@clear-capabilities/agentic-security-scanner 0.78.0 → 0.80.0

package/src/posture/realtime-cve-monitor.js

@@ -0,0 +1,214 @@
+// Real-time emergent vulnerability monitor — Recommendation #8 of the
+// world-class+2 plan.
+//
+// Polls / streams the OSV.dev feed (and optionally GHSA + vendor
+// advisories) and within minutes of a new CVE matching the customer's
+// SBOM, delivers a push notification via Slack / Discord / generic
+// webhook.
+//
+// Design:
+//   1. Pre-index the project's SBOM by (ecosystem, package) so a new
+//      CVE checks against the index in O(1)
+//   2. Poll OSV's published-vulns endpoint every POLL_INTERVAL_MS
+//      (default 60s)
+//   3. On match: emit a structured alert + write to a state file so
+//      duplicate alerts aren't sent
+//   4. Push delivery via existing webhook utilities; configurable
+//      destinations: AGENTIC_SECURITY_ALERT_WEBHOOK, _SLACK, _DISCORD
+//
+// Out of scope: persistent daemon. This module exposes a `cycle()`
+// function the caller invokes from a cron-shape loop (the existing
+// commands/cve-alerts.md skill is the user-facing wrapper).
+
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import { statePath, safeWriteState } from './state-dir.js';
+
+const STATE_FILE = 'cve-monitor-state.json';
+const DEFAULT_POLL_MS = 60_000;
+const DEFAULT_WINDOW_HOURS = 24;
+const SLA_MINUTES = 5;
+
+const OSV_FEED_URL = 'https://api.osv.dev/v1/vulns/recent';
+
+function _statePath(scanRoot) { return statePath(scanRoot, STATE_FILE); }
+
+function _loadState(scanRoot) {
+  const fp = _statePath(scanRoot);
+  if (!fs.existsSync(fp)) return { alertedIds: [], lastPolledAt: null };
+  try { return JSON.parse(fs.readFileSync(fp, 'utf8')); }
+  catch { return { alertedIds: [], lastPolledAt: null }; }
+}
+function _saveState(scanRoot, state) {
+  safeWriteState(_statePath(scanRoot), JSON.stringify(state, null, 2));
+}
+
+/**
+ * Build a name→component index from the project's most recent SBOM.
+ * Accepts the SBOM as parsed from the engine output's `components`
+ * array (each entry: { ecosystem, name, version, … }).
+ */
+export function indexSbom(components) {
+  const index = new Map();   // key: `${ecosystem}:${name}` → component
+  for (const c of components || []) {
+    if (!c || !c.ecosystem || !c.name) continue;
+    const key = `${String(c.ecosystem).toLowerCase()}:${String(c.name).toLowerCase()}`;
+    if (!index.has(key)) index.set(key, []);
+    index.get(key).push(c);
+  }
+  return index;
+}
+
+/**
+ * Pull the recent-published-vulnerabilities feed from OSV. Returns an
+ * array of OSV vulnerability records.
+ */
+export async function pollOsv(opts = {}) {
+  if (process.env.AGENTIC_SECURITY_OFFLINE === '1') return [];
+  const url = opts.feedUrl || OSV_FEED_URL;
+  const since = opts.sinceIso || new Date(Date.now() - DEFAULT_WINDOW_HOURS * 3600_000).toISOString();
+  try {
+    const u = new URL(url);
+    u.searchParams.set('time', since);
+    const res = await fetch(u.toString(), {
+      headers: { 'User-Agent': 'agentic-security/0.1' },
+    });
+    if (!res.ok) return [];
+    const body = await res.json();
+    return Array.isArray(body.vulns) ? body.vulns : (Array.isArray(body) ? body : []);
+  } catch { return []; }
+}
+
+/**
+ * Match a single OSV record against the SBOM index. Returns an array of
+ * (component, vuln) tuples that match.
+ */
+export function matchOsvToSbom(vuln, sbomIndex) {
+  const out = [];
+  for (const aff of (vuln.affected || [])) {
+    const eco = (aff.package?.ecosystem || '').toLowerCase();
+    const name = (aff.package?.name || '').toLowerCase();
+    if (!eco || !name) continue;
+    const key = `${_normalizeEco(eco)}:${name}`;
+    const matches = sbomIndex.get(key);
+    if (!matches) continue;
+    for (const comp of matches) {
+      // Best-effort version-range check — many advisories don't carry
+      // semver ranges in OSV affected objects; we conservatively
+      // include all matches and let the customer triage.
+      out.push({ component: comp, vuln });
+    }
+  }
+  return out;
+}
+
+function _normalizeEco(eco) {
+  const e = eco.toLowerCase();
+  if (e === 'npm' || e === 'node') return 'npm';
