dual-brain 0.2.30 → 0.3.0

package/src/cost-tracker.mjs

@@ -1,184 +0,0 @@
-// cost-tracker.mjs — Lightweight cost estimation and efficiency tracking for .dualbrain/costs.jsonl.
-
-import { readFileSync, appendFileSync, mkdirSync, existsSync } from 'node:fs';
-import { join } from 'node:path';
-
-const TOKEN_COSTS = {
-  'claude-opus-4-6':           0.03,
-  'claude-sonnet-4-6':         0.006,
-  'claude-haiku-4-5-20251001': 0.001,
-  'gpt-5.5':                   0.04,
-  'o3':                        0.03,
-  'gpt-4o':                    0.005,
-  'gpt-4o-mini':               0.0003,
-  'default':                   0.01,
-};
-
-export function estimateTokenCost(model, tokens) {
-  const rate = TOKEN_COSTS[model] ?? TOKEN_COSTS['default'];
-  return (tokens / 1000) * rate;
-}
-
-export function trackCost(action, cwd = process.cwd()) {
-  try {
-    const dir = join(cwd, '.dualbrain');
-    mkdirSync(dir, { recursive: true });
-    const entry = {
-      timestamp:      new Date().toISOString(),
-      action:         action.action         ?? 'execute',
-      model:          action.model          ?? 'default',
-      tokensEstimated: action.tokensEstimated ?? 0,
-      costEstimated:  estimateTokenCost(action.model ?? 'default', action.tokensEstimated ?? 0),
-      tier:           action.tier           ?? 'standard',
-      wasCacheHit:    action.wasCacheHit    ?? false,
-      tokensSaved:    action.tokensSaved    ?? 0,
-    };
-    appendFileSync(join(dir, 'costs.jsonl'), JSON.stringify(entry) + '\n', 'utf8');
-    return entry;
-  } catch {
-    return null;
-  }
-}
-
-function readCostLines(cwd) {
-  const p = join(cwd, '.dualbrain', 'costs.jsonl');
-  if (!existsSync(p)) return [];
-  try {
-    return readFileSync(p, 'utf8').trim().split('\n').filter(Boolean).flatMap(line => {
-      try { return [JSON.parse(line)]; } catch { return []; }
-    });
-  } catch { return []; }
-}
-
-export function getCostSummary(cwd = process.cwd(), days = 7) {
-  const cutoff = new Date(Date.now() - days * 86400000).toISOString();
-  const all    = readCostLines(cwd).filter(e => e.timestamp >= cutoff);
-
-  if (all.length === 0) {
-    return {
-      period: `${days} days`,
-      totalCost: 0, totalTokens: 0, totalActions: 0,
-      cacheHits: 0, tokensSaved: 0, costSaved: 0, savingsRate: 0,
-      byTier: {}, byModel: {}, trend: 'stable',
-    };
-  }
-
-  let totalCost = 0, totalTokens = 0, cacheHits = 0, tokensSaved = 0;
-  const byTier  = {};
-  const byModel = {};
-
-  for (const e of all) {
-    totalCost    += e.costEstimated   ?? 0;
-    totalTokens  += e.tokensEstimated ?? 0;
-    if (e.wasCacheHit) { cacheHits++; tokensSaved += e.tokensSaved ?? 0; }
-
-    const tier = e.tier ?? 'standard';
-    if (!byTier[tier]) byTier[tier] = { count: 0, tokens: 0, cost: 0 };
-    byTier[tier].count  += 1;
-    byTier[tier].tokens += e.tokensEstimated ?? 0;
-    byTier[tier].cost   += e.costEstimated   ?? 0;
-
-    const model = e.model ?? 'default';
-    if (!byModel[model]) byModel[model] = { count: 0, tokens: 0, cost: 0 };
-    byModel[model].count  += 1;
-    byModel[model].tokens += e.tokensEstimated ?? 0;
