ship-safe 6.1.1 → 6.3.0

package/cli/commands/agent.js

@@ -1,608 +1,608 @@
-/**
- * Agent Command
- * =============
- *
- * AI-powered autonomous security audit.
- * Scans for secrets AND code vulnerabilities, classifies findings with Claude,
- * remediates confirmed secrets, and provides specific fix suggestions for
- * confirmed code vulnerabilities.
- *
- * USAGE:
- *   npx ship-safe agent [path]           Full AI-powered audit
- *   npx ship-safe agent . --dry-run      Preview without writing files
- *   npx ship-safe agent . --model sonnet Use a more capable model
- *
- * REQUIRES:
- *   ANTHROPIC_API_KEY in your environment or .env file.
- *   Falls back to pattern-only remediation if no key is found.
- *
- * FLOW:
- *   scan (secrets + vulns)
- *     → classify secrets (REAL/FP) → remediate confirmed secrets
- *     → classify vulns (REAL/FP + fix suggestion) → print fix table
- *     → re-scan to verify secrets clean
- */
-
-import fs from 'fs';
-import path from 'path';
-import fg from 'fast-glob';
-import chalk from 'chalk';
-import ora from 'ora';
-import {
-  SECRET_PATTERNS,
-  SECURITY_PATTERNS,
-  SKIP_DIRS,
-  SKIP_EXTENSIONS,
-  SKIP_FILENAMES,
-  TEST_FILE_PATTERNS,
-  MAX_FILE_SIZE
-} from '../utils/patterns.js';
-import { isHighEntropyMatch, getConfidence } from '../utils/entropy.js';
-import { remediateCommand } from './remediate.js';
-import * as output from '../utils/output.js';
-
-const DEFAULT_MODEL = 'claude-haiku-4-5-20251001';
-const ANTHROPIC_API_URL = 'https://api.anthropic.com/v1/messages';
-
-// =============================================================================
-// MAIN COMMAND
-// =============================================================================
-
-export async function agentCommand(targetPath = '.', options = {}) {
-  const absolutePath = path.resolve(targetPath);
-
-  if (!fs.existsSync(absolutePath)) {
-    output.error(`Path does not exist: ${absolutePath}`);
-    process.exit(1);
-  }
-
-  const model = options.model || DEFAULT_MODEL;
-
-  console.log();
-  output.header('Ship Safe — AI Security Agent');
-  console.log();
-
-  // ── 1. Load API key ────────────────────────────────────────────────────────
-  const apiKey = loadApiKey(absolutePath);
-
-  // ── 2. Scan (secrets + vulnerabilities) ────────────────────────────────────
-  const scanSpinner = ora({ text: 'Scanning for secrets and vulnerabilities...', color: 'cyan' }).start();
-  const allResults = await scanProject(absolutePath);
-
-  // Separate findings by category
-  const secretResults = [];
-  const vulnResults = [];
-  for (const { file, findings } of allResults) {
-    const secrets = findings.filter(f => f.category !== 'vulnerability');
-    const vulns = findings.filter(f => f.category === 'vulnerability');
-    if (secrets.length > 0) secretResults.push({ file, findings: secrets });
-    if (vulns.length > 0) vulnResults.push({ file, findings: vulns });
-  }
-
-  const secretCount = secretResults.reduce((n, r) => n + r.findings.length, 0);
-  const vulnCount = vulnResults.reduce((n, r) => n + r.findings.length, 0);
-  scanSpinner.stop();
-
-  // ── 3. Nothing found ───────────────────────────────────────────────────────
-  if (secretCount === 0 && vulnCount === 0) {
-    output.success('No secrets or vulnerabilities detected — your project is clean!');
-    console.log();
-    return;
-  }
-
-  if (secretCount > 0) {
-    console.log(chalk.red(`\n  Found ${secretCount} potential secret(s) in ${secretResults.length} file(s)`));
-  }
-  if (vulnCount > 0) {
-    console.log(chalk.yellow(`  Found ${vulnCount} code vulnerability/vulnerabilities in ${vulnResults.length} file(s)`));
-  }
-  console.log();
-
-  // ── 4. Fallback: no API key ────────────────────────────────────────────────
-  if (!apiKey) {
-    console.log(chalk.yellow('  ⚠  No ANTHROPIC_API_KEY found.')); // ship-safe-ignore — env var name in user-facing message, no key value
-    console.log(chalk.gray('     Set it in your environment or .env to enable AI classification.'));
-    if (secretCount > 0) {
-      console.log(chalk.gray('     Falling back to pattern-based remediation for secrets...\n'));
-      await remediateCommand(targetPath, { yes: true, dryRun: options.dryRun });
-    }
-    if (vulnCount > 0) {
-      console.log(chalk.gray('\n     Code vulnerabilities require manual review.'));
-      console.log(chalk.gray('     Run: ') + chalk.cyan('npx ship-safe scan .') + chalk.gray(' to see details.'));
-    }
-    return;
-  }
-
-  // ── 5. Classify secrets ────────────────────────────────────────────────────
-  if (secretCount > 0) {
-    const classifySpinner = ora({ text: `Classifying ${secretCount} secret(s) with ${model}...`, color: 'cyan' }).start();
-    let classifiedSecrets;
-
-    try {
-      classifiedSecrets = await classifyWithClaude(secretResults, absolutePath, apiKey, model);
-    } catch (err) {
-      classifySpinner.stop();
-      console.log(chalk.yellow(`  ⚠  Claude secret classification failed: ${err.message}`));
-      console.log(chalk.gray('     Treating all findings as real secrets (safe fallback).\n'));
-      classifiedSecrets = secretResults.map(({ file, findings }) => ({
-        file,
-        findings: findings.map(f => ({ ...f, classification: 'REAL', reason: 'Classification unavailable' }))
-      }));
-    }
-
-    classifySpinner.stop();
-
-    // ── 6. Print secret classification table ──────────────────────────────
-    printClassificationTable(classifiedSecrets, absolutePath);
-
-    const realSecretCount = classifiedSecrets.reduce(
-      (n, { findings }) => n + findings.filter(f => f.classification === 'REAL').length, 0
-    );
-    const fpCount = secretCount - realSecretCount;
-
-    console.log();
-    if (realSecretCount === 0) {
