/** * The names the built shell document travels under, shared by the build-time * emitter (`./vite`) and the runtime loaders (`./shell-html`, `./worker`). * * In its own module so runtime loaders can import the constants without * dragging in the build-time plugin machinery (esbuild, child-process Vite). */ /** * Where `mcpAppsShellAsset` emits the STABLE-NAMED copy of the built shell * document, alongside the content-hashed one. * * The hashed name exists for browsers: the console's artifact page loads the * shell by URL, and a content hash is what lets that URL be cached forever. A * Worker serving the `ui://executor/shell.html` MCP resource has the opposite * problem — it fetches the document server-side through the ASSETS binding, * at runtime, from a bundle that cannot know the hash the client build * produced. A second copy at a name both sides agree on ahead of time is the * seam. No staleness risk: the binding serves the current deployment's * assets, not a browser cache keyed on the URL. */export const MCP_APPS_SHELL_STABLE_ASSET_PATH = "/assets/executor-mcp-apps-shell-stable.html"; /** * A fragment every built shell document contains (the `<meta>` in * `src/shell/mcp-app.html`), and no other document does. * * Exists because a fetch through a Workers ASSETS binding cannot be trusted * by status alone: with `not_found_handling: "single-page-application"` * (host-cloudflare), a miss returns the console's `index.html` with a 200. * Serving THAT as the shell resource would be the exact silent-hang failure * this seam exists to kill — the client loads a real document that never * speaks the MCP-Apps protocol. Loaders verify the marker and fail loudly. */export const MCP_APPS_SHELL_DOCUMENT_MARKER = 'name="executor-mcp-apps-shell"';