[FEATURE]: Plugin UI surface for the web/desktop app, mirroring the TUI plugin API

The gap

V2 has a rich client-side plugin surface for the TUI (@opencode-ai/plugin/tui): Plugin.define({ id, setup }) with keymap.layer (slash commands with run(input) that execute before any model dispatch), ui.dialog (show/select/prompt/confirm), ui.slot (e.g. sidebar.content), host-reactive storage, and client. The loader resolves package plugins through the exports["./tui"] subpath.

The web/desktop app (packages/app) has no equivalent: no plugin runtime, no slot system, no client-side command hooks. The Extensions settings tab (packages/app/src/components/settings-v2/extensions.tsx) is a read-only list of server config. So any plugin UX built for the TUI simply cannot exist in the web or desktop clients.

Concrete use case

opencode-waits schedules a prompt for later (/wait 1hour implement this). Its whole point is working while rate limited, so scheduling must not cost a model turn. In the TUI that works: a plugin slash command intercepts before submission, plus a sidebar section with live countdowns and a manager dialog (list/search/add/edit/delete/send-now) built from ui.dialog.show + keymap layers.

In the web/desktop app the only available fallback is a server-side command template that instructs the agent to call a scheduling tool — one model turn per use, and unusable exactly when rate limited, which is the scenario the feature exists for.

The ask

A plugin UI surface for packages/app, even a small subset of the TUI's:

1. Client-side slash commands with a run(input) executed before prompt submission — this alone unlocks model-free plugin commands.
2. Dialogs (the select/prompt/confirm primitives).
3. A sidebar slot.

A natural shape would be an exports["./web"] entrypoint mirroring the ./tui loader convention, with a context exposing the same client/storage/ui capabilities adapted to solid-on-DOM.

I understand the V2 plugin API is beta and moving; happy to adapt the plugin to whatever shape this takes, and to test early iterations.