agentaudit 3.9.40 → 3.9.42

package/cli.mjs

@@ -525,12 +525,20 @@ function extractJSON(text) {
 
 const MAX_FILE_SIZE = 50_000;
 const MAX_TOTAL_SIZE = 300_000;
+// Directories safe to skip: dependencies, caches, build artifacts, editor config.
+// SECURITY RULE: If a directory can contain source code, workflow files, or
+// prose (prompt injection), it MUST be scanned.
+// Reviewed 2026-02-17:
+//   - RESTORED: test/tests/__tests__/spec/specs/e2e (malware in test hooks)
+//   - RESTORED: .github (workflow injection, supply chain attacks)
+//   - RESTORED: examples/example (hidden backdoors in "example" code)
+//   - RESTORED: docs/doc (prompt injection in documentation)
+//   - RESTORED: fixtures (test data can contain malicious payloads)
 const SKIP_DIRS = new Set([
-  'node_modules', '.git', '__pycache__', '.venv', 'venv', 'dist', 'build',
-  '.next', '.nuxt', 'coverage', '.pytest_cache', '.mypy_cache', 'vendor',
-  'test', 'tests', '__tests__', 'spec', 'specs', 'docs', 'doc',
-  'examples', 'example', 'fixtures', '.github', '.vscode', '.idea',
-  'e2e', 'benchmark', 'benchmarks', '.tox', '.eggs', 'htmlcov',
+  'node_modules', '.git', '__pycache__', '.venv', 'venv', 'vendor',
+  'dist', 'build', '.next', '.nuxt',
+  'coverage', '.pytest_cache', '.mypy_cache', '.tox', '.eggs', 'htmlcov',
+  '.vscode', '.idea',
 ]);
 const SKIP_EXTENSIONS = new Set([
   '.lock', '.png', '.jpg', '.jpeg', '.gif', '.svg', '.ico', '.woff',
@@ -539,31 +547,33 @@ const SKIP_EXTENSIONS = new Set([
   '.dylib', '.dll', '.exe', '.bin', '.dat', '.db', '.sqlite',
   '.snap', '.patch', '.diff', '.log', '.csv', '.tsv', '.parquet',
 ]);
-// Files that are never security-relevant (pure text/metadata only — NO executable files!)
-// Rule: if it CAN execute code (.js, .ts, .mjs, .py etc.), it MUST be scanned.
-// An attacker could hide malware in any executable config file.
+// Files safe to skip: ONLY inert line-based config that cannot execute code
+// and cannot contain prompt injections (no prose, no markdown, no scripts).
+// SECURITY RULE: When in doubt, SCAN IT. False positives > false negatives.
+// Reviewed 2026-02-17: All .md files removed from skip list (prompt injection vector).
+// All executable configs removed (malware vector). Only line-based dotfiles remain.
 const SKIP_FILES = new Set([
-  'license', 'license.md', 'license.txt', 'licence', 'licence.md',
-  'changelog.md', 'changelog', 'changes.md', 'history.md',
-  'contributing.md', 'contributors.md', 'authors', 'authors.md',
-  'code_of_conduct.md', 'security.md', 'funding.yml',
   '.gitignore', '.gitattributes', '.npmignore', '.dockerignore',
   '.editorconfig', '.browserslistrc', '.nvmrc', '.node-version',
   '.prettierignore', '.eslintignore',
 ]);
 
-function collectFiles(dir, basePath = '', collected = [], totalSize = { bytes: 0 }) {
-  if (totalSize.bytes >= MAX_TOTAL_SIZE) return collected;
+function collectFiles(dir, basePath = '', collected = [], totalSize = { bytes: 0, truncated: false, skippedPaths: [] }) {
+  if (totalSize.bytes >= MAX_TOTAL_SIZE) { totalSize.truncated = true; return collected; }
   let entries;
   try { entries = fs.readdirSync(dir, { withFileTypes: true }); }
   catch { return collected; }
   entries.sort((a, b) => a.name.localeCompare(b.name));
   for (const entry of entries) {
-    if (totalSize.bytes >= MAX_TOTAL_SIZE) break;
+    if (totalSize.bytes >= MAX_TOTAL_SIZE) { totalSize.truncated = true; totalSize.skippedPaths.push(relPath); continue; }
     const relPath = basePath ? `${basePath}/${entry.name}` : entry.name;
     const fullPath = path.join(dir, entry.name);
+    // SECURITY: Never follow symlinks — attacker could link to /etc/passwd or ~/.ssh/
+    if (entry.isSymbolicLink()) continue;
     if (entry.isDirectory()) {
-      if (SKIP_DIRS.has(entry.name) || entry.name.startsWith('.')) continue;
+      // Allow .github (workflow security), skip other dot-dirs (editor/system config)
+      if (SKIP_DIRS.has(entry.name)) continue;
+      if (entry.name.startsWith('.') && entry.name !== '.github') continue;
       collectFiles(fullPath, relPath, collected, totalSize);
     } else {
       const ext = path.extname(entry.name).toLowerCase();
@@ -1451,8 +1461,15 @@ async function auditRepo(url) {
   
