Alibaba
Qwen Code CLI
2 days since its last release · 0.27.2 ·
Qwen Code is an AI coding tool from Alibaba. Its latest release is 0.27.2, released on 10 Nov 2026. It shipped 15 releases in the last 30 days, and we have tracked 95 since 1 Aug 2025.
- Latest
- 0.27.2
- 10 Nov 2026
- Last 30 days
- 15
- releases
- Usual gap
- ~3d
- between releases
- Tracked
- 95
- since 1 Aug 2025
- Homepage
- qwenlm.github.io/qwen-code-docs/en/users/overview
Releases per month
last 12 months- 6 Dec
- 8 Jan
- 10 Feb
- 11 Mar
- 13 Apr
- 7 May
- Jun
- 1 Jul
- Aug
- Sep
- 8 Oct
- 8 Nov
Release history
95 releasesReleases 80
-
WeChat and Email channel setup in Web Shell, plus Mods upgrades
- Web Shell can now set up the WeChat channel with an in-browser QR login, and the Email channel from the channel manager
- Mods can run model completions, rewrite reply rows in the Ink terminal and keep each plugin's values between sessions
- Managed agents can import Skills from the original Runtime and gain private BTW and session recap
- Web Shell shows session speed and cache hit rate under the composer, and managed sessions can reuse the host shell
- Auto session titles and Goal verifier verdicts now use the conversation's language
-
Mem0 personal memory and session diagnostics for managed agents
- Managed agents add opt-in personal memory backed by Mem0, plus archiving and self-service MCP installation
- Managed agent Sessions now save model statistics, request context and execution diagnostics
- Web shell adds sidebar multi-select with bulk pin and archive, plus collapsible, drag-sortable workspaces
- LSP can now be turned on with a user-scoped setting and environment variable pair
- Default background execution for agents is now configurable, and model switching works again for subscriptions
-
FileChanged hooks and GitHub event filters for channels
- New FileChanged hook notifications for local sessions
- Channels gain a GitHub repository event filter with allowlist and blocklist, and idle channel sessions are now reclaimed
- Web shell shows tool arguments and results in expanded cards and surfaces task Q&A activity above the composer
- Pending AskUser questions in the web shell can spin off a details side task
- Fixes for cancelled shells reported as completed, Anthropic refusal details and workflow agent approvals in the TUI
-
Skill archives can now be uploaded and installed from the web shell
- Skills can be installed from file-backed archives, with SDK support, receipt lookup and uploads from the web shell
- Managed-agent automation is now on by default, and managed agents can ask durable structured questions
- Image input is now accepted for Llama 3.2 Vision, Llama 4, LLaVA, Pixtral, Mistral Small 3.1/3.2 and Qwen VL models via Ollama names
- CLI reports interactive program status with OSC 7501, and the QQ bot can send proactive loop messages to groups
- Fixes for subscription sign-in recovery, session list titles, session pagination and web shell draft editing
-
Qwen subscription support arrives, and project hooks now need approval
- Added Qwen subscription support across the CLI, VS Code and Desktop
- Project hooks now need approval in a dialog before they run, and headless project-hook trust requires the folder-trust feature
- Mods can submit their own prompts in headless runs and in the Ink terminal, and prompt.submit hooks now run in the OpenTUI terminal
- GitHub bound-session delivery gets role-aware action templates you can edit, and Web Shell channel events link to PRs, issues and sessions
- You can now set the Conversations working-directory root, and images now work with Qwen2-VL, Qwen2.5-VL and QVQ models missing from the catalog
-
GitHub events reach PR-bound sessions, plus stricter project trust checks
- Channels can now deliver GitHub events to sessions bound to a pull request, and the web shell shows a Channel event timeline
- GitHub repository settings can be edited from the web shell's channel editor
- Project frontmatter, LSP and settings-file project hooks now require explicit trust before they run
- Remote access can be turned on from settings without restarting the server
- Managed agent sessions accept full model configuration at creation and allow concurrent workspace uploads
-
Web shell gets a conversation activity view and richer token usage stats
- Web shell adds a conversation activity view and an Add to Split View option in session menus
- Token usage now shows request success rate, per-request speed and failures, and distinct colors per model
- Desktop app now ships rpm installers, and GitHub channel configuration in the web shell is now usable
- Headless runs now trigger prompt.submit hooks, and Kimi K3 models keep a 1M context window
