The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and present output from an external helper tool (referred to as 'Pi'). This ingestion of data from another process constitutes a potential surface for indirect prompt injection if the helper tool's output were to contain malicious instructions intended to influence the agent's behavior.
Ingestion points: SKILL.md defines how the agent must handle 'Pi output' including findings, verdicts, and summaries.
Boundary markers: The instructions explicitly mandate preserving 'evidence boundaries' and keeping specific structures for findings and residual risks, which helps maintain the integrity of the data.
Capability inventory: The skill significantly restricts capabilities by including a critical directive that forbids the agent from making any code changes or auto-applying fixes based on the review output without explicit user consent.
Sanitization: While no specific character escaping is mentioned, the skill enforces a highly structured output format and requires human-in-the-loop verification, which serves as a primary defense against unintended command execution.

pi-result-handling