+  if (e === 'pypi' || e === 'pip' || e === 'python') return 'pypi';
+  if (e === 'maven') return 'maven';
+  if (e === 'gem' || e === 'rubygems') return 'rubygems';
+  if (e === 'cargo' || e === 'crates.io') return 'cargo';
+  if (e === 'go' || e === 'golang') return 'golang';
+  if (e === 'packagist' || e === 'composer') return 'packagist';
+  if (e === 'pub' || e === 'dart') return 'pub';
+  return e;
+}
+
+/**
+ * Compose the alert payload for delivery.
+ */
+function _composeAlert(matchTuple, ts = new Date().toISOString()) {
+  const { component, vuln } = matchTuple;
+  return {
+    schema: 'agentic-security/cve-alert/v1',
+    ts,
+    cve: (vuln.aliases || []).find(a => /^CVE-/.test(a)) || null,
+    osvId: vuln.id,
+    summary: vuln.summary || vuln.details?.slice(0, 200) || '(no summary)',
+    component: {
+      ecosystem: component.ecosystem,
+      name: component.name,
+      version: component.version,
+    },
+    references: (vuln.references || []).slice(0, 4).map(r => r.url),
+    severity: _inferSeverity(vuln),
+    sla: { promised: `${SLA_MINUTES}m`, ts },
+  };
+}
+
+function _inferSeverity(vuln) {
+  if (Array.isArray(vuln.severity) && vuln.severity.length) {
+    const cvss = vuln.severity.find(s => /^CVSS_V/.test(s.type || ''));
+    if (cvss && cvss.score) return cvss.score;
+  }
+  if (vuln.database_specific?.severity) return String(vuln.database_specific.severity);
+  return 'unknown';
+}
+
+/**
+ * Deliver an alert via every configured channel. Webhook URLs come from
+ * env: AGENTIC_SECURITY_ALERT_WEBHOOK / _SLACK_WEBHOOK / _DISCORD_WEBHOOK.
+ */
+export async function deliverAlert(alert) {
+  const sinks = {
+    generic: process.env.AGENTIC_SECURITY_ALERT_WEBHOOK,
+    slack:   process.env.AGENTIC_SECURITY_SLACK_WEBHOOK,
+    discord: process.env.AGENTIC_SECURITY_DISCORD_WEBHOOK,
+  };
+  const results = {};
+  for (const [name, url] of Object.entries(sinks)) {
+    if (!url) continue;
+    try {
+      const body = name === 'slack'
+        ? { text: `*New CVE affecting your stack*\n• ${alert.cve || alert.osvId} — ${alert.severity}\n• Package: \`${alert.component.ecosystem}:${alert.component.name}@${alert.component.version}\`\n• ${alert.summary}` }
+        : name === 'discord'
+          ? { content: `**New CVE affecting your stack**\n• ${alert.cve || alert.osvId} — ${alert.severity}\n• \`${alert.component.ecosystem}:${alert.component.name}@${alert.component.version}\`\n• ${alert.summary}` }
+          : alert;
+      const res = await fetch(url, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json', 'User-Agent': 'agentic-security/0.1' },
+        body: JSON.stringify(body),
+      });
+      results[name] = { ok: res.ok, status: res.status };
+    } catch (e) {
+      results[name] = { ok: false, error: String(e && e.message || e) };
+    }
+  }
+  return results;
+}
+
+/**
+ * One polling cycle. Caller invokes from a daemon / cron / loop.
+ * - Fetches recent OSV records
+ * - Matches against the SBOM index
+ * - Deduplicates against prior alerts (by OSV id)
+ * - Delivers alerts
+ * - Persists state
+ */
+export async function cycle(scanRoot, sbomComponents, opts = {}) {
+  const state = _loadState(scanRoot);
+  const seen = new Set(state.alertedIds || []);
+  const index = indexSbom(sbomComponents);
+  const recent = await pollOsv({ sinceIso: state.lastPolledAt, ...opts });
+  const matches = [];
+  for (const v of recent) {
+    if (seen.has(v.id)) continue;
+    const m = matchOsvToSbom(v, index);
+    if (m.length === 0) continue;
+    matches.push(...m);
+  }
+  const alerts = matches.map(m => _composeAlert(m));
+  const deliveries = [];
+  for (const alert of alerts) {
+    const d = await deliverAlert(alert);
+    deliveries.push({ alert, delivery: d });
+    seen.add(alert.osvId);
+  }
+  state.alertedIds = [...seen];
+  state.lastPolledAt = new Date().toISOString();
+  _saveState(scanRoot, state);
+  return { recentCount: recent.length, matches: matches.length, alerts: alerts.length, deliveries };
+}
+
+export const _internals = { _normalizeEco, _composeAlert, _inferSeverity, OSV_FEED_URL, DEFAULT_POLL_MS, SLA_MINUTES };

