@claude-flow/cli 3.10.16 → 3.10.18

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@claude-flow/cli",
-  "version": "3.10.16",
+  "version": "3.10.18",
   "type": "module",
   "description": "Ruflo CLI - Enterprise AI agent orchestration with 60+ specialized agents, swarm coordination, MCP server, self-learning hooks, and vector memory for Claude Code",
   "main": "dist/src/index.js",

package/scripts/benchmark-pretrained-retrieval.mjs

@@ -0,0 +1,193 @@
+#!/usr/bin/env node
+// benchmark-pretrained-retrieval.mjs — proof that pretrained patterns are
+// retrievable, not just stored.
+//
+// Runs sample queries against the neural store (post-pretrain) and reports
+// the top-k matches. Demonstrates that after `pretrain-from-github.mjs`
+// runs, an agent can recall relevant past work by intent.
+//
+// Usage:
+//   1. node scripts/pretrain-from-github.mjs           # populate the store
+//   2. node scripts/benchmark-pretrained-retrieval.mjs # query + report
+
+import { writeFileSync, mkdirSync } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+import { dirname, join, resolve } from 'node:path';
+import { performance } from 'node:perf_hooks';
+
+const SCRIPT_DIR = dirname(fileURLToPath(import.meta.url));
+const CLI_ROOT = resolve(SCRIPT_DIR, '..');
+const REPO_ROOT = resolve(SCRIPT_DIR, '../../../..');
+const RUNS_DIR = join(REPO_ROOT, 'docs', 'benchmarks', 'runs');
+
+// Real ruflo-history-shaped queries. Each one targets a concept that should
+// have been seen during pretrain. `expect` is a regex matched against the
+// result's `name` field (commit subject or issue title) to score relevance.
+// A result is "relevant" if its name matches the expect regex (case-insensitive).
+const QUERIES = [
+  { q: 'how was the Opus model alias fixed', expect: /opus|2232|model.*alias|4\.8/i },
+  { q: 'self-learning wiring task-completed pretrain', expect: /self.?learning|task.?completed|pretrain|2245|074/i },
+  { q: 'deterministic codemod engine var-to-const', expect: /codemod|var.?to.?const|143|deterministic/i },
+  { q: 'MCP server orphan leak parent-death', expect: /mcp.*orphan|orphan.*mcp|parent.?death|leak/i },
+  { q: 'unified learning stats aggregator', expect: /unified|stats|aggregator|075/i },
+  { q: 'structured distillation 4-field schema', expect: /distillation|structured|076|4.?field/i },
+  { q: 'SQL injection migrate.ts table identifier', expect: /sql.?injection|migrate|table|identifier/i },
+  { q: 'recall@k HNSW benchmark harness', expect: /recall|hnsw|benchmark|harness/i },
+  { q: 'Q-learning encoder keyword block', expect: /q.?learning|encoder|keyword|2239/i },
+  { q: 'security hardening crypto random IDs', expect: /security|hardening|crypto|random/i },
+];
+
+const TOP_K = 5;
+
+async function main() {
+  const intel = await import(join(CLI_ROOT, 'dist/src/memory/intelligence.js'));
+  const neural = await import(join(CLI_ROOT, 'dist/src/mcp-tools/neural-tools.js'));
+
+  // §1 — snapshot the neural store + globalStats so we know what's there.
+  const unified = await intel.getUnifiedLearningStats();
+  const total = unified.neuralPatterns.patternCount;
+
+  if (total === 0) {
+    console.error('No patterns in neural store. Run scripts/pretrain-from-github.mjs first.');
+    process.exit(2);
+  }
+
+  // §2 — A/B hybrid vs cosine-only (HYBRID=0 forces pre-3.10.18 behaviour).
+  // Default runs hybrid (cosine + BM25 + MMR per ADR-078).
