Back to the changelog

• Security

  Outbound calls work again

  Since mid-July every call to an external system was failing instantly, without even attempting to connect: the check that validates the destination address was built with one library and used with another, incompatible with it. Affected were callbacks to external systems, alert channels, links between installations and the email providers that work over the web — Resend, Brevo, SendGrid. Traditional email sending was not affected. Now, when the AI engine does not respond, the error carries a recognisable code and is recorded with its technical cause, where before it vanished: that is why the problem stayed invisible for so long. And the "Test connection" button uses the same path as real calls, so it can no longer show green while the service is unreachable.

• Security

  History stays inside your organisation

  A request to the AI assistant connection that omitted the item identifier returned the history of every organisation on the same server. The check existed, but guarded the wrong step: without an identifier the filter disappeared entirely instead of rejecting the request. The identifier is now required and verified before any read, with a double boundary on the organisation, and sensitive fields are masked as they already are in the interface. The same review surfaced further cases of the same nature, all closed: incomplete requests were accepted and answered with whatever item came first, an event could be raised without naming it, and a key restricted to certain processes could still publish others.

• Security

  The AI engine stays out of the internal network

  Anyone configuring their own AI engine can no longer point it at the platform's internal addresses. The protection existed but never came into force, because the setting that enables it had not been declared in the environments — and enabling it would not have been enough, because it governed two different things at once: the AI engine chosen by the organisation and the n8n connection, which is instead configured by whoever administers the installation and legitimately lives on internal addresses. These are now two separate settings: the first is protected, the second keeps working. On-premise installations with the AI engine on the local network must now declare it explicitly; the configuration is documented.

• Improvement

  Work on items no longer duplicates

  Forcing a status change towards the phase an item was already in recreated its work items every time: completing the original one was no longer enough and the item stayed in progress forever, with no way to close it. An open work item is now one per type, and forcing a change towards the current status is refused with an explanation, because it unblocks nothing. In the same area: a processing loop can no longer bounce indefinitely — the limit existed but reset every time work passed through the internal queue — and one attempt timing out no longer closes the next one.

• Improvement

  Reconciliation does not touch what you have not seen

  Between the preview and the confirmation the list of occurrences could change, and reconciliation rewrote even the ones never shown. The confirmation is now tied to the preview in front of you: if anything changes in the meantime the system stops and shows the updated figures. "Create user and reconcile" can also be resumed: if the second part fails you retry only that, instead of recreating the user and hitting a duplicate email conflict. And a loading error no longer disguises itself as "no names to fix": it is stated, with a retry.

• Improvement

  The dashboard follows your organisation's timezone

  "Completed today" and the seven-day trend now use the organisation's timezone, which has become a real setting in preferences. Before, two figures on the same screen relied on different references, and after Italian midnight the current day simply did not appear: anyone working evening shifts never saw their own day.

• Improvement

  Forms save data of the right type

  A dropdown with numeric values always returned text: with type checking on, saving was rejected; without it, the value was stored in the wrong format. Multi-value fields now have a real editor, instead of a text box that turned the list into a single string. Required fields inside nested sections are checked, naming the missing one. And starting the same process twice after a network error no longer creates two: the second attempt is recognised as the same one.

• Improvement

  Errors no longer get lost along the way

  AI-assisted generation in the designer reports its failures, and the text you wrote is not lost. Saving in the designer distinguishes what was actually saved: if the design goes through and the settings do not, it says so, and the next attempt no longer fails with a version conflict. The keys panel tells a failure apart from an absence of keys, with a retry, and the agent policy draft is no longer overwritten by an update arriving in the background. "Copied" appears only when the copy really happened: on a credential shown only once, that difference matters.

• Improvement

  Session, waiting and sign-out

  Returning to a tab left open for a long time produced five or six technical error messages in English: there is now a single, understandable notice that the session has expired. A request that gets no answer no longer blocks the page indefinitely — after a reasonable wait it is treated as a network problem — while AI generation and downloads keep longer allowances, as they should. And signing out while the network is unresponsive no longer claims a sign-out that did not happen.

• Design

  Language, wording and accessibility

  The reassignment confirmation in English is readable again: names appeared twice and the words "from" and "to" vanished. Counts of one no longer read "1 users", because the translation system now has separate singular and plural forms. Thirty Italian texts with accents written as apostrophes have been corrected, with an automatic check that stops the problem from coming back. The exported CSV columns follow the chosen language instead of always being Italian, and fourteen texts left in Italian inside the English interface have been translated — among them the confirmation for sending to a production environment, which was a safety barrier unreadable to non-Italian speakers. The "copy" controls in the keys panel are now one single button with the same icon, and the "New role" and "New connector" forms can be filled in with screen readers.

• Improvement

  Sturdier contracts and Process Maps

  Renaming a process really does change the custom entity names: four fields were accepted and silently ignored. An invalid role or scope returns a clear error instead of a server failure or a silent correction, and the scope of an AI assistant key survives the publication of a new process version. On Process Maps: PDF export warns you if the browser blocks the window instead of closing silently, the "Time" lens shows the estimate for manual procedures instead of declaring insufficient data — the "Cost" lens already used the same values — and the help text on quiet hours no longer contradicts the selected timezone.