   // Step 2: Collect files
   process.stdout.write(`  ${c.dim}[2/4]${c.reset} Collecting source files...`);
-  const files = collectFiles(repoPath);
+  const _collectMeta = { bytes: 0, truncated: false, skippedPaths: [] };
+  const files = collectFiles(repoPath, '', [], _collectMeta);
   console.log(` ${c.green}${files.length} files${c.reset}`);
+  if (_collectMeta.truncated) {
+    console.log(`  ${c.yellow}⚠  Size limit reached (${(MAX_TOTAL_SIZE / 1000).toFixed(0)}KB) — ${_collectMeta.skippedPaths.length} files NOT collected:${c.reset}`);
+    const shown = _collectMeta.skippedPaths.slice(0, 5);
+    for (const p of shown) console.log(`  ${c.dim}    • ${p}${c.reset}`);
+    if (_collectMeta.skippedPaths.length > 5) console.log(`  ${c.dim}    ... and ${_collectMeta.skippedPaths.length - 5} more${c.reset}`);
+  }
   
   // Step 3: Build audit payload
   process.stdout.write(`  ${c.dim}[3/4]${c.reset} Preparing audit payload...`);
@@ -1463,13 +1480,24 @@ async function auditRepo(url) {
   // ~15k tokens per chunk for code → fits comfortably in 32k+ context models
   // with room for system prompt (~2k tokens) + output (4k tokens)
   const MAX_CHUNK_CHARS = 60_000;
+  // Sort files by directory to keep related files in the same chunk.
+  // This preserves cross-file context (imports, shared modules) within each pass.
+  const sortedFiles = [...files].sort((a, b) => {
+    const dirA = a.path.includes('/') ? a.path.substring(0, a.path.lastIndexOf('/')) : '';
+    const dirB = b.path.includes('/') ? b.path.substring(0, b.path.lastIndexOf('/')) : '';
+    return dirA.localeCompare(dirB) || a.path.localeCompare(b.path);
+  });
   const chunks = []; // array of code block strings
+  const chunkFileNames = []; // track which files are in each chunk for error reporting
   let currentChunk = '';
   let currentChars = 0;
-  for (const file of files) {
+  let currentFiles = [];
+  for (const file of sortedFiles) {
     const entry = `\n### FILE: ${file.path}\n\`\`\`\n${file.content}\n\`\`\`\n`;
     if (currentChars + entry.length > MAX_CHUNK_CHARS && currentChars > 0) {
       chunks.push(currentChunk);
+      chunkFileNames.push([...currentFiles]);
+      currentFiles = [];
       currentChunk = '';
       currentChars = 0;
     }
@@ -1482,8 +1510,9 @@ async function auditRepo(url) {
       currentChunk += entry;
       currentChars += entry.length;
     }
+    currentFiles.push(file.path);
   }
-  if (currentChunk) chunks.push(currentChunk);
+  if (currentChunk) { chunks.push(currentChunk); chunkFileNames.push([...currentFiles]); }
   
   const needsMultiPass = chunks.length > 1;
   if (needsMultiPass) {
@@ -1604,7 +1633,7 @@ async function auditRepo(url) {
         const modelInfo = modelData.data?.find(m => m.id === actualModel);
         if (modelInfo?.context_length) {
           const ctx = modelInfo.context_length;
-          outputTokenBudget = ctx >= 128_000 ? 8192 : ctx >= 64_000 ? 4096 : ctx >= 32_000 ? 2048 : 1024;
+          outputTokenBudget = ctx >= 128_000 ? 8192 : ctx >= 64_000 ? 4096 : ctx >= 32_000 ? 2048 : 2048;
           if (process.argv.includes('--debug')) {
             console.log(`  ${c.dim}  Model context: ${ctx.toLocaleString()} tokens → max_tokens: ${outputTokenBudget}${c.reset}`);
           }
@@ -1744,8 +1773,11 @@ async function auditRepo(url) {
       const result = await callLLM(chunks[i], `pass ${i + 1}`);
       
       if (result.error) {
+        const failedFiles = chunkFileNames[i] || [];
         console.log(` ${c.red}failed${c.reset} ${c.dim}(${result.error.slice(0, 80)})${c.reset}`);
-        // Don't abort on individual pass failures — continue with remaining chunks
+        if (failedFiles.length > 0) {
+          console.log(`  ${c.yellow}⚠  ${failedFiles.length} files NOT analyzed:${c.reset} ${c.dim}${failedFiles.slice(0, 5).join(', ')}${failedFiles.length > 5 ? ` (+${failedFiles.length - 5} more)` : ''}${c.reset}`);
+        }
         continue;
       }
       

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentaudit",
-  "version": "3.9.40",
+  "version": "3.9.42",
   "description": "Security scanner for AI packages — MCP server + CLI",
   "type": "module",
   "bin": {