Qwen Code Weekly: /loop Autonomous Mode, Desktop Voice Dictation, Tabbed Settings View

Qwen Code shipped v0.19.4 this week with 80+ PRs merged across three directions:

/loop keeps working while you’re away. A bare /loop launches autonomous mode — the agent advances established work while you’re gone, without inventing new tasks or performing irreversible operations. Add a .qwen/loop.md task file and the loop re-reads the editable checklist each tick; delete it and the loop fires no-op ticks to stay alive. Claude Code recently published a getting-started guide on agent loops  about agents repeatedly executing an established plan — /loop autonomous mode does something similar, with one addition: the task list is editable at any time and the loop picks up changes automatically.

Desktop voice dictation is here. Voice dictation was already available in CLI and Web Shell, but the desktop app was missing it. This release closes the gap — all three surfaces now share the same experience.

Terminal interaction catches up. /settings is now a tabbed view instead of a flat list — search, status, and stats all on one screen. Virtualized History mode now supports mouse clicks and hover on menus, completions, and the input box.

New Features

/loop Autonomous Mode + loop.md Task File

/loop now supports two new modes:

Autonomous mode: A bare /loop (no prompt, no interval) no longer prints usage hints — it launches a self-paced autonomous loop. The agent advances established work while you’re away — maintaining open PRs, continuing half-finished tasks, delivering on commitments — but only what the transcript has already established. It won’t invent new tasks, won’t perform irreversible actions (push/delete/send) without explicit authorization, and stops when things go quiet. The autonomous preamble fires once at the start; each subsequent tick sends only a short reminder to keep token costs low.

loop.md task file: Create .qwen/loop.md at the project root or under ~/.qwen/, set the wakeup/cron prompt to a sentinel (<<loop.md-dynamic>> or <<loop.md>>), and the loop re-reads the file on every fire. The full content is delivered on first read; subsequent ticks send a short reminder if unchanged; editing the file triggers full re-delivery. When the file is missing, the loop fires no-op ticks (the agent does nothing, just re-arms to keep the loop alive), falling back to autonomous-mode base behavior. Recreate loop.md and the next tick auto-detects the change and re-delivers the full content.

Channel loops: /loop isn’t limited to the terminal — DingTalk and Feishu channels work too. Use /loop add "0 9 * * 1-5" check CI status in a group chat to create a scheduled loop, /loop list to see all loops in the current channel, /loop inspect <id> to check run count and recent results, and /loop cancel <id> to stop one. Loops execute through the channel session router, waiting for the current active turn to finish before starting; results are written back through the channel’s proactive send. One-shot loops auto-disable after success or failure.

What you can do with it:

• Bare /loop lets the agent watch and advance work while you’re away
• Use .qwen/loop.md to maintain a persistent, editable task list without re-describing it every tick
• Delete loop.md and the loop fires no-op ticks to stay alive; recreate it to seamlessly resume
• Use /loop add in DingTalk/Feishu group chats to schedule recurring agent work

See PR #5890 , #5991 , #6073 

Desktop Voice Dictation: All Three Surfaces Aligned

Voice dictation was already available in CLI and Web Shell, but missing from the desktop app — users who configured voice input lost the capability when switching to the desktop. This release closes the last gap.

A microphone button has been added to the composer toolbar. During recording, the toolbar transforms into a recording bar (dotted guide line, live waveform, timer). When stopped, the transcription result is inserted into the input box for review. Transcription runs server-side, reusing the CLI voice pipeline (batch qwen3-asr-flash and realtime *-realtime models); provider credentials never enter the renderer. Voice can be toggled and the model selected in Settings > Input and the composer model dropdown.