-    byModel[model].cost   += e.costEstimated   ?? 0;
-  }
-
-  const costSaved    = estimateTokenCost('default', tokensSaved);
-  const savingsRate  = (tokensSaved + totalTokens) > 0
-    ? tokensSaved / (tokensSaved + totalTokens)
-    : 0;
-
-  // Trend: compare first half vs second half savings rate
-  const mid   = Math.floor(all.length / 2);
-  const first = all.slice(0, mid);
-  const second = all.slice(mid);
-  const halfSavings = (half) => {
-    const ts = half.reduce((s, e) => s + (e.tokensSaved ?? 0), 0);
-    const tt = half.reduce((s, e) => s + (e.tokensEstimated ?? 0), 0);
-    return (ts + tt) > 0 ? ts / (ts + tt) : 0;
-  };
-  let trend = 'stable';
-  if (all.length >= 4) {
-    const delta = halfSavings(second) - halfSavings(first);
-    if (delta > 0.05)       trend = 'improving';
-    else if (delta < -0.05) trend = 'degrading';
-  }
-
-  return {
-    period: `${days} days`,
-    totalCost, totalTokens, totalActions: all.length,
-    cacheHits, tokensSaved, costSaved, savingsRate,
-    byTier, byModel, trend,
-  };
-}
-
-export function formatCostReport(summary) {
-  const {
-    period, totalCost, totalTokens, totalActions,
-    cacheHits, tokensSaved, costSaved, savingsRate,
-    byTier, byModel, trend,
-  } = summary;
-
-  const lines = [`COST EFFICIENCY (${period})`];
-
-  const fmtK = (n) => n >= 1000 ? `${Math.round(n / 1000)}K` : String(Math.round(n));
-  const fmtD = (n) => `~$${n.toFixed(2)}`;
-
-  lines.push(`  Total: ${fmtD(totalCost)} (${fmtK(totalTokens)} tokens, ${totalActions} actions)`);
-
-  if (cacheHits > 0) {
-    const pct = Math.round(savingsRate * 100);
-    lines.push(`  Saved: ${fmtD(costSaved)} (${fmtK(tokensSaved)} tokens from ${cacheHits} cache hits)`);
-    lines.push(`  Savings rate: ${pct}%`);
-  }
-
-  const tierOrder = ['recall', 'quick', 'standard', 'deep', 'ultra'];
-  const tierKeys  = [...new Set([...tierOrder, ...Object.keys(byTier)])].filter(k => byTier[k]);
-  if (tierKeys.length > 0) {
-    lines.push('');
-    lines.push('  Tier breakdown:');
-    for (const tier of tierKeys) {
-      const t = byTier[tier];
-      const isRecall = tier === 'recall' && (t.cost < 0.001 || t.tokens === 0);
-      const costStr  = isRecall ? '$0.00  (cache hits!)' : fmtD(t.cost);
-      lines.push(`    ${tier.padEnd(10)}${String(t.count).padStart(4)} actions  ${costStr}`);
-    }
-  }
-
-  const trendIcon = trend === 'improving' ? '↗' : trend === 'degrading' ? '↘' : '→';
-  lines.push('');
-  if (trend !== 'stable') {
-    const pct = Math.round(Math.abs(savingsRate) * 100);
-    lines.push(`  Trend: ${trendIcon} ${trend} (savings rate ${trend === 'improving' ? 'up' : 'down'} vs last half)`);
-  } else {
-    lines.push(`  Trend: ${trendIcon} stable`);
-  }
-
-  return lines.join('\n');
-}
-
-export function getEfficiencyScore(cwd = process.cwd()) {
-  const summary = getCostSummary(cwd, 7);
-  if (summary.totalActions === 0) return 50;
-
-  const TIER_WEIGHTS = { recall: 0, quick: 1, standard: 2, deep: 4, ultra: 6 };
-  const totalTierCost = Object.entries(summary.byTier).reduce((s, [tier, v]) => {
-    return s + (TIER_WEIGHTS[tier] ?? 2) * v.count;
-  }, 0);