-      output.success(`Claude classified all ${secretCount} secret finding(s) as false positives — nothing to fix!`);
-      if (fpCount > 0) {
-        console.log(chalk.gray('  Tip: Add # ship-safe-ignore on those lines to suppress future warnings.'));
-      }
-    } else {
-      console.log(chalk.cyan(`  ${realSecretCount} confirmed secret(s) to remediate.${fpCount > 0 ? chalk.gray(` ${fpCount} false positive(s) skipped.`) : ''}`));
-      console.log();
-
-      // ── 7. Remediate confirmed secrets ──────────────────────────────────
-      if (options.dryRun) {
-        console.log(chalk.cyan('  Dry run — secrets not modified. Remove --dry-run to apply fixes.'));
-      } else {
-        await remediateCommand(targetPath, { yes: true });
-      }
-    }
-  }
-
-  // ── 8. Classify vulnerabilities ────────────────────────────────────────────
-  if (vulnCount > 0) {
-    console.log();
-    const vulnSpinner = ora({
-      text: `Analyzing ${vulnCount} vulnerability/vulnerabilities with ${model}...`,
-      color: 'cyan'
-    }).start();
-    let classifiedVulns;
-
-    try {
-      classifiedVulns = await classifyVulnsWithClaude(vulnResults, absolutePath, apiKey, model);
-    } catch (err) {
-      vulnSpinner.stop();
-      console.log(chalk.yellow(`  ⚠  Claude vulnerability analysis failed: ${err.message}`));
-      console.log(chalk.gray('     Showing raw findings without AI fix suggestions.\n'));
-      classifiedVulns = vulnResults.map(({ file, findings }) => ({
-        file,
-        findings: findings.map(f => ({ ...f, classification: 'REAL', reason: 'Analysis unavailable', fix: null }))
-      }));
-    }
-
-    vulnSpinner.stop();
-
-    // ── 9. Print vulnerability fix table ──────────────────────────────────
-    printVulnFixTable(classifiedVulns, absolutePath);
-  }
-
-  // ── 10. Verify secrets clean ───────────────────────────────────────────────
-  if (secretCount > 0 && !options.dryRun) {
-    console.log();
-    const verifySpinner = ora({ text: 'Re-scanning to verify secrets removed...', color: 'cyan' }).start();
-    const verifyResults = await scanProject(absolutePath);
-    const remainingSecrets = verifyResults.reduce(
-      (n, r) => n + r.findings.filter(f => f.category !== 'vulnerability').length, 0
-    );
-    verifySpinner.stop();
-
-    if (remainingSecrets === 0) {
-      output.success('Secrets verified clean — 0 remain in your codebase!');
-    } else {
-      output.warning(`${remainingSecrets} secret(s) still remain. Review them manually or run npx ship-safe scan .`);
-    }
-  }
-
-  // ── 11. Next steps ─────────────────────────────────────────────────────────
-  console.log();
-  console.log(chalk.yellow.bold('  Next steps:'));
-  let step = 1;
-  if (secretCount > 0) {
-    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Rotate any exposed keys:   ') + chalk.cyan('npx ship-safe rotate'));
-    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Commit the fixes:          ') + chalk.cyan('git add . && git commit -m "fix: remove hardcoded secrets"'));
-    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Fill in .env with fresh values from your providers'));
-  }
-  if (vulnCount > 0) {
-    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Apply the code fixes shown above, then re-run: ') + chalk.cyan('npx ship-safe agent .'));
-  }
-  console.log();
-}
-
-// =============================================================================
-// API KEY LOADING
-// =============================================================================
-
-/**
- * Load ANTHROPIC_API_KEY from environment or .env file.
- * Returns the key string or null if not found.
- */
-function loadApiKey(rootPath) {
-  if (process.env.ANTHROPIC_API_KEY) { // ship-safe-ignore — reading env var at runtime, no hardcoded key value
-    return process.env.ANTHROPIC_API_KEY; // ship-safe-ignore — returning env var value, not a hardcoded secret
-  }
-
-  const envPath = path.join(rootPath, '.env');
-  if (fs.existsSync(envPath)) {
-    try {
-      const lines = fs.readFileSync(envPath, 'utf-8').split('\n');
-      for (const line of lines) {
-        const trimmed = line.trim();
-        if (trimmed.startsWith('#') || !trimmed.includes('=')) continue;
-        const eqIdx = trimmed.indexOf('=');
-        const key = trimmed.slice(0, eqIdx).trim();
-        if (key === 'ANTHROPIC_API_KEY') { // ship-safe-ignore — parsing .env file to read user's own API key from their project
-          const val = trimmed.slice(eqIdx + 1).trim().replace(/^["']|["']$/g, '');
-          if (val) return val;
-        }
-      }
-    } catch {
-      // ignore read errors
-    }
-  }
-
-  return null;
-}
-
-// =============================================================================
-// PROJECT SCANNING (secrets + vulnerabilities)
-// =============================================================================
-
-async function scanProject(rootPath) {
-  const globIgnore = Array.from(SKIP_DIRS).map(dir => `**/${dir}/**`);
-
-  const allFiles = await fg('**/*', {
-    cwd: rootPath,
-    absolute: true,
-    onlyFiles: true,
-    ignore: globIgnore,
-    dot: true
-  });
-
-  const files = allFiles.filter(file => {
-    const ext = path.extname(file).toLowerCase();
-    if (SKIP_EXTENSIONS.has(ext)) return false;
-    if (SKIP_FILENAMES.has(path.basename(file))) return false;
-    const basename = path.basename(file);
-    if (basename.endsWith('.min.js') || basename.endsWith('.min.css')) return false;
-    if (TEST_FILE_PATTERNS.some(p => p.test(file))) return false;
-    try {
-      const stats = fs.statSync(file);
-      if (stats.size > MAX_FILE_SIZE) return false;
-    } catch {
-      return false;
-    }
-    return true;
-  });
-
-  const results = [];
-  for (const file of files) {
-    const findings = scanFile(file);
-    if (findings.length > 0) {
