New AppifactRepl tool for building artifacts via a JavaScript REPL

What

• A new tool called AppifactRepl runs a full Node.js REPL (a read-eval-print loop, meaning it executes a whole JavaScript program and returns the result) against an Artifact's document store, via claude.use("db"), claude.use("files"), and claude.use("assets"), instead of making one Artifact store call at a time. It also supports claude.see() for attaching images or text to the tool's result, and can run a skill's own SDK script instead of the built-in one.
• Internally it spawns a child process (running node scripts/appifact_sdk.js --repl -) in a sandbox, and its output is scrubbed before being returned.
• The tool is wired into the permission system, the always-allowed tool list, tool-name resolution caching, and file-read tracking. A permission-check helper adds a special case for it: it is treated as disabled if there is a matching deny rule, or a permission rule whose content is only whitespace or asterisks.
• Writes made through AppifactRepl do not trigger the automatic artifact-opening behavior that normal file writes do.
• The tool is gated behind multiple conditions: it only activates when the app is launched with the remote_cowork entry point, an internal flag (tengu_buzzing_lightning, defaulting to false) is enabled, and a further enablement check (isAppifactReplToolEnabled()) whose implementation lives outside the shipped bundle and can't be fully inspected from source.

Why

This is groundwork for a more powerful way to build artifacts step by step with real code instead of individual tool calls, but it is not yet available to regular users: it's gated behind a server-side flag and only activates in the remote_cowork context, so most users won't see or be able to use it yet.