-  const maxPossible = summary.totalActions * (TIER_WEIGHTS['ultra'] ?? 6);
-  const tierScore   = maxPossible > 0 ? 1 - (totalTierCost / maxPossible) : 0.5;
-
-  const cacheScore = summary.savingsRate;
-  const trendBonus = summary.trend === 'improving' ? 10 : summary.trend === 'degrading' ? -10 : 0;
-
-  const raw = Math.round(
-    tierScore   * 40 +
-    cacheScore  * 40 +
-    20          +
-    trendBonus
-  );
-
-  return Math.max(1, Math.min(100, raw));
-}

package/src/debrief.mjs

@@ -1,228 +0,0 @@
-const STATUSES = ['success', 'partial', 'blocked', 'pivoted'];
-
-function emptyDebrief() {
-  return {
-    status: 'partial',
-    findings: [],
-    blockers: [],
-    scopeChange: 'same',
-    confidence: 0.5,
-    recommendations: [],
-    artifacts: { filesChanged: [], filesRead: [], testsRun: 0 },
-    pivotReason: undefined,
-    unexpectedFindings: [],
-  };
-}
-
-// ─── Parse debrief from raw agent output ─────────────────────────────────────
-
-const FILE_PATH_RE = /(?:^|\s)([\w./-]+\.(?:mjs|js|ts|tsx|json|md|py|sh|yaml|yml|css|html))/gm;
-const TEST_COUNT_RE = /(\d+)\s*(?:tests?|specs?|assertions?)\s*(?:passed|ran|run|succeeded|ok)/i;
-
-function extractByPatterns(text, patterns, minLen, maxLen, limit) {
-  const results = [];
-  for (const re of patterns) {
-    for (const m of text.matchAll(re)) {
-      const v = m[1].trim();
-      if (v.length > minLen && v.length < maxLen) results.push(v);
-    }
-  }
-  return [...new Set(results)].slice(0, limit);
-}
-
-const extractFindings = t => extractByPatterns(t, [
-  /(?:found|discovered|identified|noticed|see that|observed)\s+(?:that\s+)?(.+?)(?:\.|$)/gim,
-  /^[-•*]\s*(.+)$/gm,
-], 10, 300, 20);
-
-const extractBlockers = t => extractByPatterns(t, [
-  /(?:blocked by|couldn't|cannot|unable to|failed to|can't)\s+(.+?)(?:\.|$)/gim,
-  /(?:blocker|obstacle|issue|problem):\s*(.+?)(?:\.|$)/gim,
-], 5, 200, 10);
-
-const extractRecommendations = t => extractByPatterns(t, [
-  /(?:recommend|suggest|should|next step|consider)\s+(.+?)(?:\.|$)/gim,
-  /(?:TODO|ACTION):\s*(.+?)(?:\.|$)/gim,
-], 5, 200, 10);
-
-const extractUnexpected = t => extractByPatterns(t, [
-  /(?:also noticed|unexpected|surprisingly|interestingly|aside:)\s+(.+?)(?:\.|$)/gim,
-], 10, 200, 5);
-
-function inferStatus(text, blockers, findings) {
-  const lower = text.toLowerCase();
-  if (/pivoted|changed approach|switched to/i.test(lower)) return 'pivoted';
-  if (blockers.length > 0 && findings.length === 0) return 'blocked';
-  if (blockers.length > 0) return 'partial';
-  if (/complete|done|success|finished|all (?:tests )?pass/i.test(lower)) return 'success';
-  return 'partial';
-}
-
-function inferScopeChange(text) {
-  const lower = text.toLowerCase();
-  if (/scope.{0,20}(?:grew|larger|expanded|more than expected)/i.test(lower)) return 'larger';
-  if (/scope.{0,20}(?:smaller|reduced|simpler than)/i.test(lower)) return 'smaller';
-  if (/(?:different approach|completely different|pivoted to)/i.test(lower)) return 'different';
-  return 'same';
-}
-
-function inferConfidence(text, findings, blockers) {
-  if (/(?:very confident|high confidence|certain)/i.test(text)) return 0.9;
