registry / code-brain / 0.9.0

code-brain@0.9.0

Turn any codebase into searchable knowledge for Claude Code. LLM Wiki + AST Index + MCP Server.

AI Security Review

scanned 3h ago · by lpm-firewall-ai

LPM blocks this version under the AI-agent control-surface policy. The package performs unconsented install-time mutation of Claude Code control files in the installing project. It persists package-owned hooks that inject agent context and run npx code-brain on Claude sessions/tool use.

Static reason

High-risk behavior combination matched malicious policy.

Trigger

npm postinstall on package install or upgrade

Impact

Alters Claude Code behavior for the project without an explicit user command, steering agent tool choice and executing package CLI hooks in future sessions.

Mechanism

postinstall AI-agent control surface mutation

Policy narrative

Installing code-brain runs scripts/postinstall.cjs, which locates the consumer project root and writes Claude Code skill/instruction files plus .claude/settings.json hooks. Those hooks later execute npx code-brain session-start and pre-tool-nudge to inject context into Claude sessions and Grep tool use. The behavior is package-aligned and documented, but it is still unconsented postinstall mutation of a broad AI-agent control surface.

Rationale

Source inspection confirms concrete install-time writes into Claude Code control files, including persistent hooks that affect later agent behavior. Under the provided policy, unconsented npm postinstall mutation of a foreign/broad AI-agent control surface is blocking even without exfiltration. Product guard normalized a concrete AI-agent control hijack publish_block to the blockable dangerous-capability shape.

Evidence

package.jsonscripts/postinstall.cjsdist/hooks/adoption-hooks.jsskill/SKILL.mdbin/code-brain.jsdist/db/better-sqlite-driver.js.claude/skills/code-brain/SKILL.mdCLAUDE.md.claude/settings.json

Decision evidence

public snapshot

AI called this Malicious at 93.0% confidence as Dangerous Capability with low false-positive risk.

Evidence for policy block

package.json runs postinstall: node scripts/postinstall.cjs || true
scripts/postinstall.cjs writes .claude/skills/code-brain/SKILL.md during install
scripts/postinstall.cjs creates/appends CLAUDE.md with code-brain instructions
scripts/postinstall.cjs mutates .claude/settings.json hooks to run npx code-brain session-start and pre-tool-nudge
dist/hooks/adoption-hooks.js injects SessionStart and PreToolUse context steering Claude tool choice

Evidence against

README.md openly documents the auto-wired Claude hooks
Hook commands are package-aligned and call local code-brain CLI behavior
No credential harvesting or exfiltration endpoints found in inspected files
Dynamic require in dist/db/better-sqlite-driver.js only loads optional better-sqlite3
Shell use is limited to user-invoked CLI features such as graph opener and git helpers

Behavioral surface

Source

ChildProcessCryptoDynamicRequireEnvironmentVarsFilesystemShell

Supply chain

HighEntropyStringsUrlStrings

ManifestNo manifest risk signals triggered.

scanned 48 file(s), 236 KB of source, external domains: github.com

Source & flagged code

5 flagged

package.jsonView file

Lines 1-32json

2 "name": "code-brain",

3 "version": "0.9.0",

4 "description": "Turn any codebase into searchable knowledge for Claude Code. LLM Wiki + AST Index + MCP Server.",

5 "type": "module",

6 "bin": {

Source & flagged code

5 flagged · loading source

package.jsonView file

•scripts.postinstall = node scripts/postinstall.cjs || true

High

Install Time Lifecycle Scripts

Package defines install-time lifecycle scripts.

package.jsonView on unpkg

•scripts.postinstall = node scripts/postinstall.cjs || true

Medium

Ambiguous Install Lifecycle Script

Install-time lifecycle script is not statically allowlisted and needs review.

package.jsonView on unpkg

dist/db/better-sqlite-driver.jsView file

4import path from "node:path"; L5: const require = createRequire(import.meta.url); L6: /** Soft warning threshold — log when DB exceeds this */

Medium

Dynamic Require

Package source references dynamic require/import behavior.