-      results.push({ file, findings });
-    }
-  }
-  return results;
-}
-
-function scanFile(filePath) {
-  const findings = [];
-  try {
-    const content = fs.readFileSync(filePath, 'utf-8');
-    const lines = content.split('\n');
-
-    for (let lineNum = 0; lineNum < lines.length; lineNum++) {
-      const line = lines[lineNum];
-      if (/ship-safe-ignore/i.test(line)) continue;
-
-      for (const pattern of [...SECRET_PATTERNS, ...SECURITY_PATTERNS]) {
-        pattern.pattern.lastIndex = 0;
-        let match;
-        while ((match = pattern.pattern.exec(line)) !== null) {
-          if (pattern.requiresEntropyCheck && !isHighEntropyMatch(match[0])) continue;
-          findings.push({
-            line: lineNum + 1,
-            column: match.index + 1,
-            matched: match[0],
-            patternName: pattern.name,
-            severity: pattern.severity,
-            confidence: getConfidence(pattern, match[0]),
-            description: pattern.description,
-            category: pattern.category || 'secret'
-          });
-        }
-      }
-    }
-  } catch {
-    // Skip unreadable files
-  }
-
-  // Deduplicate by (line, matched)
-  const seen = new Set();
-  return findings.filter(f => {
-    const key = `${f.line}:${f.matched}`;
-    if (seen.has(key)) return false;
-    seen.add(key);
-    return true;
-  });
-}
-
-// =============================================================================
-// CLAUDE CLASSIFICATION — SECRETS
-// =============================================================================
-
-async function classifyWithClaude(scanResults, rootPath, apiKey, model) {
-  const items = [];
-  for (const { file, findings } of scanResults) {
-    let lines = [];
-    try {
-      lines = fs.readFileSync(file, 'utf-8').split('\n');
-    } catch {
-      // include without context
-    }
-
-    for (const finding of findings) {
-      const startLine = Math.max(0, finding.line - 3);
-      const endLine = Math.min(lines.length - 1, finding.line + 1);
-      const context = lines.slice(startLine, endLine + 1).join('\n');
-
-      // Truncate matched value — don't send real secrets to the API
-      const matchedPrefix = finding.matched.length > 12
-        ? finding.matched.slice(0, 12) + '...'
-        : finding.matched;
-
-      items.push({
-        id: `${path.relative(rootPath, file)}:${finding.line}`,
-        file: path.relative(rootPath, file),
-        line: finding.line,
-        patternName: finding.patternName,
-        severity: finding.severity,
-        matchedPrefix,
-        codeContext: context
-      });
-    }
-  }
-
-  const prompt = `You are a security expert reviewing potential secret leaks in source code.
-
-For each finding below, classify it as REAL or FALSE_POSITIVE:
-- REAL: a genuine hardcoded secret, credential, or API key that should be moved to environment variables
-- FALSE_POSITIVE: a placeholder, example value, test fixture, documentation sample, or non-sensitive identifier
-
-Respond with a JSON array ONLY — no markdown, no explanation, just the JSON:
-[{"id":"<id>","classification":"REAL"|"FALSE_POSITIVE","reason":"<brief one-line reason>"}]
-
-Findings to classify:
-${JSON.stringify(items, null, 2)}`;
-
-  const response = await fetch(ANTHROPIC_API_URL, {
-    method: 'POST',
-    headers: {
-      'x-api-key': apiKey,
-      'anthropic-version': '2023-06-01',
-      'content-type': 'application/json',
-    },
-    body: JSON.stringify({
-      model,
-      max_tokens: 2048,
-      messages: [{ role: 'user', content: prompt }]
-    })
-  });
-
-  if (!response.ok) {
-    const body = await response.text();
-    throw new Error(`Anthropic API error ${response.status}: ${body.slice(0, 200)}`);
-  }
-
-  const data = await response.json();
-  const text = data.content?.[0]?.text || '[]';
-  const jsonText = text.replace(/^```(?:json)?\s*/i, '').replace(/\s*```\s*$/i, '').trim();
-  let classifications;
-  try {
-    classifications = JSON.parse(jsonText);
-  } catch {
-    throw new Error('Claude returned non-JSON response');
-  }
-
-  return scanResults.map(({ file, findings }) => ({
-    file,
-    findings: findings.map(f => {
-      const id = `${path.relative(rootPath, file)}:${f.line}`;
-      const cl = classifications.find(c => c.id === id);
-      return {
-        ...f,
-        classification: cl?.classification ?? 'REAL',
-        reason: cl?.reason ?? ''
-      };
-    })
-  }));
-}
-
-// =============================================================================
-// CLAUDE CLASSIFICATION — VULNERABILITIES
-// =============================================================================
-
-/**
- * Send vulnerability findings to Claude for classification + specific fix suggestions.
- * Unlike secrets, code context is NOT masked — the pattern itself is the finding.
- */
-async function classifyVulnsWithClaude(vulnResults, rootPath, apiKey, model) {
-  const items = [];
-  for (const { file, findings } of vulnResults) {
-    let lines = [];
-    try {
-      lines = fs.readFileSync(file, 'utf-8').split('\n');
-    } catch {
-      // include without context
-    }
-
-    for (const finding of findings) {
-      const startLine = Math.max(0, finding.line - 3);
-      const endLine = Math.min(lines.length - 1, finding.line + 1);
-      const context = lines.slice(startLine, endLine + 1).join('\n');
-
-      items.push({
-        id: `${path.relative(rootPath, file)}:${finding.line}`,
-        file: path.relative(rootPath, file),
-        line: finding.line,
-        type: finding.patternName,
-        severity: finding.severity,
-        codeContext: context  // Not masked — it's a code pattern, not a secret
-      });
-    }
-  }
-
-  const prompt = `You are a security expert reviewing code vulnerabilities.
