@sienklogic/plan-build-run 2.11.0 → 2.12.1

package/plugins/pbr/scripts/audit-checks/error-analysis.js

@@ -0,0 +1,895 @@
+'use strict';
+
+/**
+ * Error & Failure Analysis Check Module
+ *
+ * Implements all 7 EF error and failure analysis dimensions
+ * for the PBR audit system. Each check returns a structured result:
+ * { dimension, status, message, evidence }.
+ *
+ * Checks:
+ *   EF-01: Tool failure rate analysis (PostToolUseFailure events by tool type)
+ *   EF-02: Agent failure/timeout detection (missing completion markers, null durations)
+ *   EF-03: Hook false positive detection (blocks on legitimate writes)
+ *   EF-04: Hook false negative detection (bad actions not blocked)
+ *   EF-05: Retry/repetition pattern detection (same tool called 3+ times consecutively)
+ *   EF-06: Cross-session interference detection (overlapping sessions, .active-skill races)
+ *   EF-07: Session cleanup verification (SessionEnd cleanup and stale file removal)
+ *
+ * Config dependencies:
+ *   - config.audit.thresholds.tool_failure_rate_warn (EF-01, default 0.10)
+ *   - config.audit.thresholds.agent_timeout_ms (EF-02, default 600000)
+ *   - config.audit.thresholds.retry_pattern_min_count (EF-05, default 3)
+ */
+
+const fs = require('fs');
+const path = require('path');
+
+// ---------------------------------------------------------------------------
+// Result helper
+// ---------------------------------------------------------------------------
+
+/**
+ * Build a structured check result.
+ * @param {string} dimCode - Dimension code (e.g. "EF-01")
+ * @param {'pass'|'warn'|'fail'} status
+ * @param {string} message
+ * @param {string[]} [evidence]
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function result(dimCode, status, message, evidence) {
+  return {
+    dimension: dimCode,
+    status,
+    message,
+    evidence: evidence || [],
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Shared helpers
+// ---------------------------------------------------------------------------
+
+/**
+ * Get the logs directory path from a planningDir.
+ * @param {string} planningDir - Path to .planning directory
+ * @returns {string}
+ */
+function getLogsDir(planningDir) {
+  return path.join(planningDir, 'logs');
+}
+
+/**
+ * Read all JSONL files matching a prefix from a logs directory.
+ * E.g., readJsonlFiles(logsDir, 'hooks') reads all hooks-*.jsonl files.
+ * @param {string} logsDir - Path to logs directory
+ * @param {string} prefix - File prefix (e.g. 'hooks', 'events')
+ * @returns {object[]} Array of parsed JSON entries
+ */
+function readJsonlFiles(logsDir, prefix) {
+  const entries = [];
+  if (!fs.existsSync(logsDir)) return entries;
+
+  let files;
+  try {
+    files = fs.readdirSync(logsDir);
+  } catch (_e) {
+    return entries;
+  }
+
+  const pattern = new RegExp(`^${prefix}-.*\\.jsonl$`);
+  for (const file of files) {
+    if (!pattern.test(file)) continue;
+    try {
+      const content = fs.readFileSync(path.join(logsDir, file), 'utf8');
+      const lines = content.split('\n');
+      for (const line of lines) {
+        const trimmed = line.trim();
+        if (!trimmed) continue;
+        try {
+          entries.push(JSON.parse(trimmed));
+        } catch (_e) {
+          // Skip malformed lines
+        }
+      }
+    } catch (_e) {
+      // Skip unreadable files
+    }
+  }
+
+  return entries;
+}
+
+// ---------------------------------------------------------------------------
+// Shared: test-source entry filter
+// ---------------------------------------------------------------------------
+
+/**
+ * Return true if the entry appears to originate from a test run.
+ * Test entries have a "cat" field that is a temp directory path
+ * (e.g. C:\Users\...\Temp\... or /tmp/...).
+ * @param {object} entry - Parsed JSON log entry
+ * @returns {boolean}
+ */
+function isTestSourced(entry) {
+  const cat = entry.cat || '';
+  if (typeof cat !== 'string') return false;
+  return /[/\\]temp[/\\]|[/\\]tmp[/\\]/i.test(cat);
+}
+
+// ---------------------------------------------------------------------------
+// EF-01: Tool Failure Rate Analysis
+// ---------------------------------------------------------------------------
+
+/**
+ * Analyze tool failure rates from event logs only (single source of truth).
