Qwen Code Weekly: AI Remembers Across Sessions, Auto Chat Titles, Batch Multi-File Operations

v0.15.0 is here.

This release focuses on making Qwen Code smarter about context and workflow: it now remembers what you’ve told it across sessions and cleans up stale memories; chat sessions get auto-generated titles you can rename or delete; /batch lets you modify multiple files in one go; and Hook expansion plus subagent background execution make automation and enterprise integrations more flexible.

✨ New Features

AI Remembers Across Sessions

Memory used to require manual upkeep. Now Qwen Code has auto-memory and auto-dream: the AI extracts important details from your conversations into memory, then periodically reorganizes and prunes it — like tidying a bookshelf — so context flows naturally between sessions.

What you can do with it:

• No more repeating yourself: tech stack, coding style, project structure — the AI remembers it next time
• Auto-organize and merge: background scanning merges duplicates, updates stale entries, and keeps the index lean
• Less manual maintenance: things you used to write into QWEN.md or rule files are now handled automatically

See PR #3087 

Auto Chat Titles, with Rename and Delete

Session titles used to be your first message — long and hard to browse. Now the AI generates meaningful titles automatically, and you can rename or delete sessions at any time.

What you can do with it:

• AI auto-generates titles: no need to name sessions manually
• Rename by project: tweak AI-generated titles to match your workflow
• Clean up: delete throwaway or test sessions, keep what matters

See PR #3093 

/batch — Process Multiple Files in One Command

Editing multiple files used to mean running tasks one by one. Now a single /batch command processes them all in parallel — great for bulk lint fixes, multi-document updates, or refactoring.

What you can do with it:

• Bulk lint fixes: fix the same error across many files at once
• Multi-document sync: add the same section or bump version numbers across files in one command
• Batch refactoring: rename variables or extract functions across multiple files simultaneously

See PR #3079 

Hook Expansion: Auto-Trigger Actions During AI Execution

Previously Qwen Code could only run single tasks on demand. Now you can set up “auto-trigger” actions via Hooks. Three new Hook types: HTTP Hooks send notifications to external services (like Feishu or DingTalk); Function Hooks execute your own code; Async Hooks run long tasks in the background without blocking your current conversation. VSCode extension supports Hooks too.

What you can do with it:

• Team notifications: AI automatically pings Feishu or DingTalk when it modifies key files
• Compliance auditing: every AI action is logged to your company’s audit system automatically
• Deep toolchain integration: register your own code via SDK — it runs automatically when events fire

See PR #2827  + #3248 

Subagent Background Execution

Subagents previously had to run in the foreground, blocking the interface. Now they support headless background execution, and the SDK has full support — perfect for CI/CD pipelines and automation scripts.

What you can do with it:

• Background parallelism: launch multiple subagents for different tasks without blocking the main UI
• SDK integration: call subagent capabilities through the SDK API, embedding them into your own toolchain
• CI/CD code review: trigger subagent-based code review in PR pipelines without opening a terminal

See PR #3076 

Directory-Specific Rules

Previously all directories shared the same rules. Now context rules are automatically injected based on file path, read from .qwen/rules/, making AI behavior more precise per directory.

What you can do with it:

• Module-specific conventions: React rules for frontend, Node.js rules for backend — auto-applied
• Project-level customization: drop special style rules into .qwen/rules/ and they take effect automatically
• Less manual config: no need to specify rules in every conversation — path matching handles it

See PR #3339 

/doctor — One-Click Environment Diagnostics

Troubleshooting used to mean manually checking configs and connections. Now /doctor automatically detects environment issues, config problems, network errors, and more — with fix suggestions.

What you can do with it:

• Quick diagnostics: MCP server down? Model won’t load? One command checks everything
• Config validation: automatically verifies config file syntax and API Key validity
• Automation-friendly: non-interactive JSON output for scripts and CI/CD pipelines

See PR #3404 

Read PDFs and Analyze Jupyter Notebooks Directly

Text models previously errored on PDFs. Now system tools extract plain text automatically, with page-range support. Jupyter Notebooks no longer return raw JSON — code and outputs are displayed cell by cell.