- Fixes for MCP tool name collisions, OAuth refresh locks, and shell permission checks on 2>&1 redirects
-
Generated UI on by default, plus PR autofix retry limits
- Generated UI is now on by default, and you can keep chatting with interactive card values attached as removable context
- /autofix-pr hands a check to a human after three failed fix pushes and posts a PR comment when a round is blocked
- PR-watch skill gains an --in-place mode that works in your checkout, pushes the branch and replies on the PR
- Web Shell adds a Workflow switch, Session Overview and Split View shortcuts, and background agent counts in the status bar
- Built-in reasoning effort now covers DeepSeek-V4 and new Qwen model prefixes
-
Generated UI views follow the host theme and become interactive
- Generated UI views now follow the host theme, with a localized, reshaped card frame and interactive views for adjusting values
- Managed main sessions can run in independent worktrees
- Extensions can be installed and updated from unscoped npm packages
- Web shell shows model timing inline in Token Usage and pauses requests offline, recovering automatically
- Higher limits for workspace skill installs, plus fixes for session branch replay and provider menu overlap
-
Generated UI cards keep their state, plus persistent Workflow mode
- Generated UI cards keep inputs and state across navigation and split views, and appear only for clients that can render MCP Apps
- Persistent session Workflow mode can be turned on with --workflow-mode on, /workflows mode on, the SDK or ACP
- Tool results shrink to fit remaining headroom before auto-compaction, and qwen4 and qwen-latest-series aliases get 1M-token context windows
- Valid XML tool calls are now recovered when parameters quote other call markup, and quoted examples are no longer dispatched
- /autofix-pr start now stops if the PR branch has unpushed commits or uncommitted changes
-
Auto approval and permission rules for managed sessions, plus OpenTUI session resume
- Managed agent sessions get permission rules and an Auto approval mode
- Resumed OpenTUI sessions keep text, images and tool results; rewind restores transcript and draft, with edit-approval checkpoints
- Experimental shared agent loop now runs headless and ACP prompts
- Web Shell can clone a Git URL as a workspace, copy replies as Markdown, and show or stop /autofix-pr watches
- Breaking: Web Shell controls simplified and cwd submissions retired in managed agent
-
Daemon token saving, Managed Agent sessions and OpenTUI parity
- Daemon can save its bearer token to an owner-only file with qwen daemon start --save-token, managed by qwen daemon token
- Managed Agent gains MCP server management without recreating sessions, webfetch, subagent delegation by default, and one-command local start
- OpenTUI input, editing, Vim, history and /statusline behavior now align with Ink, with configurable status lines
- Scheduled tasks can be gated on a shell command run outside the model, with PR watch probes and an /autofix-pr skill
- Breaking: QQ bot restores per-group session isolation under thread scope; DeepSeek support now uses deepseek-flash
-
Managed-agent daemon tooling, Web Shell previews and loop protection
- New qwen daemon start, restart and status commands, plus local daemon discovery and control
- Managed-agent setup gets a --check preflight, credential bootstrap, systemd units and a config template
- Web Shell adds PDF and DOCX previews, per-model speed in Token Usage and trajectory follow and pause controls
- Loop protection is now configurable via model.loopProtection and reminds instead of stopping by default
- Ctrl+V pastes clipboard text on Linux; WeCom application callback support added for channels
-
Managed agent sessions, Web Shell file tools and QwenCloud sign-in
- QwenCloud is now the first entry in the /auth provider list
- Managed sessions join the Web Shell home and sidebar, with workspace, approval mode and model selection plus context usage
- Hosted parent agents can launch snapshot-isolated child agents that read a frozen copy of the workspace
- Files panel gains New file, New folder and Delete actions, with confirmation before anything is permanently removed
- Web Shell trajectory panel adds a Recent executions section showing live requests and tools with elapsed time
-
Managed-agent teams, scriptable Mods and a richer Web Shell
- Managed-agent teams gain member messaging, shutdown requests, plan approval and workflow child tasks that can be cancelled
- Opt-in headless Mods get TypeScript authoring, middleware, and tool-call interception to rewrite, refuse or answer calls