+ * Groups failures by tool type, calculates rates, caps at 100%.
+ * Filters out test-sourced entries before analysis.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object (may have audit.thresholds.tool_failure_rate_warn)
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkToolFailureRate(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const threshold = config?.audit?.thresholds?.tool_failure_rate_warn ?? 0.10;
+
+  // Use ONLY event logs as the single source of truth (no hook log merging)
+  const eventEntries = readJsonlFiles(logsDir, 'events')
+    .filter(e => !isTestSourced(e));
+
+  // Only count PostToolUseFailure events as real tool failures —
+  // PreToolUse blocks are intentional enforcement, not tool execution failures
+  const eventFailures = eventEntries.filter(
+    e => e.cat === 'tool' && e.event === 'failure' && e.type !== 'PreToolUse'
+  );
+  const totalToolEvents = eventEntries.filter(e => e.cat === 'tool');
+
+  // Count failures per tool
+  const failuresByTool = {};
+  for (const entry of eventFailures) {
+    const tool = entry.tool || entry.data?.tool || 'unknown';
+    failuresByTool[tool] = (failuresByTool[tool] || 0) + 1;
+  }
+
+  // Count total calls per tool
+  const totalByTool = {};
+  for (const entry of totalToolEvents) {
+    const tool = entry.tool || entry.data?.tool || 'unknown';
+    totalByTool[tool] = (totalByTool[tool] || 0) + 1;
+  }
+
+  // Separately tally PreToolUse enforcement blocks (informational only)
+  const hookEntries = readJsonlFiles(logsDir, 'hooks')
+    .filter(e => !isTestSourced(e));
+  const enforcementBlocks = hookEntries.filter(
+    e => (e.action === 'block' || e.decision === 'block')
+  );
+  const enforcementByTool = {};
+  for (const entry of enforcementBlocks) {
+    const tool = entry.details?.tool || entry.data?.tool || entry.tool || 'unknown';
+    enforcementByTool[tool] = (enforcementByTool[tool] || 0) + 1;
+  }
+  const totalEnforcement = Object.values(enforcementByTool).reduce((a, b) => a + b, 0);
+
+  const totalFailures = Object.values(failuresByTool).reduce((a, b) => a + b, 0);
+  if (totalFailures === 0 && totalEnforcement === 0) {
+    return result('EF-01', 'pass', 'No tool failures detected in logs', []);
+  }
+
+  // Build evidence and determine status
+  const evidence = [];
+  let anyAboveThreshold = false;
+
+  for (const [tool, rawCount] of Object.entries(failuresByTool)) {
+    const total = totalByTool[tool];
+    if (total && total > 0) {
+      // Cap failure count at total to prevent >100% rates
+      const count = Math.min(rawCount, total);
+      const rate = count / total;
+      const pct = (rate * 100).toFixed(1);
+      evidence.push(`${tool}: ${count} failures (${pct}% rate of ${total} calls)`);
+      if (rate > threshold) anyAboveThreshold = true;
+    } else {
+      evidence.push(`${tool}: ${rawCount} failures (total calls unknown)`);
+    }
+  }
+
+  // Report enforcement blocks as informational (not counted toward failure rate)
+  if (totalEnforcement > 0) {
+    const blockDetails = Object.entries(enforcementByTool)
+      .map(([tool, count]) => `${tool}: ${count}`)
+      .join(', ');
+    evidence.push(`Enforcement blocks (informational, not failures): ${blockDetails}`);
+  }
+
+  if (totalFailures === 0) {
+    return result('EF-01', 'pass',
+      `No real tool failures (${totalEnforcement} enforcement block(s) excluded)`,
+      evidence);
+  }
+
+  const status = anyAboveThreshold ? 'fail' : 'warn';
+  const message = anyAboveThreshold
+    ? `Tool failure rate exceeds ${(threshold * 100).toFixed(0)}% threshold for some tools`
+    : `${totalFailures} tool failure(s) detected but within threshold`;
+
+  return result('EF-01', status, message, evidence);
+}
+
+// ---------------------------------------------------------------------------
+// EF-02: Agent Failure/Timeout Detection
+// ---------------------------------------------------------------------------
+
+/** Known completion markers that agents should output. */
+const COMPLETION_MARKERS = [
+  '## EXECUTION COMPLETE',
+  '## PLANNING COMPLETE',
+  '## VERIFICATION COMPLETE',
+  '## PLAN COMPLETE',
+  '## PLAN FAILED',
+  '## CHECKPOINT:',
+];
+
+/**
+ * Detect agents that failed, timed out, or are missing completion markers.
