Qwen Code Weekly: Smart Tool Parallelism, Fork Sub-Agent Context Sharing, CJK Word Segmentation

This week we released three feature versions: v0.14.3, v0.14.4, and v0.14.5.

After v0.14.3 went live, we received plenty of community feedback and focused on input experience and sub-agent issues this week. v0.14.4 brings smart tool parallel execution — read-only tools are automatically batched in parallel, significantly improving response speed in multi-tool scenarios. CJK word segmentation finally makes cursor navigation work properly for Chinese, Japanese, and Korean users. v0.14.5 adds fork sub-agents, allowing subtasks to share the parent agent’s context. Additionally, Qwen OAuth free tier has been discontinued as of April 15 — we recommend migrating to Alibaba Cloud Bailian, OpenRouter, or Fireworks.

✨ New Features

Smart Tool Parallel Execution

Qwen Code now automatically determines which tools can run in parallel based on tool type: read-only operations like search and file reading are automatically batched in parallel, while shell write operations remain serial for safety. Wait times are significantly reduced in multi-tool scenarios. See PR #2864 

What you can do with it:

• Search multiple files and read multiple code segments simultaneously — no more waiting in queue
• The agent automatically identifies which tool calls can be parallelized — no configuration needed
• Write operations still execute serially, so no need to worry about file conflicts

Fork Sub-Agent Shared Context

Fork sub-agents allow you to spawn child agents that inherit the parent agent’s full context. Sub-agents can independently complete subtasks and return results to the parent. Ideal for scenarios requiring parallel processing of multiple related tasks. See PR #2936 

What you can do with it:

• Spawn subtasks from the main task to handle independent modules
• Sub-agents inherit the parent’s full context without redundant information passing
• Process multiple related tasks in parallel for improved efficiency

CJK Word Segmentation + Ctrl+Arrow Navigation

Optimized terminal input for Chinese, Japanese, and Korean users. Cursor now moves by word boundaries instead of character-by-character, and Ctrl+Arrow key navigation works properly. See PR #2942 

What you can do with it:

• Move cursor by word, making long sentence editing more efficient
• Use Ctrl+Left/Right to quickly jump between words
• Backspace deletes by word, no more deleting one character at a time

ACP Message Rewrite Middleware

LLM-based message rewrite middleware that reformats model output through a second LLM pass. Configure a custom prompt to automatically restructure responses into your preferred format. See PR #3191 

What you can do with it:

• Automatically rewrite AI responses into structured formats (summary + change table + risk assessment)
• Customize rewrite prompts for different scenarios
• Works as ACP middleware, transparent to the user

Skill Model Override

Skills can now specify which model to use via the model field in Skill frontmatter. Different Skills can use different models. See PR #2949 

What you can do with it:

• Configure the optimal model for each Skill independently
• Code review uses a strong model, simple tasks use a fast model — balancing quality and cost
• No need to manually switch models — the Skill automatically selects the right one

Context-Aware Tips

After a conversation ends, the system automatically recommends next steps based on the current context. For example, after modifying code it might suggest “want to run the tests?”. See PR #2904 

What you can do with it:

• After completing code changes, automatically get suggestions for related follow-up actions
• New users can more easily discover and learn features
• Relevant slash commands are recommended based on conversation content

/review Enhancement: Deterministic Analysis & Auto-Fix

/review now supports deterministic analysis (static checks independent of LLM), auto-fix suggestions, and security hardening checks. Review results are more reliable with direct fix suggestions. See PR #2932 

What you can do with it:

• Code review results are more stable — same code produces consistent conclusions every time
• Issues found come with fix suggestions — no need to figure out the fix yourself
• Automatically checks for common security vulnerabilities

Idle Context Micro-Compaction

Automatically cleans up redundant context during conversation idle time, reducing token consumption. Particularly effective in long conversations. See PR #3006 

What you can do with it:

• Long conversations no longer hit context window limits as quickly
• Automatically cleans up intermediate results no longer needed, saving tokens
• No manual action required — runs automatically in the background

More New Features

• Customizable Status Bar: /statusline command to customize terminal bottom status bar (#2923 )
• /context detail Subcommand: View detailed context usage information (#3042 )
• SDK Context Usage Data Exposure: Access /context usage data in non-interactive mode and SDK API (#2916 )
• Sub-Agent Approval Inheritance: Sub-agents automatically inherit parent’s approval settings (#3066 )
• Sub-Agent Tool Restrictions: New disallowedTools field in Agent definitions (#3064 )
• Unicode Agent Names: Agent names can now use Chinese and other Unicode characters (#3194 )
• Telegram Voice Messages: TelegramAdapter now supports voice message send/receive (#3150 )
• French Localization: Added French (fr-FR) language support (#3126 )
• StopFailure and PostCompact Hook Events: Two new Hook event points for extended automation (#2825 )
• Startup Profiler: --profile parameter to identify startup performance bottlenecks (#3232 )
• Sandbox Image Configuration: tools.sandboxImage setting for custom sandbox images (#3146 )
• Workspace Config Override Warning: Warning when workspace config overrides global modelProviders (#3148 )
• VSCode /account Command: View account information in the VSCode extension (#2984 )
• Direct exit/quit to Exit CLI: No need to type /exit anymore, just type exit (#3201 )

📊 Improvements

🔧 Important Fixes

📢 Important Announcement: Qwen OAuth Free Tier Discontinued

Starting April 15, 2026, the Qwen OAuth free tier has been discontinued. Existing users can continue using their remaining quota. We recommend migrating to the following channels:

• Alibaba Cloud Bailian  (Recommended): Supports latest models including qwen3-coder-plus and qwen3.6-plus, pay-as-you-go. See Bailian documentation  for configuration
• OpenRouter : Aggregates multiple model providers with various payment options
• Fireworks : High-performance inference platform

See PR #3291 , #3299 , #3207 

🎈 Other Improvements

👋 Welcome New Contributors

• @doudouOUC — Fixed checkpointing initialization in non-repo directories, added sandbox image configuration and workspace config override warning
• @BZ-D — Fixed Shift+Tab incorrectly accepting placeholder suggestions
• @YuchenLiang00 — Added /context detail subcommand
• @Apophis3158 — Implemented CJK word segmentation and Ctrl+Arrow navigation optimization
• @sibylassana95 — Added French localization support
• @evan70 — Upgraded normalize-package-data to fix DEP0169 warning
• @B-A-M-N — Fixed crash when OAuth refresh response body is empty
• @feyclaw — Added Telegram voice message support
• @dreamWB — Fixed VSCode session tab title overflow
• @reidliu41 — Improved /bug command experience in terminals without hyperlink support
• @Shenean — Fixed literal Tab input issue in input box

If you have questions or suggestions, feel free to provide feedback on GitHub Issues !