- Web Shell adds a workspace file browser with previews, Excel viewing, file drop uploads and copy as plain text or Markdown
- Web Shell adds task reminders for approvals and questions, plus a managed child agent panel
- New repeatable --disable-mcp-server flag excludes MCP servers per run, and MCP configs saved with a UTF-8 BOM now load
-
Managed-agent runtime expands, plus Web Shell and fixes
- Managed Agent runtime gains background Shell and Monitor, an experimental Kubernetes CSI runtime, durable tool records and child agent sessions
- Web Shell adds Send now for queued messages, fullscreen history tables, trajectory search and simultaneous saved daemon connections
- Agent and Goal coordination tools are now deferred by default and discovered on demand, cutting tool declarations in ordinary sessions
- Hosted approval requests can include bounded previews of file and shell input
- Fixes include QWEN.md and AGENTS.md discovery stopping at the Git root, MCP permission rule collisions and Windows vim-mode paste
-
Auto Recall profile, longer /verify runs and Web Shell task panels
- Opt-in Auto Recall profile searches external context when you submit a prompt, with built-in security limits
- /verify gets a 120-minute budget (up from 25), evidence screenshots in reports and new verification techniques
- Web Shell adds contextual task panels, theme-aware visuals, a retry action for failed prompts and a Queued on server status
- Channels: list, approve and revoke pairing approvals per workspace, plus a GitLab polling adapter and DingTalk interactive cards
- GitHub channels post one final comment per event and show an eyes reaction while the agent works
-
Auto approval mode, /goal command, worktrees and NotebookEdit
- New Auto approval mode uses an LLM classifier to decide which tool calls to approve
- New session-scoped /goal command keeps turns going until a judge decides the goal is met
- Worktree support with EnterWorktree and ExitWorktree tools, plus agent isolation and session resume
- New NotebookEdit tool for editing Jupyter notebooks
- MCP no longer blocks first input, and /rewind can now restore files
-
Structured JSON output in headless mode, standalone installer, lighter prompts
- New --json-schema option returns structured output in headless mode
- Standalone archive installation added
- Low-frequency built-in tools are deferred to shrink the initial prompt; new tools.toolSearch.enabled setting
- Ctrl+B keybind to promote a task, and runtime.json sidecar written for active sessions
- Legacy qwen auth subcommand removed; use the /auth dialog instead
-
New /diff command, ToolSearch and reactive context compression
- New /diff command shows git diff statistics
- ToolSearch loads deferred tool schemas on demand
- Reactive compression kicks in when the context overflows
- autoSkill extracts project skills in the background
- QWENHOME env var sets a custom config directory
-
Fork conversations with /branch, plus commit attribution for AI edits
- New /branch command forks the current conversation
- Commit attribution tracks per-file AI contributions
- Searchable /resume picker with focus-aware modes
- Idealab added as a third-party provider
- Foreground tasks can be promoted to background; slash commands reload when skills change
-
Live agent panel, light theme for HTML export, skills symlink fix
- Inline agent display replaced by an always-on live agent panel
- /export HTML now has a light theme and a toggle
- Skills can now be symlinks pointing outside the skills directory
- Fixed a finished foreground agent lingering in the status bar and ESC in the background tasks dialog cancelling the running request
- Spinner now uses tmux-safe dots to reduce redraw pressure
-
Background agent resume, event monitor tool and read-before-edit safety
- Background agents can now be resumed and continued
- New event monitor tool streams throttled stdout, with monitors shown in /tasks and the Background tasks dialog
- Edit and WriteFile now require a prior read of the file; unchanged re-reads are short-circuited by a cache
- Expanded review pipeline with new qwen review CLI subcommands
- Switch models non-interactively from the CLI; reasoning effort gains a 'max' tier for DeepSeek
-
Background shells in task stop and Background tasks dialog
- Background shells can now be stopped with the taskstop tool
- Background shells appear in the combined Background tasks dialog
- MCP configuration is now available through the CLI
- Static header refreshes when you switch models
- Fix: reasoningcontent is now included when replaying DeepSeek tool calls
-
Background agent UI, /tasks shell pool and OpenRouter auth
- Background agents get a status pill, combined dialog and detail view, and the model can now stop agents or message them
