Kitewell 0.22.1 — Public beta

Kitewell 0.22.1 is a beta update.

What's in this release

• Project settings → Screen steps picks the model that reads the screen
  for a project's desktop steps: Kitewell vision, which is included, or one
  of your own API models. Before saving your model, Kitewell checks that it
  reads a test screenshot, and says why when it cannot. The choice syncs
  with the project, and a workflow or step that names its own model keeps
  it. An API model's page in Agents & models offers Use for screen
  steps, and deleting that model returns desktop steps to Kitewell vision.
  A computer without the chosen model's key says so before the workflow
  runs.
• Desktop automations work on every plan. On the free plan, Kitewell vision
  reads the screen within your workspace's credits.
• The assistant learns an app from its window's buttons, fields, and labels,
  as 0.21.0 described. That needed the engine in this release; with the
  engine 0.21.0 bundled, the assistant could not look at the app at all.
• A check of a desktop workflow never presses a button that saves, sends,
  deletes, or confirms, in Japanese or English, even when the model working
  the screen tries to. Desktop steps that run after one the check only shows
  are shown too, instead of failing on the form the save would have closed.
• Checks of desktop workflows take minutes, not half an hour. A look at an
  app closes what it opened, even when it fails; a check waits 30 seconds
  for a window instead of five minutes; and a model lost on the screen stops
  after 20 actions, and the assistant is told to try a model with a
  computer-use tool.
• Checks blame the assistant's draft less often for what is not its fault.
  "Do not press 保存" no longer turns the step that fills in the form into
  one the check only shows, and a step that reads a shown step's output no
  longer fails the check. The assistant is told why a step was only shown.
• The assistant keeps its place through long builds. Your answers on question
  cards, and your leave to work the screen, open sites, or upload files, hold
  for the whole conversation, and work that goes on by itself knows the draft
  is its own. It used to ask the same questions again ten minutes in and
  explore the app a second time.
• An address you type into a question card opens in the browser like one
  written in the chat. The assistant used to refuse it as a guess and ask
  again.
• Trying one step on its own no longer replaces the assistant's draft.
• Desktop workflows the assistant writes fail less. It reads each step of a
  loop item that failed, opens a combo box or pop-up to read its choices,
  and is warned before a check when a step asks its model about a value the
  model never sees.
• An imported API's field that takes a number or null is sent as null when
  its value is empty, instead of the request being refused.
• When the window's page stops, it loads again instead of staying blank, and
  Quit and Restart to Update no longer ask to discard changes nobody made.
• The bundled engine is now Dagu 2.25.1. Desktop steps stay inside the app
  they work in and pick a list's option by name, and a recorded act replays
  only on the rows it fits. OpenRouter failures sent with a 200 status are
  retried.

Plans

From December 1, 2026, the free plan holds no projects or workflows.
Signing in to Dagu Cloud starts a 14-day trial of Kitewell Team, and running
with nobody signed in needs Kitewell Team or its trial. What a plan no
longer holds stays to open, read, and export. Until then nothing changes.
From three days before a trial ends, the plan banner says when it ends and
links to subscribe.

Upgrading

Nothing to do. Kitewell 0.21.0 and 0.20.2 update themselves to 0.22.1 at a
quiet moment, and running workflows go on across the restart.