The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local Elixir and Phoenix development tools such as mix compile, mix test, mix hex.audit, and mix sobelow. It also utilizes standard shell utilities like grep, find, and wc to analyze the project's structure and code quality. These commands are executed within the local environment to generate reports and do not pose a security risk in the context of a project audit.
[DATA_EXFILTRATION]: The skill accesses codebase files (e.g., lib/, test/) and writes audit results to the .claude/audit/ directory. No network-based exfiltration or access to sensitive system directories (e.g., .ssh, .aws) was detected. The data processing is strictly local and serves the intended auditing functionality.
[PROMPT_INJECTION]: The skill operates as an audit tool and is naturally exposed to an indirect prompt injection surface when scanning untrusted project code.
Ingestion points: The skill reads source code and comments from the lib/ and test/ directories via grep and subagent tools.
Boundary markers: There are no explicit delimiters used in subagent prompts to separate the analyzed codebase content from the agent's internal instructions.
Capability inventory: The subagents possess the ability to run shell commands and write to the local filesystem to perform their auditing tasks.
Sanitization: No specific sanitization or filtering of the ingested source code is performed before it is analyzed by the specialist subagents. While this presents a theoretical attack surface, it is a standard characteristic of codebase analysis tools.

audit