package/src/posture/rule-overrides.js

@@ -11,11 +11,10 @@ import * as fs from 'node:fs';
 import * as path from 'node:path';
 import * as yaml from 'js-yaml';
 import { verifyLastScan } from './integrity.js';
-
-const OVERRIDES_PATH = '.agentic-security/rules.yml';
+import { statePath } from './state-dir.js';
 
 function _path(scanRoot) {
-  return path.join(scanRoot || process.cwd(), OVERRIDES_PATH);
+  return statePath(scanRoot, 'rules.yml');
 }
 
 export function loadOverrides(scanRoot) {

package/src/posture/rule-pack-signing.js

@@ -30,8 +30,7 @@
 import * as fs from 'node:fs';
 import * as path from 'node:path';
 import * as crypto from 'node:crypto';
-
-const TRUSTED_KEYS_FILE = '.agentic-security/trusted-keys.json';
+import { statePath } from './state-dir.js';
 
 // Built-in trust root. These are the keys the maintainers of agentic-security
 // use to sign official rule packs. Production deployment requires the
@@ -49,7 +48,7 @@ export const BUNDLED_OFFICIAL_KEYS = [
 ];
 
 function _trustedKeysPath(scanRoot) {
-  return path.join(scanRoot || process.cwd(), TRUSTED_KEYS_FILE);
+  return statePath(scanRoot, 'trusted-keys.json');
 }
 
 // Load the EFFECTIVE trusted-key set. Composition:

package/src/posture/rule-synthesis.js

@@ -13,14 +13,12 @@
 
 import * as fs from 'node:fs';
 import * as path from 'node:path';
-
-const TRIAGE_PATH = path.join('.agentic-security', 'triage-feedback.json');
-const PROPOSED_DIR = path.join('.agentic-security', 'rules-proposed');
+import { statePath, isSafeStateDir } from './state-dir.js';
 
 const DEFAULT_FP_THRESHOLD = 5;
 
 function _readTriage(scanRoot) {
-  const fp = path.join(scanRoot || process.cwd(), TRIAGE_PATH);
+  const fp = statePath(scanRoot, 'triage-feedback.json');
   if (!fs.existsSync(fp)) return null;
   try { return JSON.parse(fs.readFileSync(fp, 'utf8')); } catch { return null; }
 }
@@ -87,7 +85,8 @@ export function synthesizeRules(scanRoot, opts = {}) {
     groups.get(k).push(e);
   }
   const proposals = [];
-  const dir = path.join(scanRoot || process.cwd(), PROPOSED_DIR);
+  const dir = statePath(scanRoot, 'rules-proposed');
+  if (!opts.dryRun && !isSafeStateDir(path.dirname(dir))) return [];
   for (const [, group] of groups) {
     if (group.length < threshold) continue;
     const summary = _summarizeGroup(group);
@@ -105,4 +104,4 @@ export function synthesizeRules(scanRoot, opts = {}) {
   return proposals;
 }
 