- Managed background shell pool with a new /tasks command
- OpenRouter authentication added, plus an API Key option in the qwen auth menu
- LLM-generated summary labels for batches of tool calls
- Slash command argument hints, API preconnect for faster first calls, and terminal notifications in iTerm2, Kitty and Ghostty
-
Conversation rewind, Python SDK and Traditional Chinese UI
- Rewind a conversation with double-ESC or the new /rewind command
- Python SDK implementation added
- Traditional Chinese (zh-TW) available as a UI language
- Sticky todo panel in app layouts and Space-to-preview in the /resume session picker
- VS Code extension: native copy actions in chat and /skills as a slash command with a picker
-
Auto-titled sessions, built-in web search removed, ACP and resume fixes
- Sessions are now titled automatically by a fast model, and /rename --auto can regenerate a title
- Built-in websearch tool removed in favour of an MCP-based approach
- ACP mode now supports SSE and HTTP MCP servers, and Agent tool calls run concurrently
- Reasoning content is now kept when resuming a session and during active sessions
- VS Code companion supports /export; fixes for garbled output after Ctrl+C and queued slash commands
-
Shell time indicator shows elapsed and timeout together, plus two fixes
- Shell commands now show elapsed time and timeout combined in a single time indicator
- Fixed an endless render loop when completing slash commands
- Fixed streaming tool-call parsing so each stream gets its own parser
-
Background subagents, auto-memory, /recap and /doctor commands
- Background subagents with headless and SDK support
- Managed auto-memory and auto-dream system
- New /recap session recap, /doctor diagnostics and /batch skill for parallel batch operations
- Hooks gain HTTP, function and async types, plus full hooks support over ACP
- Path-based context rules from .qwen/rules/ and multi-line status line with periodic refresh
-
Qwen OAuth free tier ends; /context data now in SDK and non-interactive mode
- Qwen OAuth free tier is discontinued as of 2026-04-15, and discontinued qwen-oauth models can no longer be selected in the model dialog
- /context usage data is now available in non-interactive mode and through the SDK API
- New fork subagent shares context with the parent session
- VS Code companion gains an /account command to show account details
- Fixes for rate-limit detection in streamed responses, custom Gemini baseUrl, DingTalk allowed-user matching and Tab input handling
-
Smarter tool batching, skill model overrides and new hooks
- Tools now run in parallel in smarter batches, with read-only shell commands detected automatically
- Skills can set a model override in their frontmatter; agent definitions gain a disallowedTools field
- Sub-agents inherit the approval mode, and idle context is cleaned up with microcompaction
- New StopFailure and PostCompact hook events, plus contextual tips after responses
- Better CJK word navigation, French locale, queued messages editable with up/Esc, and a model picker for Fast Model in /settings
-
Customizable status line, richer /review and plan-mode restore
- New /statusline command lets you customize the status line
- /review now includes deterministic analysis, autofix and security hardening
- Exiting plan mode offers a 'Yes, restore previous mode' option
- Env-prefixed shell commands now match saved permission rules
- Fixes for checkpointing outside repos, Shift+Tab accepting suggestions, and subagent input conflicts
-
New /plan command for plan mode, plus qwen3.6-plus model support
- New /plan command to use plan mode from the CLI
- Adds the qwen3.6-plus model to the ModelStudio Coding Plan
- Thinking blocks are kept across turns, with cleanup when idle
- Output limit starts at 8K tokens and retries at 64K when a response is cut off
- Fixes a blank VS Code webview screen; follow-up suggestions are now off by default
-
New /btw side questions, follow-up suggestions and Ctrl+O view toggle
- Improved /btw side question with a better prompt, and Ctrl+C or Ctrl+D to cancel it
- Follow-up suggestions added in the CLI and web UI
- Ctrl+O switches between verbose and compact display modes
- Ctrl+Y skips the rate-limit retry delay immediately
- Queued messages are drained mid-turn during agent execution; system prompt adds guidance for dangerous actions
-
Qwen3.6-Plus support, Channels for Telegram/WeChat/DingTalk, and cron loops
- Adds support for the Qwen3.6-Plus model
- New Channels platform with a plugin system and Telegram, WeChat and DingTalk channels
- In-session loop scheduling with cron tools
- Subagents can now use models from a different provider