+ *
+ * Examines event logs for agent spawn/complete lifecycle events, and hook logs
+ * for check-subagent-output warnings about missing artifacts.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object (may have audit.thresholds.agent_timeout_ms)
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkAgentFailureTimeout(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const timeoutMs = config?.audit?.thresholds?.agent_timeout_ms ?? 600000;
+
+  const eventEntries = readJsonlFiles(logsDir, 'events');
+  const hookEntries = readJsonlFiles(logsDir, 'hooks');
+
+  // Collect agent spawn and complete events from event logs
+  const spawns = eventEntries.filter(e => e.cat === 'agent' && e.event === 'spawn');
+  const completions = eventEntries.filter(e => e.cat === 'agent' && e.event === 'complete');
+
+  // Also collect from hook logs (log-subagent entries)
+  const hookSpawns = hookEntries.filter(e => e.hook === 'log-subagent' && e.action === 'spawned');
+  const hookCompletions = hookEntries.filter(e => e.hook === 'log-subagent' && e.action === 'completed');
+
+  // Build a map of agent_id -> spawn info
+  const agentMap = new Map();
+  for (const s of [...spawns, ...hookSpawns]) {
+    const id = s.agent_id || s.details?.agent_id || null;
+    if (!id) continue;
+    if (agentMap.has(id)) continue; // first seen wins
+    agentMap.set(id, {
+      type: s.agent_type || s.details?.agent_type || 'unknown',
+      ts: s.ts || null,
+      completed: false,
+      duration_ms: null,
+    });
+  }
+
+  // Mark completions
+  for (const c of [...completions, ...hookCompletions]) {
+    const id = c.agent_id || c.details?.agent_id || null;
+    if (!id || !agentMap.has(id)) continue;
+    const agent = agentMap.get(id);
+    agent.completed = true;
+    agent.duration_ms = c.duration_ms || c.details?.duration_ms || null;
+  }
+
+  // Check for check-subagent-output warnings (agents that completed without expected artifacts)
+  const subagentWarnings = hookEntries.filter(
+    e => e.hook === 'check-subagent-output' && e.action === 'warned'
+  );
+
+  const evidence = [];
+  let failCount = 0;
+  let warnCount = 0;
+
+  // Check each agent for failure conditions
+  for (const [id, agent] of agentMap) {
+    const label = `${agent.type} agent (${id})`;
+    const timeStr = agent.ts ? ` (started ${agent.ts.substring(11, 16)})` : '';
+
+    if (!agent.completed) {
+      evidence.push(`${label}: no completion event found${timeStr}`);
+      failCount++;
+    } else if (agent.duration_ms === null || agent.duration_ms === undefined) {
+      evidence.push(`${label}: completed with null duration${timeStr}`);
+      warnCount++;
+    } else if (agent.duration_ms > timeoutMs) {
+      const mins = (agent.duration_ms / 60000).toFixed(1);
+      evidence.push(`${label}: duration ${mins}min exceeds ${(timeoutMs / 60000).toFixed(0)}min timeout${timeStr}`);
+      failCount++;
+    }
+  }
+
+  // Add warnings from check-subagent-output
+  for (const w of subagentWarnings) {
+    const desc = w.details?.description || w.details?.message || 'missing expected artifact';
+    const agentType = w.details?.agent_type || 'unknown';
+    evidence.push(`${agentType} agent: ${desc}`);
+    warnCount++;
+  }
+
+  if (failCount === 0 && warnCount === 0) {
+    return result('EF-02', 'pass', 'All agents completed successfully', []);
+  }
+
+  const status = failCount > 0 ? 'fail' : 'warn';
+  const message = failCount > 0
+    ? `${failCount} agent(s) failed or timed out`
+    : `${warnCount} agent warning(s) detected but all completed`;
+
+  return result('EF-02', status, message, evidence);
+}
+
+// ---------------------------------------------------------------------------
+// EF-05: Retry/Repetition Pattern Detection
+// ---------------------------------------------------------------------------
+
+/**
+ * Detect consecutive same-tool call patterns that suggest retries or repetition.