-
-For each finding below, classify it and provide a specific fix if it's a real issue.
-
-- REAL: genuinely exploitable as written (user-controlled input reaches a dangerous sink)
-- FALSE_POSITIVE: safe in this context (static/hardcoded input, internal tool, test code, build script)
-
-For REAL findings: provide a concise, specific 1-line code fix showing what to change.
-For FALSE_POSITIVE: briefly explain why it's safe.
-
-Respond with a JSON array ONLY — no markdown, no explanation:
-[{"id":"<id>","classification":"REAL"|"FALSE_POSITIVE","reason":"<brief reason>","fix":"<specific fix code, or null>"}]
-
-Vulnerabilities to analyze:
-${JSON.stringify(items, null, 2)}`;
-
-  const response = await fetch(ANTHROPIC_API_URL, {
-    method: 'POST',
-    headers: {
-      'x-api-key': apiKey,
-      'anthropic-version': '2023-06-01',
-      'content-type': 'application/json',
-    },
-    body: JSON.stringify({
-      model,
-      max_tokens: 4096,
-      messages: [{ role: 'user', content: prompt }]
-    })
-  });
-
-  if (!response.ok) {
-    const body = await response.text();
-    throw new Error(`Anthropic API error ${response.status}: ${body.slice(0, 200)}`);
-  }
-
-  const data = await response.json();
-  const text = data.content?.[0]?.text || '[]';
-  const jsonText = text.replace(/^```(?:json)?\s*/i, '').replace(/\s*```\s*$/i, '').trim();
-  let classifications;
-  try {
-    classifications = JSON.parse(jsonText);
-  } catch {
-    throw new Error('Claude returned non-JSON response');
-  }
-
-  return vulnResults.map(({ file, findings }) => ({
-    file,
-    findings: findings.map(f => {
-      const id = `${path.relative(rootPath, file)}:${f.line}`;
-      const cl = classifications.find(c => c.id === id);
-      return {
-        ...f,
-        classification: cl?.classification ?? 'REAL',
-        reason: cl?.reason ?? '',
-        fix: cl?.fix ?? null
-      };
-    })
-  }));
-}
-
-// =============================================================================
-// OUTPUT
-// =============================================================================
-
-function printClassificationTable(classified, rootPath) {
-  const SEVERITY_COLOR = {
-    critical: chalk.red.bold,
-    high: chalk.yellow,
-    medium: chalk.blue
-  };
-
-  console.log(chalk.cyan('  Secret Classification'));
-  console.log(chalk.cyan('  ' + '─'.repeat(58)));
-  console.log();
-
-  for (const { file, findings } of classified) {
-    const relPath = path.relative(rootPath, file);
-    for (const f of findings) {
-      const isReal = f.classification === 'REAL';
-      const icon = isReal ? chalk.red('✗') : chalk.gray('~');
-      const label = isReal ? chalk.red('REAL') : chalk.gray('SKIP');
-      const sevColor = SEVERITY_COLOR[f.severity] || chalk.white;
-      const matchedShort = f.matched.length > 16 ? f.matched.slice(0, 16) + '…' : f.matched;
-
-      console.log(
-        `  ${icon}  ${label.padEnd(8)} ${chalk.white(`${relPath}:${f.line}`).padEnd(40)}  ` +
-        `${sevColor(f.patternName.padEnd(24))}  ` +
-        chalk.gray(matchedShort)
-      );
-      if (f.reason) {
-        console.log(chalk.gray(`            → ${f.reason}`));
-      }
-    }
-  }
-}
-
-function printVulnFixTable(classifiedVulns, rootPath) {
-  const SEVERITY_COLOR = {
-    critical: chalk.red.bold,
-    high: chalk.yellow,
-    medium: chalk.blue
-  };
-
-  const totalCount = classifiedVulns.reduce((n, { findings }) => n + findings.length, 0);
-  const realCount = classifiedVulns.reduce(
-    (n, { findings }) => n + findings.filter(f => f.classification === 'REAL').length, 0
-  );
-  const fpCount = totalCount - realCount;
-
-  console.log(chalk.yellow('  Code Vulnerability Analysis'));
-  console.log(chalk.yellow('  ' + '─'.repeat(58)));
-  console.log();
-
-  for (const { file, findings } of classifiedVulns) {
-    const relPath = path.relative(rootPath, file);
-    for (const f of findings) {
-      const isReal = f.classification === 'REAL';
-      const icon = isReal ? chalk.red('✗') : chalk.gray('~');
-      const label = isReal ? chalk.red('REAL') : chalk.gray('SKIP');
-      const sevColor = SEVERITY_COLOR[f.severity] || chalk.white;
-      const snippet = f.matched.length > 55 ? f.matched.slice(0, 55) + '…' : f.matched;
-
-      console.log(
-        `  ${icon}  ${label.padEnd(8)} ${chalk.white(`${relPath}:${f.line}`).padEnd(38)}  ` +
-        sevColor(`[${f.severity.toUpperCase()}]`)
-      );
-      console.log(chalk.gray(`            ${f.patternName}`));
-      console.log(chalk.gray('     Code:   ') + chalk.cyan(snippet));
-      if (f.reason) {
-        console.log(chalk.gray(`     Reason: ${f.reason}`));
-      }
-      if (isReal && f.fix) {
-        console.log(chalk.gray('     Fix:    ') + chalk.green(f.fix));
-      }
-      console.log();
-    }
-  }
-
-  if (realCount === 0) {
-    output.success(`Claude classified all ${totalCount} vulnerability/vulnerabilities as false positives!`);
-    console.log(chalk.gray('  Tip: Add # ship-safe-ignore on those lines to suppress future warnings.'));
-  } else {
-    console.log(chalk.yellow(`  ${realCount} real vulnerability/vulnerabilities require manual fixes.${fpCount > 0 ? chalk.gray(` ${fpCount} false positive(s) skipped.`) : ''}`));
-  }
-}
+/**
+ * Agent Command
+ * =============
+ *
+ * AI-powered autonomous security audit.
