Back to the changelog

• Security

  Private-network and MCP channel security fixes

  Protection against requests to internal networks (SSRF) now also blocks the Alibaba Cloud and Oracle Cloud Infrastructure metadata addresses that had slipped through, and a domain name that resolves to a metadata address under the on-premise exemption. The block on uploaded procedure PDFs is now case-insensitive. A preview requested via MCP (connected assistants) that would exceed the size useful to an agent is now rejected with a hint on how to shrink it, instead of returning a huge document.

• Improvement

  The Kanban view also flags a case waiting for an operational flow

  The "waiting for an operational flow" notice (a scheduled case whose version is no longer published, and no later one is either) now also appears in the single-flow view and the Kanban board, not just the cases list; the same notice is now also available to assistants connected via MCP.

• Improvement

  Solution import: more accurate messages

  The report of an interrupted Solution import now shows the real cause of an unresolved connector, instead of a generic code; the warning about a connector skipped for a single task no longer disappears if it's the map, not the already-created flow, that fails later on; the notice about demo data for a flow shared by two maps in the same import now shows the correct text. Ingesting a document from an address that serves it compressed (gzip) no longer fails the import.

• Improvement

  Jev diagnostics visible at runtime too

  When a Jev question comes back below the confidence threshold, with a fallback to the expected score, or with no answer at all, the reason now shows up in the task's log (the "AI Execution" section) during an actual case run too — previously this was only visible when trying the question from the Designer. When Jev answers a Score question with more levels than the question declares, the extra keys are now flagged in the same log instead of being silently dropped.

• Feature

  Retry automation on a fallback open for an error

  A task whose human fallback is open because of a technical error (for example, a momentarily unreachable AI provider) and that no one has decided yet now shows, to whoever manages the flow, a «Retry automation» button in the task detail: one click sends the task back to automation, without waiting for a human decision. A fallback open for an outcome or low confidence remains reserved for a person (or «Skip»): the button does not appear in that case.

• Security

  Process Maps: finer permissions on wiring and unwiring

  Wiring or unwiring a link between two flows in a Process Map now requires management permission on the actually affected target flow, not just the ability to see the map. Someone who can see the map but doesn't manage that flow now gets an explicit denial, instead of being able to change an automation on a flow they don't administer. Saving a map design that links (or re-links) a node to a flow you cannot read is also rejected; nodes untouched by the save are not re-checked.

• Feature

  Process Map import preview

  The Process Map import modal now has a «Preview» button: it shows the same report as a real import — created, linked or skipped flows, warnings — without writing anything. From there you can go back to the options or proceed with «Import» to run the real import with the same choices. For a package over the importable limit, the preview gives the same rejection a real import would.

• Improvement

  Process Map import safely repeatable

  Anyone importing a Process Map package via the API can now pass a request key (requestId): a second submission with the same key and the same package returns the map already created instead of creating a second one; the same key with a different package is rejected; a submission still in progress with the same key is rejected until the first one finishes, not made to wait. Without that key, the import behaves as before. In preview (dryRun: true) any request key is ignored: the preview never consumes it.