- Extensions can be installed from an npm registry; hooks are no longer experimental
-
Fixes for Windows shell, CRLF line endings, and slow /clear
- Added a bundled qc-helper skill and a qwen-code-claw reference with a README guide
- Editing files now preserves the original line endings (CRLF or LF)
- Fixed /clear and ESC key lag caused by the hooks system
- Windows: Git Bash path is now resolved correctly for the shell
- Fixed terminal response text leaking in on high-latency SSH sessions
-
Hooks UI and telemetry, Alibaba Cloud API key auth, and fixes
- Support for Alibaba Cloud Standard API Key authentication
- Redesigned hooks interface, plus telemetry for hook execution and new hooks documentation
- Permission prompts now use human-readable labels, with feedback when a deny rule blocks an action
- Extensions can be installed from non-GitHub git URLs
- Fixes for Git Bash/MSYS2 shell detection on Windows, VS Code companion loading hangs, and C++/Java/Python language server support
-
Hooks system, agent Arena, /context and /btw commands, plus permissions
- Hooks extension mechanism with 10 core event hooks for session lifecycle and tool execution
- Agent collaboration Arena runs multiple models competitively on the same task
- New /context command shows a context window token usage breakdown; /btw asks ephemeral side questions
- New Explore agent (TaskTool renamed AgentTool), and task tools now run concurrently
- Skills load from .agents directories; new auth CLI command; VS Code gets image paste and fuzzy file completion
-
New built-in /review skill for code review, plus many fixes
- Bundled /review skill gives out-of-the-box code review
- Default output token limit raised from 8K to 16K, and maxtokens is auto-detected from the model when the provider doesn't set it
- deepseek-r1 added to output token limit patterns
- Fixes for PDF files corrupting sessions, shell output capture in interactive mode, and API errors on retry
- Slash command descriptions localized; extension install on Windows and macOS sandbox terminal access fixed
-
Smoother MCP OAuth sign-in plus Windows and IDE fixes
- MCP OAuth sign-in gets better feedback after authenticating, translated messages and a way to clear saved auth
- Fixes IDE connection problems in some VS Code clients
- GPT-5.x input token limit corrected to 272K
- Windows fixes for shell argument quoting, and node-pty is disabled on older builds with broken ConPTY
- qwen mcp add now handles array options correctly, and Ctrl+F no longer leaks a ^F artifact into the terminal
-
VS Code sidebar view and chat layout options, plus MCP and CLI fixes
- VS Code companion gains a sidebar view and multi-position chat layout
- MCP OAuth now uses scopes from protected resource metadata (RFC 9728)
- MCP allowed/excluded server lists merge with a new concat strategy, and the MCP server display is cleaner
- Editor diff no longer opens in VS Code when running in YOLO mode
- Fixes a Windows terminal resize race condition and clears stale error messages when a new query starts
-
Hooks system, interactive MCP and extension management, PTY shell by default
- New hooks system with CLI and UI management
- Interactive TUI for managing extensions, plus MCP management that enables or disables servers at runtime
- New AskUserQuestion tool lets the agent ask you questions interactively
- Shell commands now run in a PTY by default
- Windows fixes for CRLF line endings; non-UTF-8 file encoding is preserved on read and write
-
AGENTS.md read as default context file, Ctrl+Y retry, Zed fixes
- AGENTS.md is now supported as a default context file
- New Ctrl+Y shortcut retries a failed request
- Fixes for ACP compatibility with the Zed editor, including changing config options from Zed's UI
- Detects and guards against truncated tool call output
- Rate limit error code 1305 and custom retry error codes are now handled
-
Clipboard image paste, a new /insight command and merged Qwen models
- Paste images from the clipboard into the CLI, with a new attachment display
- New /insight command gives personalized programming insights, surfaced through tips
- coder-model and qwen3.5-plus are merged, and automatic vision model switching is removed
- MCP tools marked readOnlyHint are now allowed in plan mode; plan mode restrictions are enforced in ACP sessions
- settings.json is backed up automatically before auth changes; ESC cancel no longer leaves input blocked
-
New third-party models added to Coding Plan
- Coding Plan now includes third-party models: glm-4.7, kimi-k2.5 and qwen3-coder-next
- Bailian Coding Plan model list updated
- Model configuration in the runner now supports authtype