+ * Scans for secrets AND code vulnerabilities, classifies findings with Claude,
+ * remediates confirmed secrets, and provides specific fix suggestions for
+ * confirmed code vulnerabilities.
+ *
+ * USAGE:
+ *   npx ship-safe agent [path]           Full AI-powered audit
+ *   npx ship-safe agent . --dry-run      Preview without writing files
+ *   npx ship-safe agent . --model sonnet Use a more capable model
+ *
+ * REQUIRES:
+ *   ANTHROPIC_API_KEY in your environment or .env file.
+ *   Falls back to pattern-only remediation if no key is found.
+ *
+ * FLOW:
+ *   scan (secrets + vulns)
+ *     → classify secrets (REAL/FP) → remediate confirmed secrets
+ *     → classify vulns (REAL/FP + fix suggestion) → print fix table
+ *     → re-scan to verify secrets clean
+ */
+
+import fs from 'fs';
+import path from 'path';
+import fg from 'fast-glob';
+import chalk from 'chalk';
+import ora from 'ora';
+import {
+  SECRET_PATTERNS,
+  SECURITY_PATTERNS,
+  SKIP_DIRS,
+  SKIP_EXTENSIONS,
+  SKIP_FILENAMES,
+  TEST_FILE_PATTERNS,
+  MAX_FILE_SIZE
+} from '../utils/patterns.js';
+import { isHighEntropyMatch, getConfidence } from '../utils/entropy.js';
+import { remediateCommand } from './remediate.js';
+import * as output from '../utils/output.js';
+
+const DEFAULT_MODEL = 'claude-haiku-4-5-20251001';
+const ANTHROPIC_API_URL = 'https://api.anthropic.com/v1/messages';
+
+// =============================================================================
+// MAIN COMMAND
+// =============================================================================
+
+export async function agentCommand(targetPath = '.', options = {}) {
+  const absolutePath = path.resolve(targetPath);
+
+  if (!fs.existsSync(absolutePath)) {
+    output.error(`Path does not exist: ${absolutePath}`);
+    process.exit(1);
+  }
+
+  const model = options.model || DEFAULT_MODEL;
+
+  console.log();
+  output.header('Ship Safe — AI Security Agent');
+  console.log();
+
+  // ── 1. Load API key ────────────────────────────────────────────────────────
+  const apiKey = loadApiKey(absolutePath);
+
+  // ── 2. Scan (secrets + vulnerabilities) ────────────────────────────────────
+  const scanSpinner = ora({ text: 'Scanning for secrets and vulnerabilities...', color: 'cyan' }).start();
+  const allResults = await scanProject(absolutePath);
+
+  // Separate findings by category
+  const secretResults = [];
+  const vulnResults = [];
+  for (const { file, findings } of allResults) {
+    const secrets = findings.filter(f => f.category !== 'vulnerability');
+    const vulns = findings.filter(f => f.category === 'vulnerability');
+    if (secrets.length > 0) secretResults.push({ file, findings: secrets });
+    if (vulns.length > 0) vulnResults.push({ file, findings: vulns });
+  }
+
+  const secretCount = secretResults.reduce((n, r) => n + r.findings.length, 0);
+  const vulnCount = vulnResults.reduce((n, r) => n + r.findings.length, 0);
+  scanSpinner.stop();
+
+  // ── 3. Nothing found ───────────────────────────────────────────────────────
+  if (secretCount === 0 && vulnCount === 0) {
+    output.success('No secrets or vulnerabilities detected — your project is clean!');
+    console.log();
+    return;
+  }
+
+  if (secretCount > 0) {
+    console.log(chalk.red(`\n  Found ${secretCount} potential secret(s) in ${secretResults.length} file(s)`));
+  }
+  if (vulnCount > 0) {
+    console.log(chalk.yellow(`  Found ${vulnCount} code vulnerability/vulnerabilities in ${vulnResults.length} file(s)`));
+  }
+  console.log();
+
+  // ── 4. Fallback: no API key ────────────────────────────────────────────────
+  if (!apiKey) {
+    console.log(chalk.yellow('  ⚠  No ANTHROPIC_API_KEY found.')); // ship-safe-ignore — env var name in user-facing message, no key value
+    console.log(chalk.gray('     Set it in your environment or .env to enable AI classification.'));
+    if (secretCount > 0) {
+      console.log(chalk.gray('     Falling back to pattern-based remediation for secrets...\n'));
+      await remediateCommand(targetPath, { yes: true, dryRun: options.dryRun });
+    }
+    if (vulnCount > 0) {
+      console.log(chalk.gray('\n     Code vulnerabilities require manual review.'));
+      console.log(chalk.gray('     Run: ') + chalk.cyan('npx ship-safe scan .') + chalk.gray(' to see details.'));
+    }
+    return;
+  }
+
+  // ── 5. Classify secrets ────────────────────────────────────────────────────
+  if (secretCount > 0) {
+    const classifySpinner = ora({ text: `Classifying ${secretCount} secret(s) with ${model}...`, color: 'cyan' }).start();
+    let classifiedSecrets;
+
+    try {
+      classifiedSecrets = await classifyWithClaude(secretResults, absolutePath, apiKey, model);
+    } catch (err) {
+      classifySpinner.stop();
+      console.log(chalk.yellow(`  ⚠  Claude secret classification failed: ${err.message}`));
+      console.log(chalk.gray('     Treating all findings as real secrets (safe fallback).\n'));
+      classifiedSecrets = secretResults.map(({ file, findings }) => ({
+        file,
+        findings: findings.map(f => ({ ...f, classification: 'REAL', reason: 'Classification unavailable' }))
+      }));
+    }
+
+    classifySpinner.stop();
+
+    // ── 6. Print secret classification table ──────────────────────────────
+    printClassificationTable(classifiedSecrets, absolutePath);
+
+    const realSecretCount = classifiedSecrets.reduce(
+      (n, { findings }) => n + findings.filter(f => f.classification === 'REAL').length, 0
+    );
+    const fpCount = secretCount - realSecretCount;
+
+    console.log();
+    if (realSecretCount === 0) {
+      output.success(`Claude classified all ${secretCount} secret finding(s) as false positives — nothing to fix!`);
+      if (fpCount > 0) {
