_skill = ?').all(res.id).map(r => r.from_skill);
-}
-
-export function getCallees(nameOrId) {
-  const db = getDb();
-  const res = resolveId(db, nameOrId);
-  if (res.error) return res;
-  return db.prepare('SELECT to_skill FROM edges WHERE from_skill = ?').all(res.id).map(r => r.to_skill);
-}
-
-export function getImpact(nameOrId) {
-  const db = getDb();
-  const res = resolveId(db, nameOrId);
-  if (res.error) return res;
-  const visited = new Set();
-  const queue = [res.id];
-  while (queue.length) {
-    const cur = queue.shift();
-    if (visited.has(cur)) continue;
-    visited.add(cur);
-    const callers = db.prepare('SELECT from_skill FROM edges WHERE to_skill = ?').all(cur).map(r => r.from_skill);
-    queue.push(...callers);
-  }
-  visited.delete(res.id);
-  return [...visited];
-}
-
-export function listAll() {
-  const db = getDb();
-  return db.prepare('SELECT id, name, description, source, version, author, license, updated_at, downloads, verified FROM skills ORDER BY source, name').all();
-}
+import { embed, cosineSimilarity } from './embedder.js';
+import { getDb, blobToVec } from './db.js';
+import { annSearch } from './ann.js';
+import { TRUST_LEVEL_BOOST } from './marketplace.js';
+
+function getHybridWeights(query) {
+  const hasTechTerms = /[A-Z]|\d/.test(query);
+  if (hasTechTerms) return { embedWeight: 0.5, bm25Weight: 0.5 }
+  return { embedWeight: 0.7, bm25Weight: 0.3 }
+}
+
+function applyRatingBoost(db, id, score) {
+  const r = db.prepare('SELECT success, fail FROM ratings WHERE skill_id = ?').get(id);
+  let boost = 1.0;
+  if (r && (r.success + r.fail) > 3) {
+    boost *= (0.85 + 0.15 * (r.success / (r.success + r.fail)));
+  }
+
+  // Trust level boost (verified/official/trusted rank higher)
+  const re = db.prepare('SELECT trust_level, popularity FROM registry_entries WHERE id = ?').get(id);
+  if (re) {
+    boost *= (TRUST_LEVEL_BOOST[re.trust_level] || 1.0);
+    // Popularity bump: top 20% of skills get +5%, rest unchanged
+    if (re.popularity > 0) {
+      const top20 = db.prepare('SELECT popularity FROM registry_entries ORDER BY popularity DESC LIMIT 1').get();
+      if (top20 && re.popularity >= top20.popularity * 0.2) {
+        boost *= 1.05;
+      }
+    }
+  }
+
+  return score * boost;
+}
+
+function normalizeBM25(raw) {
+  return Math.max(0, 1 + raw / 10);
+}
+
+async function runEmbeddingSearch(db, queryVec, topK) {
+  // Try ANN index first (fast)
+  const annResults = await annSearch(queryVec, topK);
+  if (annResults && annResults.length > 0) {
+    const bestBySkill = new Map();
+    for (const r of annResults) {
+      const prev = bestBySkill.get(r.skill_id);
+      if (!prev || r.score > prev) bestBySkill.set(r.skill_id, r.score);
+    }
+    return [...bestBySkill.entries()]
+      .map(([id, score]) => ({ id, score: Math.max(0, score) }))
+      .sort((a, b) => b.score - a.score);
+  }
+
+  // Fallback: brute force cosine (used before first reindex)
+  const chunks = db.prepare('SELECT skill_id, embedding FROM chunks').all();
+  if (chunks.length > 0) {
+    const bestBySkill = new Map();
+    for (const chunk of chunks) {
+      const score = cosineSimilarity(queryVec, blobToVec(chunk.embedding));
+      const prev = bestBySkill.get(chunk.skill_id);
+      if (!prev || score > prev) bestBySkill.set(chunk.skill_id, score);