+  const listTool = neural.neuralTools.find((t) => t.name === 'neural_patterns');
+  const mode = process.env.HYBRID === '0' ? 'cosine' : 'hybrid';
+
+  // top-1-uniqueness — fraction of queries whose top-1 result is NOT the
+  // same pattern ID as another query's top-1. Catches the "everyone gets
+  // the same generic top-1" failure mode.
+  const top1Ids = new Map();
+
+  const tQuery0 = performance.now();
+  const results = [];
+  for (const { q, expect } of QUERIES) {
+    const r = await listTool.handler({ action: 'search', query: q, mode, limit: TOP_K });
+    const matches = (r.patterns || r.results || r.matches || []).slice(0, TOP_K);
+    if (matches.length > 0) {
+      const top1 = matches[0].id;
+      top1Ids.set(top1, (top1Ids.get(top1) ?? 0) + 1);
+    }
+    // ADR-078 relevance: does the top-1/top-3 name match the expect regex?
+    const top1Name = matches[0]?.name ?? '';
+    const top1Relevant = expect.test(top1Name);
+    let top3Relevant = false;
+    let firstRelevantRank = -1;
+    for (let i = 0; i < Math.min(matches.length, 3); i++) {
+      if (expect.test(matches[i]?.name ?? '')) {
+        top3Relevant = true;
+        if (firstRelevantRank === -1) firstRelevantRank = i + 1;
+        break;
+      }
+    }
+    results.push({
+      query: q,
+      matched: matches.length > 0,
+      top1Relevant,
+      top3Relevant,
+      firstRelevantRank,
+      topK: matches.map((m) => ({
+        id: m.id,
+        name: m.name?.slice(0, 100),
+        type: m.type,
+        score: m.score ?? m.similarity,
+        cosineScore: m.cosineScore,
+        bm25Score: m.bm25Score,
+        mmrScore: m.mmrScore,
+      })),
+    });
+  }
+  const queryMs = performance.now() - tQuery0;
+
+  const matchedQueries = results.filter((r) => r.matched).length;
+  const top1Hits = results.filter((r) => r.top1Relevant).length;
+  const top3Hits = results.filter((r) => r.top3Relevant).length;
+  const ranks = results.filter((r) => r.firstRelevantRank > 0).map((r) => r.firstRelevantRank);
+  // MRR over the top-3 window (rank-of-first-relevant). Queries with no
+  // relevant result in top-3 contribute 0 to the mean.
+  const mrr3 = QUERIES.length > 0
+    ? Number((ranks.reduce((s, r) => s + 1 / r, 0) / QUERIES.length).toFixed(4))
+    : 0;
+  // top-1 collision: number of distinct top-1 IDs over the query count.
+  const uniqueTop1 = top1Ids.size;
+  const top1Diversity = Number((uniqueTop1 / QUERIES.length).toFixed(4));
+
+  // top-3 redundancy: average fraction of top-3 results that are duplicates
+  // of the same pattern within a single query (should be 0 — but if the same
+  // ID appears twice in top-3 we surface it).
