The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and branches logic based on output from the gbrain CLI tool (status and doctor reports). Because these reports may reflect data stored within the database pages or facts, there is a risk that malicious data could attempt to influence the agent's branching logic.
Ingestion points: Output from gbrain engine status --json and gbrain doctor in SKILL.md.
Boundary markers: Absent; tool output is processed directly without specific delimiters or isolation instructions.
Capability inventory: The skill uses the exec tool for database initialization and data migration.
Sanitization: While the skill notes that CLI output is redacted, it does not specify sanitization or escaping of the ingested data for the agent's prompt context.
[CREDENTIALS_UNSAFE]: The skill manages database access via environment variables such as SUPABASE_ACCESS_TOKEN and PGPASSWORD. Although it includes safety instructions to avoid echoing these secrets in the agent's response, the requirement to handle these high-privilege credentials within the agent's environment constitutes a risk surface.
[COMMAND_EXECUTION]: The skill utilizes the exec tool to run commands that modify the system environment, including Docker container management and local database creation. These operations are correctly gated by explicit opt-in flags (--allow-docker, --allow-create-db) to ensure user oversight.

postgres-adopt