tribunal-kit 9.2.6 → 9.2.7

package/src/system1/decision_engine.js

@@ -0,0 +1,498 @@
+'use strict';
+
+/**
+ * decision_engine.js — System-1 Capability 2: Cross-Capability Decision Engine
+ * ============================================================================
+ * Unified System-1 decision layer that combines:
+ * - Impact (Laya ONNX + change volume heuristic)
+ * - Risk (diff patterns, sensitive file types, credential leaks)
+ * - Evidence Intelligence (relevance, coverage, confidence L1-L4)
+ * - Browser / Interaction Intelligence (UI risk, route impact, CDP availability)
+ * - Adaptive Reviewer Orchestration (minimal covering reviewer set)
+ *
+ * Core Directives:
+ * 1. Monotonic Safety: Uncertainty maintains or increases scrutiny; never silently reduces scrutiny.
+ * 2. Precedence: High-risk signals can never be downgraded by low-confidence capabilities.
+ * 3. Explainability: Every decision contains an explicit audit trail of why the tier was chosen,
+ *    which signals fired, which capabilities were consulted or unavailable, and whether fallback was used.
+ */
+
+const crypto = require('crypto');
+const {
+  HIGH_RISK_PATTERNS,
+  HIGH_RISK_EXTENSIONS,
+  TIER_NAMES,
+  ROUTING_NAMES,
+} = require('./constants');
+const { collectAndRankEvidence } = require('./evidence_engine');
+const { evaluateBrowserRequirement } = require('./browser_intelligence');
+const { orchestrateReviewers } = require('./reviewer_orchestrator');
+
+function resolveMonotonicImpactTier(options = {}) {
+  const opts = options && typeof options === 'object' ? options : {};
+  const rawFiles = opts.files;
+  const files = (
+    Array.isArray(rawFiles) ? rawFiles : typeof rawFiles === 'string' ? rawFiles.split(',') : []
+  )
+    .map(f =>
+      typeof f === 'string'
+        ? f
+            .replace(/[\u200B-\u200D\uFEFF]/g, '')
+            .trim()
+            .replace(/\\/g, '/')
+        : '',
+    )
+    .filter(Boolean);
+  const rawLines =
+    typeof opts.lines === 'number' && !isNaN(opts.lines)
+      ? opts.lines
+      : parseInt(opts.lines, 10) || 0;
+  const lines = Math.max(0, rawLines);
+  const task = typeof opts.task === 'string' ? opts.task.replace(/[\u200B-\u200D\uFEFF]/g, '') : '';
+  const diff = typeof opts.diff === 'string' ? opts.diff.replace(/[\u200B-\u200D\uFEFF]/g, '') : '';
+  const layaTier = opts.layaTier;
+
+  const fileCount = files.length;
+
+  // 1. Base tier from change volume
+  const isFastPass =
+    fileCount <= 2 &&
+    lines <= 10 &&
+    files.length > 0 &&
+    files.every(f => /\.(css|scss|less|md|txt|svg|png|jpg)$/i.test(f));
+
+  let baseTier = 0;
+  if (isFastPass || (fileCount === 0 && lines <= 5)) baseTier = 0;
+  else if (fileCount <= 1 && lines <= 50) baseTier = 1;
+  else if (fileCount <= 5 && lines <= 200) baseTier = 2;
+  else baseTier = 3;
+
+  // 2. Concrete risk evidence analysis
+  const isPureDocFiles =
+    fileCount > 0 &&
+    files.every(f => /\.(md|txt|markdown|rst|adoc)$/i.test(f) || f.includes('docs/'));
+
+  // Concrete secret or code patterns that make even documentation dangerous
+  const secretOrExecPatterns =
+    /(?:-----BEGIN [A-Z ]+KEY-----|eyJ[A-Za-z0-9-_=]+\.[A-Za-z0-9-_=]+|AKIA[0-9A-Z]{16}|(?:aws[_-]?)?(?:api[_-]?key|secret(?:[_-]?key)?|password|bearer|auth[_-]?token|access[_-]?key(?:[_-]?id)?)\s*[:=]\s*['"]?[^'"\s\n]{6,}['"]?)/i;
+
+  const diffHasCode =
+    /^\+[ \t]*(?:const|let|var|function|def|import|require|class|export|db\.|SELECT|UPDATE|DELETE|INSERT|jwt\.|auth\.)/im.test(
+      diff,
+    );
+  const diffMatchesHighRiskCode = diffHasCode && HIGH_RISK_PATTERNS.test(diff);