+    }
+    return [...bestBySkill.entries()]
+      .map(([id, score]) => ({ id, score: Math.max(0, score) }))
+      .sort((a, b) => b.score - a.score);
+  }
+
+  return [];
+}
+
+function runBM25Search(db, query, topK) {
+  try {
+    const terms = query.replace(/[^\w\s]/g, ' ').trim().split(/\s+/).filter(Boolean).join(' OR ');
+    const rows = db.prepare(
+      `SELECT s.id, bm25(skills_fts) AS score FROM skills_fts
+       JOIN skills s ON skills_fts.id = s.id
+       WHERE skills_fts MATCH ? ORDER BY score LIMIT ?`
+    ).all(terms, topK);
+    return rows.map(r => ({ id: r.id, score: normalizeBM25(r.score) }));
+  } catch {
+    return [];
+  }
+}
+
+export async function search(query, topK = 5) {
+  const db = getDb();
+  const { embedWeight, bm25Weight } = getHybridWeights(query);
+  const queryVec = await embed(query);
+
+  const embedResults = await runEmbeddingSearch(db, queryVec, topK * 4);
+  const bm25Results = runBM25Search(db, query, topK * 4);
+
+  // Fallback: pure BM25 if embeddings unavailable
+  if (!embedResults.length) {
+    return bm25Results.slice(0, topK)
+      .map(r => skillWithSnippet(db, r.id, applyRatingBoost(db, r.id, r.score)))
+      .filter(Boolean);
+  }
+
+  // Fallback: pure embedding if BM25 returns nothing
+  if (!bm25Results.length) {
+    return embedResults.slice(0, topK)
+      .map(({ id, score }) => skillWithSnippet(db, id, applyRatingBoost(db, id, score)))
+      .filter(Boolean);
+  }
+
+  // Hybrid: combine normalized scores from both signals
+  const combined = new Map();
+  for (const { id, score } of embedResults) {
+    combined.set(id, { embedScore: score, bm25Score: 0 });
+  }
+  for (const { id, score } of bm25Results) {
+    const entry = combined.get(id);
+    if (entry) {
+      entry.bm25Score = score;
+    } else {
+      combined.set(id, { embedScore: 0, bm25Score: score });
+    }
+  }
+
+  const ordered = [...combined.entries()]
+    .map(([id, s]) => ({
+      id,
+      score: embedWeight * s.embedScore + bm25Weight * s.bm25Score,
+    }))
+    .sort((a, b) => b.score - a.score)
+
+  const rerankerEnabled = process.env.PG_RERANKER !== '0'
+
+  if (rerankerEnabled) {
+    const { Reranker } = await import('./src/reranker/reranker.js')
+    const reranker = new Reranker()
+    const topN = ordered.slice(0, Math.max(50, topK * 6))
+      .map(({ id, score }) => {
+        const s = skillWithSnippet(db, id, score)
+        return s ? { id, text: s.snippet, score } : null
+      })
+      .filter(Boolean)
+    const reranked = await reranker.rerank(query, topN, topK)
+    return reranked.map(r => skillWithSnippet(db, r.id, applyRatingBoost(db, r.id, r.score))).filter(Boolean)
+  }
+
+  return ordered.slice(0, topK)
+    .map(({ id, score }) => skillWithSnippet(db, id, applyRatingBoost(db, id, score)))
+    .filter(Boolean);
+}
+
+function skillWithSnippet(db, id, score) {
+  const skill = db.prepare('SELECT id, name, description, path, source, content, version, author, license, updated_at, downloads, verified FROM skills WHERE id = ?').get(id);
+  if (!skill) return null;
+  const { content, ...rest } = skill;
+  const snippet = content?.replace(/^---[\s\S]*?---\n?/, '').trim().slice(0, 400) || '';
+  return { ...rest, score, snippet };
+}
+
+export function getContext(nameOrId) {
+  const db = getDb();
