Conversation export

Conversation export adds a More options menu to the chat header with three actions:

• Copy as Markdown puts a transcript on the clipboard.
• Download as Markdown saves the same transcript as a .md file.
• Download as JSON saves the raw message array as a .json file.

Visitors use it to share an answer with a teammate, attach a transcript to a support ticket, or keep a record. It is off by default. Everything happens in the visitor’s browser, and nothing is sent to your worker.

Enabling it

<ChatWidget apiUrl="https://api.example.com" conversationExport />

<script>
  window.ClaudiusConfig = {
    apiUrl: "https://api.example.com",
    conversationExport: true,
  };
</script>

<claudius-chat api-url="https://api.example.com" conversation-export></claudius-chat>

In a client config, set "widget": { "conversationExport": true } and regenerate the snippet.

Only the literal true enables it. A value such as the string "false", which a CMS template can easily produce, leaves it off. The web component attribute is stricter than the others for the same reason: it enables export only when conversation-export is present with no value or set to "true" (case and surrounding spaces do not matter). Every other value leaves it off, including "False", which a Python or Jinja template writes for a false value, and "0" and "no".

The actions are greyed out while the conversation is empty and while a reply is still arriving, so a transcript never ends in a half-finished answer.

The Markdown transcript

# Chat transcript

Exported Sep 19, 2026, 3:32 PM (GMT+01:00)

## User · Sep 19, 2026, 3:03 PM

What does the error in this screenshot mean?

Attachments:

- `error_v2.png` (image/png, 48 KB)

## Assistant · Sep 19, 2026, 3:03 PM

The response body is not valid JSON. Guard the parse:

```js
const data = JSON.parse(text);
```

Sources:

1. [Parsing JSON](https://example.com/docs/json)

• The transcript is written in the widget’s language, and dates follow the visitor’s regional format and time zone. The zone is given once, as a UTC offset.
• Message text is exported as written, so code blocks, lists, and links survive. A code block that a stopped reply left open is closed, so it cannot swallow the messages after it. Line breaks are kept as the visitor saw them.
• Citations become a numbered list of links. A source whose URL is not http: or https: keeps its title and loses its link.
• Attachments are listed by name, type, and size. The file itself is never included, and neither is its download link.
• Tools the assistant used are named. Their inputs and results are not.
• Messages from a conversation that began before you upgraded have no timestamp, and are exported without one.

The JSON file

A pretty-printed array of ChatMessage objects, exactly as the widget keeps them in sessionStorage:

[
  {
    "id": "msg-1",
    "role": "user",
    "content": "What are your prices?",
    "createdAt": "2026-09-19T14:03:00.000Z"
  },
  {
    "id": "msg-2",
    "role": "assistant",
    "content": "Plans start at $10.",
    "createdAt": "2026-09-19T14:03:04.120Z",
    "sources": [
      { "url": "https://example.com/pricing", "title": "Pricing", "type": "page" }
    ]
  }
]

Inline attachment bytes are never included. createdAt is new in this release; it is sent to the worker along with the rest of each message, and the worker ignores it.

Privacy posture

• Export runs entirely in the browser. No request is made, and your worker never learns that a conversation was exported.
• A transcript contains whatever the conversation contains. That includes anything a visitor typed that the PII plugin did not redact. Once it is a file or on the clipboard, it is outside the widget’s control.
• The JSON can include attachment storage details. With the R2 attachment backend, each stored attachment carries its storage key and a signed URL that works until it expires. The Markdown transcript leaves both out.
• A transcript is not evidence. It is an editable text file, and a visitor can type text into a message that looks like an assistant heading. If you need a reliable record of what the assistant said, keep it on the worker side.

This is why the option exists and is off by default: leave it off for deployments where conversations should not leave the chat window.

Limitations

• Copying needs a secure (HTTPS) page. In an <iframe>, the frame also needs allow="clipboard-write". Where copying is refused, the widget says so and suggests downloading instead.
• Some in-app browsers (the ones inside social media apps) silently block downloads. The page cannot detect that. Copy still works there.
• A Markdown viewer decides how raw HTML in a message is shown. The transcript escapes HTML that could hide later messages, and leaves the rest as written.