+  let dupCount = 0, top3Slots = 0;
+  for (const r of results) {
+    const ids = r.topK.slice(0, 3).map((m) => m.id);
+    top3Slots += ids.length;
+    const seen = new Set();
+    for (const id of ids) {
+      if (seen.has(id)) dupCount++;
+      seen.add(id);
+    }
+  }
+  const top3DupRate = top3Slots > 0 ? Number((dupCount / top3Slots).toFixed(4)) : 0;
+
+  const summary = {
+    runAt: new Date().toISOString(),
+    benchmark: 'pretrained-retrieval',
+    mode,                              // ADR-078: which retrieval path was used
+    storeSize: total,
+    queries: QUERIES.length,
+    matchedQueries,
+    matchRate: Number((matchedQueries / QUERIES.length).toFixed(4)),
+    top1HitRate: Number((top1Hits / QUERIES.length).toFixed(4)),  // ADR-078: relevance, not just any-match
+    top3HitRate: Number((top3Hits / QUERIES.length).toFixed(4)),
+    mrr3,                              // mean reciprocal rank over top-3
+    top1Diversity,                     // 1.0 = every query gets a distinct top-1
+    top3DupRate,                       // 0.0 = no duplicate IDs inside any top-3
+    avgQueryLatencyMs: Number((queryMs / QUERIES.length).toFixed(2)),
+    totalQueryMs: Number(queryMs.toFixed(2)),
+    results,
+    passed: matchedQueries === QUERIES.length,
+  };
+
+  if (process.env.BENCH_JSON) {
+    console.log(JSON.stringify(summary, null, 2));
+  } else {
+    console.log(`# Pretrained-retrieval benchmark — proof of learning`);
+    console.log(`Mode: ${mode}${mode === 'hybrid' ? ' (cosine + BM25 + MMR, ADR-078)' : ' (cosine-only, pre-3.10.18)'}`);
+    console.log(`Store size: ${total} patterns`);
+    console.log(`Queries: ${QUERIES.length}`);
+    console.log(`Match rate: ${(summary.matchRate * 100).toFixed(0)}% (${matchedQueries}/${QUERIES.length})`);
+    console.log(`Top-1 hit rate (RELEVANCE): ${(summary.top1HitRate * 100).toFixed(0)}% (${top1Hits}/${QUERIES.length})`);
+    console.log(`Top-3 hit rate (RELEVANCE): ${(summary.top3HitRate * 100).toFixed(0)}% (${top3Hits}/${QUERIES.length})`);
+    console.log(`MRR@3: ${summary.mrr3}`);
+    console.log(`Top-1 diversity: ${(summary.top1Diversity * 100).toFixed(0)}% (${uniqueTop1} distinct top-1 IDs across ${QUERIES.length} queries)`);
+    console.log(`Top-3 dup rate: ${(summary.top3DupRate * 100).toFixed(0)}%`);
+    console.log(`Avg query latency: ${summary.avgQueryLatencyMs} ms`);
+    console.log('');
+    for (const r of results) {
+      console.log(`Q: "${r.query}"`);
+      if (r.topK.length === 0) {
+        console.log(`   → no matches`);
+      } else {
+        for (const m of r.topK.slice(0, 3)) {
+          console.log(`   → ${m.score?.toFixed?.(3) ?? '—'}  ${m.name}`);
+        }
+      }
+    }
+    console.log('');
+    console.log(`Overall: ${summary.passed ? '✅ PASSED' : '⚠️  partial'}`);
+  }
+
+  if (!process.env.BENCH_NO_WRITE) {
+    mkdirSync(RUNS_DIR, { recursive: true });
+    const stamp = summary.runAt.replace(/[:.]/g, '-');
+    writeFileSync(join(RUNS_DIR, `pretrained-retrieval-${stamp}.json`), JSON.stringify(summary, null, 2));
+    writeFileSync(join(RUNS_DIR, 'pretrained-retrieval-latest.json'), JSON.stringify(summary, null, 2));
+    if (!process.env.BENCH_JSON) console.log(`\nWrote ${join(RUNS_DIR, `pretrained-retrieval-${stamp}.json`)}`);
+  }
+
+  // ONNX runtime keeps a worker thread alive — force exit.
+  process.exit(summary.passed ? 0 : 1);
+}
+
+main().catch((err) => { console.error(err); process.exit(1); });

package/scripts/pretrain-from-github.mjs

@@ -0,0 +1,352 @@
+#!/usr/bin/env node
+// pretrain-from-github.mjs — pretrain ruflo's self-learning system from its
+// own GitHub history (commits + issues). Each commit/issue becomes one
+// trajectory through the SONA + EWC++ pipeline; Structured Distillation
+// (ADR-076) compresses each into the 4-field schema before embedding.
+//
+// What this proves:
+//   - globalStats.{trajectoriesRecorded, patternsLearned, signalsProcessed}
+//     all move from a measured before to a measured after.
+//   - neural_patterns.patternCount grows.
+//   - memory-bridge entries grow.
+//   - The unified-stats aggregator's consistency block stays clean.