+  let skill = db.prepare('SELECT * FROM skills WHERE id = ?').get(nameOrId);
+  if (!skill) {
+    const res = resolveId(db, nameOrId);
+    if (res.error) return res;
+    skill = db.prepare('SELECT * FROM skills WHERE id = ?').get(res.id);
+  }
+  if (!skill) return null;
+  const callees = db.prepare('SELECT to_skill FROM edges WHERE from_skill = ?').all(skill.id).map(r => r.to_skill);
+  const callers = db.prepare('SELECT from_skill FROM edges WHERE to_skill = ?').all(skill.id).map(r => r.from_skill);
+
+  return { ...skill, callees, callers };
+}
+
+function resolveId(db, nameOrId) {
+  const byId = db.prepare('SELECT id FROM skills WHERE id = ?').get(nameOrId);
+  if (byId) return { id: byId.id };
+  const byName = db.prepare('SELECT id FROM skills WHERE name = ?').all(nameOrId);
+  if (byName.length === 1) return { id: byName[0].id };
+  if (byName.length > 1) {
+    const candidates = byName.map(r => r.id).join(', ');
+    return { error: `Ambiguous name "${nameOrId}" — multiple skills match. Use a full id: ${candidates}` };
+  }
+  return { error: `Skill not found: ${nameOrId}` };
+}
+
+export function getCallers(nameOrId) {
+  const db = getDb();
+  const res = resolveId(db, nameOrId);
+  if (res.error) return res;
+  return db.prepare('SELECT from_skill FROM edges WHERE to_skill = ?').all(res.id).map(r => r.from_skill);
+}
+
+export function getCallees(nameOrId) {
+  const db = getDb();
+  const res = resolveId(db, nameOrId);
+  if (res.error) return res;
+  return db.prepare('SELECT to_skill FROM edges WHERE from_skill = ?').all(res.id).map(r => r.to_skill);
+}
+
+export function getImpact(nameOrId) {
+  const db = getDb();
+  const res = resolveId(db, nameOrId);
+  if (res.error) return res;
+  const visited = new Set();
+  const queue = [res.id];
+  while (queue.length) {
+    const cur = queue.shift();
+    if (visited.has(cur)) continue;
+    visited.add(cur);
+    const callers = db.prepare('SELECT from_skill FROM edges WHERE to_skill = ?').all(cur).map(r => r.from_skill);
+    queue.push(...callers);
+  }
+  visited.delete(res.id);
+  return [...visited];
+}
+
+export function listAll() {
+  const db = getDb();
+  return db.prepare('SELECT id, name, description, source, version, author, license, updated_at, downloads, verified FROM skills ORDER BY source, name').all();
+}

package/src/filter/classifier.js

@@ -1,88 +1,88 @@
-import { cosineSimilarity } from '../../embedder.js';
-
-const SKILL_THRESHOLD = 0.35;
-const UNSURE_THRESHOLD = 0.15;
-
-const DOC_FIRST_HEADERS = /^(overview|introduction|about|background|welcome|getting started|what is|why |table of contents|toc|foreword|preface|readme)/i;
-const INSTRUCTION_HEADERS = /^#{1,3}\s+(steps?|usage|instructions?|how\s+to|when\s+to\s+use|workflow|process|procedure|example|examples?|commands?|output|result)/i;
-const IMPERATIVE_HEADERS = /\b(run|use|apply|execute|check|debug|fix|create|add|remove|deploy|test|write|generate|analyze|review|refactor|optimize|configure|setup|install|scan|audit|validate|search|find|extract|parse)\b/i;
-
-export function getFeatureVector(raw) {
-  const lines = raw.split('\n');
-  const nonEmpty = lines.filter(l => l.trim());
-  const headers = nonEmpty.filter(l => /^#{1,3}\s/.test(l));
-  const paragraphs = raw.split(/\n\n+/).filter(p => p.trim() && !p.trim().startsWith('#'));