+        console.log(chalk.gray('  Tip: Add # ship-safe-ignore on those lines to suppress future warnings.'));
+      }
+    } else {
+      console.log(chalk.cyan(`  ${realSecretCount} confirmed secret(s) to remediate.${fpCount > 0 ? chalk.gray(` ${fpCount} false positive(s) skipped.`) : ''}`));
+      console.log();
+
+      // ── 7. Remediate confirmed secrets ──────────────────────────────────
+      if (options.dryRun) {
+        console.log(chalk.cyan('  Dry run — secrets not modified. Remove --dry-run to apply fixes.'));
+      } else {
+        await remediateCommand(targetPath, { yes: true });
+      }
+    }
+  }
+
+  // ── 8. Classify vulnerabilities ────────────────────────────────────────────
+  if (vulnCount > 0) {
+    console.log();
+    const vulnSpinner = ora({
+      text: `Analyzing ${vulnCount} vulnerability/vulnerabilities with ${model}...`,
+      color: 'cyan'
+    }).start();
+    let classifiedVulns;
+
+    try {
+      classifiedVulns = await classifyVulnsWithClaude(vulnResults, absolutePath, apiKey, model);
+    } catch (err) {
+      vulnSpinner.stop();
+      console.log(chalk.yellow(`  ⚠  Claude vulnerability analysis failed: ${err.message}`));
+      console.log(chalk.gray('     Showing raw findings without AI fix suggestions.\n'));
+      classifiedVulns = vulnResults.map(({ file, findings }) => ({
+        file,
+        findings: findings.map(f => ({ ...f, classification: 'REAL', reason: 'Analysis unavailable', fix: null }))
+      }));
+    }
+
+    vulnSpinner.stop();
+
+    // ── 9. Print vulnerability fix table ──────────────────────────────────
+    printVulnFixTable(classifiedVulns, absolutePath);
+  }
+
+  // ── 10. Verify secrets clean ───────────────────────────────────────────────
+  if (secretCount > 0 && !options.dryRun) {
+    console.log();
+    const verifySpinner = ora({ text: 'Re-scanning to verify secrets removed...', color: 'cyan' }).start();
+    const verifyResults = await scanProject(absolutePath);
+    const remainingSecrets = verifyResults.reduce(
+      (n, r) => n + r.findings.filter(f => f.category !== 'vulnerability').length, 0
+    );
+    verifySpinner.stop();
+
+    if (remainingSecrets === 0) {
+      output.success('Secrets verified clean — 0 remain in your codebase!');
+    } else {
+      output.warning(`${remainingSecrets} secret(s) still remain. Review them manually or run npx ship-safe scan .`);
+    }
+  }
+
+  // ── 11. Next steps ─────────────────────────────────────────────────────────
+  console.log();
+  console.log(chalk.yellow.bold('  Next steps:'));
+  let step = 1;
+  if (secretCount > 0) {
+    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Rotate any exposed keys:   ') + chalk.cyan('npx ship-safe rotate'));
+    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Commit the fixes:          ') + chalk.cyan('git add . && git commit -m "fix: remove hardcoded secrets"'));
+    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Fill in .env with fresh values from your providers'));
+  }
+  if (vulnCount > 0) {
+    console.log(chalk.white(`  ${step++}.`) + chalk.gray(' Apply the code fixes shown above, then re-run: ') + chalk.cyan('npx ship-safe agent .'));
+  }
+  console.log();
+}
+
+// =============================================================================
+// API KEY LOADING
+// =============================================================================
+
+/**
+ * Load ANTHROPIC_API_KEY from environment or .env file.
+ * Returns the key string or null if not found.
+ */
+function loadApiKey(rootPath) {
+  if (process.env.ANTHROPIC_API_KEY) { // ship-safe-ignore — reading env var at runtime, no hardcoded key value
+    return process.env.ANTHROPIC_API_KEY; // ship-safe-ignore — returning env var value, not a hardcoded secret
+  }
+
+  const envPath = path.join(rootPath, '.env');
+  if (fs.existsSync(envPath)) {
+    try {
+      const lines = fs.readFileSync(envPath, 'utf-8').split('\n');
+      for (const line of lines) {
+        const trimmed = line.trim();
+        if (trimmed.startsWith('#') || !trimmed.includes('=')) continue;
+        const eqIdx = trimmed.indexOf('=');
+        const key = trimmed.slice(0, eqIdx).trim();
+        if (key === 'ANTHROPIC_API_KEY') { // ship-safe-ignore — parsing .env file to read user's own API key from their project
+          const val = trimmed.slice(eqIdx + 1).trim().replace(/^["']|["']$/g, '');
+          if (val) return val;
+        }
+      }
+    } catch {
+      // ignore read errors
+    }
+  }
+
+  return null;
+}
+
+// =============================================================================
+// PROJECT SCANNING (secrets + vulnerabilities)
+// =============================================================================
+
+async function scanProject(rootPath) {
+  const globIgnore = Array.from(SKIP_DIRS).map(dir => `**/${dir}/**`);
+
+  const allFiles = await fg('**/*', {
+    cwd: rootPath,
+    absolute: true,
+    onlyFiles: true,
+    ignore: globIgnore,
+    dot: true
+  });
+
+  const files = allFiles.filter(file => {
+    const ext = path.extname(file).toLowerCase();
+    if (SKIP_EXTENSIONS.has(ext)) return false;
+    if (SKIP_FILENAMES.has(path.basename(file))) return false;
+    const basename = path.basename(file);
+    if (basename.endsWith('.min.js') || basename.endsWith('.min.css')) return false;
+    if (TEST_FILE_PATTERNS.some(p => p.test(file))) return false;
+    try {
+      const stats = fs.statSync(file);
+      if (stats.size > MAX_FILE_SIZE) return false;
+    } catch {
+      return false;
+    }
+    return true;
+  });
+
+  const results = [];
+  for (const file of files) {
+    const findings = scanFile(file);
+    if (findings.length > 0) {
+      results.push({ file, findings });