dist/db/better-sqlite-driver.jsView on unpkg · L4

scripts/postinstall.cjsView file

4Install-time AI-agent control hijack evidence: L4: * 1. Install/UPGRADE the skill (version-stamped; refreshes on new versions) L5: * 2. Refresh the code-brain block in CLAUDE.md (version-stamped, marker-bounded) L6: * 3. Wire the deterministic adoption hooks into .claude/settings.json L7: * (SessionStart banner + PreToolUse Grep nudge) — merged, never clobbering ... L33: const src = path.join(pkgDir, "skill", "SKILL.md"); L34: const dest = path.join(projectRoot, ".claude", "skills", "code-brain", "SKILL.md"); L35: if (!fs.existsSync(src)) { console.log(" ⚠️ skill/SKILL.md missing in package"); return; } L36: L37: fs.mkdirSync(path.dirname(dest), { recursive: true }); L38: const srcBody = fs.readFileSync(src, "utf-8"); ... L40: if (!fs.existsSync(dest)) { L41: fs.writeFileSync(dest, srcBody); Payload evidence from skill/SKILL.md: L90: code_brain_recent_activity(module="auth", failures_only=true) L91: → ❌ [Apr 7] Tried WebSocket approach → too complex L92: ``` ... L168: Use when: finished editing, before commit. Catches regressions. L169: Exit code 1 if breaking changes found — safe to use in CI. L170: ... L217: domain: "backend/schedule", L218: body: "Finding: ...\nEvidence:…

Critical

Ai Agent Control Hijack

Install-time source drops package-supplied AI-agent/MCP control files or instructions.

scripts/postinstall.cjsView on unpkg · L4

16Manifest entrypoint (scripts.postinstall) carries capability families absent from dist/build output: environment+network L16: L17: const pkgDir = path.resolve(__dirname, ".."); L18: const projectRoot = findProjectRoot(pkgDir); ... L96: if (fs.existsSync(settingsPath)) { L97: try { settings = JSON.parse(fs.readFileSync(settingsPath, "utf-8")); } L98: catch { console.log(" ⚠️ .claude/settings.json is not valid JSON — skipping hook wiring"); return; } ... L126: L127: This project is indexed by [code-brain](https://github.com/kitara2005/code-brain). Prefer its MCP tools over blind Glob/Grep/Read. L128: ... L156: function readPkgVersion(dir) { L157: try { return JSON.parse(fs.readFileSync(path.join(dir, "package.json"), "utf-8")).version || "0.0.0"; } L158: catch { return "0.0.0"; }

High

Entrypoint Build Divergence

Manifest entrypoint contains risky behavior absent from dist/build output.

scripts/postinstall.cjsView on unpkg · L16

Findings

1 Critical2 High4 Medium4 Low

CriticalAi Agent Control Hijackscripts/postinstall.cjs

HighInstall Time Lifecycle Scriptspackage.json

HighEntrypoint Build Divergencescripts/postinstall.cjs

MediumAmbiguous Install Lifecycle Scriptpackage.json

MediumDynamic Requiredist/db/better-sqlite-driver.js

MediumEnvironment Vars

MediumStructural Risk Force Deep Review

LowScripts Present

LowFilesystem

LowHigh Entropy Strings

LowUrl Strings

scripts/postinstall.cjsView file

Lines 1-36javascript

1#!/usr/bin/env node

2/**

3 * Postinstall: runs after `pnpm add -D code-brain` (and on upgrade).

4 * 1. Install/UPGRADE the skill (version-stamped; refreshes on new versions)

5 * 2. Refresh the code-brain block in CLAUDE.md (version-stamped, marker-bounded)

6 * 3. Wire the deterministic adoption hooks into .claude/settings.json

7 * (SessionStart banner + PreToolUse Grep nudge) — merged, never clobbering

8 * 4. Print next steps

9 *

10 * The whole point of (1)–(3): make code-brain actually get used, and stay

11 * correct across upgrades, WITHOUT relying on the agent choosing to read prose.

12 * CommonJS (.cjs), zero deps, for max package-manager compatibility.

13 */

14const fs = require("node:fs");

15const path = require("node:path");

17const pkgDir = path.resolve(__dirname, "..");

18const projectRoot = findProjectRoot(pkgDir);

High

Entrypoint Build Divergence

Manifest entrypoint contains risky behavior absent from dist/build output.

scripts/postinstall.cjsView on unpkg · L16

19if (!projectRoot) process.exit(0); // dev mode (not inside node_modules) — skip

21const VERSION = readPkgVersion(pkgDir);

22console.log(`code-brain: Setting up (v${VERSION})...`);

24installSkill();

25upsertClaudeMdBlock();

26wireSettingsHooks();

27printNextSteps();

29// ---------------------------------------------------------------------------

31/** 1. Install or upgrade the skill, version-stamped so upgrades aren't frozen. */

32function installSkill() {

33 const src = path.join(pkgDir, "skill", "SKILL.md");

34 const dest = path.join(projectRoot, ".claude", "skills", "code-brain", "SKILL.md");

35 if (!fs.existsSync(src)) { console.log(" ⚠️ skill/SKILL.md missing in package"); return; }

Critical

Ai Agent Control Hijack

Install-time source drops package-supplied AI-agent/MCP control files or instructions.

scripts/postinstall.cjsView on unpkg · L4