The changelog

A selection of visible changes, newest first. Internal operations and check-only adjustments are omitted to keep the history readable.

  • A project can now span several repositories, each with its own settings, tracking and grouped release
  • A deployment’s progress is visible at a glance, and the project list is lighter on phones
  • Sending a video or file in the conversation now shows its upload progress
  • Claude accounts are now used in order, spreading the load instead of draining the first one

  • Each agent step — question, memory, tool — now appears in its own illustrated box, easier to read than a wall of text
  • The in-depth plan becomes the default setting for a new card

  • Monitors can now be handed to an agent, with their own recipe, rhythm and 24-hour history
  • The client space now sends notifications and has a drawer dedicated to a task’s detail

  • The client space became an installable app, “Haiko Chat”
  • The quota panel shows a double ring, a help bubble and the two Cursor engine reserves
  • On phones, extra buttons now fold into a dropdown menu

  • Deployment tracking now lives in the same panel as going to production

  • Recap emails were redesigned in a messaging style
  • A badge signals unread messages, even with the window closed

  • The client space now shows site backups, opens without a login wall and is easier to read on phones
  • Each new plan version is added to the previous ones, so the iteration history stays readable
  • Progress bars follow the work, and a dedicated banner accompanies going to production
  • The client space also sends notifications, including on phones

  • Site backups now use that name, with the same action for files, development and production
  • The plan keeps numbered versions, and the card’s messages stay available while you review it

  • The client space was rebuilt with kanban requests, detailed records, comments, files and a phone-ready view
  • A card plan now opens directly on its task list and stays readable in a more compact drawer
  • Backups are now readable records that clearly separate files, development and production

  • The client space now has four columns — To do, In progress, Done and Approved — with drag-and-drop and focused alerts
  • On phones, the plan fills the screen while its decisions remain within reach at the bottom
  • Request and Plan columns gained their characters and now follow the real six-stage journey

  • The board now has six columns: Request, Plan, Work, Report, To deploy and Archived
  • The Work story has three readable levels — stage, action and detail — without mixing in internal traces
  • Decision buttons and alerts now open exactly where an action is expected

  • The Codex catalogue shows every GPT model actually offered by the engine
  • A known account limit now selects an available reserve and resumes work without losing the task list
  • The board loads cards in batches of twenty to stay responsive on computers and phones

  • The Beluga name replaces the former name throughout services and machine paths
  • A card now follows six stages: Setup, Request, Understanding, Plan, Work and Report
  • The offline interface keeps its content and explains the interruption without shifting the whole page

  • Claude accounts show their identity, can be removed and remain configurable on phones
  • Older plans collapse, the active plan returns to its position and a start can be scheduled
  • Memory limits every answer, says what it leaves out and serves the continuation in sections

  • A card conversation brings request, understanding, plan, work and report into one flow
  • Memory is now a shared tree of topics, branches and details with project, foundation and central scopes
  • Two accounts sharing one quota are recognised as a single reserve

  • The request appears in the card immediately, before the engine gives its first answer
  • Deployment follows the dev branch and production follows main, with a warning when the project diverges
  • A visible question pauses the work until it is answered, without starting two turns at once

  • Starting a card immediately moves it to Work and shows its preparation before the first answer
  • The complete log joins scoping, execution and report in one exportable timeline
  • Server load is monitored so starts can be delayed without interrupting essential services

  • Agents answer in the language selected in the application
  • The final report continues the card flow instead of opening a separate view
  • Screenshots appear in the conversation and suggestions can be moved into the reply field
  • A saturated model waits and retries instead of turning a limit into a failure

  • The vault became central and available to every project
  • Dictation gained its own engine and clearly reports a server refusal
  • The public repository took its current name, belugabuildtool
  • Cards suggested overnight now include their scoping and remain open for discussion before launch

  • All five languages gained a complete dictionary and stable technical markers

  • Complete data export and import arrived with three conflict policies and duplicate-free replay
  • The backup archive covers nine data categories and the Git state of every project

  • A Claude account can be selected directly in an agent’s settings

  • A Language menu makes the whole interface available in French, English, German, Spanish and Chinese

  • Production now requires explicit confirmation before it starts

  • Voice conversation arrived with click-to-listen, voice and text replies, and interruption
  • A configurable keyboard shortcut can start voice listening