-export const _internals = { DEFAULT_FP_THRESHOLD, TRIAGE_PATH, PROPOSED_DIR };
+export const _internals = { DEFAULT_FP_THRESHOLD };

package/src/posture/runtime-correlation.js

@@ -0,0 +1,174 @@
+// eBPF runtime instrumentation correlation — Recommendation #5 of the
+// world-class roadmap.
+//
+// The hardest false-positive class is "this code path is technically
+// reachable but is DEAD in production." Static analysis can't distinguish
+// dead code from reachable code; runtime observation can. This module
+// consumes an eBPF trace dataset (produced by an out-of-band collector
+// running in the customer's prod environment) and demotes findings
+// whose call-graph paths were unobserved.
+//
+// Trace format (JSONL, one record per observation):
+//
+//   { "ts": "2026-05-28T...", "host": "prod-host-1",
+//     "kind": "function-call", "qid": "com.acme.UserController.getUser",
+//     "fileRel": "src/main/java/com/acme/UserController.java", "line": 42,
+//     "count": 1234, "lastSeen": "2026-05-28T..." }
+//
+// Common record kinds:
+//   - "function-call":   QID was invoked at least once in the trace window
+//   - "route-hit":       HTTP route received traffic
+//   - "syscall":         filesystem / network / process syscall fired
+//   - "file-touch":      file was read / written
+//
+// The trace file lives at one of:
+//   .agentic-security/runtime-trace.jsonl (per-project, committed)
+//   $AGENTIC_SECURITY_RUNTIME_TRACE_PATH    (override)
+//
+// Output: every finding gets a `runtimeObserved: true|false|unknown` field.
+//   true   — at least one node on the finding's call-graph path appears in trace
+//   false  — none of the nodes appear (finding's path is dead in observed window)
+//   unknown — no trace data available
+
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import * as readline from 'node:readline';
+import { createReadStream } from 'node:fs';
+
+const DEFAULT_TRACE_NAMES = ['runtime-trace.jsonl', 'runtime.jsonl', 'ebpf-trace.jsonl'];
+const DEFAULT_OBSERVATION_WINDOW_DAYS = 30;
+
+export async function loadTrace(scanRoot, opts = {}) {
+  const explicit = opts.tracePath || process.env.AGENTIC_SECURITY_RUNTIME_TRACE_PATH;
+  const candidates = explicit ? [explicit] : DEFAULT_TRACE_NAMES.map(n => path.join(scanRoot, '.agentic-security', n));
+  let chosen = null;
+  for (const c of candidates) {
+    if (fs.existsSync(c)) { chosen = c; break; }
+  }
+  if (!chosen) return null;
+  const trace = {
+    path: chosen,
+    qidsObserved: new Set(),
+    routesObserved: new Set(),
+    filesObserved: new Set(),
+    fileLinesObserved: new Map(), // file → Set<line>
+    syscallsObserved: new Set(),
+    recordCount: 0,
+  };
+  const window = (opts.windowDays || DEFAULT_OBSERVATION_WINDOW_DAYS) * 86400_000;
+  const now = Date.now();
+  const stream = createReadStream(chosen, { encoding: 'utf8' });
+  const rl = readline.createInterface({ input: stream, crlfDelay: Infinity });
+  for await (const line of rl) {
+    if (!line.trim()) continue;
+    let r;
+    try { r = JSON.parse(line); } catch { continue; }
+    if (r.ts) {
+      const tsMs = Date.parse(r.ts);
+      if (Number.isFinite(tsMs) && now - tsMs > window) continue;
+    }
+    trace.recordCount++;
+    switch (r.kind) {
+      case 'function-call':
+        if (r.qid) trace.qidsObserved.add(r.qid);
+        if (r.fileRel && typeof r.line === 'number') {
+          let set = trace.fileLinesObserved.get(r.fileRel);
+          if (!set) { set = new Set(); trace.fileLinesObserved.set(r.fileRel, set); }
+          set.add(r.line);
+        }
+        if (r.fileRel) trace.filesObserved.add(r.fileRel);
+        break;
+      case 'route-hit':