+    }
+  }
+  return results;
+}
+
+function scanFile(filePath) {
+  const findings = [];
+  try {
+    const content = fs.readFileSync(filePath, 'utf-8');
+    const lines = content.split('\n');
+
+    for (let lineNum = 0; lineNum < lines.length; lineNum++) {
+      const line = lines[lineNum];
+      if (/ship-safe-ignore/i.test(line)) continue;
+
+      for (const pattern of [...SECRET_PATTERNS, ...SECURITY_PATTERNS]) {
+        pattern.pattern.lastIndex = 0;
+        let match;
+        while ((match = pattern.pattern.exec(line)) !== null) {
+          if (pattern.requiresEntropyCheck && !isHighEntropyMatch(match[0])) continue;
+          findings.push({
+            line: lineNum + 1,
+            column: match.index + 1,
+            matched: match[0],
+            patternName: pattern.name,
+            severity: pattern.severity,
+            confidence: getConfidence(pattern, match[0]),
+            description: pattern.description,
+            category: pattern.category || 'secret'
+          });
+        }
+      }
+    }
+  } catch {
+    // Skip unreadable files
+  }
+
+  // Deduplicate by (line, matched)
+  const seen = new Set();
+  return findings.filter(f => {
+    const key = `${f.line}:${f.matched}`;
+    if (seen.has(key)) return false;
+    seen.add(key);
+    return true;
+  });
+}
+
+// =============================================================================
+// CLAUDE CLASSIFICATION — SECRETS
+// =============================================================================
+
+async function classifyWithClaude(scanResults, rootPath, apiKey, model) {
+  const items = [];
+  for (const { file, findings } of scanResults) {
+    let lines = [];
+    try {
+      lines = fs.readFileSync(file, 'utf-8').split('\n');
+    } catch {
+      // include without context
+    }
+
+    for (const finding of findings) {
+      const startLine = Math.max(0, finding.line - 3);
+      const endLine = Math.min(lines.length - 1, finding.line + 1);
+      const context = lines.slice(startLine, endLine + 1).join('\n');
+
+      // Truncate matched value — don't send real secrets to the API
+      const matchedPrefix = finding.matched.length > 12
+        ? finding.matched.slice(0, 12) + '...'
+        : finding.matched;
+
+      items.push({
+        id: `${path.relative(rootPath, file)}:${finding.line}`,
+        file: path.relative(rootPath, file),
+        line: finding.line,
+        patternName: finding.patternName,
+        severity: finding.severity,
+        matchedPrefix,
+        codeContext: context
+      });
+    }
+  }
+
+  const prompt = `You are a security expert reviewing potential secret leaks in source code.
+
+For each finding below, classify it as REAL or FALSE_POSITIVE:
+- REAL: a genuine hardcoded secret, credential, or API key that should be moved to environment variables
+- FALSE_POSITIVE: a placeholder, example value, test fixture, documentation sample, or non-sensitive identifier
+
+Respond with a JSON array ONLY — no markdown, no explanation, just the JSON:
+[{"id":"<id>","classification":"REAL"|"FALSE_POSITIVE","reason":"<brief one-line reason>"}]
+
+Findings to classify:
+${JSON.stringify(items, null, 2)}`;
+
+  const response = await fetch(ANTHROPIC_API_URL, {
+    method: 'POST',
+    headers: {
+      'x-api-key': apiKey,
+      'anthropic-version': '2023-06-01',
+      'content-type': 'application/json',
+    },
+    body: JSON.stringify({
+      model,
+      max_tokens: 2048,
+      messages: [{ role: 'user', content: prompt }]
+    })
+  });
+
+  if (!response.ok) {
+    const body = await response.text();
+    throw new Error(`Anthropic API error ${response.status}: ${body.slice(0, 200)}`);
+  }
+
+  const data = await response.json();
+  const text = data.content?.[0]?.text || '[]';
+  const jsonText = text.replace(/^```(?:json)?\s*/i, '').replace(/\s*```\s*$/i, '').trim();
+  let classifications;
+  try {
+    classifications = JSON.parse(jsonText);
+  } catch {
+    throw new Error('Claude returned non-JSON response');
+  }
+
+  return scanResults.map(({ file, findings }) => ({
+    file,
+    findings: findings.map(f => {
+      const id = `${path.relative(rootPath, file)}:${f.line}`;
+      const cl = classifications.find(c => c.id === id);
+      return {
+        ...f,
+        classification: cl?.classification ?? 'REAL',
+        reason: cl?.reason ?? ''
+      };
+    })
+  }));
+}
+
+// =============================================================================
+// CLAUDE CLASSIFICATION — VULNERABILITIES
+// =============================================================================
+
+/**
+ * Send vulnerability findings to Claude for classification + specific fix suggestions.
+ * Unlike secrets, code context is NOT masked — the pattern itself is the finding.
+ */
+async function classifyVulnsWithClaude(vulnResults, rootPath, apiKey, model) {
+  const items = [];
+  for (const { file, findings } of vulnResults) {
+    let lines = [];
+    try {
+      lines = fs.readFileSync(file, 'utf-8').split('\n');
+    } catch {
+      // include without context
+    }
+
+    for (const finding of findings) {
+      const startLine = Math.max(0, finding.line - 3);
+      const endLine = Math.min(lines.length - 1, finding.line + 1);
+      const context = lines.slice(startLine, endLine + 1).join('\n');
+
+      items.push({
+        id: `${path.relative(rootPath, file)}:${finding.line}`,
+        file: path.relative(rootPath, file),
+        line: finding.line,
+        type: finding.patternName,
+        severity: finding.severity,
+        codeContext: context  // Not masked — it's a code pattern, not a secret
+      });
+    }
+  }
+
+  const prompt = `You are a security expert reviewing code vulnerabilities.
