The Agent Skills Directory

[COMMAND_EXECUTION]: The skill directs the agent to perform repository diagnostics using commands such as git log, git blame, and grep, as well as running existing test suites. These operations are intended for the primary purpose of auditing technical items. Security risk is significantly mitigated by explicit instructions that all commands must be read-only and that modifying code is strictly forbidden.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted technical descriptions (e.g., bug reports, feature requests) and uses them to guide its analysis. A malicious description could attempt to manipulate the agent's diagnostic commands. However, the risk is categorized as low due to the restricted scope of the commands and the requirement for evidence-based conclusions.
Ingestion points: User-provided technical item descriptions in SKILL.md.
Boundary markers: No explicit boundary markers or delimiters for untrusted data are defined in the output templates.
Capability inventory: Local shell execution for git, grep, and diagnostic test runners.
Sanitization: No specific sanitization or escaping of the user-provided descriptions is performed before the verification phase.
[SAFE]: The skill incorporates strong behavioral constraints, including a mandatory verification phase and a read-only rule, which align with AI safety best practices for technical decision-making assistants.

decision-gate