+        if (r.route) trace.routesObserved.add(r.route);
+        break;
+      case 'syscall':
+        if (r.name) trace.syscallsObserved.add(r.name);
+        break;
+      case 'file-touch':
+        if (r.fileRel) trace.filesObserved.add(r.fileRel);
+        break;
+    }
+  }
+  return trace;
+}
+
+/**
+ * Test whether a finding's call-graph path was observed in trace.
+ * Checks:
+ *   1. The finding's file appears in filesObserved (necessary condition)
+ *   2. AT LEAST ONE line on the finding's chain (source, sink, intermediate
+ *      nodes) was observed in that file
+ *   3. The finding's containing function's qid appears in qidsObserved
+ */
+export function findingObservedInRuntime(finding, trace) {
+  if (!trace) return 'unknown';
+  // qid match (most specific)
+  if (finding.scope && trace.qidsObserved.has(finding.scope)) return true;
+  if (finding.functionQid && trace.qidsObserved.has(finding.functionQid)) return true;
+  // file + any-line on the chain match
+  const file = finding.file || (finding.sink && finding.sink.file);
+  if (file && trace.filesObserved.has(file)) {
+    const linesObserved = trace.fileLinesObserved.get(file);
+    if (linesObserved) {
+      const chainLines = [
+        finding.line,
+        ...(finding.chain || []).map(s => s.line),
+        ...(finding.taintPath || []).map(s => s.line),
+      ].filter(n => typeof n === 'number');
+      for (const ln of chainLines) {
+        // Allow ±2 line tolerance for compiler reordering / minor inlining.
+        for (let off = -2; off <= 2; off++) {
+          if (linesObserved.has(ln + off)) return true;
+        }
+      }
+    }
+    // File appears but no line match — partial evidence. Don't claim
+    // observed; don't claim dead.
+    return 'unknown';
+  }
+  // route match — every finding inside a route handler whose route was hit.
+  if (finding._inRoute && finding._inRoute.path && trace.routesObserved.has(finding._inRoute.path)) return true;
+  if (finding.routeRooted && finding._inRoute && trace.routesObserved.size > 0) {
+    for (const r of trace.routesObserved) {
+      if (finding._inRoute.path && r.includes(finding._inRoute.path)) return true;
+    }
+  }
+  return false;
+}
+
+/**
+ * Annotate all findings with runtimeObserved + demote unobserved findings.
+ * Demotion: critical → high, high → medium, medium → low.
+ *
+ * Findings classified as unknown are left alone (no demotion). This is
+ * the principled position — absence of observation in a partial trace
+ * is not evidence of dead code.
+ */
+export async function annotateRuntimeCorrelation(scanRoot, findings, opts = {}) {
+  if (!Array.isArray(findings)) return { observed: 0, dead: 0, unknown: 0 };
+  const trace = await loadTrace(scanRoot, opts);
+  if (!trace) {
+    for (const f of findings) f.runtimeObserved = 'unknown';
+    return { observed: 0, dead: 0, unknown: findings.length, trace: null };
+  }
+  let observed = 0, dead = 0, unknown = 0;
+  const demoteLadder = { critical: 'high', high: 'medium', medium: 'low', low: 'info' };
+  for (const f of findings) {
+    const v = findingObservedInRuntime(f, trace);
+    f.runtimeObserved = v;
+    if (v === true) observed++;
+    else if (v === false) {
+      dead++;
+      // Demote one tier — the finding is real but unobserved in 30 days
+      // of production traffic, so it's not P0.
+      const next = demoteLadder[f.severity];
+      if (next) {
+        f._runtimeDemoted = f.severity;
+        f.severity = next;
+      }
+    } else unknown++;
+  }
+  return { observed, dead, unknown, trace: { recordCount: trace.recordCount, path: trace.path } };
+}
+
+export const _internals = { DEFAULT_TRACE_NAMES, DEFAULT_OBSERVATION_WINDOW_DAYS };

