Harden Confluence search and return its sources and let trusted providers continue chats that hold confidential tool results

app/MindWork AI Studio/wwwroot/changelog/v26.9.1.md

@@ -1,5 +1,7 @@
 # v26.9.1, build 256 (2026-09-xx xx:xx UTC)
-- Added a Confluence search tool for your company's wiki. Set its wiki address in the tool settings, then select it alongside Read Web Page so the AI can find relevant pages and open them. Search uses your Windows sign-in and is available only with a provider trusted to receive your company's content.
+- Added a Search Confluence tool for your company's wiki. Set your wiki's address in the tool settings, then select the tool alongside Read Web Page so the AI can find relevant pages and open them. The Confluence logo shows you the tool in tool lists and in your chat's tool activity.
+- Added sign-in with your operating system account to the Search Confluence tool, so you do not have to enter a password. When your wiki does not accept that sign-in, AI Studio tells you so instead of reporting an empty search.
+- Added safeguards to the Search Confluence tool: it works only with a High-confidence provider or one your organization trusts, and it never follows a redirect that leads away from your wiki. Each search appears in the sources of the answer.
 - Added tools that AI models can use on their own, starting with Web Search and Read Web Page. When you ask something a model cannot answer from what it knows, it now searches the web, reads the pages it found, and answers with the sources it used. You decide which tools a model may use, right below the message field, and you can watch it work: AI Studio shows which tool is running and, afterward, every call it made with its result. Whether tools are offered at all depends on the model because it has to support them. Read Web Page works right away; for Web Search you pick a search service in the app settings — Tavily or Staan with a free API key, or a SearXNG instance you run yourself. Set up more than one, and they can take turns when one of them finds nothing, or be asked all at once with their results combined. Many thanks to Peer Schütt (`peerschuett`) and Nils Kruthoff (`nilskruthoff`) for building this feature.
 - Added answers that appear word by word even while the AI uses its tools. You read along as the model writes, including the short note it puts down before it looks something up, and the answer that follows a tool call arrives the same way instead of all at once at the end.
 - Added safeguards around everything these tools bring back. Anything fetched from the web is treated as untrusted: AI Studio removes instructions hidden in a page before a model reads it and tells you when it did, exactly as it already does for the documents and web pages you load yourself. A model can never point a tool at your own network. Each tool states how much you have to trust a provider before it may be used with it, so your questions do not travel further than you allow. You can adjust that requirement per tool in the app settings.
@@ -35,7 +37,6 @@
 - Added an optional API key to every server you host yourself, among them LM Studio, llama.cpp, and whisper.cpp. Such a server may ask for one itself or sit behind a login your organization placed in front of it. So far, only Ollama and vLLM could be given a key.
 - Added a setting for the audio quality used when your speech and your audio and video files are transcribed. AI Studio prepares every recording before it goes to your transcription provider, and you now decide how much detail it keeps: a lower quality travels faster, a higher one gives the transcription model more to work with. You find it in the app settings, right below your transcription provider. Thanks, Dominic Neuburg (`donework`), for this contribution.
 - Added organization-wide management for the audio quality used when transcribing. IT departments can set the quality their organization works with and lock it, or leave it as a default their colleagues are free to change.
-- Improved the Search Confluence tool with the Confluence logo, so you can recognize it in tool lists and in your chat's tool activity. The Information page identifies the logo's owner.
 - Improved loading web content in the assistants: it now uses the same reader as the Read Web Page tool, which extracts the main content of a page more reliably and skips navigation and boilerplate. Pages from your own network, including local servers, keep working as before. When a page cannot be read, AI Studio now says why instead of leaving the field empty.
 - Improved the app icon. The previous one was generated by an image model; the new one was created based on it and keeps the familiar green landscape with the chat bubble. Because it is now a vector drawing, it stays sharp everywhere it appears: in your taskbar or dock, in the window list, and on the start screen while AI Studio is loading.
 - Improved how AI Studio works out what a model can do. Every model family now stands on its own, together with the page it was read from, and our build refuses rules which contradict each other or name no source. That way, mistakes are caught before they ever reach you.
@@ -50,8 +51,8 @@
 - Improved what happens when you ask for web content to be cleaned up and no model is available for it. AI Studio loads the page and tells you it arrived uncleaned, instead of quietly handing you the raw page with its navigation and advertising still in it.
 - Changed what a tile that opens a chat directly starts with: when it uses a chat template that brings its own tools or data sources, that template decides them. The Assistant Builder says so while you build such a tile.
 - Changed which model a security check of an assistant plugin may fall back on. When you have set none aside for these checks, AI Studio uses the model you were working with, but only when that model meets the trust your organization requires for a check. One ruled out for that purpose no longer gets to read a plugin's code.
-- Changed how provider trust and provider confidence work together. Marking a provider as trustworthy in a configuration no longer also satisfies a required confidence level: one says who runs the provider, the other how confidential it is. Organizations raise a provider's level in their own confidence scheme instead. This applies beyond local data sources, for example, when a model reads a page from your intranet.
-- Fixed Confluence search failing with a 401 error on wikis that reject the REST search request. Search now uses the wiki's normal search page and the same sign-in path as Read Web Page.
+- Changed how provider trust and provider confidence work together for local data sources. Marking a provider as trustworthy in a configuration no longer also satisfies the confidence level a data source requires: one says who runs the provider, the other how confidential it is. Organizations raise a provider's level in their own confidence scheme instead.
+- Changed which providers may continue a chat that holds confidential tool results, for example a page from your intranet or a search in your company wiki. A provider your organization marked as trustworthy in its configuration may continue such a chat, just as it may use the tools which brought those results in.
 - Fixed the abilities AI Studio assumed for many models. We checked the families against their documentation: some models gained image input, reasoning, or tool calling, others lost an ability they never had.
 - Fixed the Document Analysis assistant freezing while you edited a policy. It needed a change of yours to be saved in the background just as you were making the next one — picking a provider, for instance — which is why it hit some of you again and again and others never at all.
 - Fixed a renamed policy losing its new name in the Document Analysis assistant. The name was kept only when you happened to change something else afterward.