+ *
+ * Scans event logs for tool-use events sorted by timestamp. A "repetition" is
+ * when the same tool appears N+ times consecutively. If failures occurred in
+ * the sequence, it's classified as a "retry" pattern.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object (may have audit.thresholds.retry_pattern_min_count)
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkRetryRepetitionPattern(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const minCount = config?.audit?.thresholds?.retry_pattern_min_count ?? 3;
+
+  const eventEntries = readJsonlFiles(logsDir, 'events')
+    .filter(e => !isTestSourced(e));
+
+  // Filter for tool-use events and sort by timestamp
+  const toolEvents = eventEntries
+    .filter(e => e.cat === 'tool')
+    .sort((a, b) => (a.ts || '').localeCompare(b.ts || ''));
+
+  if (toolEvents.length === 0) {
+    return result('EF-05', 'pass', 'No tool events found in logs', []);
+  }
+
+  // Scan for consecutive runs of the same tool
+  const patterns = [];
+  let currentTool = null;
+  let runStart = 0;
+  let runCount = 0;
+  let runFailures = 0;
+
+  for (let i = 0; i <= toolEvents.length; i++) {
+    const entry = i < toolEvents.length ? toolEvents[i] : null;
+    const tool = entry ? (entry.tool || entry.data?.tool || 'unknown') : null;
+
+    if (tool === currentTool && i < toolEvents.length) {
+      runCount++;
+      if (entry.event === 'failure') runFailures++;
+    } else {
+      // End of a run — check if it meets the threshold
+      if (currentTool && runCount >= minCount) {
+        const startTs = toolEvents[runStart]?.ts || '';
+        const endTs = toolEvents[Math.min(runStart + runCount - 1, toolEvents.length - 1)]?.ts || '';
+        const startTime = startTs.substring(11, 16) || '??:??';
+        const endTime = endTs.substring(11, 16) || '??:??';
+        const patternType = runFailures > 0 ? 'retry' : 'repetition';
+
+        patterns.push({
+          tool: currentTool,
+          count: runCount,
+          failures: runFailures,
+          type: patternType,
+          timeWindow: `${startTime}-${endTime}`,
+        });
+      }
+
+      // Start new run
+      currentTool = tool;
+      runStart = i;
+      runCount = 1;
+      runFailures = (entry && entry.event === 'failure') ? 1 : 0;
+    }
+  }
+
+  if (patterns.length === 0) {
+    return result('EF-05', 'pass', `No consecutive same-tool patterns (${minCount}+) detected`, []);
+  }
+
+  const evidence = patterns.map(p => {
+    const failStr = p.failures > 0 ? ` (${p.failures} failures)` : '';
+    return `${p.tool} called ${p.count} times consecutively${failStr} -- ${p.type} pattern at ${p.timeWindow}`;
+  });
+
+  const hasRetries = patterns.some(p => p.type === 'retry');
+  const status = hasRetries ? 'warn' : 'pass';
+  const message = hasRetries
+    ? `${patterns.filter(p => p.type === 'retry').length} retry pattern(s) detected`
+    : `${patterns.length} repetition pattern(s) detected (no failures, likely normal batch work)`;
+
+  return result('EF-05', status, message, evidence);
+}
+
+// ---------------------------------------------------------------------------
+// EF-03: Hook False Positive Detection
+// ---------------------------------------------------------------------------
+
+/**
+ * Analyze PreToolUse block decisions for potential false positives.