package/src/posture/sbom-diff.js

@@ -0,0 +1,171 @@
+// SBOM diff + dependency drift detection — Recommendation #10 of the
+// world-class+2 plan.
+//
+// Tracks SBOMs across releases (keyed by git commit hash). On each scan,
+// compares the current SBOM against the previous snapshot to surface
+// drift before a CVE-publication catches it:
+//
+//   - dependency-added              new package since previous SBOM
+//   - dependency-removed            package no longer present
+//   - dependency-version-bumped     version changed
+//   - dependency-substitution       SAME package name but different
+//                                   ecosystem / publisher / repo source
+//                                   (the SolarWinds / event-stream pattern)
+//   - dependency-deprecated         transitioned to a deprecated state
+//
+// Suspicious additions (i.e., a new package that doesn't appear in any
+// PR diff / commit message) get a higher severity tier than expected ones.
+//
+// Snapshots live at .agentic-security/sbom-history/<sha>.json. The
+// engine writes the current snapshot at the end of every scan; this
+// module produces the diff on the next scan.
+
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+import * as crypto from 'node:crypto';
+import { execSync } from 'node:child_process';
+
+const HISTORY_DIR = 'sbom-history';
+
+function _historyDir(scanRoot) {
+  return path.join(scanRoot, '.agentic-security', HISTORY_DIR);
+}
+
+function _gitHead(scanRoot) {
+  try {
+    return execSync('git rev-parse HEAD', { cwd: scanRoot, encoding: 'utf8', stdio: ['ignore', 'pipe', 'ignore'] }).trim();
+  } catch { return null; }
+}
+
+function _snapshotKey(component) {
+  return `${component.ecosystem || 'unknown'}:${component.name || ''}`;
+}
+
+/**
+ * Persist the current SBOM as a snapshot keyed by the current git HEAD.
+ * If no git, falls back to a content hash of the components list.
+ */
+export function persistSbom(scanRoot, components) {
+  const dir = _historyDir(scanRoot);
+  try { fs.mkdirSync(dir, { recursive: true }); } catch {}
+  const sha = _gitHead(scanRoot) || crypto.createHash('sha256').update(JSON.stringify(components)).digest('hex').slice(0, 12);
+  const snap = {
+    sha, ts: new Date().toISOString(),
+    componentCount: components.length,
+    components: components.map(c => ({
+      ecosystem: c.ecosystem, name: c.name, version: c.version,
+      purl: c.purl, scope: c.scope, isUnpinned: !!c.isUnpinned,
+      sha256: c.sha256 || c.integrity || null,
+    })),
+  };
+  try { fs.writeFileSync(path.join(dir, `${sha}.json`), JSON.stringify(snap, null, 2)); } catch {}
+  return snap;
+}
+
+/**
+ * Load the previous SBOM snapshot for diffing.
+ */
+export function loadPreviousSnapshot(scanRoot, currentSha) {
+  const dir = _historyDir(scanRoot);
+  if (!fs.existsSync(dir)) return null;
+  let snaps;
+  try { snaps = fs.readdirSync(dir); } catch { return null; }
+  snaps = snaps.filter(f => f.endsWith('.json') && f !== `${currentSha}.json`);
+  if (!snaps.length) return null;
+  // Sort by mtime descending; take the most recent.
+  snaps.sort((a, b) => fs.statSync(path.join(dir, b)).mtimeMs - fs.statSync(path.join(dir, a)).mtimeMs);
+  try {
+    return JSON.parse(fs.readFileSync(path.join(dir, snaps[0]), 'utf8'));
+  } catch { return null; }
+}
+
+/**
+ * Compute the structured diff between two SBOMs and emit drift findings.
+ */
+export function diffSboms(previous, current) {
+  if (!previous || !current) return { findings: [], summary: { added: 0, removed: 0, bumped: 0, substituted: 0 } };
+  const findings = [];
+  const prevByKey = new Map();
+  for (const c of previous.components || []) prevByKey.set(_snapshotKey(c), c);
+  const curByKey = new Map();
+  for (const c of current.components || []) curByKey.set(_snapshotKey(c), c);
+  let added = 0, removed = 0, bumped = 0, substituted = 0;
