Interactive Task outputs

An Agent Task can return an ordinary answer, a file, or an optional interface. Start with the smallest output that makes the work useful. A Task does not need to generate an app.

Need	Use
Answer, report, patch, or bounded file	Text / Artifact
Small status, form, or controls	Live View
Temporary executable collaborative mini-app	Generated Task Room App
Backend, arbitrary networking, durable deployment, or larger application	External app

Text / Artifact

Text and artifacts are the default and smallest surface. Return an answer in the Task conversation, or attach a bounded file such as a report or patch. Keep durable work in a repository or another participant-owned system.

Live View

A Live View is a small declarative interface published for one Task. Free4Chat validates the data and renders its own components; it does not run arbitrary Agent HTML, JavaScript, CSS, or an iframe.

Use it for a progress/status view, a small form, parameter controls, or a structured result with a few simple interactions. Button and input actions can be deterministic and browser-local, so they do not need to start another Agent turn. The Room keeps one current canonical Live View snapshot; a participant's local interaction values are not copied to another browser. Visibility does not activate an Agent.

Generated Task Room App

When a Live View is too limited but deploying a real web application would be overkill, the Task Agent can publish a bounded temporary mini-app for the Room. Examples include a calculator, temporary control panel, visualization, small game, vote, or Task-specific collaborative interface.

The Generated Task Room App is an optional escape hatch. It has one Task-owned publication, a bounded self-contained bundle, sandboxed execution, Room-scoped lifetime, bounded shared state, and realtime collaboration. In V0 it has no general network access. The publication and state disappear when the Room expires.

When a Task uses this surface, the Agent returns one complete self-contained HTML document between the terminal Task-output markers. It does not serialize an internal JSON bundle. The Runtime parses the HTML5 document and normalizes it into the existing Generated Room App bundle before the usual validation and Task publication.

The V1 source subset requires exactly one non-empty <title>, a <head>, and a <body>. Put zero or more inline <style> elements in <head> and zero or one classic inline <script> as the final meaningful child of <body>. The script may be omitted. External scripts, modules, async/defer, executable scripts in <head>, external stylesheets, and other external resource dependencies are not supported. The Runtime derives the App title from <title>, uses an empty initial state, and keeps networkOrigins empty. Author normal markup, CSS, and JavaScript directly; quotes and backslashes do not need JSON escaping.

For example, the payload between the markers can be:

<!doctype html>
<html>
<head>
  <title>Printer</title>
  <style>body { font: 16px sans-serif; }</style>
</head>
<body>
  <h1>Printer</h1>
  <p>Status: <span id="status">Loading</span></p>
  <button id="refresh">Refresh</button>
  <script>
    document.querySelector("#refresh").addEventListener("click", async () => {
      const result = await free4chat.capabilities.observe("printer_status");
      document.querySelector("#status").textContent = result.ok
        ? JSON.stringify(result.value)
        : result.error;
    });
  </script>
</body>
</html>

It is not general app hosting, a backend runtime, or a way to publish a local service. Exact publication commands and protocol limits belong in the CLI reference and MCP Room API.

Using a Runtime local capability

When the originating Runtime has projected a local semantic capability into the Task context, a Generated Task App may call free4chat.capabilities.observe(id) or free4chat.capabilities.invoke(id, action, args). A participant's --capability advertisement is discovery metadata; it is not this Runtime descriptor and grants no authority. The capability descriptor also does not authorize an operation by itself.

Each operation must come from one explicit trusted Human control click and is limited to that deterministic operation. It runs through the existing Human↔originating-Agent private reliable participant lane to the Runtime and its external Adapter; it does not start another Harness/LLM turn. The result is bounded semantic data returned to the same App. If the Runtime or Adapter has departed or is unavailable, the call fails closed. The iframe receives no Adapter configuration, device endpoint/URI, queue name, hostname/IP, credentials, or Runtime host/session/routing identity.

Room App vs Generated Task Room App

Surface	Created by	Best for	Lifecycle
Live View	Task Agent	Small declarative UI	Task / Room
Generated Task Room App	Task Agent	Temporary executable mini-app	Room
Room App	Existing curated App	Reusable shared activity or tool	Room session
External app	External developer or system	Backend, networking, or durable deployment	External

A curated Room App is an existing activity opened inside the Room. A Generated Task Room App is produced by one Task and remains discoverable through that Room-scoped publication. Neither turns the Room into a permanent workspace.