The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data which serves as a potential attack surface. * Ingestion points: The agent reads 'plan tasks', 'TEST conformance review findings', and content from 'guidelines files' into its context. * Boundary markers: The instructions specify following only one printed guidelines file per path, but they lack explicit delimiters or instructions to ignore embedded malicious commands within those files. * Capability inventory: The skill has the ability to write files (TDD loop), execute local shell scripts, and run analysis tools. * Sanitization: There is no mention of sanitizing or validating the content of the ingested tasks or guidelines before processing.
[COMMAND_EXECUTION]: The skill instructs the agent to execute specific local shell scripts and tools as part of its workflow. * Evidence: Commands include 'sh .guardrails/scripts/guidelines-file.sh TEST', 'grill-requirements', 'analyze-risks', and references to 'task-worktree.sh'. These are executed with arguments (file paths) derived from the task dispatch, which could lead to command injection if paths are maliciously crafted.

develop-tdd