The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on several Shell scripts (grok-search.sh, tavily-search.sh, web-fetch.sh, etc.) for its core functionality. These scripts are executed by the agent via Bash. User-supplied queries are passed to these scripts and generally handled safely using jq to build JSON payloads for API requests.
[EXTERNAL_DOWNLOADS]: The docker-compose.yml and setup.sh scripts pull and run container images from GitHub Container Registry (ghcr.io) for third-party services including grok2api, TavilyProxyManager, and flaresolverr.
[CREDENTIALS_UNSAFE]: The skill is designed to manage and automate the import of sensitive credentials, including Grok SSO session tokens (JWTs) and Tavily/FireCrawl API keys. The import-keys.sh script processes these from a local file (export_sso.txt) and environment variables, sending them to the locally-hosted proxy services.
[INDIRECT_PROMPT_INJECTION]: As a web search and scraping tool, the skill has a significant attack surface for indirect prompt injection via search results or scraped web content.
Ingestion points: grok-search.sh, tavily-search.sh, and web-fetch.sh ingest data from external web sources into the agent's context.
Boundary markers: The SKILL.md provides explicit instructions for the agent to treat search results as untrusted third-party suggestions and mandates cross-verification from multiple independent sources.
Capability inventory: The skill allows the agent to execute subprocesses (Shell scripts) and perform network operations via curl through the local proxy services.
Sanitization: Content is fetched primarily in Markdown format; however, no specific sanitization logic is implemented to strip potential injection vectors from the fetched text before it is processed by the agent.

ultimate-search