The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands during runtime to handle dependencies.\n
Evidence: In scripts/stock_advisor.py, the _generate_pdf_with_fpdf2 method uses subprocess.check_call to run pip install fpdf2. Executing commands to modify the system environment during runtime is an unsafe practice for agent skills.\n- [EXTERNAL_DOWNLOADS]: The system is designed to download third-party software at runtime.\n
Evidence: The skill automatically attempts to download and install the fpdf2 library from PyPI if it is missing, which bypasses static dependency declarations and introduces risks from dynamic network activity.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via external financial data.\n
Ingestion points: The NewsAnalystAgent in scripts/agents/analysts.py processes raw news content fetched from akshare.\n
Boundary markers: System prompts do not use delimiters or isolation techniques to prevent external news content from being interpreted as instructions by the LLM.\n
Capability inventory: The agents have access to disk write operations (reports), network communication (DashScope and DingTalk), and subprocess execution (pip).\n
Sanitization: No sanitization or validation is performed on the news text before it is interpolated into the agent's context.

trading-agents