+
+  // Added
+  for (const [key, cur] of curByKey) {
+    if (prevByKey.has(key)) continue;
+    added++;
+    findings.push({
+      family: 'dependency-drift', subfamily: 'dependency-added',
+      severity: 'medium', cwe: 'CWE-1357',
+      vuln: `Dependency added since previous release: ${cur.ecosystem}:${cur.name}@${cur.version}`,
+      file: cur.purl || 'package-lock.json', line: 0,
+      drift: { kind: 'added', component: cur, sinceSha: previous.sha },
+      remediation: 'Confirm this addition appears in a reviewed PR. Unexplained additions are the standard supply-chain-attack pattern.',
+    });
+  }
+  // Removed
+  for (const [key, prev] of prevByKey) {
+    if (curByKey.has(key)) continue;
+    removed++;
+    findings.push({
+      family: 'dependency-drift', subfamily: 'dependency-removed',
+      severity: 'low', cwe: 'CWE-1357',
+      vuln: `Dependency removed since previous release: ${prev.ecosystem}:${prev.name}@${prev.version}`,
+      file: prev.purl || 'package-lock.json', line: 0,
+      drift: { kind: 'removed', component: prev, sinceSha: previous.sha },
+      remediation: 'Confirm the removal was intentional. Silent removal of a vulnerable dep is fine; silent removal of a fix-receiving dep means CVEs may re-introduce.',
+    });
+  }
+  // Bumped
+  for (const [key, cur] of curByKey) {
+    const prev = prevByKey.get(key);
+    if (!prev) continue;
+    if (prev.version !== cur.version) {
+      bumped++;
+      const isMajor = _isMajorBump(prev.version, cur.version);
+      findings.push({
+        family: 'dependency-drift', subfamily: 'dependency-version-bumped',
+        severity: isMajor ? 'medium' : 'low',
+        cwe: 'CWE-1357',
+        vuln: `${cur.ecosystem}:${cur.name} bumped ${prev.version} → ${cur.version}${isMajor ? ' (MAJOR)' : ''}`,
+        file: cur.purl || 'package-lock.json', line: 0,
+        drift: { kind: 'bumped', from: prev.version, to: cur.version, isMajor, sinceSha: previous.sha },
+        remediation: 'Major-version bumps are breaking-change candidates AND attack pivot points. Verify the changelog signals match what your dependency intended.',
+      });
+    }
+    // Substitution check: integrity / hash changed but version stayed the same
+    if (prev.sha256 && cur.sha256 && prev.sha256 !== cur.sha256 && prev.version === cur.version) {
+      substituted++;
+      findings.push({
+        family: 'dependency-drift', subfamily: 'dependency-substitution',
+        severity: 'critical', cwe: 'CWE-1357',
+        vuln: `Suspicious substitution: ${cur.ecosystem}:${cur.name}@${cur.version} hash changed without version change`,
+        file: cur.purl || 'package-lock.json', line: 0,
+        drift: { kind: 'substituted', component: cur, oldHash: prev.sha256, newHash: cur.sha256, sinceSha: previous.sha },
+        remediation: 'Same package + same version + DIFFERENT content hash = the registry served a different artifact under the same identity. Investigate immediately; rotate the lockfile via fresh install.',
+      });
+    }
+  }
+  return { findings, summary: { added, removed, bumped, substituted } };
+}
+
+function _isMajorBump(prev, cur) {
+  const pa = (prev || '').match(/^(\d+)/);
+  const pc = (cur || '').match(/^(\d+)/);
+  if (!pa || !pc) return false;
+  return parseInt(pa[1], 10) !== parseInt(pc[1], 10);
+}
+
+/**
+ * Convenience entry — run the full pipeline: persist current, diff
+ * against previous, return findings.
+ */
+export function runSbomDiff(scanRoot, components) {
+  const current = persistSbom(scanRoot, components);
+  const previous = loadPreviousSnapshot(scanRoot, current.sha);
+  if (!previous) return { findings: [], summary: { added: 0, removed: 0, bumped: 0, substituted: 0 }, first: true };
+  return diffSboms(previous, current);
+}
+
+export const _internals = { _historyDir, _gitHead, _isMajorBump, _snapshotKey };