-  const words = raw.toLowerCase().split(/\s+/).filter(w => w.length > 3);
-
-  const firstHeader = headers[0]?.replace(/^#+\s*/, '') || '';
-  const longProse = paragraphs.filter(p => p.split(' ').length > 60 && !/```/.test(p));
-  const wordRatio = words.length > 80 ? new Set(words).size / words.length : 1;
-
-  return [
-    Number(raw.length >= 150),
-    Number(headers.length >= 1),
-    Number(lines.some(l => INSTRUCTION_HEADERS.test(l))),
-    Number(headers.some(h => IMPERATIVE_HEADERS.test(h))),
-    Number(raw.includes('```') || raw.includes('    ')),
-    Number(nonEmpty.some(l => /^\d+\.\s/.test(l))),
-    Number(nonEmpty.some(l => /^[-*+]\s/.test(l))),
-    Number(headers.length >= 2),
-    Number(headers.length >= 4),
-    Number(DOC_FIRST_HEADERS.test(firstHeader)) ? 1 : 0,
-    Number(longProse.length > paragraphs.length * 0.6 && paragraphs.length > 3) ? 1 : 0,
-    Number(wordRatio >= 0.22) ? 0 : 1,
-    Number(raw.length < 150) ? 1 : 0,
-    Number(headers.length < 1) ? 1 : 0,
-  ];
-}
-
-/**
- * Rule A override: if centroid says reject/unsure but file clearly looks like
- * a skill by content (code + instructions + list + verbs, no SKILL in name/path),
- * override to skill. Catches adversarial wrong-heading cases with 0 FP cost.
- */
-function ruleA(rawVec, raw, filePath) {
-  const fv = getFeatureVector(raw);
-  const hasCode = fv[4];
-  const hasInstructions = fv[2];
-  const hasNumberedList = fv[5];
-  const hasVerb = fv[3];
-  const filename = filePath ? filePath.split(/[/\\]/).pop().replace(/\.md$/i, '').toUpperCase() : '';
-  const pathUpper = filePath ? filePath.toUpperCase() : '';
-  const hasSkillName = filename.includes('SKILL');
-  const hasSkillPath = /SKILL/.test(pathUpper);
-
-  if (hasCode && hasInstructions && hasNumberedList && hasVerb && !hasSkillName && !hasSkillPath) {
-    return { label: 'skill', score: 0.6, method: 'ruleA' };
-  }
-  return null;
-}
-
-export function classify(rawVec, centroids, raw, filePath) {
-  if (!centroids) {
-    return { label: 'skill', score: 1, method: 'fallback' };
-  }
-
-  const goodSim = centroids.good ? cosineSimilarity(rawVec, centroids.good) : 0;
-  const badSim = centroids.bad ? cosineSimilarity(rawVec, centroids.bad) : 0;
-
-  const rawScore = goodSim - badSim;
-  const score = (rawScore + 1) / 2;
-
-  if (score >= SKILL_THRESHOLD) {
-    return { label: 'skill', score, goodSim, badSim, method: 'centroid' };
-  }
-
-  // Below threshold — try Rule A content override
-  if (raw && filePath) {
-    const override = ruleA(rawVec, raw, filePath);
-    if (override) return override;
-  }
-
-  if (score >= UNSURE_THRESHOLD) {
-    return { label: 'unsure', score, goodSim, badSim, method: 'centroid' };
-  }
-  return { label: 'reject', score, goodSim, badSim, method: 'centroid' };
-}
-
-export { SKILL_THRESHOLD, UNSURE_THRESHOLD };
+import { cosineSimilarity } from '../../embedder.js';
+
+const SKILL_THRESHOLD = 0.35;
+const UNSURE_THRESHOLD = 0.15;
+
+const DOC_FIRST_HEADERS = /^(overview|introduction|about|background|welcome|getting started|what is|why |table of contents|toc|foreword|preface|readme)/i;
+const INSTRUCTION_HEADERS = /^#{1,3}\s+(steps?|usage|instructions?|how\s+to|when\s+to\s+use|workflow|process|procedure|example|examples?|commands?|output|result)/i;