+
+  const diffHasHighRisk =
+    HIGH_RISK_EXTENSIONS.test(diff) ||
+    diffMatchesHighRiskCode ||
+    (!isPureDocFiles && HIGH_RISK_PATTERNS.test(diff)) ||
+    secretOrExecPatterns.test(diff);
+
+  const fileHasHighRisk = files.some(
+    f =>
+      HIGH_RISK_EXTENSIONS.test(f) ||
+      (!isPureDocFiles && HIGH_RISK_PATTERNS.test(f)) ||
+      f.includes('docker') ||
+      f.includes('.github/'),
+  );
+
+  let riskEvidenceTier = 0;
+  if (diffHasHighRisk || fileHasHighRisk) {
+    riskEvidenceTier = 3;
+  } else if (!isPureDocFiles && HIGH_RISK_PATTERNS.test(task + ' ' + files.join(' '))) {
+    riskEvidenceTier = 3;
+  }
+
+  // 3. Negative controls (Documentation changes)
+  // A doc change remains low risk (Tier 0) UNLESS concrete high-risk implementation evidence exists in files or diff
+  if (isPureDocFiles && !diffHasHighRisk && !fileHasHighRisk) {
+    return 0;
+  }
+
+  // 4. Ambiguity tier: short vague tasks (<= 2 words) touching non-doc code files
+  let ambiguityTier = 0;
+  const wordCount = task.trim().length > 0 ? task.trim().split(/\s+/).length : 0;
+  if (wordCount > 0 && wordCount <= 2 && fileCount > 0 && !isPureDocFiles) {
+    ambiguityTier = 2;
+  }
+
+  // 5. Monotonic aggregation: finalTier = max(baseTier, layaTier, riskEvidenceTier, ambiguityTier)
+  const candidateTiers = [baseTier, riskEvidenceTier, ambiguityTier];
+  if (typeof layaTier === 'number' && !isNaN(layaTier) && layaTier >= 0 && layaTier <= 3) {
+    candidateTiers.push(layaTier);
+  }
+
+  return Math.max(...candidateTiers);
+}
+
+/**
+ * Evaluate cross-capability System-1 decision
+ * @param {Object} options
+ * @param {string} [options.task]
+ * @param {string} [options.diff]
+ * @param {string[]|string} [options.files]
+ * @param {number} [options.lines]
+ * @param {Object} [options.evidence] - Optional pre-computed evidence
+ * @param {Object} [options.browserSignal] - Optional pre-computed browser signal
+ * @param {number|null} [options.layaTier] - Optional Laya inference output
+ * @param {string} [options.repoRoot] - Optional repository root
+ * @param {Array} [options.memory] - Optional memory items
+ * @param {Array} [options.previousFindings] - Optional previous findings
+ * @param {Object} [options.environment] - Optional environment overrides (e.g. browserFound)
+ * @returns {Object} Comprehensive Decision Payload
+ */
+function evaluateDecision(options = {}) {
+  const rawTask = typeof options.task === 'string' ? options.task : '';
+  const task = rawTask.replace(/[\u200B-\u200D\uFEFF]/g, '');
+  const rawDiff = typeof options.diff === 'string' ? options.diff : '';
+  const diff = rawDiff.replace(/[\u200B-\u200D\uFEFF]/g, '');
+  const rawFiles = options.files;
+  const files = (
+    Array.isArray(rawFiles) ? rawFiles : typeof rawFiles === 'string' ? rawFiles.split(',') : []
+  )
+    .map(f =>
+      typeof f === 'string'
+        ? f
+            .replace(/[\u200B-\u200D\uFEFF]/g, '')
+            .trim()
+            .replace(/\\/g, '/')
+        : '',
+    )
+    .filter(Boolean);
+  const rawLines =
+    typeof options.lines === 'number' && !isNaN(options.lines)
+      ? options.lines
+      : parseInt(options.lines, 10) || (diff ? diff.split('\n').length : 0);
+  const lines = Math.max(0, rawLines);
+  const repoRoot = options.repoRoot || process.cwd();
+  const layaTier = options.layaTier ?? null;
+
+  const capabilitiesConsulted = [];
+  const capabilitiesUnavailable = [];
+  const firedRiskSignals = [];
+  const monotonicSafeguards = [];
+  let fallbackUsed = false;
+
+  // 1. Evidence Intelligence Consultation
