The Agent Skills Directory

[PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection through its knowledge management system.
Ingestion points: Data ingested from the database via db_query.py and fetch_structure.py (such as table comments or stored procedure definitions) is stored as domain knowledge in markdown files within the references/ directory.
Boundary markers: The 'Selective Loading Protocol' in SKILL.md loads these files into the agent's context without using boundary markers or safety instructions to ignore potentially malicious content embedded in the ingested data.
Capability inventory: The skill utilizes several Python scripts for database operations and file management, which could be abused if the agent is manipulated by injected instructions.
Sanitization: While SQL commands are validated against a read-only whitelist, the actual content retrieved from the database is not sanitized before being persisted as knowledge.
[COMMAND_EXECUTION]: Local shell command execution is used for administrative and lifecycle tasks.
scripts/setup.py uses subprocess.run to initialize the knowledge system by calling decay_engine.py.
scripts/tests/test_decay_engine.py uses subprocess.run to execute CLI tests for the engine.
scripts/setup.py creates a behavior-modifying rule file at .claude/rules/db-investigator-evolution.md to ensure the agent follows the knowledge capture protocol across sessions.
[EXTERNAL_DOWNLOADS]: The skill depends on the pymysql library for database connectivity. The scripts/setup.py script checks for its presence and directs the user to install it from official package registries.

db-investigator