agent-devkit 0.3.2 → 0.3.3

package/runtime/vendor/plugins/ruflo/ruflo-cost-tracker/scripts/burn.mjs

@@ -1,200 +0,0 @@
-#!/usr/bin/env node
-// cost-burn — burn-rate trend over time + acceleration alert.
-//
-// Iters 34 + 50 of ADR-149 router work added windowed-trend visibility
-// and drift-alert exit codes. This is the cost-tracker plugin equivalent
-// for session spend: bin sessions into daily (or custom-duration) buckets,
-// surface trend, and optionally fail builds on rate acceleration.
-//
-// METHOD
-//   1. Read `cost-tracking` namespace session records.
-//   2. Bin into buckets of `--bucket` duration (default 1d) over `--lookback`
-//      window (default 14d). Each bucket gets {n, spend, avgPerSession}.
-//   3. Compute window-over-window delta: latest bucket's spend vs the mean
-//      of the prior N buckets (default: all prior buckets).
-//   4. If `--alert-on-acceleration-pct N` is set AND the latest bucket
-//      spend is > (prior mean × (1 + N/100)), exit 1 with an alert message.
-//
-// Distinct from `cost-trend` (which surfaces BENCHMARK drift across
-// docs/benchmarks/runs/*.json). This tracks PRODUCTION spend over time.
-//
-// Pairs with:
-//   - cost-budget-check    (reactive: "have we crossed?")
-//   - cost-projection      (predictive: "when will we cross?")
-//   - cost-counterfactual  (comparative: "is routing earning its keep?")
-//   - cost-burn            (trend: "is daily burn accelerating?")  ← this
-//
-// USAGE
-//   node scripts/burn.mjs                                           # default 14d / 1d bins
-//   node scripts/burn.mjs --bucket 1d --lookback 30d                # daily bins, last month
-//   node scripts/burn.mjs --bucket 1w --lookback 90d                # weekly bins, last quarter
-//   node scripts/burn.mjs --alert-on-acceleration-pct 50            # exit 1 if latest > +50% vs prior mean
-//   node scripts/burn.mjs --format json
-//
-// Env: BURN_NAMESPACE (default cost-tracking), BURN_QUIET=1.
-
-// iter 73 — shared session-loader (was duplicated across 6 scripts).
-import { loadSessions, sessionTs, parseDurationMs } from './_sessions.mjs';
-
-const NS = process.env.BURN_NAMESPACE || 'cost-tracking';
-
-const ARGS = (() => {
-  const a = { bucket: '1d', lookback: '14d', alertPct: null, format: 'table' };
-  if (process.env.BURN_QUIET === '1') a.format = 'json';
-  for (let i = 2; i < process.argv.length; i++) {
-    const v = process.argv[i];
-    if (v === '--bucket') a.bucket = process.argv[++i];
-    else if (v === '--lookback') a.lookback = process.argv[++i];
-    else if (v === '--alert-on-acceleration-pct') a.alertPct = parseFloat(process.argv[++i]);
-    else if (v === '--format') a.format = process.argv[++i];
-  }
-  return a;
-})();
-
-function main() {
-  const bucketMs = parseDurationMs(ARGS.bucket);
-  const lookbackMs = parseDurationMs(ARGS.lookback);
-  if (!bucketMs) {
-    console.error(`burn: --bucket must be N(h|d|w|m); got "${ARGS.bucket}"`);
-    process.exit(2);
-  }
-  if (!lookbackMs) {
-    console.error(`burn: --lookback must be N(h|d|w|m); got "${ARGS.lookback}"`);
-    process.exit(2);
-  }
-  if (bucketMs > lookbackMs) {
-    console.error(`burn: --bucket (${ARGS.bucket}) cannot exceed --lookback (${ARGS.lookback})`);
-    process.exit(2);
-  }
-
-  const now = Date.now();
-  const cutoffMs = now - lookbackMs;
-
-  // Read + filter sessions (shared loader, iter 73).
-  const records = loadSessions(NS);
-  const windowed = records.filter((r) => sessionTs(r) >= cutoffMs);
-
-  // Bin: bucket index counts BACKWARDS from now, so the LAST bucket is
-  // "the most recent <bucket>". Operators read newest-first.
-  // bucket index 0 = newest (now-bucketMs .. now), index 1 = (now-2*bucketMs .. now-bucketMs), etc.