+  let evidenceResult = options.evidence;
+  if (!evidenceResult) {
+    try {
+      evidenceResult = collectAndRankEvidence({
+        task,
+        diff,
+        files,
+        repoRoot,
+        memory: options.memory,
+        previousFindings: options.previousFindings,
+      });
+      capabilitiesConsulted.push('evidence_engine');
+    } catch (_) {
+      capabilitiesUnavailable.push('evidence_engine');
+      evidenceResult = {
+        evidence: [],
+        summary: { confidence: 'L4', critical_evidence_count: 0, has_security_boundary: false },
+      };
+    }
+  } else {
+    capabilitiesConsulted.push('evidence_engine (provided)');
+  }
+
+  // 2. Browser Intelligence Consultation
+  let browserSignal = options.browserSignal;
+  if (!browserSignal) {
+    try {
+      browserSignal = evaluateBrowserRequirement({
+        files,
+        diff,
+        task,
+        environment: options.environment,
+      });
+      capabilitiesConsulted.push('browser_intelligence');
+      if (browserSignal.browser_available === 'UNAVAILABLE') {
+        capabilitiesUnavailable.push('browser_cdp');
+      }
+    } catch (_) {
+      capabilitiesUnavailable.push('browser_intelligence');
+      browserSignal = {
+        requires_browser_verification: false,
+        validation_level: 'NONE',
+        browser_available: 'UNAVAILABLE',
+        escalate: false,
+      };
+    }
+  } else {
+    capabilitiesConsulted.push('browser_intelligence (provided)');
+  }
+
+  // 3. Impact & Risk Analysis
+  if (layaTier !== null && typeof layaTier === 'number') {
+    capabilitiesConsulted.push('laya_onnx');
+  } else {
+    capabilitiesConsulted.push('deterministic_heuristic');
+    capabilitiesUnavailable.push('laya_onnx');
+    fallbackUsed = true;
+  }
+
+  // Scan concrete risk signals
+  const combinedContext = `${task}\n${diff}\n${files.join('\n')}`;
+  if (HIGH_RISK_PATTERNS.test(combinedContext)) {
+    firedRiskSignals.push('HIGH_RISK_PATTERNS (auth/crypto/secrets/tokens/database)');
+  }
+  if (files.some(f => HIGH_RISK_EXTENSIONS.test(f)) || HIGH_RISK_EXTENSIONS.test(diff)) {
+    firedRiskSignals.push('HIGH_RISK_EXTENSIONS (.sql, .prisma, .tf)');
+  }
+  if (
+    /(?:-----BEGIN [A-Z ]+KEY-----|eyJ[A-Za-z0-9-_=]+\.[A-Za-z0-9-_=]+|(?:api[_-]?key|secret|password)\s*[:=])/i.test(
+      diff,
+    )
+  ) {
+    firedRiskSignals.push('SECRET_OR_KEY_PATTERN (possible exposed credentials)');
+  }
+  if (evidenceResult.summary?.has_security_boundary) {
+    firedRiskSignals.push(
+      'SECURITY_BOUNDARY_EVIDENCE (direct touch of security policy / boundary)',
+    );
+  }
+
+  // Calculate baseline monotonic tier with fail-closed production safeguard
+  let calculatedTier = 3;
+  try {
+    const resolver =
+      typeof options.tierResolver === 'function'
+        ? options.tierResolver
+        : resolveMonotonicImpactTier;
+    calculatedTier = resolver({
+      files,
+      lines,
+      task,
+      diff,
+      layaTier,
+    });
+  } catch (err) {
+    calculatedTier = 3;
+    monotonicSafeguards.push(
+      `Tier calculation fallback to Tier 3 due to exception: ${err.message}`,
+    );
+  }
+
+  if (
+    typeof calculatedTier !== 'number' ||
+    isNaN(calculatedTier) ||
+    calculatedTier < 0 ||
+    calculatedTier > 3
+  ) {
+    calculatedTier = 3;
+    monotonicSafeguards.push(
+      `Tier calculation fallback to Tier 3 due to invalid output from resolver`,
+    );
+  }
+
+  let finalTier = calculatedTier;
+  let escalated = false;
+  let escalationReason = null;
+
+  // 4. Cross-Capability Monotonic Rules & Conflict Resolution
