Qwen Code Weekly: /fork Background Agents, /skills Management Panel, Cross-Project Memory

Qwen Code releases v0.18.0-preview (preview.0 / preview.1 / preview.2) this week with 50+ PRs merged, covering three major directions:

Agents can finally work in parallel. The /fork command lets you dispatch a background agent mid-conversation — inheriting full context, tools, and model config while you continue working in the main thread. Fork notifies you via the background task panel when done.

Skill management has a panel. Previously Qwen Code’s skill system had no unified entry — you had to remember each skill name to use it. /skills now opens a picker dialog for browsing, searching, enabling/disabling, one-click execution, and workspace-level visibility management.

Memory goes cross-project. Previously auto-memory only accumulated in the current project directory — switching repos meant the agent forgot everything. This week adds ~/.qwen/memories/ user-level memory directory. Preferences, work style, and cross-project feedback automatically route to user-level, while project-level deadlines and events stay in the project directory. Uses Qwen Code’s existing four memory categories for routing, no new config needed.

✨ New Features

/fork: Dispatch Background Agents Mid-Conversation Without Waiting

/fork lets you run multiple things simultaneously in one session — no need to open a new terminal or switch sessions.

Type /fork <instruction> and a background agent starts immediately, inheriting the current conversation’s full context (system prompt, history, tools, model, prompt cache), executing your instruction independently in the background. The main conversation is not blocked — you continue working on other things. When the fork completes, you get notified via the background task panel, with results written to a JSONL transcript.

This differs from /branch: /branch copies the current conversation to a new session for exploring different directions, while /fork dispatches a background worker within the current session.

What you can do with it:

• Mid-way through coding, /fork help me add tests for this module — tests run in the background while you continue writing the next module
• Real-time fork progress visible in the background task panel (main / ○ fork: …), terminal notification on completion
• Multiple forks can run in parallel, constrained by concurrency limits
• Clear error messages for empty instructions, missing config, or streaming output — no silent failures

See PR #4780 

/skills: Skill Management Finally Has a Panel

Previously using Qwen Code skills required remembering skill names and typing /<skill-name> directly — no browsing, no search, no toggles. Want to disable a noisy skill? Manually edit settings.json.

Now type /skills to open a picker dialog. All registered skills at a glance: name, description, source (project-level / user-level / built-in). Keyboard search for quick navigation, Space to toggle enable/disable, Enter to fill the selected skill command into the input box. Disabled state writes to workspace-level skills.disabled config, taking effect immediately in the current session — the model’s <available_skills> list and slash completions sync instantly, no restart needed.

Cross-scope config is handled: if a skill is disabled in user-level settings, the workspace-level picker shows [locked: User] and cannot be overridden at workspace level.

What you can do with it:

• /skills opens the panel, browse all available skills and descriptions
• Type keywords to search skills
• Space to enable/disable skills, Esc to save and exit
• Enter to select a skill and fill it into the input box for execution
• Disabled skills no longer appear in the model’s available skills list or slash completions

See PR #4533 

User-Level Cross-Project Memory: No More Amnesia When Switching Repos

Previously auto-memory only accumulated in the current project directory — switching repos meant the agent forgot everything. This week adds ~/.qwen/memories/ user-level memory directory, using existing four memory categories (user / feedback / project / reference) for routing, no new config options needed.

New ~/.qwen/memories/ directory alongside project-level .qwen/memories/. Routing rules are decided in the memory extraction agent’s prompt, not hardcoded:

• user type: Always routes to user-level (it’s about you as a person)
• feedback type: Defaults to user-level (most feedback is personal style), but explicitly project-level conventions (testing strategy, build specs) route to project-level
• project type: Always project-level (deadlines, events, ongoing work)
• reference type: Defaults to project-level (this project’s Linear, Grafana), but company-wide common resources route to user-level

Both scope indexes render in the system prompt — user-level first (background info), project-level after (more specific, naturally overrides). Even if user-level MEMORY.md is empty, a placeholder renders to ensure the agent knows the directory exists and can write to it.