What you can do with it:

• Analyze PDFs: let AI read PDF reports, papers, or tech docs directly — no manual conversion
• Page-range reading: load specific page ranges without loading the entire file
• Review Notebooks: AI analyzes .ipynb files with code and outputs clearly laid out

See PR #3160 

📊 Improvements

• Real-time Token usage: Token consumption shown during input and output phases, so you always know your session usage (#3329 , #2742 )
• Session recap /recap: returning to a session shows a summary automatically, or run /recap manually for a quick review (#3434 )
• Tool execution progress: progress info shown during tool execution — no more staring at a blinking cursor (#3155 )
• Loop detection enhanced: retry-loop and stall detection prevent the AI from spinning in ineffective loops (#3178 , #3236 )
• Context >100% display: shows “>100%” when context window is exceeded instead of hiding it (#2766 )
• /btw uses live conversation context: /btw now works with current context — no need to re-explain background (#3429 )
• Compact Mode UX: keyboard shortcuts, settings sync, and safety guards for smoother compact mode operation (#3100 )
• Shell output width constraint: output box width limited to prevent long lines from breaking the UI (#2857 )
• Inline shell output limits: configurable line cap for shell output to prevent flooding (#3508 )
• Terminal theme auto-detection: UI colors adapt to your terminal theme automatically (#3460 )
• Multi-line status line: status line supports multi-line display for more information (#3311 )
• Periodic status line refresh: refreshInterval config for auto-refreshing status line display (#3383 )
• M-d key binding: M-d shortcut (Emacs-style) deletes the next word in the input box (#3358 )
• Early input capture: keystrokes captured at startup to prevent missed input during launch (#3319 )
• Clickable MCP OAuth URL: OAuth authorization URLs are now clickable — no manual copy needed (#3489 )
• OSC 52 copy shortcut: OAuth URLs support OSC 52 copy shortcut (#3393 )
• VSCode long conversation input lag fixed: typing no longer stutters during long conversations (#2550 )
• VSCode Plan Mode toggle and approval UI: VSCode Companion supports Plan Mode switching and approval UI (#2551 )
• VSCode /insight command: /insight now available in VSCode (#2593 )
• VSCode agent execution display: agent execution status now visible in VSCode (#2590 )
• WebUI markdown rendering: tool output and WebFetch results render as markdown in WebUI (#3469 )
• Dual-output sidecar mode: TUI supports dual-output sidecar mode for viewing two output streams simultaneously (#3352 )
• /stats lines attributed to subagent: /stats now correctly shows lines consumed by each subagent (#3229 )
• WebFetch supports Markdown for Agents: web content fetching is better optimized for Agent scenarios (#2734 )
• ACP message rewrite timeout configurable: customize ACP message rewrite timeout to avoid long waits (#3475 )
• slashCommands.disabled setting: disable specific slash commands via configuration (#3445 )
• Slash Command multi-mode expansion (Phase 2): ACP fixes and UX improvements for slash commands (#3377 )
• OAuth flags supported by mcp add: mcp add now accepts OAuth configuration parameters (#3442 )
• SDK API exposes /context usage data: /context usage data accessible via SDK (#2916 )
• Bare startup mode: new bare startup mode skips default config loading for lightweight scenarios (#3448 )
• CLI startup profiler: new startup profiling to diagnose slow launch times (#3232 )

🔧 Key Fixes

Windows Platform Fixes

🎈 Other Improvements

• Fixed Windows install command for CMD and PowerShell compatibility (#3252 )
• Updated auth method docs to reflect OAuth deprecation (#3325 )
• New CI stale policy: PRs with 60 days of inactivity marked stale, closed after 30 more days (#3375 )
• Build script optimized to call tsx directly instead of via npx (#3237 )
• Tool registration lazy factory supports inflight concurrency deduplication (#3297 )
• DingTalk channel fixes: continuation message suffix, empty @mention text preservation, reaction context leak (#2977 , #2978 , #2979 )
• Sandbox image tag fallback to latest (#2962 )
• Integration test stdinDoesNotEnd option fix (#2966 )
• JSON schema generation fix: no more “undefined Options” when enum settings lack descriptions (#2963 )
• AskUserQuestionDialog auto-submits on numeric key press (#3407 )
• /recap pinned above input box, aligned with fastModel default (#3478 )
• plan/subagent/arena system reminders injected into ACP (#3479 )
• Graceful handling when xdg-open is missing to prevent crashes (#1675 )
• reasoning-only assistant content uses empty string instead of null (#3499 )

👋 Welcome New Contributors

• @Skyline-9 — Fixed Windows install command compatibility (#3252 )
• @pedrormjunior — M-d key binding (#3358 )
• @szepeviktor — Fixed class name typo (#2189 )
• @sharziki — Fixed /clear dismissing /btw dialog (#3431 )
• @gin-lsl — WebFetch supports Markdown for Agents (#2734 )
• @ihubanov — Added slashCommands.disabled setting (#3445 )
• @gy1016 — Zed.app detection on macOS (#3303 )
• @yeelam-gordon — Fixed OpenAI samplingParams passing (#3458 )
• @Ojhaharsh — Graceful handling for missing xdg-open (#1675 )

Questions or suggestions? Feel free to open an issue on GitHub !