The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes arbitrary user-defined commands specified in the loop.run.yaml configuration file to drive the hardening loop.
Evidence: Executes <target_cmd>, <oracle_cmd>, and <gate_cmd> to interact with the system under test.
Evidence: Utilizes git and gh CLI commands to manage code changes and create pull requests for the generated patches.
[EXTERNAL_DOWNLOADS]: The skill references and suggests the installation of external sibling skills provided by the same author.
Evidence: Instructs the user to run npx skills add gaasher/agent-loop-skills if required dependencies are missing. These are identified as vendor-owned resources.
[PROMPT_INJECTION]: The skill's primary function involves processing adversarial data, which creates an inherent surface for indirect prompt injection.
Ingestion points: Adversarial inputs generated during the red-team phase (stored in failures.cycle<N>.jsonl) are processed by the agent during the blue-team (fix) phase and passed to the <target_cmd>.
Boundary markers: The instructions do not specify explicit delimiters or "ignore instructions" wrappers for the adversarial payloads being analyzed.
Capability inventory: The skill possesses the ability to execute shell commands (target_cmd, gh) and perform file system writes (target_files).
Sanitization: No explicit sanitization or filtering of the adversarial input is described in the orchestration instructions.

purple-team