What shipped to production, newest first. Every entry has its own link — click a title to copy a permalink straight to that change.
The same 32-probe corpus that ranks models on the TrustBench leaderboard can now be run against your own deployed release — prompt, moderation, RAG, and tools included. One attack class per request so a full sweep cannot time out after charging for the work. Check measured before citing a 0.0; that number is a perfect score on an empty run.
A hosted Hermes agent's proactive loop can now run on an interval — every 3 hours, every 6, up to 24 — instead of a single daily slot. The scheduled time becomes the anchor for the first wake, and each wake is measured from the last one, so a missed window produces one wake rather than a backlog. Every wake is a billed turn, so the cost scales with the cadence you choose.
Pinecone serverless can now be selected as the vector store for a knowledge base, alongside Qdrant, and it is configurable from the web app rather than only through the API. It accepts every embedding model the platform supports. Content you have just added is visible to the very next search, so a file you upload can be asked about immediately — the SDK waits for the write to be readable rather than leaving you to guess.
Anyone using a chat can decline to have their questions counted in the operator's "what people are asking" insights. The choice is honoured server-side and applies retroactively — questions already asked are removed from those insights the moment someone opts out. Browsers sending Global Privacy Control are opted out automatically, with no click required.
Marking a workspace HIPAA-scoped turns on privacy protections that cannot then be switched off: personal details are stripped from every conversation, and a turn is refused unless the model provider is covered by a business associate agreement. Every access to protected health information is logged.
The daily upload cap — 50 files on free, 100 on premium — now applies to bulk uploads as well as single ones. Previously the bulk route carried no limit at all.
If a caller sent a photo after speaking their question, the assistant used to reply by asking what they wanted to know. It now answers using the image and the question already asked.
Error toasts and the insufficient-funds notice rendered dark-on-dark on the dark and high-contrast themes, making them impossible to read. They are now legible on every theme.
Whitelabel admins can write a Terms of Service in markdown and attach it to a release. Publishing bumps the version, and every user must accept the latest version before using any release that references it. Acceptance is recorded per person per version, and the terms are enforced everywhere a message can be sent — anonymous chat, signed-in chat, the API, and the CLI. Drafts never trigger re-acceptance; only publishing does. The feature is opt-in, so existing releases stay untouched until an admin attaches a document.
When you import website content, navigation junk like "Skip to Content" and "Open Menu Close Menu" links, and blocked-image placeholders, used to be stored as knowledge and could surface in answers. That junk is now filtered out by default on every import path — the web app, the API, and connectors alike. You can still switch the filter off per import if you want the raw text.
Speech-to-text is now provider-configurable instead of fixed. Deepgram Nova-3 is available as a provider, and you can bring your own key for it. Choose a provider for your whole workspace, or override it for a single release — so one workspace can run different releases on different providers.
A converted site can be published to your own GitHub account and your own Cloudflare account. Divinci never holds a Cloudflare credential for it: Cloudflare's own Deploy button creates the repository under your account, and Divinci pushes the site into it. Every later push redeploys automatically, including edits made by an agent. All site assets are committed into your repository, so the published site does not depend on Divinci's storage to keep working.
The agent joins your workspace session as a visible participant with its own cursor, so it can point at the exact control it is talking about instead of only describing it. Everyone on the page sees the same guidance, not just the person who asked. A setup journey alongside it shows what your workspace still needs: connect a website, upload files, connect other sources, and deploy.