-  const numBuckets = Math.ceil(lookbackMs / bucketMs);
-  const buckets = Array.from({ length: numBuckets }, (_, i) => ({
-    index: i,
-    startMs: now - (i + 1) * bucketMs,
-    endMs: now - i * bucketMs,
-    n: 0,
-    spendUsd: 0,
-  }));
-  for (const rec of windowed) {
-    const ts = sessionTs(rec);
-    const offsetMs = now - ts;
-    const idx = Math.floor(offsetMs / bucketMs);
-    if (idx >= 0 && idx < numBuckets) {
-      buckets[idx].n++;
-      buckets[idx].spendUsd += rec.total_cost_usd || 0;
-    }
-  }
-
-  // Window-over-window: latest bucket (index 0) vs mean of prior buckets
-  // that have at least one session (avoid divide-by-zero on sparse history).
-  const latest = buckets[0];
-  const priorNonEmpty = buckets.slice(1).filter((b) => b.n > 0);
-  const priorMean = priorNonEmpty.length > 0
-    ? priorNonEmpty.reduce((s, b) => s + b.spendUsd, 0) / priorNonEmpty.length
-    : 0;
-  const deltaUsd = latest.spendUsd - priorMean;
-  const deltaPct = priorMean > 0 ? (deltaUsd / priorMean) * 100 : (latest.spendUsd > 0 ? Infinity : 0);
-
-  // Alert check.
-  let alertTriggered = false;
-  let alertReason = null;
-  if (ARGS.alertPct !== null) {
-    if (!isFinite(ARGS.alertPct) || ARGS.alertPct <= 0) {
-      console.error(`burn: --alert-on-acceleration-pct must be > 0; got ${ARGS.alertPct}`);
-      process.exit(2);
-    }
-    if (priorNonEmpty.length === 0) {
-      alertReason = `skipped (no prior non-empty buckets to compare — need ≥1)`;
-    } else if (priorMean === 0) {
-      alertReason = `skipped (prior buckets are all $0)`;
-    } else if (deltaPct > ARGS.alertPct) {
-      alertTriggered = true;
-      alertReason = `latest bucket $${latest.spendUsd.toFixed(6)} is ${deltaPct.toFixed(1)}% above prior mean $${priorMean.toFixed(6)} (threshold +${ARGS.alertPct}%)`;
-    } else {
-      alertReason = `latest bucket within +${ARGS.alertPct}% of prior mean (actual delta: ${deltaPct.toFixed(1)}%) — OK`;
-    }
-  }
-
-  const payload = {
-    namespace: NS,
-    config: { bucket: ARGS.bucket, lookback: ARGS.lookback, alertOnAccelerationPct: ARGS.alertPct },
-    bucketsConsidered: buckets.length,
-    sessionsInLookback: windowed.length,
-    latest: {
-      windowStart: new Date(latest.startMs).toISOString(),
-      windowEnd: new Date(latest.endMs).toISOString(),
-      sessions: latest.n,
-      spendUsd: Math.round(latest.spendUsd * 1e6) / 1e6,
-    },
-    priorMean: {
-      bucketsConsidered: priorNonEmpty.length,
-      meanSpendUsd: Math.round(priorMean * 1e6) / 1e6,
-    },
-    delta: {
-      deltaUsd: Math.round(deltaUsd * 1e6) / 1e6,
-      deltaPct: isFinite(deltaPct) ? Math.round(deltaPct * 100) / 100 : null,
-    },
-    series: buckets.map((b) => ({
-      bucketIndex: b.index,
-      windowStart: new Date(b.startMs).toISOString(),
-      windowEnd: new Date(b.endMs).toISOString(),
-      sessions: b.n,
-      spendUsd: Math.round(b.spendUsd * 1e6) / 1e6,
-    })),
-    alert: ARGS.alertPct !== null ? {
-      triggered: alertTriggered,
-      reason: alertReason,
-      thresholdPct: ARGS.alertPct,
-    } : null,
-    generatedAt: new Date().toISOString(),
-  };
-
-  if (ARGS.format === 'json') {
-    console.log(JSON.stringify(payload, null, 2));
-  } else {
-    console.log(`# cost-burn (bucket=${ARGS.bucket}, lookback=${ARGS.lookback})`);
-    console.log('');
-    console.log(`| Metric | Value |`);
-    console.log(`|---|---:|`);
-    console.log(`| Buckets considered | ${buckets.length} |`);
-    console.log(`| Sessions in lookback | ${windowed.length} |`);