+//   - Every item gets a learningPath of 'trajectory-pipeline' (not 'recorded-only').
+//
+// Usage:
+//   node scripts/pretrain-from-github.mjs                   # 50 commits + 30 issues
+//   COMMITS=200 ISSUES=100 node scripts/pretrain-from-github.mjs
+//   SOURCE=git node scripts/pretrain-from-github.mjs        # git only, skip gh
+//   BENCH_JSON=1 node scripts/pretrain-from-github.mjs      # machine-readable
+//   BENCH_NO_WRITE=1 node scripts/pretrain-from-github.mjs  # don't write a run JSON
+//
+// Repro from a fresh checkout:
+//   git clone https://github.com/ruvnet/ruflo && cd ruflo
+//   npm install && ( cd v3/@claude-flow/cli && npx tsc -b )
+//   node v3/@claude-flow/cli/scripts/pretrain-from-github.mjs
+
+import { execSync } from 'node:child_process';
+import { writeFileSync, mkdirSync } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+import { dirname, join, resolve } from 'node:path';
+import { performance } from 'node:perf_hooks';
+
+const SCRIPT_DIR = dirname(fileURLToPath(import.meta.url));
+const CLI_ROOT = resolve(SCRIPT_DIR, '..');
+const REPO_ROOT = resolve(SCRIPT_DIR, '../../../..');
+const RUNS_DIR = join(REPO_ROOT, 'docs', 'benchmarks', 'runs');
+
+const COMMITS = Number(process.env.COMMITS) || 50;
+const ISSUES  = Number(process.env.ISSUES)  || 30;
+const SOURCE  = process.env.SOURCE || 'all'; // 'all' | 'git' | 'issues'
+
+// ---------------------------------------------------------------------------
+// Harvesters
+// ---------------------------------------------------------------------------
+
+// ADR-078 outcome signal — classify each commit by whether it was reverted,
+// hotfix-followed, or stuck. Operates on the harvested-commits window plus a
+// wider lookahead window for revert/hotfix detection.
+//
+// Verdicts emitted:
+//   success — landed cleanly, no later commit reverted or fixed it
+//   reverted — a later commit subject starts with `Revert "<this subject>"`
+//   hotfixed — a later commit (within HOTFIX_WINDOW_COMMITS) shares >=50%
+//              of the same touched files AND subject contains fix|hotfix|patch
+//
+// The "later" direction is git-log order (newest first → we look at older
+// indices in the lookahead, which are NEWER commits chronologically).
+const HOTFIX_WINDOW_COMMITS = Number(process.env.HOTFIX_WINDOW_COMMITS) || 20;
+const HOTFIX_FILE_OVERLAP = Number(process.env.HOTFIX_FILE_OVERLAP) || 0.5;
+const HOTFIX_KEYWORDS = /\b(fix|hotfix|patch|revert|bugfix|fixup)\b/i;
+
+function harvestCommits(n) {
+  if (SOURCE === 'issues') return [];
+  const fmt = '%H%x00%s%x00%b%x01';
+  // Pull a wider window so we have lookahead for revert/hotfix detection
+  // without changing the trained set size n.
+  const lookahead = HOTFIX_WINDOW_COMMITS;
+  const raw = execSync(
+    `git log --pretty=format:'${fmt}' -n ${n + lookahead} 2>/dev/null`,
+    { cwd: REPO_ROOT, encoding: 'utf-8', maxBuffer: 64 * 1024 * 1024 },
+  );
+
+  // Parse all blocks (window + lookahead).
+  const all = [];
+  for (const block of raw.split('\x01')) {
+    if (!block.trim()) continue;
+    const [sha, subject, body] = block.split('\x00');
+    if (!sha || !subject) continue;
+    all.push({ sha, subject: subject.trim(), body: (body || '').trim() });
+  }
+
+  // For each commit in the trained slice, get its touched files.