- Improved BOM detection in file handling
- Updated the security vulnerability reporting channel
-
Adds qwen3.5-plus model support for Coding Plan
- The qwen3.5-plus model is now supported for Coding Plan users
-
Coding Plan gets Global/International region support
- Adds Coding Plan Global/Intl region support
- News banner announcing the Qwen3.5-Plus launch added to the docs
-
Session ID flag, TPM throttling retries and MCP schema fix
- New --session-id option for the CLI and SDK
- Rate-limit (TPM throttling) errors now retry automatically after a 1-minute delay
- MCP tools using JSON Schema draft-2020-12 are now supported
- Line numbers now show by default, as intended (showLineNumbers default corrected to true)
- Fixed memory leaks from abort listeners and a Node install permission issue in the shell script
-
Coding Plan sign-in mode and MCP tool progress updates
- New Coding Plan authentication mode with a unified sign-in dialog and configuration management
- MCP tools now show progress updates in the terminal UI and in SDK mode
- Fixed the model name not updating in the top-right corner of the UI
- Fixed Windows sandbox detection, extension install path handling and an empty message sent by Enter in the login UI
- Fixed a warning in the installation shell script and SDK MCP integration tests
-
Session export, resume by ID, stable Agent Skills; free quota halved
- Free Qwen OAuth quota is cut from 2000 to 1000 requests per day
- New command exports session history as Markdown or HTML
- Sessions can be resumed by session ID, in the CLI query and in the SDK (new resume and continue options)
- Agent Skills are promoted from experimental to stable
- New settings.env field for environment variables; disable settings renamed to positive enable names
-
MCP cleanup on exit, skill symlinks and safer file editing
- MCP server subprocesses are now properly cleaned up when the CLI exits
- Skill manager supports symlinks
- Editing files now preserves UTF-8 BOM and trailing whitespace in replacement text
- WebFetch and WebSearch tools are auto-enabled in Plan mode
- Subagents now enforce tool restrictions
-
Multi-modal input, LSP support and a shell command injection fix
- Multi-modal input (image, PDF, audio) now works across all content generators
- Experimental LSP support adds code intelligence
- Security fix: newline bypass in shell command validation could allow command injection
- Better retry logic for 429 and 5xx errors; extrabody support for OpenAI-compatible providers
- Concurrent runner for batch CLI execution and a new Zed extension for the Qwen Code agent
-
Adds Japanese and Portuguese (pt-BR) language support
- Japanese language support added, along with fixes to menu labels in other languages
- Portuguese (pt-BR) added with full translations; language management reworked to make adding languages easier
- Extension list now shows the skills and agents each extension provides
- Proxy handling no longer prints an experimental-feature warning
-
Redesigned welcome screen and settings dialog, plus extension upgrades
- Redesigned CLI welcome screen and settings dialog
- Extensions get a detail command, better validation and a plugin selection UI for the Claude marketplace
- Security fix for command injection in awk and sed read-only command checks
- The /bug command now works in non-interactive mode
- Diff display respects the showLineNumbers setting, and YOLO mode borders use dimmer colors
-
User feedback dialog, experimental skills setting, auth and ACP fixes
- Adds a user feedback dialog to the CLI
- Adds a setting to turn on experimental skills, placed under the experimental namespace
- QWEN.md loading now filters out system files and limits its scope
- Fixes credential management and authentication flows, and updates Qwen OAuth model information
- ACP gains session/setmodel support for JetBrains, and ENOENT errors are now reported correctly
-
Multi-provider model config, /skills command and custom headers
- Models can now be configured across multiple providers
- New experimental /skills command with hot reload
- Custom headers support added
- Optional timeout for foreground shell commands and a customisable sandbox environment
- Fixes for OAuth URL display in headless mode, external editor launch on macOS and Windows, and auth switching
-
German language support and automatic LLM output language detection
- LLM output language is now auto-detected from the system locale
- German language support added; Russian translation updated
- The --experimental-acp flag graduates to a stable --acp flag
- /approval-mode now accepts a direct argument
- Non-interactive runs honor tools.core and tools.allowed, and exit non-zero on API errors in text mode
