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. While intended to facilitate LLM-based ideation, this provides a mechanism for shell command execution from within the skill.
[DYNAMIC_EXECUTION]: The script scripts/validate_candidate.py dynamically constructs a Python code snippet and executes it using uv run python -c. This involves interpolating filesystem paths and stage identifiers into a string that is subsequently executed as code.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from multiple external sources, including OHLCV parquet files, CSV/Parquet tables for news and futures, and outputs from external commands. This untrusted content is used to populate fields like hypothesis and description in generated research tickets. The mandatory evidence chain is as follows:
Ingestion points: Files provided via --ohlcv, --hints, --news-reactions, --futures-ohlcv, and stdout from --llm-ideas-cmd in scripts/auto_detect_candidates.py.
Boundary markers: None; content is inserted directly into YAML structures.
Capability inventory: The skill can execute shell commands via subprocess.run (found in auto_detect_candidates.py and validate_candidate.py) and write files to the local filesystem (export_candidate.py).
Sanitization: Identifiers are sanitized via regex, but the natural language content from external sources is not filtered for potential prompt injection patterns before being stored in tickets.

edge-candidate-agent