Skip to main content

v0.4.17 AI Conversation Input, References, and Long-Conversation Improvements

Release date: 2026-09-20

AI Conversation Input​

  • Type / to choose a work command or conversation action, or @ to pick business records and conversation messages you can access. Keep typing after selecting, then send. Analysis commands stay read-only. The option list appears only for / or @; the old always-visible shortcut buttons and draft-recovery entry are gone.
  • Drafts and queued messages are isolated by account, workspace, and conversation, and keep their commands, references, and images. Switching conversations while a send is being prepared no longer mixes in messages or clears a new draft. Legacy drafts recover safely in the background, and a recovery entry appears only when a send result is unclear.
  • Stop now works during startup. Requests that were not dispatched are cancelled directly; dispatched requests wait for the original turn to confirm the interruption, and a failed stop can be retried.
  • Switching to Plan or another command no longer drops earlier constraints. New threads keep a bounded history reference, while Plan read-only behavior and business confirmations stay the same.
  • Desktop chat sends no longer return a 400 error while creating the run record. A failed startup keeps your input and clears the Starting state in the environment panel.
  • Model switching is blocked while a turn is running or a startup result is awaiting confirmation. Drafts and history with images stay on an image-capable model instead of switching to text-only. The message context menu is no longer covered by the bottom input area, so Add to AI, Convert to Task Draft, and exporting a single answer remain usable.

References and Context​

  • Reference search groups results by exact number, current record, recent references, and other results, with up to eight per page. Recent references survive a refresh and are tracked per workspace; records you can no longer access are hidden.
  • Unsaved changes in a business workspace can be previewed and then explicitly added to an AI draft as an unsaved snapshot, without writing to the business tables.
  • Task, project, mold, product, and customer lists can be added to AI using the current filter, keeping the search, sort, and record range. Detail pages can add the current record or reference it with @Current Page, fixed to that record.
  • You can review the source status, read version, and scope limits of the current context, then validate it again. Message menus, the text-selection toolbar, and answer menus can add content to an AI draft; converting an answer to a task draft keeps the original input and does not send automatically.
  • Selections that cross bold text, links, and line breaks can be added to AI. The selection remains accurate after a refresh, the original message is checked again before sending, and invalid content is not silently replaced. Repeated text in the same answer now resolves to the correct position.
  • Reference cards now show a read-only detail view before editing. Task cards support View, Edit, Accept, Complete, Archive, Delete, and quick changes for status, priority, and assignee; without edit permission only View remains. A task card no longer overwrites a collaboration conflict automatically: it shows the latest value and your change, then rechecks the version when you confirm.

Long Conversations and Floating AI Window​

  • A Back to Top button now appears above the follow-up navigation group in the lower-right corner. It expands on hover or keyboard focus and automatically expands collapsed history when needed.
  • Reference cards flow in an adaptive multi-column grid with entity icon, type, status, name, and number. Long names are truncated and can be read on hover or keyboard focus.
  • The collapsed AI workspace buoy keeps only the model icon. Hovering expands Split, Floating, and Full Screen on an arc, shows the hovered icon's name, and uses an arrow cursor throughout.
  • Multi-question prompts now advance as soon as you pick an option. The last question still waits for Send to submit all answers, while free-form input continues with Next or Enter.

Translation, Export, and Chat​

  • The /Translate command uses the existing translation capability, with language, tone, and term-preservation options. It processes only the explicit source and keeps the original chat draft and references.
  • Events from the same turn are processed in order, so duplicate or late completion events no longer overwrite a translation cancellation. A translation cache error no longer leaves a finished translation stuck on Processing.
  • Paginated image, long-image, and PDF exports now place AI answers on the left and user questions in right-side bubbles, matching the workspace reading layout. Markdown export content is unchanged.
  • Contacts, conversation lists, and message avatars show real profile images, falling back to initials when none is set.
  • New private or group messages flash the taskbar on Windows and bounce the Dock with an unread badge on macOS. When the window is in the tray or the conversation is muted, the tray icon flashes instead; alerts stop when messages are read or the app returns to the foreground.
  • The emoji button opens a panel with categories, Chinese search, and recent emoji. Emoji insert at the cursor and the panel stays open for picking several in a row.
  • Kanban cards with multiple image attachments stack their thumbnails, and the full preview switches between images with the arrow buttons or ←/→ while showing the current position and total.
  • Recurring tasks are temporarily disabled: the navigation entry is hidden, old addresses return to the regular task list, and the backend no longer exposes recurring-task APIs or automatic generation. Data and code are retained so the feature can return later.

Work Tasks​

  • Kanban cards show commenter avatars in the lower-left corner. Hover to open a board with all member comments, newest first; click the avatar area to write a comment, paste images directly with Ctrl/⌘+V, and delete only your own comments.
  • Open task details by double-clicking a kanban card; a single click no longer opens them. Kanban cards no longer show task numbers, while the table and task details are unchanged.
  • When editing an existing task, added or removed images now submit together with Save, and Cancel restores the original state. Failed uploads remain for retry.