+  const filesCache = new Map();
+  const filesOf = (sha) => {
+    if (filesCache.has(sha)) return filesCache.get(sha);
+    try {
+      const out = execSync(
+        `git show --pretty=format: --name-only ${sha} 2>/dev/null`,
+        { cwd: REPO_ROOT, encoding: 'utf-8', maxBuffer: 8 * 1024 * 1024 },
+      ).trim().split('\n').filter(Boolean);
+      filesCache.set(sha, new Set(out));
+      return filesCache.get(sha);
+    } catch {
+      filesCache.set(sha, new Set());
+      return filesCache.get(sha);
+    }
+  };
+
+  // Build a revert-target map: any commit whose subject starts with
+  // `Revert "<X>"` flags X as reverted.
+  const revertedSubjects = new Set();
+  for (const c of all) {
+    const m = c.subject.match(/^Revert\s+"(.+?)"/);
+    if (m) revertedSubjects.add(m[1].trim());
+  }
+
+  const trained = all.slice(0, n);
+  const entries = [];
+  for (let i = 0; i < trained.length; i++) {
+    const c = trained[i];
+    const myFiles = filesOf(c.sha);
+
+    let verdict = 'success';
+    let outcomeNote = null;
+
+    // Reverted? Check if any later (lower-index = newer) commit reverted us.
+    if (revertedSubjects.has(c.subject)) {
+      verdict = 'reverted';
+      outcomeNote = 'subject reverted by a later commit';
+    } else {
+      // Hotfixed? A later commit within HOTFIX_WINDOW_COMMITS shares files
+      // AND has fix/hotfix in its subject.
+      const start = Math.max(0, i - HOTFIX_WINDOW_COMMITS);
+      for (let j = start; j < i; j++) {
+        const later = trained[j];
+        if (!HOTFIX_KEYWORDS.test(later.subject)) continue;
+        const laterFiles = filesOf(later.sha);
+        if (laterFiles.size === 0 || myFiles.size === 0) continue;
+        let overlap = 0;
+        for (const f of laterFiles) if (myFiles.has(f)) overlap++;
+        // Use min() so a small targeted fix on a big change still triggers
+        // (semantic: "≥half of the smaller commit's files overlap").
+        const overlapFrac = overlap / Math.min(laterFiles.size, myFiles.size);
+        if (overlapFrac >= HOTFIX_FILE_OVERLAP) {
+          verdict = 'hotfixed';
+          outcomeNote = `${(overlapFrac * 100).toFixed(0)}% file overlap with later fix ${later.sha.slice(0, 8)}`;
+          break;
+        }
+      }
+    }
+
+    entries.push({
+      source: 'commit',
+      id: `commit-${c.sha.slice(0, 12)}`,
+      subject: c.subject,
+      body: c.body,
+      verdict,
+      outcomeNote,
+      content: `${c.subject}\n\n${c.body}`.slice(0, 8192),
+    });
+  }
+  return entries;
+}
+
+function harvestIssues(n) {
+  if (SOURCE === 'git') return [];
+  try {
+    const raw = execSync(
+      `gh issue list --repo ruvnet/ruflo --state all --limit ${n} --json number,title,body,state,closedAt 2>/dev/null`,
+      { cwd: REPO_ROOT, encoding: 'utf-8', maxBuffer: 64 * 1024 * 1024 },
+    );
+    const items = JSON.parse(raw);
+    return items.map((i) => ({
+      source: 'issue',
+      id: `issue-${i.number}`,
+      subject: i.title,
+      body: (i.body || '').slice(0, 8192),
+      // closed = success outcome; open = partial (in-progress).
+      verdict: i.state === 'CLOSED' ? 'success' : 'partial',
+      content: `${i.title}\n\n${(i.body || '').slice(0, 8192)}`,
+    }));
+  } catch (err) {
+    console.error(`gh issue harvest skipped (${err.message?.slice(0, 60)}). Set SOURCE=git to silence.`);
+    return [];
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Main
+// ---------------------------------------------------------------------------
+
+async function main() {
+  const intel = await import(join(CLI_ROOT, 'dist/src/memory/intelligence.js'));
+  const neural = await import(join(CLI_ROOT, 'dist/src/mcp-tools/neural-tools.js'));
+  const { distillAndSerialise } = await import(join(CLI_ROOT, 'dist/src/memory/structured-distill.js'));
+
+  // §1 — record the baseline (no clear; we want to learn ON TOP of whatever
+  // history the user already has).