-  if (/(?:uncertain|not sure|low confidence|unsure)/i.test(text)) return 0.3;
-  if (blockers.length > 2) return 0.3;
-  if (blockers.length > 0) return 0.5;
-  if (findings.length > 3) return 0.8;
-  return 0.6;
-}
-
-export function parseDebrief(rawOutput) {
-  if (!rawOutput || typeof rawOutput !== 'string') return emptyDebrief();
-
-  // Try JSON-structured debrief first
-  const jsonMatch = rawOutput.match(/```(?:json)?\s*(\{[\s\S]*?"status"[\s\S]*?\})\s*```/);
-  if (jsonMatch) {
-    try {
-      const parsed = JSON.parse(jsonMatch[1]);
-      if (STATUSES.includes(parsed.status)) {
-        const d = emptyDebrief();
-        return { ...d, ...parsed, artifacts: { ...d.artifacts, ...(parsed.artifacts || {}) } };
-      }
-    } catch {}
-  }
-
-  // Best-effort extraction from prose
-  const findings = extractFindings(rawOutput);
-  const blockers = extractBlockers(rawOutput);
-  const recommendations = extractRecommendations(rawOutput);
-  const unexpectedFindings = extractUnexpected(rawOutput);
-
-  const filesAll = [...rawOutput.matchAll(FILE_PATH_RE)].map(m => m[1]);
-  const lower = rawOutput.toLowerCase();
-  const filesChanged = filesAll.filter(f => lower.includes(`wrote ${f.toLowerCase()}`) || lower.includes(`created ${f.toLowerCase()}`) || lower.includes(`modified ${f.toLowerCase()}`) || lower.includes(`edited ${f.toLowerCase()}`));
-  const filesRead = filesAll.filter(f => !filesChanged.includes(f));
-
-  const testMatch = rawOutput.match(TEST_COUNT_RE);
-  const testsRun = testMatch ? parseInt(testMatch[1], 10) : 0;
-
-  const pivotMatch = rawOutput.match(/pivoted?\s+(?:because|since|due to)\s+(.+?)(?:\.|$)/i);
-
-  return {
-    status: inferStatus(rawOutput, blockers, findings),
-    findings,
-    blockers,
-    scopeChange: inferScopeChange(rawOutput),
-    confidence: inferConfidence(rawOutput, findings, blockers),
-    recommendations,
-    artifacts: { filesChanged, filesRead, testsRun },
-    pivotReason: pivotMatch ? pivotMatch[1].trim() : undefined,
-    unexpectedFindings,
-  };
-}
-
-// ─── Generate debrief instruction for agent prompts ──────────────────────────
-
-const TIER_EMPHASIS = {
-  search: 'Focus on: findings, confidence, unexpectedFindings.',
-  execute: 'Focus on: artifacts (filesChanged, testsRun), blockers, scopeChange.',
-  think: 'Focus on: recommendations, confidence, pivotReason.',
-};
-
-export function generateDebriefInstruction(tier, contract) {
-  const emphasis = TIER_EMPHASIS[tier] || TIER_EMPHASIS.execute;
-  const scope = contract?.scope ? ` Scope: ${contract.scope}.` : '';
-  return `\n---\nRESPONSE FORMAT: End your response with a JSON block:\n\`\`\`json\n{"status":"success|partial|blocked|pivoted","findings":[],"blockers":[],"scopeChange":"same|larger|smaller|different","confidence":0.0-1.0,"recommendations":[],"artifacts":{"filesChanged":[],"filesRead":[],"testsRun":0},"pivotReason":"...if pivoted","unexpectedFindings":[]}\n\`\`\`\n${emphasis}${scope}`;
-}
-
-// ─── Integrate debrief into HEAD's situation model ───────────────────────────
-
-export function integrateDebrief(currentSituation, debrief) {
-  const sit = structuredClone(currentSituation);
