Back to the changelog

• Improvement

  Complete Jev conversion

  The “Manual review” phase created by “Convert to a Jev decision” now also receives the source phase's form, so the reviewer sees and can edit the same fields the copied rules read, and the “Save data” button appears again on the case. When the source phase is of type Human, the conversion dialog warns that the Jev task's outcome will not move the case forward on its own and offers a checkbox to convert the phase to Auto in the same save, off by default and with an explanation of what changes.

• Feature

  Starts to confirm for external assistants

  With “Require confirmation” on for the flow — the default setting — a case created by an assistant connected over MCP does not start on its own: it waits, running no tasks and no AI calls, so at no cost, until a person presses Confirm start in the case detail (in the list it shows as “Awaiting confirmation”). The assistant cannot confirm it itself, not even through the Flow Agent; a phase move made by a person counts as confirmation, with the same start permissions, and is kept in the history. To reject the request, an administrator deletes the case. If you want MCP starts on a flow to run immediately, untick “Require confirmation”, now visible even when Flow Agent start is off. Manual start, the extension, API keys and events work as before.

• Integration

  More precise errors for assistants and integrations

  MCP tools now answer with the same codes as the API: a case that does not exist or is not visible, a case already completed and a move that is not allowed now get three distinct answers, and a move on a completed case always answers “case completed”. Loading a document from a web address onto a completed case is rejected right away, before the file is downloaded, with a clear message.

• Improvement

  AI spend attributed to whoever started the case

  In the AI call log, the calls made by the AI and Jev tasks of a case started by an MCP assistant, the Flow Agent or an API key now carry who started it — the user behind the key or the conversation, and the session or key used — so every integration's spend can be traced back. Entries already recorded stay as they are.

• Improvement

  Uploads verified right away

  If more or fewer bytes arrive during an upload than were declared — a truncated file, an interrupted transfer, an integration that declares one size and sends another — FlowSharp answers right away with "Verification of the uploaded file failed" instead of a fake storage provider failure, and no longer waits over a minute when bytes are missing. The upload stays open: just send the right file again.

• Improvement

  Item documents: active documents only, and inactive uploads can be deleted

  In Library → Item documents the «Active» view lists active documents only, with no more nameless rows for cancelled, expired or failed uploads. Admins find those uploads in the new «Inactive» view, with the file name, the status and the automatic deletion date, and can delete them right away with «Delete»: the reserved quota is released and the bytes are removed. Every night FlowSharp automatically deletes inactive uploads that have been idle for more than 7 days, leaving a row in the activity log; the ones under legal hold stay.

• Improvement

  No leftovers from interrupted uploads

  Confirming an upload after the case has already closed now closes the upload as failed — quota released, bytes deleted — instead of leaving it pending forever; the closed case stays untouched. In the Library the slug of a new asset whose upload expired is available again from the moment it expires, with no more waiting for the periodic cleanup.

• Improvement

  One way to choose the AI provider

  An AI task or status without a chosen provider uses your organization's default provider, and execution, flow detail, diagnostics, publishing and import now all agree on it: the "AI statuses without a provider" warning appears only when a default is really missing, and a flow imported with a Jev task is no longer flagged as "without a provider" when the provider exists. A new warning tells you when Jev's uncertainty outcome matches one of the question's regular answers, so rules can always tell uncertain cases apart.

• Improvement

  Warnings where you configure

  When you turn on the AI-generated title and its text reads a restricted field, or a field that data minimization will remove, the warning appears right after saving in the Flow settings. Creating a new version from the version history also shows the warnings the version inherits, before the Designer opens.

• Improvement

  Data the AI will not see

  With data minimization on, if an AI or Jev task reads a field that is not in the Data Schema — for example the draft written by an earlier task — flow diagnostics, publishing and import tell you, not just the Designer. A Library document not authorized as agent context reaches the model as "document unavailable" instead of empty text, and diagnostics names it.

• Improvement

  Why an AI task failed

  The case history shows the cause of an AI failure with a clear code, for example when the provider chosen in the task no longer exists, instead of a generic "execution failed". "Retry" only works while the case is still in the task's phase: if the case has already moved on, the task is not run again out of context.

• Security

  Connection test recorded

  A provider's "Test connection" makes real calls: each one now appears in the AI call log on a dedicated channel, following the provider's logging level, and the test is recorded in the Activity log with outcome, model and response time, never the key.