Agent Reach

Agent Reach in ScanSci

Use the built-in agent_reach tool when the user asks to read or search a public internet source outside the local evidence library.

Routing contract

• Use operation=status to inspect channel readiness when the user asks what is available or a route has failed.
• Use operation=read with the complete public URL for a page, feed, GitHub repository/issue, or video page.
• Use operation=search with a focused query. Set channel to github, bilibili, v2ex, youtube, or another named channel when the user names a platform; otherwise use auto or web.
• Treat search results as discovery leads. Treat page/JSON/RSS content as the retrieved source, and preserve the returned URL and backend in the answer.
• Keep the boundary with web-access explicit: use browser_access when a public reader returns an anti-bot page, a login wall, incomplete dynamic content, or when the user asks for browser-context access.
• Do not claim that a login, browser interaction, cookie, subtitle download, or external CLI was used unless the tool result explicitly reports it.
• Do not ask the user to install Agent-Reach. ScanSci already ships this capability; optional tools are only enhancements reported by status.

Safety

The built-in route is read-only, accepts public HTTP(S) URLs only, does not inject cookies, and does not run arbitrary shell commands. If a platform requires authentication or returns an anti-bot page, use the separate read-only browser_access bridge when it is available; otherwise report that boundary and offer the returned public fallback or ask the user to provide an authorized source.