-    console.log(`| Latest bucket spend | $${latest.spendUsd.toFixed(6)} (${latest.n} sessions) |`);
-    console.log(`| Prior bucket mean | $${priorMean.toFixed(6)} (${priorNonEmpty.length} non-empty buckets) |`);
-    const deltaStr = deltaUsd >= 0 ? `+$${deltaUsd.toFixed(6)}` : `-$${Math.abs(deltaUsd).toFixed(6)}`;
-    const pctStr = isFinite(deltaPct) ? `${deltaPct.toFixed(2)}%` : (priorMean === 0 ? 'new' : '∞');
-    console.log(`| **Delta (latest vs prior mean)** | **${deltaStr} (${pctStr})** |`);
-    console.log('');
-    console.log('## Bucket series (newest first)');
-    console.log('');
-    console.log('| # | Window | Sessions | Spend |');
-    console.log('|---:|---|---:|---:|');
-    for (const b of buckets) {
-      const startShort = new Date(b.startMs).toISOString().slice(0, 16).replace('T', ' ');
-      const endShort = new Date(b.endMs).toISOString().slice(0, 16).replace('T', ' ');
-      console.log(`| ${b.index} | ${startShort} → ${endShort} | ${b.n} | $${b.spendUsd.toFixed(6)} |`);
-    }
-    console.log('');
-    if (alertReason !== null) {
-      if (alertTriggered) {
-        console.log(`⚠ **ALERT**: ${alertReason}`);
-      } else {
-        console.log(`✓ ${alertReason}`);
-      }
-      console.log('');
-    }
-  }
-
-  if (alertTriggered) process.exit(1);
-}
-
-main();

package/runtime/vendor/plugins/ruflo/ruflo-cost-tracker/scripts/compact.mjs

@@ -1,153 +0,0 @@
-#!/usr/bin/env node
-// cost-compact — wraps getTokenOptimizer().getCompactContext() so the
-// cost-compact-context skill can invoke a single command instead of an
-// inlined Node one-liner.
-//
-// Resolution strategy (#1930): pnpm-workspace marketplace installs don't
-// hoist `@claude-flow/integration` into `v3/node_modules/`, so resolving
-// from cwd alone breaks. Try, in order:
-//   1. createRequire(cwd)           — works for npm-style hoisted installs
-//   2. createRequire(script-dir)    — works when run from inside the plugin
-//   3. absolute path from this file — works on the marketplace layout
-//      (script lives at <ruflo>/plugins/ruflo-cost-tracker/scripts/, the
-//       integration package at <ruflo>/v3/@claude-flow/integration/)
-//
-// Also: agentic-flow's index.js binds `process.env.HEALTH_PORT || 8080`
-// unconditionally on import. On any host with something already on 8080
-// (k8s NodePort, forwarded dev server) the import EADDRINUSEs. Set
-// HEALTH_PORT to 0 by default so the OS picks a free port — users can
-// still override with an explicit value.
-//
-// Usage:
-//   node scripts/compact.mjs "<query>"
-//
-// Optional env:
-//   COMPACT_QUIET=1   emit JSON only (no markdown banner)
-//   HEALTH_PORT=<n>   override agentic-flow's health-server port (default 0)
-
-import { createRequire } from 'node:module';
-import { pathToFileURL, fileURLToPath } from 'node:url';
-import { join, resolve, dirname, isAbsolute } from 'node:path';
-import { existsSync } from 'node:fs';
-
-// #1930 step 4: pick a free port unless caller pinned one. Set BEFORE the
-// dynamic import below — agentic-flow reads HEALTH_PORT at module-load time.
-if (!process.env.HEALTH_PORT) process.env.HEALTH_PORT = '0';
-
-const __filename = fileURLToPath(import.meta.url);
-const __dirname = dirname(__filename);
-
-function tryResolveFrom(anchorPath) {
-  try {
-    const req = createRequire(anchorPath);
-    return req.resolve('@claude-flow/integration/token-optimizer');
-  } catch {
-    return null;
-  }
-}
-
-function tryAbsolutePath() {
-  // Walk up from this script looking for `v3/@claude-flow/integration/dist/token-optimizer.js`.
