The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from external URLs to generate literature surveys, creating a pathway for indirect instructions to influence the agent.
Ingestion points: The agent uses WebFetch to read from a user-provided source_url (e.g., arXiv papers or GitHub repositories) as defined in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands when the fetched content is interpolated into the final LLM prompt in scripts/build-survey.js.
Capability inventory: The skill can write files to the local system (fs.writeFileSync), execute local scripts (execFileSync), and make network requests to LLM provider APIs.
Sanitization: The script performs basic sanitization for markdown table safety but does not filter the core content for potential prompt injection vectors.
[COMMAND_EXECUTION]: The scripts/build-survey.js file uses execFileSync to invoke a local utility script, wiki-cli.js, to index the generated survey artifacts.
[DYNAMIC_EXECUTION]: The script performs a dynamic require() of a local database library (dist/db/store.js) using a path constructed at runtime.

survey-generator