+ *
+ * A false positive is a block that was later overridden (same file allowed
+ * afterward), or a block on a file that belongs to the plan's files_modified
+ * list, or a block from check-skill-workflow / check-doc-sprawl on a
+ * legitimate PBR artifact path.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkHookFalsePositive(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const hookEntries = readJsonlFiles(logsDir, 'hooks');
+
+  // Collect block and allow decisions from PreToolUse hooks
+  const blocks = [];
+  const allows = [];
+
+  for (const entry of hookEntries) {
+    const action = entry.action || entry.decision || '';
+    const file = entry.details?.file || entry.data?.file || entry.file || '';
+    const hook = entry.hook || '';
+    const reason = entry.reason || entry.details?.reason || '';
+    const ts = entry.ts || '';
+
+    if (action === 'block') {
+      blocks.push({ hook, file, reason, ts });
+    } else if (action === 'allow') {
+      allows.push({ file, ts });
+    }
+  }
+
+  if (blocks.length === 0) {
+    return result('EF-03', 'pass', 'No hook blocks found in logs', []);
+  }
+
+  // Build a set of allowed files for quick lookup
+  const allowedFiles = new Map();
+  for (const a of allows) {
+    if (!a.file) continue;
+    const existing = allowedFiles.get(a.file);
+    if (!existing || a.ts > existing) {
+      allowedFiles.set(a.file, a.ts);
+    }
+  }
+
+  // Known PBR artifact path patterns that are legitimate write targets
+  const pbrArtifactPatterns = [
+    /\.planning\//,
+    /PLAN.*\.md$/i,
+    /SUMMARY.*\.md$/i,
+    /VERIFICATION.*\.md$/i,
+    /STATE\.md$/i,
+    /ROADMAP\.md$/i,
+    /CONTEXT\.md$/i,
+  ];
+
+  const evidence = [];
+
+  for (const block of blocks) {
+    const reasons = [];
+
+    // Check if same file was allowed later (block was retried and succeeded)
+    if (block.file && allowedFiles.has(block.file)) {
+      const allowTs = allowedFiles.get(block.file);
+      if (allowTs > block.ts) {
+        reasons.push('allowed later');
+      }
+    }
+
+    // Check if block was from check-skill-workflow or check-doc-sprawl on a PBR artifact
+    const isFalsePositiveHook = block.hook === 'check-skill-workflow' ||
+      block.hook === 'check-doc-sprawl' ||
+      block.hook === 'pre-write-dispatch';
+    if (isFalsePositiveHook && block.file) {
+      const normalized = block.file.replace(/\\/g, '/');
+      const isPbrArtifact = pbrArtifactPatterns.some(p => p.test(normalized));
+      if (isPbrArtifact) {
+        reasons.push('block on legitimate PBR artifact path');
+      }
+    }
+
+    if (reasons.length > 0) {
+      const fileStr = block.file ? ` on ${block.file}` : '';
+      const timeStr = block.ts ? ` at ${block.ts.substring(11, 16)}` : '';
+      evidence.push(
+        `${block.hook} blocked${fileStr}${timeStr} — possible false positive (${reasons.join(', ')})`
+      );
+    }
+  }
+
+  if (evidence.length === 0) {
+    return result('EF-03', 'pass', `${blocks.length} block(s) reviewed, no false positives detected`, []);
+  }
+
+  return result('EF-03', 'warn',
+    `${evidence.length} potential false positive(s) in ${blocks.length} total blocks`,
+    evidence
+  );
+}
+
+// ---------------------------------------------------------------------------
+// EF-04: Hook False Negative Detection
+// ---------------------------------------------------------------------------
+
+/**
+ * Detect bad actions that should have been blocked but weren't.
+ *
+ * Scans for: non-conventional commits that passed validation, writes to files
+ * outside files_modified, and destructive git commands that were allowed.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkHookFalseNegative(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const hookEntries = readJsonlFiles(logsDir, 'hooks');
+  const eventEntries = readJsonlFiles(logsDir, 'events');
+
+  const evidence = [];
+
+  // 1. Check for commits with non-conventional format that passed validation
+  const conventionalPattern = /^(feat|fix|refactor|test|docs|chore|wip|revert|perf|ci|build)\(.+\):\s.+/;
