The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted configuration and instruction data from the repository being reviewed, which could be used to influence agent behavior. \n
Ingestion points: The skill reads .agents/code-carve-roster.md, UBIQUITOUS_LANGUAGE.md, CLAUDE.md, and instruction files specified via the file: directive in the roster from the local filesystem. \n
Boundary markers: The skill does not define specific delimiters or security warnings when parsing instructions from these repository-controlled files. \n
Capability inventory: The skill possesses the ability to invoke sub-agents, execute shell commands for linting and type-checking, and write files to the local filesystem. \n
Sanitization: No explicit sanitization or validation of the content within the roster or instruction files is mentioned. \n- [COMMAND_EXECUTION]: The skill executes project-specific type-check and lint commands and may write proposed code remedies to scratch files for compilation checks. This is a functional requirement for the skill but represents a managed execution risk. \n- [DYNAMIC_EXECUTION]: The file: directive in the roster allows the skill to dynamically load and follow instructions from arbitrary file paths, creating a mechanism for executing logic defined outside the primary skill instructions.

code-carve