The Agent Skills Directory

[CREDENTIALS_UNSAFE]: Hardcoded sensitive credentials found in configuration files:
config/mcp_config.json: Contains a hardcoded Upstash Context7 API key (ctx7sk-d78a61e2-9647-4224-9c8b-f5a679e04741).
config/settings.json: Contains a hardcoded Anthropic-compatible authentication token (0b00b813538c416fbb08ea849a4d231a.wAZH2t1Vjt9fP9zQ) used with the https://open.bigmodel.cn/api/anthropic endpoint.
[PROMPT_INJECTION]: The skill creates a major vulnerability for 'Indirect Prompt Injection'. It is designed to synchronize and overwrite the agent's global instructions (CLAUDE.md) and persona definitions (output-styles) from a user-specified remote GitHub repository. A malicious repository could use this to persistently hijack the agent's behavior or disable safety constraints.
Ingestion points: External GitHub repository cloned during setup (referenced in SKILL.md).
Boundary markers: None. The sync script directly overwrites existing behavior files without validation.
Capability inventory: scripts/sync_env.py has the ability to write to ~/.claude/CLAUDE.md, ~/.claude/output-styles/, and ~/.claude.json.
Sanitization: None. Data from the remote source is copied directly into the local environment.
[COMMAND_EXECUTION]: The skill includes Python scripts (sync_env.py, backup_env.py, restore_env.py) that perform extensive file system operations, including recursive directory removal and file overwriting in sensitive application paths.
[EXTERNAL_DOWNLOADS]: The skill instructions and configuration files encourage cloning external code and installing multiple third-party MCP servers via npx from public registries, which can lead to the execution of unverified code.

env-setup