The Agent Skills Directory

[SAFE]: Technical analysis of the provided markdown and configuration files shows no evidence of malicious intent, unauthorized network communication, or credential harvesting. All external links point to the official Reactor library documentation on hexdocs.pm.
[NO_CODE]: This skill contains no executable code or scripts. It is composed entirely of instructional documentation and prompt templates that guide an agent's behavior during a review process.
[PROMPT_INJECTION]: The skill defines a workflow where external code is ingested and interpolated into a prompt for a subagent (references/reviewer-prompt.md). This creates an indirect prompt injection surface. However, the risk is assessed as low because the subagent is restricted to generating text verdicts and does not have access to sensitive tools or the ability to execute code.
Ingestion points: The {{TARGET_CONTENT_OR_PATHS}} placeholder in references/reviewer-prompt.md is used to load external code artifacts.
Boundary markers: The content is organized under markdown headers, though it lacks explicit 'ignore embedded instructions' delimiters for the ingested target content.
Capability inventory: Across the skill, there are no subprocess calls, network operations, or dynamic execution (exec/eval) capabilities exposed to the subagent.
Sanitization: No sanitization or verification of the external code input is performed prior to prompt interpolation.

adversarial-reactor