+  // Base tier strictly from change volume (ignoring risk for safeguard comparison)
+  const isFastPassVol =
+    files.length <= 2 &&
+    lines <= 10 &&
+    files.every(f => /\.(css|scss|less|md|txt|svg|png|jpg)$/i.test(f));
+  const baseVolumeTier =
+    isFastPassVol || (files.length === 0 && lines <= 5)
+      ? 0
+      : files.length <= 1 && lines <= 50
+        ? 1
+        : files.length <= 5 && lines <= 200
+          ? 2
+          : 3;
+
+  // Rule A: High Risk + Low Base Tier -> Enforce Tier 3
+  if (firedRiskSignals.length > 0) {
+    const isPureDoc = files.length > 0 && files.every(f => /\.(md|txt)$/i.test(f));
+    const hasDiffRisk = firedRiskSignals.some(
+      s =>
+        s.includes('SECRET') ||
+        s.includes('HIGH_RISK_EXTENSIONS') ||
+        s.includes('SECURITY_BOUNDARY') ||
+        diff.includes('jwt') ||
+        diff.includes('auth'),
+    );
+    if (!isPureDoc || hasDiffRisk) {
+      if (finalTier < 3) {
+        finalTier = 3;
+      }
+      if (baseVolumeTier < 3) {
+        monotonicSafeguards.push(
+          'Enforced Tier 3: Concrete risk signals detected despite low change volume',
+        );
+      }
+    }
+  }
+
+  // Rule B: Low Evidence Confidence (L4 / unavailable) forbids Fast-Pass (Tier 0)
+  const evidenceConfidence = evidenceResult.summary?.confidence || 'L4';
+  if ((evidenceConfidence === 'L4' || evidenceResult.evidence.length === 0) && finalTier === 0) {
+    const isTrivialDoc = files.length > 0 && files.every(f => /\.(md|txt)$/i.test(f)) && lines <= 5;
+    if (!isTrivialDoc) {
+      finalTier = 1; // Cannot fast-pass when evidence is unavailable/unverified
+      monotonicSafeguards.push(
+        'Elevated from Tier 0 to Tier 1: Low evidence confidence forbids Fast-Pass assumption',
+      );
+    }
+  }
+
+  // Rule C: Laya Downgrade Prevention
+  if (layaTier !== null && layaTier < finalTier) {
+    monotonicSafeguards.push(
+      `Prevented Laya downgrade: Laya suggested Tier ${layaTier}, but deterministic evidence mandated Tier ${finalTier}`,
+    );
+  }
+
+  // Rule D: Browser Intelligence Escalation
+  // If browser validation is required (SMOKE/FUNCTIONAL/HIGH_RISK) but browser is UNAVAILABLE, escalate
+  if (browserSignal.requires_browser_verification && browserSignal.escalate) {
+    escalated = true;
+    escalationReason =
+      browserSignal.safety_notice ||
+      'Browser verification is required but browser runtime is unavailable.';
+    monotonicSafeguards.push(
+      'Escalation triggered: Missing browser execution capability for UI change',
+    );
+  }
+
+  // If browser risk is HIGH_RISK, elevate tier to at least Tier 2
+  if (browserSignal.validation_level === 'HIGH_RISK' && finalTier < 2) {
+    finalTier = 2;
+    monotonicSafeguards.push(
+      'Elevated to Tier 2: High-risk browser verification (payment/auth UI)',
+    );
+  }
+
+  // If Tier 2 and Browser risk is HIGH_RISK + browser unavailable -> Escalate to human gate
+  if (
+    finalTier >= 2 &&
+    browserSignal.validation_level === 'HIGH_RISK' &&
+    browserSignal.browser_available === 'UNAVAILABLE'
+  ) {
+    escalated = true;
+    escalationReason =
+      'High-risk payment/auth browser verification required but browser environment is unavailable.';
+  }
+
+  // 5. Adaptive Reviewer Orchestration
+  let reviewerResult;
+  try {
+    reviewerResult = orchestrateReviewers({
+      tier: finalTier,
+      files,
+      diff,
+      task,
+      evidence: evidenceResult.evidence,
+      previousFindings: options.previousFindings,
+    });
+    capabilitiesConsulted.push('reviewer_orchestrator');