-  // Handles the marketplace layout (script at plugins/ruflo-cost-tracker/scripts/) and the
-  // dev-repo layout. Caps depth at 6 to avoid runaway traversal.
-  let dir = __dirname;
-  for (let i = 0; i < 6; i++) {
-    const candidate = join(dir, 'v3', '@claude-flow', 'integration', 'dist', 'token-optimizer.js');
-    if (existsSync(candidate)) return candidate;
-    const parent = dirname(dir);
-    if (parent === dir) break;
-    dir = parent;
-  }
-  return null;
-}
-
-async function main() {
-  const query = process.argv[2] || '';
-  if (!query) {
-    console.error('usage: compact.mjs "<query>"');
-    process.exit(2);
-  }
-
-  // Resolution chain — cwd → script dir → absolute fallback (#1930 step 5)
-  const candidates = [
-    tryResolveFrom(join(process.cwd(), 'package.json')),
-    tryResolveFrom(join(__dirname, 'package.json')),
-    tryAbsolutePath(),
-  ].filter(Boolean);
-
-  let mod;
-  let lastErr = null;
-  for (const resolved of candidates) {
-    try {
-      // Skip dangling resolutions that point at a non-existent file (a
-      // common pnpm-workspace symptom when dist/ wasn't built — see
-      // #1930 step 1).
-      const abs = isAbsolute(resolved) ? resolved : resolve(resolved);
-      if (!existsSync(abs)) {
-        lastErr = new Error(`token-optimizer.js missing at ${abs} — run \`cd v3/@claude-flow/integration && pnpm run build\` (#1930 step 1)`);
-        continue;
-      }
-      mod = await import(pathToFileURL(abs).href);
-      break;
-    } catch (err) {
-      lastErr = err;
-    }
-  }
-
-  if (!mod) {
-    const reason = lastErr
-      ? String(lastErr.message || lastErr).slice(0, 300)
-      : 'Cannot find module @claude-flow/integration/token-optimizer';
-    const out = {
-      bridgeUnavailable: true,
-      reason,
-      memoriesRetrieved: 0,
-      tokensSaved: 0,
-      agenticFlowAvailable: false,
-      hint: 'Tried cwd, script dir, and absolute path from this script — see #1930 for full troubleshooting.',
-    };
-    if (process.env.COMPACT_QUIET === '1') return console.log(JSON.stringify(out));
-    console.log(`# cost-compact-context\n\nbridge unavailable: ${reason}`);
-    console.log('');
-    console.log('Tried:');
-    console.log('  - resolve from cwd');
-    console.log('  - resolve from script dir');
-    console.log('  - absolute path walk from this script (looking for v3/@claude-flow/integration/dist/token-optimizer.js)');
-    console.log('');
-    console.log('If `@claude-flow/integration` is installed but `dist/` is missing, run:');
-    console.log('  cd v3/@claude-flow/integration && pnpm run build');
-    return;
-  }
-
-  const { getTokenOptimizer } = mod;
-  const opt = await getTokenOptimizer();
-  const ctx = await opt.getCompactContext(query);
-  const stats = opt.getStats();
-
-  const out = {
-    query,
-    memoriesRetrieved: ctx.memories?.length ?? 0,
-    tokensSaved: ctx.tokensSaved ?? 0,
-    agenticFlowAvailable: !!stats?.agenticFlowAvailable,
-    cacheHitRate: stats?.cacheHitRate || '0%',
-    upstreamReported: 'tokensSaved is bridge-reported (heuristic), not measured against a no-RAG baseline',
-  };
-
-  if (process.env.COMPACT_QUIET === '1') return console.log(JSON.stringify(out));
-  console.log(`# cost-compact-context — query: "${query}"`);
-  console.log('');
-  console.log('| Metric | Value |');
-  console.log('|---|---:|');
-  console.log(`| Memories retrieved | ${out.memoriesRetrieved} |`);
-  console.log(`| Tokens saved (bridge-reported) | ${out.tokensSaved} |`);
-  console.log(`| agentic-flow bridge available | ${out.agenticFlowAvailable} |`);
-  console.log(`| Cache hit rate | ${out.cacheHitRate} |`);
-  console.log('');
-  console.log(`> ${out.upstreamReported}`);
-  if (!out.agenticFlowAvailable) {
-    console.log('');
-    console.log('agentic-flow not installed — bridge returns inert results. No compact-context savings.');
-  }
-}
-
-main().catch((e) => { console.error('compact.mjs failed:', e.message || e); process.exit(1); });

package/runtime/vendor/plugins/ruflo/ruflo-cost-tracker/scripts/conversation.mjs

@@ -1,76 +0,0 @@
-#!/usr/bin/env node
-// cost-conversation — list all session records in the cost-tracking namespace
-// and show cost-per-conversation. Until now reports were per-agent / per-model;
-// this gives a per-conversation timeline.
