The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface for indirect prompt injection by processing untrusted code input and passing it to specialized subagents via the Task tool.
Ingestion points: The $ARGUMENTS variable, which contains the code to be reviewed, is interpolated into natural language prompts for subagents across all four review phases in SKILL.md.
Boundary markers: The skill includes high-level prose instructing the agent to use read-only scopes, but it lacks technical delimiters (such as XML tags or unique markers) or specific instructions to the subagents to ignore potential commands embedded in the code samples.
Capability inventory: The skill orchestrates subagents using the Task tool (e.g., security-auditor, performance-engineer) and instructs them to interface with static analysis and security tools like SonarQube, Snyk, and GitLeaks, implying network and file system access.
Sanitization: There are no explicit sanitization, escaping, or filtering steps defined for the content of $ARGUMENTS before it is placed into the prompts for sub-execution.
[SAFE]: The skill's instructions specifically mention integration with well-known and trusted technology services and security tools, including Snyk, SonarQube, CodeQL, Trivy, and GitLeaks. These references are documented neutrally as they align with the skill's purpose of professional code auditing.
[SAFE]: The skill contains explicit safety constraints and instructions for the agent, such as requiring authorization before running security probes, freezing the diff to maintain read-only scopes, and acknowledging the limitations of static analysis.

comprehensive-review-full-review