The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/auto_detect_candidates.py uses subprocess.run to execute an external command string provided via the --llm-ideas-cmd command-line argument. This allows for the execution of arbitrary programs with data inputs derived from market anomalies.
[DYNAMIC_EXECUTION]: The script scripts/validate_candidate.py dynamically constructs a Python snippet and executes it using subprocess.run(['uv', 'run', 'python', '-c', snippet], ...). This technique is used to perform schema validation against the trade-strategy-pipeline codebase at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted quantitative market data and user-provided hints, which are used to generate research tickets and passed to external ideation loops. This creates an entry point for indirect prompt injection.
Ingestion points: Data enters the skill via the --ohlcv, --hints, --news-reactions, and --futures-ohlcv file arguments in scripts/auto_detect_candidates.py.
Boundary markers: There are no explicit boundary markers or instructions to ignore embedded malicious content within the processed market data or YAML hints.
Capability inventory: The skill has the capability to execute shell commands and generate/execute dynamic Python code via its core automation scripts.
Sanitization: While the skill uses shlex.split and repr() to mitigate command and code injection, it lacks semantic validation of the processed market data to prevent it from influencing agent behavior or ideation output.

edge-candidate-agent