What you can do with it:

• After the agent learns your preferences (“I prefer pnpm”, “use vitest not jest for tests”), it remembers them in any new project
• No need to repeat personal preferences in every repo’s .qwen/QWEN.md
• memory.enableManagedAutoMemory toggle reuses existing config, no extra setup needed

See PR #4764 

/stats Interactive Stats Panel: See Token Usage at a Glance

Previously /stats only showed current session token counts as plain text, forgotten after viewing. Now it’s an interactive three-tab panel: Session (real-time metrics for current session), Activity (cross-session usage trends), Efficiency (performance and tool analysis).

Usage data persists automatically to ~/.qwen/usage_record.jsonl, written on session exit and /clear. Tab / Shift+Tab to switch tabs, r to switch time range (All → Month → Week → Today), arrow keys to page through monthly trend charts.

Activity tab shows GitHub-style contribution heatmap, braille line chart of token trends, top-5 project rankings. Efficiency tab shows cache hit rate, tool success rate, average latency, model comparison table.

What you can do with it:

• /stats opens the panel, see current session token usage and cost at a glance
• Switch to Activity tab for cross-session usage trends and coding heatmap
• Switch to Efficiency tab to see which tools have low success rates, which models have high latency
• Filter by time range, compare usage patterns across different periods

See PR #4779 

before	after

Prevent System Sleep: Long Tasks Won’t Get Interrupted

During long-running prompts or tool executions, the system may auto-sleep, interrupting ongoing tasks. This week adds a runtime sleep inhibitor: automatically prevents system sleep during model streaming responses or tool execution. macOS uses caffeinate, Linux uses systemd-inhibit, Windows uses SetThreadExecutionState.

Enabled by default (general.preventSystemSleep: true), disable with one config if unwanted. Reference counting manages multiple parallel tasks sharing one inhibition lock, automatically released when all complete. ACP sessions also covered.

What you can do with it:

• Long-running agent tasks (computer-use, large-scale refactoring) no longer interrupted by system sleep
• Enabled by default, no config needed
• general.preventSystemSleep: false to disable

See PR #4434 

Vim Mode Overhaul: Esc No Longer Misbehaves, Enter Submits, Command Completion

Vim users in Qwen Code have been tolerating three issues: pressing Esc in INSERT mode triggers AppContainer’s handler, showing “Press Esc again to clear” or even clearing the input buffer; pressing Enter in NORMAL mode does nothing; noticeable rendering delay in the bottom indicator during mode switches.

Fixed all at once this week. Esc in INSERT mode only switches modes, no longer leaking to other handlers. Enter in NORMAL mode submits messages normally. VimModeContext split into State and Actions contexts with useMemo, 6 consumers no longer all re-render on mode switch. Bottom indicator adds -- NORMAL --, updating instantly on switch.

Also completed a batch of NORMAL mode commands: undo, replace, case toggle, line join, indent/unindent, WORD motions, find/till motions, yank/paste (system clipboard), operator-pending combinations.

What you can do with it:

• INSERT → NORMAL (Esc) no longer clears input or interrupts model streaming
• Enter in NORMAL mode submits messages, consistent with non-vim mode behavior
• -- NORMAL -- / -- INSERT -- indicators switch instantly, no delay
• Standard commands like u (undo), r (replace), ~ (case toggle), J (join) available

See PR #4677 

before	after

Standalone Installer Auto-Update: Upgrade Without npm

Previously Qwen Code installed via standalone install script (curl ... | bash) would fall back to npm global install on update — either EACCES errors (npm prefix needs root) or installing to a different location inconsistent with the standalone binary.

This week adds native auto-update for standalone installs. No sudo, no package managers needed.

What you can do with it:

• Standalone install users see update prompts, auto-download, verify, replace, effective on restart
• Auto-rollback to previous version on update failure
• Multiple terminals launching simultaneously won’t conflict (lock file protection)
• ~/.local/bin auto-written to shell rc files, ensuring PATH priority

See PR #4629 

More New Features

🔧 Important Fixes

🎉 New Contributors

Thanks to this week’s contributors:

For questions or suggestions, feel free to report on GitHub Issues !