Introduce a minimal operations console reachable on a dedicated UI port
(default 8081), kept separate from the OpenAI-compatible API port.
- ts_web/ holds the SPA shell (index.html, app.css, vanilla app.js) with
sidebar navigation for all views from readme.ui.md and a working
Overview page backed by a new /_ui/overview JSON endpoint.
- scripts/bundle-ui.ts walks ts_web/ and emits ts_bundled/bundle.ts, a
single generated module exporting every asset as base64. Mirrors the
@stack.gallery/registry pattern so deno compile binaries embed the
entire UI with no external filesystem dependency at runtime.
- ts/ui/server.ts (UiServer) serves assets from either the bundled map
(default, prod) or directly from ts_web/ on disk (dev). The source is
chosen per-config and can be overridden by UI_ASSET_SOURCE=disk|bundle.
SPA fallback routes unknown extensionless paths to index.html.
- IModelGridConfig.ui block with enabled/port/host/assetSource defaults;
config init writes the block, the normalizer fills in defaults on
load, and the daemon starts/stops the UI server alongside the API.
- deno.json gains a bundle:ui task; compile:all now depends on it so
released binaries always contain an up-to-date bundle. dev task sets
UI_ASSET_SOURCE=disk for hot edits.
- ts_bundled/ is gitignored (generated on build).
- test/ui-server.smoke.ts exercises bundle and disk modes end to end
(index, app.js, SPA fallback, /_ui/overview, 404).
Rework the implementation-notes section of readme.ui.md so the UI
delivery story matches ModelGrid's deno-compile single-binary shape.
Adopt the @stack.gallery/registry pattern: a build step bundles ts_web/
into a generated ts_bundled/bundle.ts exporting a path->bytes map that
typedserver serves at runtime. Add a dev-vs-prod asset-source switch so
UI edits stay hot-reloadable during development while release builds
embed the whole console in the binary.
Sketch a browser-based operations console for ModelGrid, served by the
daemon on a dedicated UI port. Lays out top-level IA (overview, cluster,
gpus, deployments, models, access, logs, metrics, settings), per-view
content, key user journeys, realtime/auth expectations, and open
questions. Structural idioms adapted from the dcrouter Ops dashboard and
grounded in ModelGrid's own cluster/catalog/vLLM domain.