Registry settings

Stored under the prefix imio.omnia.assistant.browser.controlpanel.IOmniaAssistantSettings:

Widget loading

The default GenericSetup profile registers one Plone resource bundle, omnia-assistant: the self-contained UMD build of @imiobe/omnia-assistant-ui (React and styles bundled inside, rendered in a shadow root).

An OmniaAssistantConfigViewlet registered in plone.htmlhead sets window.omnia_assistant_settings from the Plone registry (model, selectors, layout, dimensions, etc.). The bundle reads it and mounts itself into a <div id="omnia-assistant-root"> it appends to the page body.

The widget is not rendered when the enabled setting is False. More generally, page-level availability and server-side prompt composition are exposed through the IOmniaAssistantAdapter multi-adapter. The default implementation reads enabled and base_prompt from the registry.

Streaming chat

Chat messages are sent as POST requests to <portal_url>/@@omnia-assistant-api/chat/completions with stream: true. The assistant proxy inherits from the shared imio.omnia.core.browser.proxy.OmniaOpenAIProxyView implementation, keeps the shared auth/streaming behavior, and forwards the request to the configured OpenAI-compatible gateway in real time.

The widget sends no Authorization header. The caller must have the imio.omnia.core: Access Omnia OpenAI proxy permission, which imio.omnia.core grants to Authenticated users: the same-origin fetch() carries the session cookie. Requests whose Origin host differs from the portal’s are rejected.

Every request must also carry plone.protect’s CSRF token in an X-CSRF-TOKEN header, otherwise the proxy answers 403. The viewlet puts it in request_headers of window.omnia_assistant_settings, for anonymous visitors too, and the widget (2.9.0 or later) adds those headers to its chat and MCP requests.

Projects that need anonymous access can override that permission mapping in their own GenericSetup rolemap.xml.

The request payload follows the OpenAI Chat Completions format:

{
  "model": "<configured model>",
  "messages": [
    { "role": "user", "content": "<user message>" }
  ],
  "tools": [{ "type": "function", "function": { "name": "read_page" } }],
  "stream": true
}

When the model calls read_page, the widget runs it and calls the model again with the result as a tool message.

If base_prompt is configured, the assistant proxy prepends it server-side as the first system message before dispatching upstream. The prompt is no longer exposed in window.omnia_assistant_settings. Projects can override that behavior with a more specific IOmniaAssistantAdapter that changes availability rules or returns a custom system prompt for the current context/request.

Conversation length limit

When max_messages_per_session is greater than 0, the assistant proxy rejects (HTTP 400) any request whose payload contains more user messages than the configured limit; the widget shows the error in the reply. Assistant replies and tool results do not consume the limit. Starting a new conversation from the panel header (“Nouveau”) resets the counter.

Page reading

When include_page_content is True, the widget offers the model a read_page tool. It returns the DOM content matched by page_content_selector, as innerHTML or innerText (controlled by page_content_clean). The page is never sent up front, so:

• the configured model must support tool calling (mistral-* and zai-glm-5-2 on the Omnia gateway do);

• base_prompt should tell the model to call read_page before answering a question about the page, otherwise it answers without reading it.

If the content exceeds max_context_chars characters, read_page tells the model to ask the visitor to quote the relevant passage.

Uninstall

The uninstall profile removes the assistant settings and the omnia-assistant bundle from the registry.

1.1 (2026-10-09)

• Bump @imiobe/omnia-assistant-ui from 0.3.2 to 2.9.0 and replace the HMAC token by Plone’s CSRF token (X-CSRF-TOKEN header). [chris-adam]

• Replace the mode setting by layout (floating / sidebar), migrated by upgrade step 1003. [chris-adam]

• Require imio.omnia.core 1.3: its proxy checks the CSRF token the widget sends. [duchenean]

1.0 (2026-08-03)

• OIA-241: Follow imio.omnia.core’s httpx -> httpx2 migration in test mocks; drop unused httpx dependency. [duchenean]

• OIA-241: Bundle the UI from the published npm package @imiobe/omnia-assistant-ui (0.3.2) instead of a git submodule. [duchenean]

1.0a4 (2026-04-03)

• Load imio.omnia.core ZCML. [aduchene]

1.0a3 (2026-04-03)

• Re-release because ZCML files were lacking (2). [aduchene]

1.0a2 (2026-04-03)

• Re-release because ZCML files were lacking. [aduchene]

1.0a1 (2026-04-03)

• Initial release. [duchenean]

• Added floating/fixed AI chat assistant widget built with Vite + Preact and @assistant-ui/react. [duchenean]

• Added real-time SSE streaming of chat completions from the OpenAI-compatible Omnia gateway. [duchenean]

• Added server-side API proxy (@@omnia-assistant-api) so upstream credentials are never exposed to the browser. [duchenean]

• Added HMAC-signed bearer token authentication (2-hour expiration) for proxy requests, signed with the Plone site keyring. [duchenean]

• Added automatic page content injection as conversation context, with configurable CSS selectors, HTML or plain-text mode, and character limit. [duchenean]

• Added server-side system prompt (base_prompt) injection — no longer exposed to the frontend. [duchenean]

• Added control panel tab (@@omnia-assistant-settings) in the shared Omnia settings UI: enable/disable, model selection, display mode (floating/fixed), dimensions, disclaimer, and conversation limits. [duchenean]

• Added IOmniaAssistantAdapter extension point for context-specific availability, system prompt composition, and frontend config overrides. [duchenean]

• Added dynamic model vocabulary fetched from the IOmniaOpenAIService. [duchenean]

• Added max_messages_per_session enforced both client and server-side. [duchenean]

• Added keyboard shortcut (Alt+I) to toggle the assistant panel. [duchenean]

• Added i18n support (en, fr). [duchenean]