-
Experimental Skills feature and new Gemini and Anthropic providers
- Adds an experimental Skills feature
- Adds a Gemini provider and removes the legacy Google OAuth login
- Adds an Anthropic provider with normalized auth and environment configuration
- /compress and /summary commands now work in non-interactive and ACP modes
- Fixes CLI path parsing on Windows and OAuth client name for Figma MCP server compatibility
-
New /resume command and chat recording toggle, plus fixes
- New /resume slash command switches between sessions
- -r and -C aliases for the --resume and --continue options
- Chat recording can be turned off from the CLI or settings
- gitCoAuthor setting is now exposed in settings.json
- Fixes for Windows glob path matching and hangs, and git remote URLs with tokens
-
Russian language support, session resume hint and bell setting
- Adds Russian language support for the interface
- Shows the command to resume the session when you exit
- New setting to turn the terminal bell audio notification on or off
- The CLI is now bundled into the VS Code extension release package
- Fixes the completion menu content in the VS Code companion and removes the obsolete corgi mode
-
VS Code extension arrives, plus easier prompt copy/paste
- New VS Code extension implementation added
- Input prompt no longer has vertical borders, making text easier to copy and paste
- Shell output on Windows now prefers UTF-8 encoding when detected
- Fixed IDE connection failures when running inside Docker
- ACP now reports usage metadata, and a channel field identifies the client
-
TypeScript SDK, DeepSeek V3.2 thinking mode and session history
- Basic TypeScript SDK for using Qwen Code programmatically
- Custom tools via SDK-controlled MCP servers
- DeepSeek V3.2 thinking mode integration
- Session-level conversation history management
- Fixed shell command detection on Ubuntu; removed the /quit-confirm flow
-
Streaming JSON for headless use, plus UI and LLM output translations
- Headless mode accepts stream-json as input and output format, so programs can drive it directly
- Internationalization support for the UI and for LLM output
- Usage statistics now keep key snapshot fields as properties
- Fix: auth type settings are now reset correctly
-
Prompt completion turned off, plus Windows /copy and exec fixes
- Prompt completion feature is disabled
- Fixed garbled characters when running /copy on Windows
- Command execution now uses execFile instead of spawn for memory safety
- Removed a broken link
-
Trae editor support, ACP slash commands and terminal alerts
- Adds support for the Trae editor
- Limited slash command support for ACP integration
- Terminal attention notifications alert you when the tool needs you
- OpenAPI window is now configurable
- Better ripgrep binary detection across platforms, plus a ModelScope streamoptions fix
-
Sturdier edits, approval mode dialog and better Zed support
- Edit tool is more reliable thanks to a new fuzzy matching pipeline
- New interactive dialog for choosing the approval mode
- Zed integration now supports TodoWrite and Task tools
- Fixed stream parsing for the Zed integration on Windows
- Request errors are now only printed in debug mode; DeepSeek chat token limits fixed
-
Simpler search tools and customizable tool output management
- Search tools (glob, grep, ripgrep) are simplified and improved
- New customizable model training and tool output management feature
- The /about and /bug commands now show the same system information
- Fixed VSCode detection null check and tidied its debug message
-
Simpler search tools and customizable tool output management
- Search tools (glob, grep, ripgrep) are simplified and improved
- New options for customizable model training and tool output management
- The /about and /bug commands now show the same system information
- Fixed VS Code detection with a null check and cleaner debug messages
-
Multi-provider web search, custom OpenAI log directory, Kimi K2 fix
- Web search tool refactored to support multiple search providers
- OpenAI logging directory can now be set with custom configuration
- Fixed token limits for Kimi K2 models
- Command loading now handles abort errors gracefully instead of failing
-
Standalone bundled CLI package, plus token limit and cancellation fixes
- The CLI is now released as a standalone bundled package
- Tool output now uses a standardized format so the model reads it more reliably
- Fixed invalid tool calls caused by improper request cancellation
- Fixed token limits for qwen3-max and updated token limit handling
- Fixed being unable to quit while the auth dialog is open; synced with upstream Gemini CLI v0.8.2
-
