AI in the panel: images, form fields, an assistant and a security guard

Each feature below is a module that plugs into the panel through core extension points, without changing the existing screens. The text that steers the AI in each one is an AI mode the administrator can edit.

Image creator

A button in the file manager opens the creator: description, style (photo, illustration, icon or none), format and AI server. Before asking for the image, the server checks that the folder exists and the quota allows it, so you never pay for an image that cannot be saved. The image is converted to WebP when saved: a 1.6 MB PNG became a 183 KB WebP in the test.

AI in fields, and a writer

Next to the label of name, description, keyword and text fields there is a button to generate, improve, shorten or fix. Sensitive fields (password, token, e-mail, price, address and others) are excluded by rule, in the browser and on the server. In rich text editors, a writer discusses the text and only sends the result to the editor when the person approves it.

Panel assistant

A button on every screen opens a conversation that answers from the system documentation (148 indexed pages, refreshed on their own) and from what is on the screen. The pages it consulted are shown with the answer. The assistant only guides: it does not run anything.

Guard

The first agent of the agents module runs six checks when the administrator asks: installation password still in use, new administrators and those without a second factor, system screens open without login, new modules in profiles, sensitive files reachable from the web, and security headers. The checks decide what a finding is; the AI only explains the findings and never receives user names or paths. Without AI, the check still finishes.

Credits per profile

Every AI use goes through the same control: each profile points to a credit plan, a request is blocked before it reaches the provider when the allowance does not cover it, and each answer becomes a line in the statement.

Status

Validated with automated tests and browser scripts in both test environments, with real requests to Gemini. OpenAI and the real per-model costs have not been checked yet. Acceptance pending.