-
-  // Update confidence on related ledger entries
-  if (!sit.ledger) sit.ledger = [];
-  for (const finding of debrief.findings) {
-    const existing = sit.ledger.find(e => e.topic && finding.toLowerCase().includes(e.topic.toLowerCase()));
-    if (existing) {
-      existing.confidence = Math.min(1, (existing.confidence || 0.5) + 0.1);
-      existing.resolved = true;
-    } else {
-      sit.ledger.push({ topic: finding.slice(0, 80), confidence: debrief.confidence, resolved: false });
-    }
-  }
-
-  // Blockers create new uncertainty entries
-  for (const blocker of debrief.blockers) {
-    sit.ledger.push({ topic: blocker.slice(0, 80), confidence: 0.3, resolved: false, isBlocker: true });
-  }
-
-  // Scope delta
-  if (debrief.scopeChange !== 'same') {
-    if (!sit.material) sit.material = {};
-    sit.material.scopeTrend = debrief.scopeChange;
-    if (debrief.artifacts?.filesChanged) {
-      sit.material.touchedFiles = [
-        ...new Set([...(sit.material.touchedFiles || []), ...debrief.artifacts.filesChanged]),
-      ];
-    }
-  }
-
-  // Recommendations → next wave planning
-  if (!sit.nextActions) sit.nextActions = [];
-  sit.nextActions.push(...debrief.recommendations);
-
-  // Unexpected findings → noticings
-  if (debrief.unexpectedFindings?.length) {
-    if (!sit.noticings) sit.noticings = [];
-    sit.noticings.push(...debrief.unexpectedFindings);
-  }
-
-  // Overall confidence update
-  sit.confidence = sit.confidence
-    ? sit.confidence * 0.7 + debrief.confidence * 0.3
-    : debrief.confidence;
-
-  return sit;
-}
-
-// ─── Summarize a parallel wave of debriefs ───────────────────────────────────
-
-export function summarizeWaveOutcome(debriefs) {
-  if (!debriefs || debriefs.length === 0) {
-    return {
-      overallStatus: 'blocked',
-      aggregateConfidence: 0,
-      allFindings: [],
-      allBlockers: [],
-      scopeDelta: 'same',
-      nextActions: [],
-    };
-  }
-
-  const statusPriority = { blocked: 0, pivoted: 1, partial: 2, success: 3 };
-  const worstStatus = debriefs.reduce((worst, d) =>
-    (statusPriority[d.status] ?? 2) < (statusPriority[worst] ?? 2) ? d.status : worst,
-    'success'
-  );
-
-  const allFindings = [...new Set(debriefs.flatMap(d => d.findings || []))];
-  const allBlockers = [...new Set(debriefs.flatMap(d => d.blockers || []))];
-  const nextActions = [...new Set(debriefs.flatMap(d => d.recommendations || []))];
-
-  const aggregateConfidence = debriefs.reduce((sum, d) => sum + (d.confidence || 0.5), 0) / debriefs.length;
-
-  // Scope: any non-same scope wins, larger beats smaller, different beats all
-  const scopeDeltas = debriefs.map(d => d.scopeChange || 'same').filter(s => s !== 'same');
-  let scopeDelta = 'same';
-  if (scopeDeltas.includes('different')) scopeDelta = 'different';
-  else if (scopeDeltas.includes('larger')) scopeDelta = 'larger';
-  else if (scopeDeltas.includes('smaller')) scopeDelta = 'smaller';
-
-  return {
-    overallStatus: worstStatus,
-    aggregateConfidence: Math.round(aggregateConfidence * 100) / 100,
-    allFindings,
-    allBlockers,
-    scopeDelta,
-    nextActions,
-  };
-}