The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to trigger when it sees specific markers like GBRAIN_DB_ACCESS <reason> within MCP tool results or command output. This establishes a vulnerability surface where untrusted external data (such as a database error message or a document containing the trigger string) can influence the agent's behavior.
Ingestion points: The agent monitors gbrain command stderr and MCP tool results for trigger markers.
Boundary markers: The instructions serve as a logic boundary, explicitly stating the marker is a symptom to diagnose, not a command to execute.
Capability inventory: The skill uses the exec tool to run gbrain db-repair (diagnosis and repair) and gbrain engine status (verification).
Sanitization: The skill mandates the use of hardcoded commands and explicitly forbids parsing commands from the input marker to prevent command injection.
[COMMAND_EXECUTION]: The skill performs shell command execution to diagnose and fix database issues.
Evidence: It uses gbrain db-repair --json for diagnosis and gbrain db-repair --yes for applying fixes.
Scope: The commands are restricted to the specific gbrain utility and utilize hardcoded flags, reducing the risk of arbitrary command execution. The skill also requires user confirmation for the 'rewrite tier' which modifies configuration files.

db-repair