The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill interacts with numerous external endpoints to fetch search results, including Google APIs, Firecrawl, AI HOT (virxact.com), FxTwitter, and Weibo. These connections are legitimate for a search router and are implemented using secure practices such as rejecting redirects to prevent credential forwarding.
[COMMAND_EXECUTION]: The skill leverages local CLI tools (like gh, yt-dlp, opencli, and zhihu-cli) to perform platform-specific searches. These tools are invoked via subprocess.run using safe list-based arguments rather than shell strings. Furthermore, the skill explicitly filters the environment variables passed to these child processes to ensure that unrelated system secrets (like OpenAI or AWS keys) are not exposed to third-party tools.
[INDIRECT_PROMPT_INJECTION]: As a search tool, it naturally ingests untrusted data from web snippets and summaries. This vulnerability surface is mitigated by returning results in a structured JSON envelope, labeling content clearly as unverified 'candidates', and performing rigorous offline URL syntax checks to ensure only public HTTP(S) addresses enter the agent context, effectively blocking attempts to probe the local network or file system.
[SAFE]: Local secret management is handled carefully. API keys and internal HMAC keys (used for signing search results) are read from the environment or private local files with forced 0600 permissions and ownership checks. The planner logic is entirely offline, preventing the router itself from being manipulated by remote data during the planning phase.

yichen-unified-search