-//
-// Usage:
-//   node scripts/conversation.mjs                    # markdown table
-//   CONV_FORMAT=json node scripts/conversation.mjs   # JSON
-//   CONV_LIMIT=20 node scripts/conversation.mjs      # most recent N
-//   CONV_NAMESPACE=cost-tracking (default)
-
-// iter 73 — shared session-loader (was duplicated across 6 scripts).
-// ADR-100 / #1748 Issue 3 — CLI_CORE=1 routes to lite cli-core (~2s cold-cache).
-import { loadSessions, memoryListSessionKeys } from './_sessions.mjs';
-
-const NS = process.env.CONV_NAMESPACE || 'cost-tracking';
-
-function fmtUsd(n) { return `$${(n || 0).toFixed(4)}`; }
-function fmtTs(s) { return s ? new Date(s).toISOString().replace('T', ' ').replace(/\.\d+Z/, '') : ''; }
-function topModel(byModel) {
-  const entries = Object.entries(byModel || {}).sort((a, b) => (b[1].cost_usd || 0) - (a[1].cost_usd || 0));
-  return entries[0]?.[0] || '—';
-}
-
-function main() {
-  const keys = memoryListSessionKeys(NS);
-  if (!keys.length) {
-    console.log(`No sessions in '${NS}'. Run \`cost track\` first.`);
-    return;
-  }
-  const records = loadSessions(NS);
-  records.sort((a, b) => (a.startedAt || '').localeCompare(b.startedAt || ''));
-
-  const limit = parseInt(process.env.CONV_LIMIT || '50', 10);
-  const recent = records.slice(-limit);
-
-  const total = records.reduce((s, r) => s + (r.total_cost_usd || 0), 0);
-  const byTier = { haiku: 0, sonnet: 0, opus: 0, unknown: 0 };
-  for (const r of records) {
-    if (r.byTier) {
-      for (const [t, v] of Object.entries(r.byTier)) byTier[t] = (byTier[t] || 0) + v;
-    }
-  }
-
-  if (process.env.CONV_FORMAT === 'json') {
-    console.log(JSON.stringify({
-      conversationCount: records.length,
-      shownCount: recent.length,
-      total_cost_usd: total,
-      byTier,
-      conversations: recent,
-    }, null, 2));
-    return;
-  }
-
-  console.log(`# cost-per-conversation — ${records.length} conversations in '${NS}'`);
-  console.log('');
-  console.log(`| Metric | Value |`);
-  console.log(`|---|---:|`);
-  console.log(`| Conversations tracked | ${records.length} |`);
-  console.log(`| **Total cost (all)** | **${fmtUsd(total)}** |`);
-  for (const t of ['haiku', 'sonnet', 'opus', 'unknown']) {
-    if (byTier[t] > 0) console.log(`| Tier ${t} | ${fmtUsd(byTier[t])} |`);
-  }
-  console.log('');
-  console.log(`## ${recent.length === records.length ? 'All' : 'Most recent ' + recent.length} conversations`);
-  console.log('');
-  console.log('| Started | Session | Messages | Top model | Cost |');
-  console.log('|---|---|---:|---|---:|');
-  for (const r of recent) {
-    const sid = (r.sessionId || '').slice(0, 8);
-    console.log(`| ${fmtTs(r.startedAt)} | \`${sid}\` | ${r.messageCount || 0} | \`${topModel(r.byModel)}\` | ${fmtUsd(r.total_cost_usd)} |`);
-  }
-}
-
-main();

package/runtime/vendor/plugins/ruflo/ruflo-cost-tracker/scripts/counterfactual.mjs

@@ -1,188 +0,0 @@
-#!/usr/bin/env node
-// cost-counterfactual — actual vs hypothetical baseline cost analysis.
