Qwen Code v0.18.4 Changelog

This release is all about making the software more reliable and globally friendly. It matters most to developers and power users who use this tool to code, automate tasks, or manage AI workflows. While there are a few cool new features like the QQ Bot integration, the bulk of this update focuses on fixing annoying bugs—like misreading image dimensions or crashing on bad data—so the app runs much smoother without unexpected crashes.

What's new

• QQ Bot integration: You can now connect the software to a QQ Bot (a popular chatbot service).
• Smarter task scheduling: The /loop feature (which runs tasks on a repeat schedule) now has a second-by-second timer engine, letting it wake up and run tasks at exact times.
• Configurable timeouts: When using ACP (Agent Communication Protocol, how AI agents talk to each other), you can now customize how long the system waits before timing out and stopping a task.
• Token budget for workflows: Workflows now have a "token budget" (a limit on how much AI processing power/text is used per run) and a UI to show you how much of that budget you've used.
• Better translations (i18n): Tool names in the terminal interface and web badges can now be translated into your local language. Extensions can also have translated names and descriptions.
• Hook tracking: The system now passes an original API call ID to the "hook" system (custom scripts that run before or after actions), making it much easier to track and trigger events.
• Deep health checks: There's a new GET /health?deep=true command that checks if the background process (daemon) is actually idle or busy doing work.
• macOS 26+ support: The desktop app now includes "Liquid Glass" visual assets for the next version of macOS.

Improvements

• The README documentation got a major revamp to be clearer and easier to understand.
• Added new documentation for CLI subcommands (like qwen sessions list), the /workflows command, and a workaround for scrolling issues in tmux (a terminal multiplexer).
• Added a "circuit breaker" for tool calls. This is a smart safety limit that stops AI tools from getting stuck in endless loops, with optional heuristics (pattern rules) to detect bad loops.

Fixes

• Image reading fixes: Fixed how the app reads the height and dimensions of BMP, WebP (VP8X, VP8L), and GIF images. It also now properly checks the WebP signature instead of just the generic file header, and keeps DeepSeek AI presets as text-only.
• Terminal and UI fixes: Fixed issues with malformed terminal color sequences (like bad RGB colors) that could mess up the display. Also fixed text truncation in the session picker, handled complex text characters (graphemes) in search, and stopped the React UI from crashing during cleanup.
• Chat and messaging fixes: The app now preserves images sent in the middle of a chat turn. For WeChat (Weixin), it fixed markdown image formatting and shows allowed image directories. For DingTalk, it now splits overly long lines of text so they display correctly.
• System and file fixes: It now correctly identifies .dat files by their content rather than just the name, tracks sed (text editing) changes in file history, and respects home folder boundaries when shortening file paths (tildeifying).
• Crash preventions: Telemetry (usage tracking) file exporters can no longer crash the whole app. The app also safely handles invalid data—like bad OAuth login expiry times, broken cron (scheduled task) entries, and wrong schema lengths—by rejecting the bad data instead of breaking.
• Session and history fixes: Multiline commands are now properly saved in shell history, and stale worktree (temporary code folder) session markers are handled correctly.
• Network and sandbox fixes: Fixed how the app validates GitHub remote hosts, parses sandbox image registry ports, and enforces a byte limit on standard input to prevent overload. It also skips the sleep inhibitor when running headless over SSH so your computer doesn't block the task.
• IDE and MCP fixes: Each VS Code IDE session now gets its own independent MCP (Model Context Protocol) server instead of sharing one, preventing conflicts. Empty MCP prompt arguments and JSON results are now handled correctly.

Breaking changes

No breaking changes in this release.