Skip to content

V2.1.49

Latest

Choose a tag to compare

@IceyLiu IceyLiu released this 06 Aug 02:58
28a2a9f

Highlights

Talk and Listen: Audio Attachments and Voice Input

  • Audio files (mp3, wav, m4a, ogg, flac) can now be attached by upload or paste, and images and audio reach capable agents as native content instead of a file path. When an agent has no native support for what you attached, the chip tells you so up front — the file still goes through, just as a path.
  • Voice input stays available while a conversation is running, works in compact mobile layouts, and picks up a Cmd/Ctrl+M shortcut to start and stop recording in whichever composer is active — including Team columns.

Agents Can Ask You Real Questions

  • When an agent asks a structured question, it now renders as a proper card: single-select, multi-select, and an "Other" row you can type into. One submit answers every question in the card at once, because Claude silently drops unanswered questions on a partial reply — so the card refuses to submit until you've answered all of them, and dismissing it sends an explicit decline rather than an empty approval.
  • Reopening a conversation with an unanswered question rebuilds the same card instead of a bare permission prompt.
  • Agents that route questions through the permission channel are now readable too: request details render as formatted JSON rather than a card whose only line was the tool name.

Watch Commands Run

  • When an ACP agent hands a command to AionUi to run, you see it live: the command, its streaming output, its status, and a Stop button that kills that specific command without touching the rest of the turn.

Team Mode: Message and Task Activity

  • Every team member — the leader included — has always had a mailbox and a task board; neither was visible anywhere. There's now a read-only board view showing both, updating in real time, with dependency labels you can jump to. Messages appear alongside tasks because leaders often assign work through the mailbox directly.
  • Selecting an agent avatar and focusing a column's input are now the same act. Previously every column raced to grab focus on mount, so the focused input was effectively random and rarely the agent you'd highlighted.
  • The view appears for teams with more than one member.

Preview Panel, Per Project

  • The preview panel is now scoped to each project and remembers its open tabs, so switching away and back restores what you had open. Unsaved text edits survive; images, PDFs, and Office files are re-fetched rather than stored, so a large image doesn't bloat your local state.
  • Several ways to lose work are closed off. Oversized files are kept out of the editor instead of being silently truncated, and saves carry the file's modification time — if something changed the file behind your back, you get told instead of overwriting it.
  • A file that changes on disk flags its tab for reload, and reloading over unsaved edits asks first.

Elsewhere

  • File change cards for a single file collapse to one compact row — filename, the +/- counts, and a preview link.
  • Scheduled tasks moved from a card grid to a compact row list, keeping the assistant avatar, schedule, next run time, status, and enable switch visible per row. The overflow menu is gone since each row already opens its detail page.
  • Mobile WebUI file handling: the file and folder picker adapts to narrow screens, project explorer items get visible action menus (desktop keeps right-click), and the home composer offers both host files and an upload from the device you're holding.

Bug Fixes

  • Long-running commands no longer corrupt the conversation view. Codex can end a turn while a command keeps running, and everything that command prints afterwards arrives under a new turn ID minutes later. Two paths applied that late completion to whatever turn was current, so a finished turn's ending could land on a turn you'd just started — and the sidebar stopped showing "generating" for every subsequent turn. One user's logs carried 1,580 dropped frames across four conversations in a single day.
  • A corrupted agent process registry no longer aborts shutdown. The file is shared with the backend, and concurrent writes could tear it; a parse failure then broke cleanup, leaking orphan agent processes and, in production, breaking message sending. Reads now degrade instead of throwing, writes are atomic, and concurrently registered entries survive the cleanup pass.
  • A backend that takes its time starting up is no longer reported as a broken installation. The "reinstall / check antivirus" dialog fired whenever the backend stayed quiet for 30 seconds, including for installs that were perfectly intact — one user's backend reported in at 47 seconds and then ran fine. That case gets an honest "startup timed out" message, and the window is now 60 seconds.
  • Windows file paths display their backslashes again in the Y2K Ledger skin. The theme listed Japanese fonts first, and Meiryo and Yu Gothic draw U+005C as a yen sign per legacy JIS convention — the text was never altered, only drawn wrong.
  • The spinner in the Thinking and View Steps headers sat 3.5px below the text next to it and jumped when it swapped to a static icon. It's now pixel-aligned, fixed centrally for both icon systems rather than per component.

Engine updates (AionCore v0.1.58 → v0.1.59, Aion CLI runtime v0.2.8 → v0.2.9)

  • Native image and audio content blocks, gated by what each agent declares it can accept — the backend half of the attachment work above.
  • The read-only mailbox and task APIs plus their real-time events, behind the Team activity view.
  • Aion CLI: turns that produced only reasoning, or whose stream carried an embedded error, no longer come back empty. Stream errors and missing terminators now fail the turn instead of passing silently, tools stay available on the first empty-response retry, and the official OpenAI endpoint defaults to max_completion_tokens.

What's Changed

  • feat(chat): audio attachments + native media capability hints in the send box by @kaizhou-lab in #3849
  • Improve voice input controls and send box alignment by @IceyLiu in #3853
  • feat(webui): improve mobile file interactions by @IceyLiu in #3855
  • fix(ui): correct Arco Spin vertical alignment by @IceyLiu in #3858
  • feat(conversation): simplify file change cards by @IceyLiu in #3860
  • fix(theme): remove Japanese-first font stack from retroma-y2k skin by @piorpua in #3859
  • fix(startup): stop misclassifying listen_timeout as incomplete install by @piorpua in #3857
  • feat(team): add read-only message & task activity view for Team mode by @piorpua in #3808
  • feat(cron): refine scheduled task list layout by @IceyLiu in #3866
  • ci: gate release artifacts behind PUBLISH_RELEASE + drop Homebrew verification by @kaizhou-lab in #3868
  • feat(preview): project-scoped preview panel with edit/persist safety (v2) by @tcp404 in #3865
  • fix(conversation): scope turn-completion handling to the turn that completed by @kaizhou-lab in #3872
  • fix(web-host): harden agent process registry against corruption and concurrent writes by @piorpua in #3874
  • feat(chat): structured question card for AskUserQuestion with a dedicated answer channel by @kaizhou-lab in #3856
  • chore: untrack developer-local skill symlinks and feedback triage artifacts by @kaizhou-lab in #3873
  • feat(chat): live terminal card for client-hosted ACP terminals by @kaizhou-lab in #3861

Full Changelog: v2.1.47-final...v2.1.49