What you can do with it:

• Click the microphone and speak — see live waveform and timer
• Select a realtime model (e.g. qwen3-asr-flash-realtime) to see interim text as you speak
• Go hands-free for input on a tablet or multi-monitor setup

See PR #5856 

How to get started:

1. Clone the repo:

git clone https://github.com/QwenLM/qwen-code.git

2. Navigate to the desktop package

cd qwen-code/packages/desktop

3. Install dependencies and start dev mode

bun install && bun run dev

Mobile Sidebar Drawer + qwen serve Experience Improvements

Previously, when accessing qwen serve from a mobile browser, the sidebar (session list, search, new session) was hidden below 760px with display: none — there was no way to switch sessions. Now it uses an overlay drawer instead: a menu button appears in the chat panel header, and tapping it slides the sidebar in from the left with a semi-transparent backdrop. Selecting a session, creating a new one, or tapping the backdrop all close the drawer. The desktop layout is completely unaffected.

Behind this is a series of changes that take qwen serve from “usable” to “good”:

• Lazy session creation (#6066 ): Sessions are only created on first prompt, reducing empty-session overhead and speeding up Web Shell startup
• Workspace remember (#5884 ): Serve mode remembers the workspace, no need to re-specify after restart
• Session archive (#6058 ): Daemon supports session archival — long-lived sessions finally have a management path

What you can do with it:

• Access qwen serve from a mobile browser and tap the menu to switch sessions
• Serve mode auto-remembers the workspace when switching between projects
• Archive long-lived sessions instead of deleting them

See PR #6003 

Description	Sidebar collapsed	Sidebar expanded
Screenshot

TUI Mouse Click and Hover

The CLI previously only supported mouse wheel and scrollbar in Virtualized History mode — no pointer interaction. Now with Virtualized History enabled (ui.useTerminalBuffer), three types of interaction are available:

• Selection menus and dialogs (permission confirmations, /model, /config, themes, etc.): hover highlights the row under the pointer, click to select
• / command and @ file suggestion lists: hover highlights suggestions, click to accept
• Input box: left-click positions the text cursor (supports multi-line wrapping, scroll offsets, wide characters)

What you can do with it:

• Click to select a model in the /model dialog instead of arrow-keying through
• Click a completion item after typing /
• Click to reposition the cursor in multi-line input

See PR #6011 

Tabbed Settings Dialog: Search, Status, and Stats on One Screen

/settings has been redesigned from a flat list into a tabbed view, inspired by Claude Code’s /config. A top tab bar splits into Settings / Status / Stats pages. The Settings page adds a search box — type to filter instantly.

The Status page shows the same system info as /status. The Stats page embeds the full /stats dashboard (Session / Activity / Efficiency sub-tabs). Keyboard focus moves vertically through tab bar, search box, and settings list — highlights only appear on the region that actually holds focus.

What you can do with it:

• Open /settings and immediately type a keyword to filter settings
• Use left/right arrows to switch to the Status page for runtime info — no need to run /status separately
• In the Stats page, Tab through Session / Activity / Efficiency; the Efficiency table auto-collapses to “+N more” when it exceeds the dialog height

See PR #6044 

Chrome Extension: Chat and Operate Pages Directly from the Browser

The Chrome extension has been revived on a daemon-direct architecture — the extension is no longer a Native Messaging host, but a thin client of the local qwen serve daemon.

Sidebar chat connects directly to qwen serve via DaemonSessionProvider (HTTP+SSE, same path as Web Shell); when the daemon isn’t reachable, it prompts “run qwen serve”. Browser tools (read page, screenshot, console, navigate, click, fill forms) are exposed to the agent as client-hosted MCP servers via daemon WebSocket. Both chat and tool paths have eliminated Native Messaging entirely.

What you can do with it:

• Chat with Qwen Code directly in the Chrome sidebar — no OS-level native host installation needed
• Let the agent read the current page, take screenshots, and interact with browser elements
• Reuse the shared Web Shell UI for a consistent experience

See PR #5777 

More New Features

Key Fixes

Contributors

If you have questions or feedback, feel free to open an issue on GitHub Issues !