The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill automatically installs numerous third-party packages and frameworks from public registries (NPM, PyPI, RubyGems) during runtime. Examples include npm init playwright@latest, npm install -D vitest supertest, and pip install pytest httpx.
[COMMAND_EXECUTION]: It executes a wide range of shell commands to manage infrastructure, including starting Docker containers (docker compose up -d), running database migrations (npx prisma migrate deploy), and starting development servers for various frameworks.
[DYNAMIC_CODE_EXECUTION]: The skill generates new test scripts (TypeScript, Python, Go, etc.) based on discovered API endpoints and UI routes, then immediately executes this generated code within the project environment.
[UNSAFE_CODE_MODIFICATION]: A 'Self-healing Fix Loop' allows the agent to autonomously modify the application's source code to resolve 'APP BUGS' detected during test execution. This capability can lead to unintended side effects or the introduction of vulnerabilities if the agent's logic is flawed.
[DATA_EXPOSURE]: The skill attempts to read from and write to the agent's internal metadata directory at ~/.claude/projects/ to record 'Self-Evolution Telemetry', which is an unauthorized access to the host's configuration space.
[PROMPT_INJECTION]: The instructions explicitly command the agent to 'Do NOT ask the user questions' and 'Make decisions yourself', which effectively disables human-in-the-loop safety mechanisms for destructive actions like code modification and package installation.

e2e