+  const unified0 = await intel.getUnifiedLearningStats();
+  const before = {
+    trajectoriesRecorded: unified0.global.trajectoriesRecorded,
+    patternsLearned: unified0.global.patternsLearned,
+    signalsProcessed: unified0.global.signalsProcessed,
+    neuralPatternCount: unified0.neuralPatterns.patternCount,
+    memoryBridgeTotal: unified0.memoryBridge.totalEntries,
+  };
+
+  // §2 — harvest
+  const tHarvest0 = performance.now();
+  const commits = harvestCommits(COMMITS);
+  const issues = harvestIssues(ISSUES);
+  const items = [...commits, ...issues];
+  const harvestMs = performance.now() - tHarvest0;
+
+  if (!process.env.BENCH_JSON) {
+    console.log(`# Pretrain from ruflo GitHub history`);
+    console.log(`Harvested: ${commits.length} commits + ${issues.length} issues = ${items.length} trajectories (${harvestMs.toFixed(0)} ms)`);
+  }
+
+  // §3 — feed each item through the trajectory pipeline. The harvester
+  // emits one of: success | partial | reverted | hotfixed. We map this to
+  // the trajectory pipeline's binary verdict {success, partial} since the
+  // pipeline doesn't accept arbitrary strings — but we preserve the original
+  // outcome in metadata + the summary's verdictMix so the signal isn't lost.
+  const verdictToPipeline = (v) => {
+    if (v === 'success') return 'success';
+    if (v === 'partial') return 'partial';
+    if (v === 'reverted') return 'partial';   // a revert is the strongest "this was wrong" signal we have
+    if (v === 'hotfixed') return 'partial';   // a same-files fix-followup is "needs adjustment"
+    return 'success';
+  };
+  const verdictMix = { success: 0, partial: 0, reverted: 0, hotfixed: 0 };
+  const tFeed0 = performance.now();
+  let trained = 0;
+  let failed = 0;
+  const failures = [];
+  for (const item of items) {
+    try {
+      const distilled = distillAndSerialise(item.content);
+      verdictMix[item.verdict] = (verdictMix[item.verdict] ?? 0) + 1;
+      await intel.recordTrajectory(
+        [{
+          type: 'result',
+          content: distilled,
+          metadata: {
+            source: item.source,
+            id: item.id,
+            subject: item.subject.slice(0, 200),
+            outcomeVerdict: item.verdict,
+            outcomeNote: item.outcomeNote ?? null,
+          },
+          timestamp: Date.now(),
+        }],
+        verdictToPipeline(item.verdict),
+      );
+      trained++;
+    } catch (err) {
+      failed++;
+      if (failures.length < 5) failures.push({ id: item.id, error: String(err.message).slice(0, 120) });
+    }
+  }
+  const feedMs = performance.now() - tFeed0;
+  intel.flushIntelligenceStats();
+
+  // §4 — also seed the neural store directly from the same items so
+  // `neural_patterns list` reflects them (closes the "globalStats moved but
+  // neural_patterns didn't" consistency note from ADR-075).
+  const tSeed0 = performance.now();
+  const neuralItems = items.map((item) => ({
+    name: item.subject.slice(0, 200),
+    type: item.source === 'commit' ? 'history-commit' : 'history-issue',
+    content: distillAndSerialise(item.content),
+    metadata: { source: item.source, id: item.id, verdict: item.verdict },
+  }));
+  const seedResult = await neural.storeNeuralPatterns(neuralItems);
+  const seedMs = performance.now() - tSeed0;
+
+  // §5 — read the after-counters via the unified aggregator (this is what
+  // hooks_intelligence_unified-stats would return for a live caller).
