The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from various external sources which could contain malicious instructions intended to influence the agent's behavior during report generation.
Ingestion points: The agent reads content from dependency files (e.g., requirements.txt, package.json, go.mod), Sourcegraph search results (code snippets and comments), and GitHub pull request titles and metadata.
Boundary markers: The skill instructions provide some behavioral boundaries, such as "The LLM summarizes but never invents severity ratings" and "Don't assess vulnerability severity yourself," which help reduce the likelihood of the agent obeying embedded instructions.
Capability inventory: The skill utilizes read_file, list_files, keyword_search, nls_search, and the gh CLI.
Sanitization: No explicit data sanitization or escaping of the ingested content is defined before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill executes gh (GitHub) CLI commands to retrieve pull request data.
Evidence: The workflow specifically invokes gh pr list with various flags and JSON output formatting to gather dependency-related PR information from configured repositories.

dependency-audit