The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a local Node.js helper script located at scripts/segments.mjs to perform file packing and verification tasks. This script handles reading source files and writing report artifacts to ensure an auditable trail of original data.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze external 'independent reads', which represents an attack surface for indirect prompt injection. Specifically, the agent is instructed to 'follow the reproduction' provided in the data, which could potentially contain malicious commands or queries.
Mandatory Evidence Chain for Indirect Prompt Injection:
Ingestion points: Step 1 of the workflow receives 'round-0 reads' from external files.
Boundary markers: The scripts/segments.mjs script implements delimiters (<<<< begin ... and >>>> end ...) to encapsulate and isolate raw input data within the report.
Capability inventory: The skill allows for local script execution via node and instructions to execute reproduction steps provided in the input data.
Sanitization: The instructions include explicit safety checks, directing the agent to only follow reproductions that are 'safe and read-only' and to reject any reproduction deemed unsafe.

multithink