-//
-// Iters 32-33 of ADR-149 router work added multi-baseline counterfactual
-// analysis for routing decisions. This is the cost-tracker equivalent for
-// session-level spend: what would each session have cost if every message
-// had been routed to a fixed tier (always-haiku / always-sonnet / always-opus)?
-//
-// METHOD
-//   1. Read `cost-tracking` namespace session records (same source as
-//      budget.mjs / conversation.mjs).
-//   2. For each session: sum token totals across all `byModel[*]` entries.
-//   3. For each baseline tier (default: all three):
-//        counterfactualCost = (input_tokens / 1M × tier.input)
-//                           + (output_tokens / 1M × tier.output)
-//                           + (cache_creation / 1M × tier.cache_write)
-//                           + (cache_read     / 1M × tier.cache_read)
-//   4. Compute savings: counterfactual − actual = how much routing saved
-//      (positive) or lost (negative) vs the baseline.
-//   5. Aggregate across sessions; emit per-baseline totals + savings %.
-//
-// Pairs with `cost-budget-check` and `cost-projection`:
-//   - check:        "have we crossed a threshold?" (reactive)
-//   - projection:   "when will we cross a threshold?" (predictive)
-//   - counterfactual: "is the routing earning its keep?" (comparative)
-//
-// USAGE
-//   node scripts/counterfactual.mjs                                # all sessions
-//   node scripts/counterfactual.mjs --since 7d                     # last week
-//   node scripts/counterfactual.mjs --baseline always-haiku        # single baseline
-//   node scripts/counterfactual.mjs --format json                  # pipe-friendly
-//
-// Env: COUNTERFACTUAL_NAMESPACE (default cost-tracking), COUNTERFACTUAL_QUIET=1.
-
-// iter 68 — shared PRICING + cost helpers.
-import { costAtTier } from './_prices.mjs';
-// iter 73 — shared session-loader.
-import { loadSessions, sessionTs, parseDurationMs } from './_sessions.mjs';
-
-const NS = process.env.COUNTERFACTUAL_NAMESPACE || 'cost-tracking';
-
-const ALL_BASELINES = ['always-haiku', 'always-sonnet', 'always-opus'];
-
-const ARGS = (() => {
-  const a = { since: null, baseline: 'all', format: 'table' };
-  if (process.env.COUNTERFACTUAL_QUIET === '1') a.format = 'json';
-  for (let i = 2; i < process.argv.length; i++) {
-    const v = process.argv[i];
-    if (v === '--since') a.since = process.argv[++i];
-    else if (v === '--baseline') a.baseline = process.argv[++i];
-    else if (v === '--format') a.format = process.argv[++i];
-  }
-  return a;
-})();
-
-function tierFromBaseline(baseline) {
-  switch (baseline) {
-    case 'always-haiku':  return 'haiku';
-    case 'always-sonnet': return 'sonnet';
-    case 'always-opus':   return 'opus';
-    default: return null;
-  }
-}
-
-function sumSessionTokens(rec) {
-  const totals = { input: 0, output: 0, cache_write: 0, cache_read: 0 };
-  const byModel = rec.byModel || {};
-  for (const slot of Object.values(byModel)) {
-    totals.input       += slot.input_tokens || 0;
-    totals.output      += slot.output_tokens || 0;
-    totals.cache_write += slot.cache_creation_input_tokens || 0;
-    totals.cache_read  += slot.cache_read_input_tokens || 0;
-  }
-  return totals;
-}
-
-function main() {
-  // Validate baseline arg.
-  const baselines = ARGS.baseline === 'all'
-    ? ALL_BASELINES
-    : [ARGS.baseline];
-  for (const b of baselines) {
-    if (!tierFromBaseline(b)) {
-      console.error(`counterfactual: invalid --baseline "${b}" (must be one of: ${ALL_BASELINES.join(' | ')} | all)`);
-      process.exit(2);
-    }
-  }
-
-  // Read + filter sessions (shared loader, iter 73).
