fix(mosaic): close claudex isolation gaps S1 (Bedrock/Vertex bypass) + S2 (fail-open catches)

Addresses MS-LEAD secrev REQUEST CHANGES on #806. S1 (CRITICAL — REQ 2 defeated by provider switches + cloud creds): the env sweep did not neutralize AWS Bedrock / GCP Vertex routing. CLAUDE_CODE_USE_BEDROCK / CLAUDE_CODE_USE_VERTEX are routing switches whose mere presence makes Claude Code talk to the real Anthropic API via the ambient cloud credential chain, bypassing ANTHROPIC_BASE_URL (the loopback proxy) entirely — and the cloud credentials (AWS_SECRET_ACCESS_KEY, AWS_BEARER_TOKEN_BEDROCK, GOOGLE_APPLICATION_CREDENTIALS, …) passed straight through. - buildClaudexEnv now force-deletes the routing switches by exact name (CLAUDEX_FORCED_UNSET_ENV) regardless of value — a name pattern is the wrong model for a boolean switch. - CLAUDEX_CREDENTIAL_ENV_RE extended to the cloud-cred families (^AWS_, ^GOOGLE_APPLICATION_CREDENTIALS, ^GOOGLE_CLOUD_, ^GCP_) and the mid-string _KEY$ / _SECRET gap (STRIPE_SECRET_KEY, SSH_PRIVATE_KEY, AWS_SECRET_ACCESS_KEY). - Doc-comment updated to match reality. S2 (HIGH — fail-open contradicts documented fail-closed): defaultCanonicalizeIntended and defaultIsSymlink swallowed ANY fs error. Now they distinguish ENOENT (genuinely absent → safe to continue) from every other errno (ELOOP, EACCES, ENOTDIR, …) which rethrows and fails CLOSED. Tests (TDD red-first, 4 new specs red before the fix): Bedrock/Vertex switches + AWS/GCP creds swept while the loopback proxy stays the only route; mid-string _KEY/_SECRET closed; real-FS ELOOP (canonicalize) and ENOTDIR (isSymlink) fail closed; injected EACCES not swallowed. New-module coverage 100% stmts/lines, 93.75% branch. Full mosaic suite 1169 green; typecheck / lint / format:check green. No --no-verify. Refs #790 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-16 17:00:04 -05:00
parent ab8c9a2d4b
commit 79e2fa35d5
3 changed files with 176 additions and 18 deletions
--- a/packages/mosaic/README.md
+++ b/packages/mosaic/README.md
@@ -82,9 +82,14 @@ resolved dir can never be — or live under — the real `~/.claude`. A claudex
 session therefore cannot mutate your normal Claude Code config.

 **No token leakage.** claudex never reads the proxy's credential file. Claude
-Code is handed only `ANTHROPIC_AUTH_TOKEN=unused`; the entire credential-bearing
-env family (`ANTHROPIC_*`, `*_TOKEN`, `*_API_KEY`, `*_SECRET`, …) is stripped
-from the composed environment. The proxy holds the real OAuth credential.
+Code is handed only `ANTHROPIC_AUTH_TOKEN=unused` pointed at the loopback proxy;
+the entire credential-bearing env family (`ANTHROPIC_*`, `AWS_*`, `GOOGLE_CLOUD_*`,
+`GOOGLE_APPLICATION_CREDENTIALS`, `*_TOKEN`, `*_KEY`, `*_SECRET`, …) is stripped
+from the composed environment. The Bedrock/Vertex routing switches
+(`CLAUDE_CODE_USE_BEDROCK`, `CLAUDE_CODE_USE_VERTEX`, and the `_SKIP_*_AUTH`
+pair) are force-removed regardless of value — otherwise their mere presence
+would route Claude Code to the real Anthropic API via AWS/GCP and bypass the
+proxy. The proxy holds the real OAuth credential.