+  const unified1 = await intel.getUnifiedLearningStats();
+  const after = {
+    trajectoriesRecorded: unified1.global.trajectoriesRecorded,
+    patternsLearned: unified1.global.patternsLearned,
+    signalsProcessed: unified1.global.signalsProcessed,
+    neuralPatternCount: unified1.neuralPatterns.patternCount,
+    memoryBridgeTotal: unified1.memoryBridge.totalEntries,
+  };
+
+  const deltas = Object.fromEntries(
+    Object.keys(after).map((k) => [k, after[k] - before[k]]),
+  );
+
+  const summary = {
+    runAt: new Date().toISOString(),
+    benchmark: 'pretrain-from-github',
+    source: SOURCE,
+    config: { COMMITS, ISSUES },
+    harvest: {
+      commits: commits.length,
+      issues: issues.length,
+      total: items.length,
+      harvestMs: Number(harvestMs.toFixed(2)),
+    },
+    feed: {
+      trained,
+      failed,
+      avgLatencyMs: items.length > 0 ? Number((feedMs / items.length).toFixed(2)) : 0,
+      totalMs: Number(feedMs.toFixed(2)),
+      sampleFailures: failures,
+      verdictMix,   // ADR-078 outcome signal — counts per harvested verdict
+    },
+    seedNeuralStore: {
+      stored: seedResult.stored,
+      total: seedResult.total,
+      seedMs: Number(seedMs.toFixed(2)),
+    },
+    before,
+    after,
+    deltas,
+    consistency: unified1.consistency,
+    passed:
+      trained === items.length &&
+      deltas.trajectoriesRecorded >= items.length &&
+      deltas.neuralPatternCount >= items.length,
+  };
+
+  if (process.env.BENCH_JSON) {
+    console.log(JSON.stringify(summary, null, 2));
+  } else {
+    console.log('');
+    console.log('| Counter | Before | After | Δ |');
+    console.log('|---|---:|---:|---:|');
+    for (const k of Object.keys(after)) {
+      console.log(`| ${k} | ${before[k]} | ${after[k]} | +${deltas[k]} |`);
+    }
+    console.log('');
+    console.log(`Trained via trajectory pipeline: ${trained}/${items.length}`);
+    console.log(`Failed: ${failed}`);
+    console.log(`Avg latency per trajectory: ${summary.feed.avgLatencyMs} ms`);
+    console.log(`Neural store seeded: ${seedResult.stored}/${seedResult.total}`);
+    console.log(`Verdict mix: success=${verdictMix.success} partial=${verdictMix.partial} reverted=${verdictMix.reverted} hotfixed=${verdictMix.hotfixed}`);
+    console.log(`Overall: ${summary.passed ? '✅ PASSED' : '⚠️  partial'}`);
+    if (unified1.consistency.notes.length > 0) {
+      console.log(`\nConsistency notes:`);
+      for (const n of unified1.consistency.notes) console.log(`  • ${n}`);
+    }
+  }
+
+  if (!process.env.BENCH_NO_WRITE) {
+    mkdirSync(RUNS_DIR, { recursive: true });
+    const stamp = summary.runAt.replace(/[:.]/g, '-');
+    writeFileSync(join(RUNS_DIR, `pretrain-from-github-${stamp}.json`), JSON.stringify(summary, null, 2));
+    writeFileSync(join(RUNS_DIR, 'pretrain-from-github-latest.json'), JSON.stringify(summary, null, 2));
+    if (!process.env.BENCH_JSON) console.log(`\nWrote ${join(RUNS_DIR, `pretrain-from-github-${stamp}.json`)}`);
+  }
+
+  // ONNX runtime keeps a worker thread alive — force exit so this can be used
+  // as a CI step or chained with other scripts.
+  process.exit(summary.passed ? 0 : 1);
+}
+
+main().catch((err) => { console.error(err); process.exit(1); });