The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where the agent reads and obeys instructions found in local log files, creating a vulnerability to data-driven instruction injection.\n
Ingestion points: Processes console.N.log, .log files, and .json reports within the validator_logs/ directory (as detailed in Step 3 and the Extract Prompt appendix).\n
Boundary markers: There are no boundary markers or instructions to the agent to ignore or sanitize embedded instructions within the ingested logs.\n
Capability inventory: The skill utilizes Bash and the Task tool (subagent delegation), and instructs the agent to fix issues by following arbitrary text-based instructions or invoking named skills.\n
Sanitization: No sanitization or validation of the 'Fix Instructions' or 'Fix Skill' names is performed before they are acted upon in Step 5.\n- [DYNAMIC_EXECUTION]: In Step 5, the skill dynamically determines its next course of action (which skill to call or which instructions to follow) based on string content parsed from log files at runtime. This allows data produced by external tools to steer the agent's logic.\n- [COMMAND_EXECUTION]: The skill executes the agent-validate command via Bash and potentially other commands derived from the 'Fix Instructions' found in the logs.

validator-run