 **Model tiers (override via env).**

--- a/packages/mosaic/src/commands/claudex.spec.ts
+++ b/packages/mosaic/src/commands/claudex.spec.ts
@@ -1,5 +1,5 @@
 import { describe, it, expect, vi } from 'vitest';
-import { mkdtempSync, mkdirSync, symlinkSync, rmSync, lstatSync } from 'node:fs';
+import { mkdtempSync, mkdirSync, symlinkSync, rmSync, lstatSync, writeFileSync } from 'node:fs';
 import { tmpdir, homedir } from 'node:os';
 import { join } from 'node:path';
 import {
@@ -205,6 +205,68 @@ describe('resolveClaudexConfigDir', () => {
      rmSync(root, { recursive: true, force: true });
    }
  });
+
+  it('FAIL CLOSED: default canonicalizer rethrows a non-ENOENT error (ELOOP) instead of a literal fallback', () => {
+    // A symlink loop makes realpathSync throw ELOOP. The guard must NOT swallow
+    // it as "does not exist yet, keep walking up" and return a literal path —
+    // it must fail closed. (REQ 1: fails CLOSED on any uncertainty.)
+    const root = mkdtempSync(join(tmpdir(), 'claudex-loop-'));
+    try {
+      const a = join(root, 'a');
+      const b = join(root, 'b');
+      symlinkSync(b, a, 'dir'); // a -> b
+      symlinkSync(a, b, 'dir'); // b -> a  (loop)
+      const looped = join(a, 'home'); // canonicalizing this hits ELOOP
+      // No canonicalize dep → the real defaultCanonicalizeIntended runs.
+      expect(() =>
+        assertIsolatedConfigDir(looped, { realClaudeDir: join(root, '.claude') }),
+      ).toThrow();
+    } finally {
+      rmSync(root, { recursive: true, force: true });
+    }
+  });
+
+  it('FAIL CLOSED: default isSymlink rethrows a non-ENOENT error (ENOTDIR) rather than reporting "not a symlink"', () => {
+    // A candidate whose parent is a regular FILE makes lstat throw ENOTDIR.
+    // The post-create symlink check must fail closed, not treat it as safe.
+    const root = mkdtempSync(join(tmpdir(), 'claudex-notdir-'));
+    try {
+      const file = join(root, 'afile');
+      writeFileSync(file, 'x');
+      const candidate = join(file, 'child'); // parent is a file → ENOTDIR on lstat
+      expect(() =>
+        // Bypass the guard/mkdir side-effects; only the default isSymlink runs live.
+        resolveClaudexConfigDir(
+          { [CLAUDEX_CONFIG_DIR_ENV]: candidate },
+          {
+            realClaudeDir: join(root, '.claude'),
+            canonicalize: idCanon,
+            mkdir: () => {},
+            // isSymlink omitted → real defaultIsSymlink runs on the ENOTDIR path.
+          },
+        ),
+      ).toThrow();
+    } finally {
+      rmSync(root, { recursive: true, force: true });
+    }
+  });
+
+  it('FAIL CLOSED: an injected canonicalize throwing EACCES is not swallowed', () => {
+    const eacces = Object.assign(new Error('permission denied'), { code: 'EACCES' });
+    expect(() =>
+      resolveClaudexConfigDir(
+        { [CLAUDEX_CONFIG_DIR_ENV]: '/home/agent/custom-claudex' },
+        {
+          realClaudeDir: '/home/agent/.claude',
+          canonicalize: () => {
+            throw eacces;
+          },
+          mkdir: () => {},
+          isSymlink: () => false,
+        },
+      ),
+    ).toThrow(/permission denied/);
+  });
 });

 // ─── model-tier map (P3) ──────────────────────────────────────────────────────
@@ -290,6 +352,62 @@ describe('buildClaudexEnv — zero token leakage', () => {
    expect(env.PATH).toBe('/usr/bin');
  });

