Kitewell 0.18.4 — Public beta

Kitewell 0.18.4 is a beta update.

What's in this release

• Each project now has its own conversation with the assistant, so a chat
  about one project no longer carries another's history or acts on another
  project. The chat shows the conversation of the project the Kitewell window
  shows, and picking a project in either moves the other, unless the window
  holds unsaved work. With two or more projects, a switcher in the chat lists
  each one with what waits for you there, whether Kitewell is at work there,
  or new activity. Notifications open their card or thread in its project's
  conversation. Deleting the conversation from the chat header covers only
  that project; Settings deletes every project's conversation.
• The assistant now checks a draft with live data without asking first. Only
  steps that read run: browser steps, GET requests, mail searches, model
  calls, scratch copies of workbooks, and commands that only read. Steps that
  send or change something, such as a command that writes, an upload, or a
  browser action that commits something, are not run. While it checks, the
  chat says how many steps are done, and each proposal says whether it was
  checked with live data. The card asking to try a workflow, the trial
  question, and the trial results in the chat are gone.
• On a Mac, the menu bar now lists the runs waiting for you under Waiting
  for you, each by its workflow and the first line of what it asks.
  Choosing one opens Kitewell where the run is answered. It lists ten and
  says how many more wait.
• A human task's form now reads like a form: choices as options, yes or no
  as a checkbox, dates as date pickers, and text fields sized to what they
  hold. A blank optional field is left out, errors name the field's label,
  and a form that only asks for one of a few answers shows them as buttons
  that complete the task. Prompts show as formatted Markdown, and a long one
  scrolls so the answer stays in reach. Clicking outside a run's dialog asks
  to discard changes only once you began an answer.
• When you approve a run on a card in the chat, the card follows the run
  until it ends, also across a restart. The assistant then reads how it
  ended and tells you, with one notification.
• Workflows the assistant writes use your language, including step names
  and what each browser or desktop step is told to do. It does data work
  between steps in JavaScript steps instead of awk, sed, or text pipelines,
  which behave differently on Windows and macOS, and it keeps large data,
  such as a pull request's diff, in a file instead of a step output, so run
  lists and sheets do not slow down reading it. It no longer asks for a
  password or key among its questions; it asks for a secret, whose value it
  never sees.
• The workflow editor has a JavaScript task, Work out values with
  JavaScript, for picking, counting, reshaping, or calculating values from
  earlier results. Its panel shows what the script reads, the script with
  highlighting and common snippets, and what it returns. A workflow
  reference typed into the script, which the engine never fills in, is
  caught and moved into the input in one click. The spreadsheet starters now
  use it, so their scripts can be changed there.
• Approving a step, completing a task, and running, stopping, or saving a
  sheet no longer fail with "Kitewell is completing another operation" while
  a sheet runs; they wait their turn. A sheet keeps going when rows wait for
  people, and a scheduled time runs every row whose last run has ended
  instead of being skipped. Stop all also ends rows waiting for an
  approval or an answer, and Cancel queued rows cancels only the rows
  queued when you asked. A row taken off its queue shows as cancelled, and a
  row the engine keeps refusing fails after 15 minutes with its reason.
• A sheet's run controls say what they do. Rows that failed or were stopped
  run again together, leaving out rows a person rejected. A row waiting for
  an approval opens its run to answer it, any row at rest can run on its
  own, and the run button says when selected rows are left out. A launch
  Kitewell refuses starts nothing and leaves every choice open, and a row
  whose run is still going is not started a second time. Conflicts and
  notices are worded plainly in the page's language.
• A sheet launched again before its first launch finished shows one gauge
  for all of its launches, and Stop all and Cancel queued rows act
  on all of them. The gauge says how many rows ended, its counts open the
  rows they count, and its time left leaves out rows waiting for people.
  Runs not started yet show an outlined badge in sheets and run history, so
  a queued row no longer looks like one at work, and a row waiting for its
  automatic retry shows as retrying.
• A sheet's columns can be resized: drag the edge of a column header, use
  the arrow keys, or double-click to fit. The widths stay on this device. A
  sheet without columns offers each published value with what it held in a
  successful run of its workflow; values too long or structured for a cell
  start unticked, and one control ticks or clears them all.
• The chat header no longer counts cards you cannot reach. Applying a change
  withdraws an earlier copy of the same change, marking a thread done or
  deleting a conversation withdraws the cards it left, and choosing the
  count opens the card that has waited longest.
• In a project synced through Dagu Cloud, a scheduled workflow runs on one
  device at a time, as mail workflows already did: turning its schedule on
  claims it, and the device it leaves turns it off. A device waking from
  sleep or starting after being off holds its schedules until it has taken
  Dagu Cloud's changes, then replays only the missed times of schedules it
  still holds, so a schedule moved to another device does not run on both.
• Kitewell Free now includes a webhook for one workflow. Personal and Team
  include as many as you need. When a plan lapses, the webhook turned on
  first keeps working, and the others start nothing until the plan is back.
• In MCP clients that show Kitewell's run view, such as Claude and ChatGPT,
  the view can list a project's activity: runs going now, each waiting run
  with its step and what it waits for, then the latest finished runs.
  Starting, retrying, or approving a run draws it live. When the connection
  may run workflows, a waiting step can be settled in the view itself, with
  its approval, form, or answer. A browser or desktop step shows its
  screenshots there.
• The ChatGPT connection now lists each operation as its own tool. ChatGPT
  does not get the tools that import APIs or answer a browser or desktop
  step's question, which is often a sign-in code; Kitewell's chat, Claude,
  and other clients keep them.
• The window shows its pages while project engines start instead of waiting
  for every engine, and each page loads only the translations of its
  language. Pages no longer sit on "Getting things ready" when several parts
  of Kitewell read the same runs at once. Switching language now also
  translates weekday names, missed-schedule policies, run filters, file
  chips, and workflow sort labels.
• Background checks only look at runs from the last day, so Kitewell stays
  light with long run histories. A run already found waiting for you stays
  seen however old it is. Sheets ask the engine about each row less often,
  and launches of deleted sheets are passed over.
• In a copy of Kitewell that cannot update itself, such as one run outside
  Applications on a Mac or from a folder on Windows, Check for Updates
  now says that only Kitewell installed from kitewell.app takes updates and
  offers to open the download page.
• The bundled engine is now Dagu 2.22.0.

Upgrading

Nothing to do. Kitewell 0.18.3 updates itself to 0.18.4 at a quiet moment, and
running workflows go on across the restart.