The Agent Skills Directory

COMMAND_EXECUTION (SAFE): The skill is designed to execute SQL queries as its primary function. It uses the standard pg library and does not execute arbitrary system commands. This behavior is expected and correctly scoped to the primary purpose of the skill.
EXTERNAL_DOWNLOADS (SAFE): The skill setup script (pgtool.ps1) uses bun install to download standard dependencies from the npm registry. No untrusted third-party script downloads or piped remote executions (curl|bash) were detected.
CREDENTIALS_UNSAFE (SAFE): Database credentials are managed securely. The skill does not contain hardcoded secrets and instead relies on a user-provided .pgtool.json file or environment variables to establish connections.
INDIRECT PROMPT INJECTION (SAFE): The skill processes data from database tables which could theoretically contain untrusted instructions.
Ingestion points: Database rows and metadata returned from SQL queries in commands like query, sample, and describe.
Boundary markers: Documentation in SKILL.md includes best practices for the agent, such as using LIMIT and parameterized values, though explicit boundary delimiters are not enforced in code.
Capability inventory: SQL execution via the pg library in connection.ts across multiple command files.
Sanitization: Metadata queries are parameterized; the raw query command is the intended interface for user-directed database interaction and does not auto-execute untrusted strings from DB content.

pgtool