+
+For each finding below, classify it and provide a specific fix if it's a real issue.
+
+- REAL: genuinely exploitable as written (user-controlled input reaches a dangerous sink)
+- FALSE_POSITIVE: safe in this context (static/hardcoded input, internal tool, test code, build script)
+
+For REAL findings: provide a concise, specific 1-line code fix showing what to change.
+For FALSE_POSITIVE: briefly explain why it's safe.
+
+Respond with a JSON array ONLY — no markdown, no explanation:
+[{"id":"<id>","classification":"REAL"|"FALSE_POSITIVE","reason":"<brief reason>","fix":"<specific fix code, or null>"}]
+
+Vulnerabilities to analyze:
+${JSON.stringify(items, null, 2)}`;
+
+  const response = await fetch(ANTHROPIC_API_URL, {
+    method: 'POST',
+    headers: {
+      'x-api-key': apiKey,
+      'anthropic-version': '2023-06-01',
+      'content-type': 'application/json',
+    },
+    body: JSON.stringify({
+      model,
+      max_tokens: 4096,
+      messages: [{ role: 'user', content: prompt }]
+    })
+  });
+
+  if (!response.ok) {
+    const body = await response.text();
+    throw new Error(`Anthropic API error ${response.status}: ${body.slice(0, 200)}`);
+  }
+
+  const data = await response.json();
+  const text = data.content?.[0]?.text || '[]';
+  const jsonText = text.replace(/^```(?:json)?\s*/i, '').replace(/\s*```\s*$/i, '').trim();
+  let classifications;
+  try {
+    classifications = JSON.parse(jsonText);
+  } catch {
+    throw new Error('Claude returned non-JSON response');
+  }
+
+  return vulnResults.map(({ file, findings }) => ({
+    file,
+    findings: findings.map(f => {
+      const id = `${path.relative(rootPath, file)}:${f.line}`;
+      const cl = classifications.find(c => c.id === id);
+      return {
+        ...f,
+        classification: cl?.classification ?? 'REAL',
+        reason: cl?.reason ?? '',
+        fix: cl?.fix ?? null
+      };
+    })
+  }));
+}
+
+// =============================================================================
+// OUTPUT
+// =============================================================================
+
+function printClassificationTable(classified, rootPath) {
+  const SEVERITY_COLOR = {
+    critical: chalk.red.bold,
+    high: chalk.yellow,
+    medium: chalk.blue
+  };
+
+  console.log(chalk.cyan('  Secret Classification'));
+  console.log(chalk.cyan('  ' + '─'.repeat(58)));
+  console.log();
+
+  for (const { file, findings } of classified) {
+    const relPath = path.relative(rootPath, file);
+    for (const f of findings) {
+      const isReal = f.classification === 'REAL';
+      const icon = isReal ? chalk.red('✗') : chalk.gray('~');
+      const label = isReal ? chalk.red('REAL') : chalk.gray('SKIP');
+      const sevColor = SEVERITY_COLOR[f.severity] || chalk.white;
+      const matchedShort = f.matched.length > 16 ? f.matched.slice(0, 16) + '…' : f.matched;
+
+      console.log(
+        `  ${icon}  ${label.padEnd(8)} ${chalk.white(`${relPath}:${f.line}`).padEnd(40)}  ` +
+        `${sevColor(f.patternName.padEnd(24))}  ` +
+        chalk.gray(matchedShort)
+      );
+      if (f.reason) {
+        console.log(chalk.gray(`            → ${f.reason}`));
+      }
+    }
+  }
+}
+
+function printVulnFixTable(classifiedVulns, rootPath) {
+  const SEVERITY_COLOR = {
+    critical: chalk.red.bold,
+    high: chalk.yellow,
+    medium: chalk.blue
+  };
+
+  const totalCount = classifiedVulns.reduce((n, { findings }) => n + findings.length, 0);
+  const realCount = classifiedVulns.reduce(
+    (n, { findings }) => n + findings.filter(f => f.classification === 'REAL').length, 0
+  );
+  const fpCount = totalCount - realCount;
+
+  console.log(chalk.yellow('  Code Vulnerability Analysis'));
+  console.log(chalk.yellow('  ' + '─'.repeat(58)));
+  console.log();
+
+  for (const { file, findings } of classifiedVulns) {
+    const relPath = path.relative(rootPath, file);
+    for (const f of findings) {
+      const isReal = f.classification === 'REAL';
+      const icon = isReal ? chalk.red('✗') : chalk.gray('~');
+      const label = isReal ? chalk.red('REAL') : chalk.gray('SKIP');
+      const sevColor = SEVERITY_COLOR[f.severity] || chalk.white;
+      const snippet = f.matched.length > 55 ? f.matched.slice(0, 55) + '…' : f.matched;
+
+      console.log(
+        `  ${icon}  ${label.padEnd(8)} ${chalk.white(`${relPath}:${f.line}`).padEnd(38)}  ` +
+        sevColor(`[${f.severity.toUpperCase()}]`)
+      );
+      console.log(chalk.gray(`            ${f.patternName}`));
+      console.log(chalk.gray('     Code:   ') + chalk.cyan(snippet));
+      if (f.reason) {
+        console.log(chalk.gray(`     Reason: ${f.reason}`));
+      }
+      if (isReal && f.fix) {
+        console.log(chalk.gray('     Fix:    ') + chalk.green(f.fix));
+      }
+      console.log();
+    }
+  }
+
+  if (realCount === 0) {
+    output.success(`Claude classified all ${totalCount} vulnerability/vulnerabilities as false positives!`);
+    console.log(chalk.gray('  Tip: Add # ship-safe-ignore on those lines to suppress future warnings.'));
+  } else {
+    console.log(chalk.yellow(`  ${realCount} real vulnerability/vulnerabilities require manual fixes.${fpCount > 0 ? chalk.gray(` ${fpCount} false positive(s) skipped.`) : ''}`));
+  }
+}