+  const commitEvents = eventEntries.filter(
+    e => (e.event === 'commit-validated' && e.status === 'allow') ||
+         (e.cat === 'tool' && e.tool === 'Bash' && e.data?.command && /git commit/.test(e.data.command))
+  );
+
+  for (const ce of commitEvents) {
+    const msg = ce.data?.message || ce.details?.message || '';
+    if (msg && !conventionalPattern.test(msg)) {
+      evidence.push(
+        `Non-conventional commit passed validation: "${msg.substring(0, 60)}"`
+      );
+    }
+  }
+
+  // 2. Check for destructive git commands that were allowed
+  const destructivePatterns = [
+    /git\s+push\s+--force/,
+    /git\s+push\s+-f\b/,
+    /git\s+reset\s+--hard/,
+    /git\s+checkout\s+--\s/,
+    /git\s+clean\s+-f/,
+    /git\s+branch\s+-D/,
+  ];
+
+  const bashEvents = hookEntries.filter(
+    e => (e.hook === 'pre-bash-dispatch' || e.hook === 'check-dangerous-commands') &&
+         (e.action === 'allow' || e.decision === 'allow')
+  );
+
+  for (const be of bashEvents) {
+    const cmd = be.details?.cmd || be.data?.cmd || be.data?.command || '';
+    for (const dp of destructivePatterns) {
+      if (dp.test(cmd)) {
+        const timeStr = be.ts ? ` at ${be.ts.substring(11, 16)}` : '';
+        evidence.push(
+          `Destructive command allowed${timeStr}: "${cmd.substring(0, 80)}"`
+        );
+        break;
+      }
+    }
+  }
+
+  // 3. Check for writes outside files_modified that were allowed
+  // Look for allow decisions on writes, cross-reference with plan frontmatter if available
+  const writeAllows = hookEntries.filter(
+    e => e.hook === 'pre-write-dispatch' && (e.action === 'allow' || e.decision === 'allow')
+  );
+
+  // Try to read current plan's files_modified from any available PLAN file
+  let filesModified = null;
+  try {
+    const phasesDir = path.join(planningDir, 'phases');
+    if (fs.existsSync(phasesDir)) {
+      const phaseDirs = fs.readdirSync(phasesDir);
+      for (const pd of phaseDirs) {
+        const planFiles = fs.readdirSync(path.join(phasesDir, pd)).filter(f => /^PLAN.*\.md$/i.test(f));
+        for (const pf of planFiles) {
+          try {
+            const content = fs.readFileSync(path.join(phasesDir, pd, pf), 'utf8');
+            const fmMatch = content.match(/files_modified:\s*\n((?:\s+-\s+"[^"]+"\n?)+)/);
+            if (fmMatch) {
+              const files = fmMatch[1].match(/"([^"]+)"/g);
+              if (files) {
+                filesModified = files.map(f => f.replace(/"/g, ''));
+              }
+            }
+          } catch (_e) { /* skip unreadable */ }
+        }
+      }
+    }
+  } catch (_e) { /* no plan files available */ }
+
+  if (filesModified && filesModified.length > 0) {
+    for (const wa of writeAllows) {
+      const file = wa.details?.file || wa.data?.file || '';
+      if (!file) continue;
+      const normalized = file.replace(/\\/g, '/');
+      // Skip .planning writes — those are always allowed
+      if (normalized.includes('.planning/')) continue;
+      const isInPlan = filesModified.some(fm => normalized.endsWith(fm) || normalized.includes(fm));
+      if (!isInPlan) {
+        evidence.push(
+          `Write allowed outside files_modified: "${normalized}"`
+        );
+      }
+    }
+  }
+
+  if (evidence.length === 0) {
+    return result('EF-04', 'pass', 'No false negatives detected in hook enforcement', []);
+  }
+
+  return result('EF-04', 'warn',
+    `${evidence.length} potential false negative(s) in hook enforcement`,
+    evidence
+  );
+}
+
+// ---------------------------------------------------------------------------
+// EF-06: Cross-Session Interference Detection
+// ---------------------------------------------------------------------------
+
+/**
+ * Detect cross-session interference: overlapping sessions, .active-skill
+ * claim races, and stale lock files.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkCrossSessionInterference(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const hookEntries = readJsonlFiles(logsDir, 'hooks');