-  const records = loadSessions(NS);
-
-  let cutoffMs = null;
-  if (ARGS.since) {
-    const ms = parseDurationMs(ARGS.since);
-    if (!ms) {
-      console.error(`counterfactual: --since must be N(h|d|w|m); got "${ARGS.since}"`);
-      process.exit(2);
-    }
-    cutoffMs = Date.now() - ms;
-  }
-  const filtered = cutoffMs === null
-    ? records
-    : records.filter((r) => sessionTs(r) >= cutoffMs);
-
-  // Aggregate.
-  let actualUsd = 0;
-  const tokens = { input: 0, output: 0, cache_write: 0, cache_read: 0 };
-  for (const rec of filtered) {
-    actualUsd += rec.total_cost_usd || 0;
-    const t = sumSessionTokens(rec);
-    tokens.input += t.input;
-    tokens.output += t.output;
-    tokens.cache_write += t.cache_write;
-    tokens.cache_read += t.cache_read;
-  }
-
-  const baselineResults = baselines.map((b) => {
-    const tier = tierFromBaseline(b);
-    const counterfactualUsd = costAtTier(tokens, tier);
-    const savingsUsd = counterfactualUsd - actualUsd;
-    const savingsPct = counterfactualUsd > 0
-      ? (savingsUsd / counterfactualUsd) * 100
-      : 0;
-    return {
-      baseline: b,
-      tier,
-      counterfactualUsd: Math.round(counterfactualUsd * 1e6) / 1e6,
-      actualUsd: Math.round(actualUsd * 1e6) / 1e6,
-      savingsUsd: Math.round(savingsUsd * 1e6) / 1e6,
-      savingsPct: Math.round(savingsPct * 100) / 100,
-    };
-  });
-
-  const payload = {
-    namespace: NS,
-    filters: { since: ARGS.since, baseline: ARGS.baseline },
-    sessionsConsidered: filtered.length,
-    tokensTotal: tokens,
-    actualUsd: Math.round(actualUsd * 1e6) / 1e6,
-    baselines: baselineResults,
-    generatedAt: new Date().toISOString(),
-  };
-
-  if (ARGS.format === 'json') {
-    console.log(JSON.stringify(payload, null, 2));
-    return;
-  }
-
-  // Markdown / human output.
-  console.log(`# cost-counterfactual${ARGS.since ? ` (since ${ARGS.since})` : ' (all-time)'}`);
-  console.log('');
-  console.log(`| Metric | Value |`);
-  console.log(`|---|---:|`);
-  console.log(`| Sessions considered | ${filtered.length} |`);
-  console.log(`| Total input tokens | ${tokens.input.toLocaleString()} |`);
-  console.log(`| Total output tokens | ${tokens.output.toLocaleString()} |`);
-  console.log(`| Total cache_write tokens | ${tokens.cache_write.toLocaleString()} |`);
-  console.log(`| Total cache_read tokens | ${tokens.cache_read.toLocaleString()} |`);
-  console.log(`| **Actual spend** | **$${actualUsd.toFixed(6)}** |`);
-  console.log('');
-  if (filtered.length === 0) {
-    console.log('_No sessions in window — nothing to compare. Run `cost track` after agent work to populate `cost-tracking`._');
-    console.log('');
-    return;
-  }
-  console.log('## Counterfactual baselines (what would the same tokens have cost?)');
-  console.log('');
-  console.log('| Baseline | Hypothetical cost | Actual | Savings | % saved |');
-  console.log('|---|---:|---:|---:|---:|');
-  for (const r of baselineResults) {
-    const savingsCell = r.savingsUsd >= 0
-      ? `+$${r.savingsUsd.toFixed(6)}`
-      : `-$${Math.abs(r.savingsUsd).toFixed(6)}`;
-    const pctCell = r.savingsPct >= 0
-      ? `${r.savingsPct.toFixed(2)}%`
-      : `${r.savingsPct.toFixed(2)}%`;
-    console.log(`| \`${r.baseline}\` | $${r.counterfactualUsd.toFixed(6)} | $${actualUsd.toFixed(6)} | ${savingsCell} | ${pctCell} |`);
-  }
-  console.log('');
-  console.log('_Positive savings = routing chose cheaper models than the baseline._');
-  console.log('_Negative savings = routing chose more-expensive models than the baseline (e.g. always-haiku was cheaper)._');
-  if (baselines.length > 1) {
-    console.log('_When always-haiku shows negative savings, qualityBar may be set too high; investigate via `cost optimize`._');
-  }
-  console.log('');
-}
-
-main();