[Widget] Add auto-start-text attribute for server-driven opening messages

[teotoplak]

Motivation

PostNord (Slack thread) wants the agent to greet widget users with a personalized, server-generated first message: the client passes only a customer_hash_id dynamic variable, a workflow tool resolves it to customer context, and the agent greets with that context.

The backend already supports this: a workflow whose start path leads to a tool node forces the first agent turn on connect (RuntimeStartAgent), and entry_behavior: generate_immediately does the same for agent-node entries. Mobile/JS SDK users get this today because they control when the session starts.

The embed widget is the one place it cannot work: in text mode the widget renders the configured first message locally and only opens the connection when the user sends their first message, so the agent never gets a chance to speak first.

What this does

Adds an opt-in auto-start-text attribute:

<elevenlabs-convai
  agent-id="..."
  auto-start-text="true"
  dynamic-variables='{"customer_hash_id": "abc123xyz"}'>
</elevenlabs-convai>

When set, the widget starts a text-only session as soon as the chat panel opens (immediately with default-expanded), and the opening message arrives over the wire from the agent. Scoped to text sessions on purpose: auto-starting voice would trigger a mic permission prompt on open.

Behavior notes:

• Starts at most once per widget mount; a session the user ended does not restart on its own.
• No-op when the agent does not support text or the attribute is absent, and terms acceptance still gates the start.
• Session start goes through the existing startSession path, so the elevenlabs-convai:call event still fires and can be used to inject fresh config.

Testing

• New browser tests: session auto-starts on open and renders the server-sent greeting; without the attribute nothing connects.
• Full convai-widget-core suite passes except DismissButton.test.ts ("expandable widget should hide all elements when dismissed"), which fails identically on clean main.

Screenshot from the browser test (MSW-mocked backend): the panel just opened, no user input, and the server-sent greeting is rendered. Real-agent capture pending an agent configured with a start-to-tool workflow.

Follow-ups

• Document the attribute in elevenlabs/docs.
• Optionally expose this as a dashboard widget setting so it can be enabled without touching the embed snippet.

🤖 Generated with Claude Code