+  it('neutralizes Bedrock/Vertex provider switches so Claude cannot bypass the proxy (REQ 2)', () => {
+    // CLAUDE_CODE_USE_BEDROCK / _USE_VERTEX are ROUTING switches: their mere
+    // presence makes Claude Code route to AWS Bedrock / GCP Vertex against the
+    // ambient cloud credential chain — reaching the real Anthropic API and
+    // bypassing ANTHROPIC_BASE_URL (the loopback proxy) entirely. They MUST be
+    // gone from the composed env regardless of the launching env.
+    const env = buildClaudexEnv(
+      {
+        CLAUDE_CODE_USE_BEDROCK: '1',
+        CLAUDE_CODE_USE_VERTEX: '1',
+        CLAUDE_CODE_SKIP_BEDROCK_AUTH: '1',
+        CLAUDE_CODE_SKIP_VERTEX_AUTH: '1',
+        AWS_ACCESS_KEY_ID: 'AKIAREAL',
+        AWS_SECRET_ACCESS_KEY: 'realsecret',
+        AWS_SESSION_TOKEN: 'realsession',
+        AWS_BEARER_TOKEN_BEDROCK: 'bearer-bedrock-real',
+        AWS_REGION: 'us-east-1',
+        GOOGLE_APPLICATION_CREDENTIALS: '/home/agent/gcp.json',
+        GOOGLE_CLOUD_ACCESS_TOKEN: 'gcp-token-real',
+        PATH: '/usr/bin',
+      },
+      { configDir, models },
+    );
+    // Routing switches gone by construction.
+    expect('CLAUDE_CODE_USE_BEDROCK' in env).toBe(false);
+    expect('CLAUDE_CODE_USE_VERTEX' in env).toBe(false);
+    expect('CLAUDE_CODE_SKIP_BEDROCK_AUTH' in env).toBe(false);
+    expect('CLAUDE_CODE_SKIP_VERTEX_AUTH' in env).toBe(false);
+    // Cloud credentials swept — none of the Claude-capable creds survive.
+    expect(env.AWS_ACCESS_KEY_ID).toBeUndefined();
+    expect(env.AWS_SECRET_ACCESS_KEY).toBeUndefined();
+    expect(env.AWS_SESSION_TOKEN).toBeUndefined();
+    expect(env.AWS_BEARER_TOKEN_BEDROCK).toBeUndefined();
+    expect(env.AWS_REGION).toBeUndefined();
+    expect(env.GOOGLE_APPLICATION_CREDENTIALS).toBeUndefined();
+    expect(env.GOOGLE_CLOUD_ACCESS_TOKEN).toBeUndefined();
+    // The proxy routing is still the only path.
+    expect(env.ANTHROPIC_BASE_URL).toBe(CLAUDEX_PROXY_URL);
+    expect(env.ANTHROPIC_AUTH_TOKEN).toBe('unused');
+    expect(env.PATH).toBe('/usr/bin');
+  });
+
+  it('closes the mid-string _KEY / _SECRET gap (STRIPE_SECRET_KEY, SSH_PRIVATE_KEY)', () => {
+    const env = buildClaudexEnv(
+      {
+        STRIPE_SECRET_KEY: 'sk-live-real',
+        SSH_PRIVATE_KEY: '-----BEGIN OPENSSH PRIVATE KEY-----',
+        HARMLESS: 'kept',
+      },
+      { configDir, models },
+    );
+    expect(env.STRIPE_SECRET_KEY).toBeUndefined();
+    expect(env.SSH_PRIVATE_KEY).toBeUndefined();
+    expect(env.HARMLESS).toBe('kept');
+  });
+
  it('no credential-NAMED key in the composed env carries a real-looking value', () => {
    const env = buildClaudexEnv(
      {
--- a/packages/mosaic/src/commands/claudex.ts
+++ b/packages/mosaic/src/commands/claudex.ts
@@ -78,7 +78,12 @@ function defaultCanonicalizeIntended(p: string): string {
    try {
      const real = realpathSync(existing);
      return tail.length > 0 ? join(real, ...tail) : real;
-    } catch {
+    } catch (err) {
+      // ONLY a genuine "does not exist yet" (ENOENT) justifies walking up to an
+      // existing ancestor. Any other errno (ELOOP, EACCES, ENOTDIR, …) means we
+      // cannot establish the canonical form — fail CLOSED rather than fall back
+      // to a possibly-wrong literal path.
+      if ((err as NodeJS.ErrnoException).code !== 'ENOENT') throw err;
      const parent = dirname(existing);
      if (parent === existing) return abs; // reached root without an existing prefix
      tail.unshift(existing.slice(parent.length + 1));
@@ -94,8 +99,11 @@ function defaultMkdir(p: string): void {
 function defaultIsSymlink(p: string): boolean {
  try {
    return lstatSync(p).isSymbolicLink();
-  } catch {
-    return false;
+  } catch (err) {
+    // A missing path is genuinely "not a symlink"; anything else (EACCES, ELOOP,
+    // ENOTDIR, …) is uncertainty the guard must not swallow — fail CLOSED.
+    if ((err as NodeJS.ErrnoException).code === 'ENOENT') return false;
+    throw err;
  }
 }

@@ -200,16 +208,38 @@ export function resolveClaudexModels(env: NodeJS.ProcessEnv = process.env): Clau

