Claude Code Sdk

WarnAudited by ClawScan on May 10, 2026.

Overview

This is a coherent Claude Code wrapper, but it gives an external coding agent broad file-editing and shell authority while strongly encouraging autonomous use, so it should be reviewed carefully before installation.

Install this only if you are comfortable letting a delegated Claude Code agent read and modify project files and run shell commands. Use it in a specific workspace, review changes before accepting them, avoid broad prompts, and start a fresh session when switching projects or tasks.

Findings (4)

Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.

High

#ASI02: Tool Misuse and Exploitation

What this means

If invoked on the wrong directory or with an overly broad prompt, the delegated agent could modify files or run commands in the user's environment.

Why it was flagged

The wrapper delegates broad read, write, edit, shell, search, and skill-use tools to Claude Code. That is relevant to a coding agent, but the artifacts do not show approval gates, path restrictions, dry-run behavior, or rollback controls for high-impact local actions.

Skill content

allowedTools: [
        "Read",
        "Edit",
        "Bash",
        "Write",
        "Glob",
        "Grep",
        "Skill",
      ],

Recommendation

Restrict execution to an explicitly approved project directory, require confirmation before Bash/Edit/Write actions, and prefer patch review or dry-run workflows for destructive or broad changes.

Medium

#ASI01: Agent Goal Hijack

What this means

The agent may route coding questions to Claude Code even when the user expected a simple explanation, review, or small direct edit.

Why it was flagged

The skill instructs the host agent to always rely completely on Claude Code for software-development tasks instead of doing the work directly, which can force a tool choice beyond a user's explicit request.

Skill content

在任何与软件开发相关的任务场景中，你都必须完全依赖Claude Code来完成任务，而不是直接进行代码编写。

Recommendation

Change the instruction to make Claude Code optional and user-directed, and ask the user before invoking it for broad or mutating coding work.

Low

#ASI04: Agentic Supply Chain Vulnerabilities

What this means

Future or different global versions of the SDK could change behavior outside what is visible in these artifacts.

Why it was flagged

The skill depends on a globally installed npm package without a pinned version in the artifact, so the reviewed skill does not fully define the code that will run.

Skill content

npm install -g @anthropic-ai/claude-agent-sdk

Recommendation

Pin the dependency version, document the trusted source, and prefer an install spec or lockfile so users know exactly what code is being installed.

Low

#ASI06: Memory and Context Poisoning

What this means

Context from one project or task could influence later work if the same session is reused too broadly.

Why it was flagged

The skill encourages resuming prior Claude Code sessions, which intentionally reuses context and can also reuse stale, unrelated, or poisoned instructions if sessions are not scoped carefully.

Skill content

除非你必须要开始一个全新的开发项目，否则都要使用--resume参数，让Claude Code在连续的上下文中进行工作

Recommendation

Resume only within the same project and task, start fresh sessions for unrelated work, and avoid sharing session IDs unnecessarily.