+const IMPERATIVE_HEADERS = /\b(run|use|apply|execute|check|debug|fix|create|add|remove|deploy|test|write|generate|analyze|review|refactor|optimize|configure|setup|install|scan|audit|validate|search|find|extract|parse)\b/i;
+
+export function getFeatureVector(raw) {
+  const lines = raw.split('\n');
+  const nonEmpty = lines.filter(l => l.trim());
+  const headers = nonEmpty.filter(l => /^#{1,3}\s/.test(l));
+  const paragraphs = raw.split(/\n\n+/).filter(p => p.trim() && !p.trim().startsWith('#'));
+  const words = raw.toLowerCase().split(/\s+/).filter(w => w.length > 3);
+
+  const firstHeader = headers[0]?.replace(/^#+\s*/, '') || '';
+  const longProse = paragraphs.filter(p => p.split(' ').length > 60 && !/```/.test(p));
+  const wordRatio = words.length > 80 ? new Set(words).size / words.length : 1;
+
+  return [
+    Number(raw.length >= 150),
+    Number(headers.length >= 1),
+    Number(lines.some(l => INSTRUCTION_HEADERS.test(l))),
+    Number(headers.some(h => IMPERATIVE_HEADERS.test(h))),
+    Number(raw.includes('```') || raw.includes('    ')),
+    Number(nonEmpty.some(l => /^\d+\.\s/.test(l))),
+    Number(nonEmpty.some(l => /^[-*+]\s/.test(l))),
+    Number(headers.length >= 2),
+    Number(headers.length >= 4),
+    Number(DOC_FIRST_HEADERS.test(firstHeader)) ? 1 : 0,
+    Number(longProse.length > paragraphs.length * 0.6 && paragraphs.length > 3) ? 1 : 0,
+    Number(wordRatio >= 0.22) ? 0 : 1,
+    Number(raw.length < 150) ? 1 : 0,
+    Number(headers.length < 1) ? 1 : 0,
+  ];
+}
+
+/**
+ * Rule A override: if centroid says reject/unsure but file clearly looks like
+ * a skill by content (code + instructions + list + verbs, no SKILL in name/path),
+ * override to skill. Catches adversarial wrong-heading cases with 0 FP cost.
+ */
+function ruleA(rawVec, raw, filePath) {
+  const fv = getFeatureVector(raw);
+  const hasCode = fv[4];
+  const hasInstructions = fv[2];
+  const hasNumberedList = fv[5];
+  const hasVerb = fv[3];
+  const filename = filePath ? filePath.split(/[/\\]/).pop().replace(/\.md$/i, '').toUpperCase() : '';
+  const pathUpper = filePath ? filePath.toUpperCase() : '';
+  const hasSkillName = filename.includes('SKILL');
+  const hasSkillPath = /SKILL/.test(pathUpper);
+
+  if (hasCode && hasInstructions && hasNumberedList && hasVerb && !hasSkillName && !hasSkillPath) {
+    return { label: 'skill', score: 0.6, method: 'ruleA' };
+  }
+  return null;
+}
+
+export function classify(rawVec, centroids, raw, filePath) {
+  if (!centroids) {
+    return { label: 'skill', score: 1, method: 'fallback' };
+  }
+
+  const goodSim = centroids.good ? cosineSimilarity(rawVec, centroids.good) : 0;
+  const badSim = centroids.bad ? cosineSimilarity(rawVec, centroids.bad) : 0;
+
+  const rawScore = goodSim - badSim;
+  const score = (rawScore + 1) / 2;
+
+  if (score >= SKILL_THRESHOLD) {
+    return { label: 'skill', score, goodSim, badSim, method: 'centroid' };
+  }
+
+  // Below threshold — try Rule A content override
+  if (raw && filePath) {
+    const override = ruleA(rawVec, raw, filePath);
+    if (override) return override;
+  }
+
+  if (score >= UNSURE_THRESHOLD) {
+    return { label: 'unsure', score, goodSim, badSim, method: 'centroid' };
+  }
+  return { label: 'reject', score, goodSim, badSim, method: 'centroid' };
+}
+
+export { SKILL_THRESHOLD, UNSURE_THRESHOLD };