The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a report and diff generated by an external persona named 'comment-sicko'. This creates an indirect prompt injection surface where the contents of the codebase being analyzed could potentially influence the agent's decision-making process.
Ingestion points: The agent reads and acts upon a report and diff generated by 'comment-sicko' which is based on potentially untrusted caller files.
Boundary markers: The instructions lack explicit boundary markers or warnings to ignore instructions that might be embedded in the code or report being processed.
Capability inventory: The skill is capable of deleting code, modifying files, and invoking several other skills such as 'architect', 'how', 'why', and 'pstack-harness'.
Sanitization: No sanitization or validation logic is defined for the external data being processed.
[PROMPT_INJECTION]: The instructions direct the agent to ignore or delete comments containing specific constraints such as 'do not remove', 'do not change wording', or 'talk to X before changing'. This is a form of instruction override that encourages the agent to bypass intentional developer protections or safety markers embedded in the code.
[COMMAND_EXECUTION]: The skill automates the construction and execution of CLI commands to spawn new roles using the 'pstack-harness' utility. This involves dynamic resolution of roles and briefings into shell commands based on external configuration files.

no-comments