The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests user input via the $ARGUMENTS variable and passes it as a 'Semantic Query' to the refactor:code-simplifier subagent. There are no boundary markers or sanitization steps to prevent the user input from overriding the subagent's instructions.
Ingestion points: SKILL.md and the prompt template in references/agent-configuration.md use unvalidated user input.
Boundary markers: None; user input is enclosed in quotes but not otherwise delimited or protected.
Capability inventory: The subagent has Edit permissions, allowing it to modify files in the repository based on the (potentially manipulated) query.
Sanitization: No filtering or validation of the semantic query is performed before interpolation.
[COMMAND_EXECUTION]: The skill is configured to 'Execute immediately without user confirmation.' This allows the agent to autonomously apply changes to the codebase via the Edit tool. While this is the intended functionality, the lack of human-in-the-loop oversight combined with the potential for injection increases the risk of harmful code modifications.
[SAFE]: The 'Aggressive Mode' refactoring instructions include deleting try-catch blocks and removing defensive null checks. Although these are intended for code cleanup, they represent a reliability risk if the agent's assessment of code safety is incorrect. These actions are documented as primary features of the skill.

refactor