+  const eventEntries = readJsonlFiles(logsDir, 'events');
+
+  const evidence = [];
+  let hasFail = false;
+
+  // 1. Extract session IDs and detect overlapping sessions
+  const sessionTimestamps = new Map(); // sessionId -> [timestamps]
+
+  for (const entry of [...hookEntries, ...eventEntries]) {
+    const sessionId = entry.session_id || entry.sessionId || entry.details?.session_id || null;
+    const ts = entry.ts || '';
+    if (!sessionId || !ts) continue;
+
+    if (!sessionTimestamps.has(sessionId)) {
+      sessionTimestamps.set(sessionId, []);
+    }
+    sessionTimestamps.get(sessionId).push(ts);
+  }
+
+  // Check for overlapping sessions: two sessions with interleaved timestamps
+  const sessions = Array.from(sessionTimestamps.entries());
+  for (let i = 0; i < sessions.length; i++) {
+    for (let j = i + 1; j < sessions.length; j++) {
+      const [idA, tsA] = sessions[i];
+      const [idB, tsB] = sessions[j];
+      if (tsA.length === 0 || tsB.length === 0) continue;
+
+      const minA = tsA.reduce((a, b) => a < b ? a : b);
+      const maxA = tsA.reduce((a, b) => a > b ? a : b);
+      const minB = tsB.reduce((a, b) => a < b ? a : b);
+      const maxB = tsB.reduce((a, b) => a > b ? a : b);
+
+      // Overlap: session A's range intersects session B's range
+      if (minA <= maxB && minB <= maxA) {
+        const overlapStart = minA > minB ? minA : minB;
+        evidence.push(
+          `Two sessions active simultaneously at ${overlapStart.substring(11, 16)} (session ${idA.substring(0, 8)} and ${idB.substring(0, 8)})`
+        );
+        hasFail = true;
+      }
+    }
+  }
+
+  // 2. Look for .active-skill conflict entries in hook logs
+  const activeSkillConflicts = hookEntries.filter(e => {
+    const details = JSON.stringify(e.details || e.data || {}).toLowerCase();
+    return details.includes('active-skill') || details.includes('active_skill');
+  }).filter(e => {
+    const action = e.action || e.decision || '';
+    return action === 'block' || action === 'warn' || action === 'warned';
+  });
+
+  for (const conflict of activeSkillConflicts) {
+    const hook = conflict.hook || 'unknown';
+    const action = conflict.action || conflict.decision || '';
+    const ts = conflict.ts ? ` at ${conflict.ts.substring(11, 16)}` : '';
+    evidence.push(
+      `${hook} ${action} on active-skill conflict${ts}`
+    );
+    if (action === 'block') hasFail = true;
+  }
+
+  // 3. Check if stale session files exist (leftover from crashed session)
+  const staleSessionFiles = [
+    { file: '.active-skill', desc: '.active-skill' },
+    { file: '.session.json', desc: '.session.json (consolidated session state)' },
+  ];
+
+  for (const sf of staleSessionFiles) {
+    const filePath = path.join(planningDir, sf.file);
+    try {
+      if (fs.existsSync(filePath)) {
+        const stat = fs.statSync(filePath);
+        const ageMs = Date.now() - stat.mtimeMs;
+        const ageMin = Math.round(ageMs / 60000);
+        evidence.push(
+          `${sf.desc} exists (${ageMin}min old) — may be stale from crashed session`
+        );
+      }
+    } catch (_e) { /* best-effort */ }
+  }
+
+  if (evidence.length === 0) {
+    return result('EF-06', 'pass', 'No cross-session interference detected', []);
+  }
+
+  const status = hasFail ? 'fail' : 'warn';
+  const message = hasFail
+    ? `Cross-session interference detected: ${evidence.length} issue(s)`
+    : `${evidence.length} potential cross-session issue(s) found`;
+
+  return result('EF-06', status, message, evidence);
+}
+
+// ---------------------------------------------------------------------------
+// EF-07: Session Cleanup Verification
+// ---------------------------------------------------------------------------
+
+/**
+ * Verify that session-cleanup.js fires at SessionEnd and removes stale files.
+ *
+ * Checks for: cleanup log entries matching SessionEnd events, orphaned
+ * artifacts (.active-skill, .active-operation, .auto-next, .context-tracker,
+ * stale .checkpoint-manifest.json), and old hook log files.