Plan Mode for safe code planning, plus Qwen3-VL-Plus limits
- New Plan Mode lets you plan code changes safely before anything is executed
- Added Qwen3-VL-Plus token limits: 256K input and 32K output
- Improved DashScope cache control
- Removed the editCorrector that injected extra escape characters into edits
- Fixed Qwen3-VL-Plus high-resolution image handling and TaskTool dynamic updates
-
Zed integration gains OpenAI and Qwen OAuth sign-in
- Zed ACP integration now supports OpenAI and Qwen OAuth authentication
- Fixed auth hang when selecting qwen-oauth in Zed
- Automatic vision model switching now works in YOLO mode
- Fixed output token limit for Qwen models
- Fixed ripgrep loading issue and Markdown list display on Windows
-
Vision model support for Qwen-OAuth and a safer /init command
- Vision model support added for Qwen-OAuth
- /init now asks for confirmation before running
- New skipLoopDetection configuration option
- Tool call type coercion enabled, and system prompt switched to avoid malformed tool calls
- Fixes for Windows multi-line paste and arrow keys, plus subagent and UI fixes
-
Subagents, welcome-back dialog and fixes for auth and token limits
- Subagents feature added
- New Welcome Back dialog with project summary and enhanced quit options
- Token limits re-implemented to work correctly for Qwen models
- Fixed unexpected re-auth when the auth token expires, and saved credentials now clear when switching auth type
- Fixed slow SharedTokenManager that caused 20-minute delays; all Gemini CLI branding replaced with Qwen Code
-
Todo Write tool for task tracking, plus Gemini CLI 0.2.1 sync
- Adds a Todo Write tool for task management and progress tracking
- Syncs with upstream Gemini CLI v0.2.1, described as a major feature update
- Adds Homebrew install instructions to the docs
-
Parallel tool use fixes and a gemini-cli upstream sync
- Fixes parallel tool use and a race condition that stopped tool responses from continuing the conversation
- Syncs with upstream gemini-cli v0.1.21
- Shell tool gets an explicit isbackground parameter
- Qwen login tokens now stay in sync across multiple sessions
- Memory operations switch between project and global scope using a sub-command
-
Project-level save option, grep limits and .geminiignore support
- Choose whether settings are saved to the project or globally
- Grep tool now respects .geminiignore files
- Grep results are now limited in size
- Deterministic cache control added for more predictable caching
- Synced with upstream gemini-cli v0.1.19; fixes prompt re-submission and a reverted trimEnd on model responses
-
Web search moves to Tavily, plus OpenAI-compatible API fixes
- Web search now uses the Tavily API instead of Google/Gemini
- Web-fetch tool no longer depends on the Google GenAI library
- Fixes for OpenAI tool calls and for custom APIs with trailing spaces or empty tool IDs
- New query submissions are blocked while another is still running
- Synced with upstream gemini-cli v0.1.18, with a GitHub Actions build for the sandbox image
-
GEMINI.md renamed to QWEN.md, plus sandbox and login fixes
- Context file is now QWEN.md instead of GEMINI.md across the codebase, and /init respects the configured context filename
- Added usage statistics logging for the Qwen integration
- Fixed EPERM error when running qwen --sandbox on macOS
- Fixed terminal flicker while waiting for login
- Fixed the OpenAI content generator and README docs (/stats replaces /status)
-
Qwen OAuth sign-in, OpenRouter support and custom system prompts
- Adds Qwen OAuth integration for signing in
- Adds OpenRouter support
- Adds a systemPromptMappings configuration option
- Adds a telemetry service and syncs with Gemini CLI 0.1.15 and 0.1.17
- Fixes streaming token usage, default model fallback, sandbox Docker mode and an auth dialog crash
Smaller updates 15
-
Bug fixes for tooluse/toolresult errors, notebooks and terminal links
-
Fixes for DeepSeek reasoning, model precedence and CLI display flicker
-
Better handling of maxtokens with safer defaults
-
Fix for garbled non-ASCII shell output on Windows
-
Fixes for session export, DeepSeek API and OAuth screen flicker
-
Documentation and README updates only
-
Maintenance release with no listed changes
-
Faster quit and a startup logging timing fix
-
No user-facing changes listed in this release
-
Small robustness fix for process info lookup
-
Fixes for macOS sandbox packaging, Windows IDE install and VS Code connection
-
Bug fixes for help text, /bug, chat compression and settings
-
Bug fixes, including an input filter fix
-
Syncs with upstream Gemini CLI 0.1.15 and fixes system prompt file
-
Release with no notes beyond a link to the full changelog