The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/auto_detect_candidates.py allows the execution of an arbitrary external command via the --llm-ideas-cmd parameter. Evidence: command_parts = shlex.split(llm_command) followed by subprocess.run(command_parts, ...) in scripts/auto_detect_candidates.py.
[REMOTE_CODE_EXECUTION]: The validate_with_pipeline_uv function in scripts/validate_candidate.py generates and executes Python code at runtime. Evidence: A multi-line Python script is constructed as a string snippet and executed via subprocess.run(["uv", "run", "python", "-c", snippet], ...).
[PROMPT_INJECTION]: The skill ingests untrusted data from multiple file formats which are then used in analysis or passed to external commands, creating a surface for indirect prompt injection. Ingestion points: scripts/auto_detect_candidates.py (OHLCV, hints, news, and futures files) and scripts/export_candidate.py (ticket YAML files). Boundary markers: None identified in the processing logic for external data. Capability inventory: subprocess.run for external commands and dynamic Python execution. Sanitization: Employs yaml.safe_load for parsing and repr() for escaping string literals in generated scripts.

edge-candidate-agent