# XWUISuperAgentStudio

The authoring and review surface for LLM agents. `agent`, `prompt`, `tool`, `model`,
`evaluation` and `run` items sit in one hub so a prompt change, the evaluation that graded it and
the run that used it are one click apart.

## When to use

- A product that lets users build agents and needs a place to version prompts and tools.
- Evaluation review, where a run's transcript is mounted into the detail pane via `mountDetail`.
- Model and tool inventories that operators must be able to inspect and disable.

Agent actions are exactly the kind that must not run on a whim: mark anything that executes a
model or a tool with `requiresConfirmation` and a `capability`, in line with the estate's
autonomy-slider rule (default `suggest`).

## Configuration surface

The constructor is `new XWUISuperAgentStudio(container, data, config)`. `data` is the state
you already hold; `config` is how the hub talks to your backend.

`config.provider` is the only integration point that matters:

- `list(query, signal)` -> `{ items, total?, nextCursor? }`. `query` carries
  `text`, `sectionId`, `cursor`, `limit` and `sort`; honour `signal`, because
  every new search, section switch or refresh aborts the in-flight request.
- `execute(request)` -> `{ message?, items?, refresh? }`, required by any entry
  in `config.actions`. An action with no `execute` fails with
  `XWUI_SUPER_ACTION_PROVIDER_REQUIRED` rather than silently doing nothing.
- `subscribe(listener)` is optional; when present the hub reloads on every
  push and unsubscribes on destroy.

The rest of `config`: `title` / `subtitle` / `className` for chrome;
`columns` (each `key` indexes the row built from the item's own fields plus its
`values` bag, so a custom column needs a matching `values` entry); `actions`
scoped `global` | `selection` | `item`, optionally `requiresConfirmation` and
`capability`; `confirmAction(action, items)` to gate confirmations (without it
a confirming action only emits `confirmationRequired`); `mountDetail` to own
the detail pane; `context` for the `XWUISuperContext` that authorizes
capabilities; plus `searchable`, `selectable`, `selectionMode`
(`single` | `multiple`), `autoLoad` (default true, loads on connect when a
provider exists), `refreshAfterAction` (default true), `pageSize` (default 50),
`density` (`compact` | `comfortable` | `spacious`) and a `labels` bag for every
piece of built-in copy.

Authorization is fail-closed: an action carrying a `capability` with no
`config.context` authorizer is **denied**, not allowed. Leave `capability` off
actions that genuinely need no permission.

`data` seeds and mirrors the view: `items`, `sections` (the tab strip),
`metrics` (the tile row), `query`, `activeSectionId`, `activeItemId`,
`selectedIds`, `loading`, `loadingMore`, `actionPendingId`, `error`, `total`,
`nextCursor` and `lastActionMessage`. Omit `columns` and the table falls back to
Title / Status / Updated.

## Agent studio

Agents, prompts, tools, models, evaluations, and runs with explicit execution boundaries.

```example
file: examples/BasicUsage.ts
html: examples/BasicUsage.html
title: Agent studio
description: Agents, prompts, tools, models, evaluations, and runs with explicit execution boundaries.
```

```api
```