+ *
+ * @param {string} planningDir - Path to .planning directory
+ * @param {object} config - Config object
+ * @returns {{ dimension: string, status: string, message: string, evidence: string[] }}
+ */
+function checkSessionCleanupVerification(planningDir, config) {
+  const logsDir = getLogsDir(planningDir);
+  const hookEntries = readJsonlFiles(logsDir, 'hooks');
+  const eventEntries = readJsonlFiles(logsDir, 'events');
+
+  const evidence = [];
+
+  // 1. Find SessionEnd events and verify matching cleanup entries
+  const sessionEndEvents = eventEntries.filter(
+    e => e.event === 'SessionEnd' || (e.cat === 'lifecycle' && e.event === 'session-end')
+  );
+
+  const cleanupEntries = hookEntries.filter(
+    e => e.hook === 'session-cleanup' && (e.action === 'cleaned' || e.decision === 'cleaned' || e.action === 'nothing')
+  );
+
+  // For each SessionEnd, check if there's a cleanup entry near the same time
+  for (const se of sessionEndEvents) {
+    const seTs = se.ts || '';
+    if (!seTs) continue;
+
+    const hasMatchingCleanup = cleanupEntries.some(ce => {
+      const ceTs = ce.ts || '';
+      if (!ceTs) return false;
+      // Within 60 seconds of each other
+      try {
+        const diff = Math.abs(new Date(ceTs).getTime() - new Date(seTs).getTime());
+        return diff < 60000;
+      } catch (_e) {
+        return false;
+      }
+    });
+
+    if (!hasMatchingCleanup) {
+      evidence.push(
+        `SessionEnd at ${seTs.substring(11, 16)} has no matching cleanup entry`
+      );
+    }
+  }
+
+  // 2. Check for orphaned artifacts that cleanup should have removed
+  const orphanChecks = [
+    { file: '.active-skill', desc: '.active-skill (should not exist between sessions)' },
+    { file: '.active-operation', desc: '.active-operation (should not exist between sessions)' },
+    { file: '.active-plan', desc: '.active-plan (should not exist between sessions)' },
+    { file: '.session.json', desc: '.session.json (consolidated session state)' },
+    { file: '.auto-next', desc: '.auto-next (transient signal file)' },
+    { file: '.context-tracker', desc: '.context-tracker (session-scoped)' },
+  ];
+
+  for (const check of orphanChecks) {
+    const filePath = path.join(planningDir, check.file);
+    try {
+      if (fs.existsSync(filePath)) {
+        evidence.push(`${check.desc} exists — cleanup may have missed it`);
+      }
+    } catch (_e) { /* best-effort */ }
+  }
+
+  // 3. Check for stale .checkpoint-manifest.json (>24 hours old)
+  const STALE_MS = 24 * 60 * 60 * 1000;
+  try {
+    const phasesDir = path.join(planningDir, 'phases');
+    if (fs.existsSync(phasesDir)) {
+      const dirs = fs.readdirSync(phasesDir);
+      for (const dir of dirs) {
+        const manifestPath = path.join(phasesDir, dir, '.checkpoint-manifest.json');
+        try {
+          if (fs.existsSync(manifestPath)) {
+            const stat = fs.statSync(manifestPath);
+            const ageMs = Date.now() - stat.mtimeMs;
+            if (ageMs > STALE_MS) {
+              const ageHrs = Math.round(ageMs / 3600000);
+              evidence.push(
+                `.checkpoint-manifest.json in ${dir} is ${ageHrs}h old — cleanup should have removed it`
+              );
+            }
+          }
+        } catch (_e) { /* skip */ }
+      }
+    }
+  } catch (_e) { /* best-effort */ }
+
+  // 4. Check hook log rotation — if hooks-*.jsonl has files older than 30 days
+  const THIRTY_DAYS_MS = 30 * 24 * 60 * 60 * 1000;
+  try {
+    if (fs.existsSync(logsDir)) {
+      const logFiles = fs.readdirSync(logsDir).filter(f => /^hooks-.*\.jsonl/.test(f));
+      for (const lf of logFiles) {
+        const stat = fs.statSync(path.join(logsDir, lf));
+        const ageMs = Date.now() - stat.mtimeMs;
+        if (ageMs > THIRTY_DAYS_MS) {
+          const ageDays = Math.round(ageMs / (24 * 60 * 60 * 1000));
+          evidence.push(
+            `${lf} is ${ageDays} days old — log rotation may not be running`
+          );
+        }
+      }
+    }
+  } catch (_e) { /* best-effort */ }
+
+  if (evidence.length === 0) {
+    const cleanupCount = cleanupEntries.length;
+    return result('EF-07', 'pass',
+      `Session cleanup verified (${cleanupCount} cleanup event(s) found)`, []);
+  }
+
+  return result('EF-07', 'warn',
+    `${evidence.length} session cleanup gap(s) found`,
+    evidence
+  );
+}
+
+// ---------------------------------------------------------------------------
+// Exports
+// ---------------------------------------------------------------------------
+
+module.exports = {
+  checkToolFailureRate,
+  checkAgentFailureTimeout,
+  checkRetryRepetitionPattern,
+  checkHookFalsePositive,
+  checkHookFalseNegative,
+  checkCrossSessionInterference,
+  checkSessionCleanupVerification,
+};