+  } catch (_) {
+    capabilitiesUnavailable.push('reviewer_orchestrator');
+    reviewerResult = {
+      selected: [],
+      rejected: [],
+      total_selected: 0,
+      coverage: { all_risks_covered: false },
+    };
+  }
+
+  // 6. Routing Decision & Socratic Gate
+  const routing = escalated ? 'ESCALATE' : ROUTING_NAMES[finalTier] || 'TARGETED_REVIEW';
+
+  const socraticGate = escalated
+    ? 'required'
+    : finalTier >= 3
+      ? 'required'
+      : finalTier >= 2
+        ? 'conditional'
+        : 'bypass';
+
+  // 7. Explanations & Provenance
+  const influencingEvidence = evidenceResult.evidence
+    .filter(e => e.score >= 0.8 || e.is_mandatory)
+    .map(e => `${e.id} (score: ${e.score})`);
+
+  let whyTierSelected = `Classified as ${TIER_NAMES[finalTier]} (Tier ${finalTier}). `;
+  if (firedRiskSignals.length > 0) {
+    whyTierSelected += `Risk signals fired: [${firedRiskSignals.join(', ')}]. `;
+  } else if (finalTier === 0) {
+    whyTierSelected +=
+      'Change is low-volume styling/documentation with zero risk signals (Fast-Pass eligible). ';
+  } else if (finalTier === 1) {
+    whyTierSelected +=
+      'Single-file or bounded component edit with isolated impact (Express Pass). ';
+  } else if (finalTier === 2) {
+    whyTierSelected += 'Multi-file feature edit requiring targeted domain verification. ';
+  } else {
+    whyTierSelected +=
+      'High-impact architectural or security change requiring full gauntlet review. ';
+  }
+
+  if (escalated) {
+    whyTierSelected += `ESCALATED: ${escalationReason}`;
+  }
+
+  // Overall Decision Confidence
+  let decisionConfidence = 'L1';
+  if (fallbackUsed && evidenceConfidence !== 'L1') {
+    decisionConfidence = 'L2';
+  } else if (evidenceConfidence === 'L3' || evidenceConfidence === 'L4') {
+    decisionConfidence = evidenceConfidence;
+  }
+
+  return {
+    decision_id: `DEC-${Date.now()}-${crypto.randomBytes(4).toString('hex')}`,
+    timestamp: new Date().toISOString(),
+    tier: finalTier,
+    tier_name: TIER_NAMES[finalTier],
+    routing,
+    socratic_gate: socraticGate,
+    escalated,
+    escalation_reason: escalationReason,
+    confidence: decisionConfidence,
+    explanation: {
+      why_tier_selected: whyTierSelected.trim(),
+      influencing_evidence: influencingEvidence,
+      fired_risk_signals: firedRiskSignals,
+      capabilities_consulted: capabilitiesConsulted,
+      capabilities_unavailable: capabilitiesUnavailable,
+      fallback_used: fallbackUsed,
+      monotonic_safeguards_applied: monotonicSafeguards,
+    },
+    signals: {
+      change_lines: lines,
+      change_files_count: files.length,
+      impact_tier_raw: calculatedTier,
+      laya_tier: layaTier,
+      evidence_confidence: evidenceConfidence,
+      browser_validation_level: browserSignal.validation_level,
+      browser_available: browserSignal.browser_available,
+    },
+    evidence_summary: {
+      total_evaluated: evidenceResult.summary?.total_items_evaluated || 0,
+      total_selected: evidenceResult.summary?.total_items_selected || 0,
+      critical_count: evidenceResult.summary?.critical_evidence_count || 0,
+      confidence: evidenceConfidence,
+    },
+    reviewers: {
+      selected: reviewerResult.selected,
+      rejected: reviewerResult.rejected,
+      total_selected: reviewerResult.total_selected,
+      all_risks_covered: reviewerResult.coverage?.all_risks_covered ?? true,
+    },
+  };
+}
+
+module.exports = {
+  evaluateDecision,
+  resolveMonotonicImpactTier,
+  HIGH_RISK_PATTERNS,
+  HIGH_RISK_EXTENSIONS,
+  TIER_NAMES,
+  ROUTING_NAMES,
+};