The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The instructions in references/agent-dependency.md permit the runtime installation of audit tools such as pip-audit or safety via the package manager if they are missing from the environment.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to perform repository reconnaissance (e.g., using find, ls, and wc commands) and to execute various package-specific audit utilities like npm audit, pip audit, and cargo audit.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits a significant attack surface for indirect prompt injection because its core functionality requires reading and analyzing the entire contents of a potentially untrusted codebase.
Ingestion points: The repository source code is ingested during the Phase 1 reconnaissance process and subsequently by each specialized sub-agent (Security, Performance, Accessibility, etc.) as described in the references/ directory files.
Boundary markers: The sub-agent briefs do not include specific instructions, delimiters, or ignore-rules to prevent the agents from following malicious instructions that might be embedded within the source files they are auditing.
Capability inventory: The skill environment includes the Bash, Agent (for spawning sub-agents), Write, and Read tools, providing a wide range of actions that could be triggered by successful injection.
Sanitization: The instructions provide no guidance for filtering or sanitizing the content of the files read from the repository before they are processed by the LLM.

overnight-repo-auditor