 /**
 * Names of env vars considered credential-bearing. The whole family is stripped
- * from the composed env so no real Anthropic key, OAuth token, or third-party
- * secret can reach the local proxy or be used by Claude Code to bypass it. We
- * then re-add ONLY the safe claudex vars (`ANTHROPIC_MODEL`, `_SMALL_FAST_MODEL`,
- * `_BASE_URL`, `_AUTH_TOKEN=unused`). A name-pattern sweep can't miss a specific
- * var a denylist forgot, while still preserving the arbitrary non-credential env
- * the harness/MCP/hooks require (PATH, HOME, XDG, terminal, proxies, …), which a
- * strict allowlist would fragilely drop.
+ * from the composed env so no real Anthropic key, OAuth token, or third-party /
+ * cloud credential can reach the local proxy or be used by Claude Code to bypass
+ * it. We then re-add ONLY the safe claudex vars (`ANTHROPIC_MODEL`,
+ * `_SMALL_FAST_MODEL`, `_BASE_URL`, `_AUTH_TOKEN=unused`). A name-pattern sweep
+ * can't miss a specific var a short denylist forgot, while still preserving the
+ * arbitrary non-credential env the harness/MCP/hooks require (PATH, HOME, XDG,
+ * terminal, proxies, …), which a strict allowlist would fragilely drop.
+ *
+ * The cloud-provider families (`AWS_*`, `GOOGLE_APPLICATION_CREDENTIALS`,
+ * `GOOGLE_CLOUD_*`, `GCP_*`) are included because Claude Code can route to the
+ * real Anthropic API via AWS Bedrock / GCP Vertex using the ambient cloud
+ * credential chain — a Claude-capable credential that must never survive into a
+ * claudex launch. `_KEY$` / `_SECRET` (not just the `_API_KEY$` tail) close the
+ * mid-string gap (`STRIPE_SECRET_KEY`, `SSH_PRIVATE_KEY`, `AWS_SECRET_ACCESS_KEY`).
 */
 export const CLAUDEX_CREDENTIAL_ENV_RE =
-  /^ANTHROPIC_|^CLAUDE_CODE_OAUTH|_API_?KEY$|_TOKEN$|_SECRET$/i;
+  /^ANTHROPIC_|^CLAUDE_CODE_OAUTH|^AWS_|^GOOGLE_APPLICATION_CREDENTIALS$|^GOOGLE_CLOUD_|^GCP_|_API_?KEY$|_KEY$|_TOKEN$|_SECRET/i;
+
+/**
+ * Provider ROUTING switches whose mere PRESENCE (independent of any credential)
+ * makes Claude Code bypass `ANTHROPIC_BASE_URL` (the loopback proxy) and talk to
+ * the real Anthropic API via Bedrock/Vertex. A name-pattern is the wrong model
+ * for a boolean switch, so these are force-deleted by exact name — REGARDLESS of
+ * value — after the credential sweep. (REQ 2: isolation must hold for any
+ * launching env, including a Bedrock/Vertex-configured enterprise host.)
+ */
+export const CLAUDEX_FORCED_UNSET_ENV = [
+  'CLAUDE_CODE_USE_BEDROCK',
+  'CLAUDE_CODE_USE_VERTEX',
+  'CLAUDE_CODE_SKIP_BEDROCK_AUTH',
+  'CLAUDE_CODE_SKIP_VERTEX_AUTH',
+] as const;

 export interface BuildClaudexEnvOptions {
  configDir: string;
@@ -220,9 +250,10 @@ export interface BuildClaudexEnvOptions {

 /**
 * Compose the launch env for Claude Code. Returns a FRESH object (never mutates
- * the base env). Strips the entire credential-bearing family (REQ 2), then sets
- * the isolated config dir and the proxy routing. Claude Code sees only
- * `ANTHROPIC_AUTH_TOKEN=unused`; the proxy holds the real OAuth credential, which
+ * the base env). Strips the entire credential-bearing family AND force-deletes
+ * the Bedrock/Vertex routing switches (REQ 2), then sets the isolated config dir
+ * and the proxy routing. Claude Code sees only `ANTHROPIC_AUTH_TOKEN=unused`
+ * pointed at the loopback proxy; the proxy holds the real OAuth credential, which
 * this module never reads.
 */
 export function buildClaudexEnv(
@@ -235,6 +266,10 @@ export function buildClaudexEnv(
    env[name] = value;
  }

+  // Force-delete routing switches by exact name — their presence (not their
+  // value) is what would route Claude Code off the proxy to the real API.
+  for (const name of CLAUDEX_FORCED_UNSET_ENV) delete env[name];
+
  env['CLAUDE_CONFIG_DIR'] = opts.configDir;
  env['ANTHROPIC_BASE_URL'] = opts.baseUrl ?? CLAUDEX_PROXY_URL;
  env['ANTHROPIC_AUTH_TOKEN'] = 'unused';