Documentation Index Fetch the complete documentation index at: /docs/llms.txt Use this file to discover all available pages before exploring further. ​ v3.1.1 — 2026-10-04 macOS (Apple silicon) Windows (64-bit) Usage resets: Subscribers can view available reset cards and the past 30 days of grant and usage history in Usage settings. Select a card and confirm to reset the applicable limits. Worktree selection: Projects opened from an existing Git worktree now also support choosing a branch and where to run the task. Usage and subscription reminders: M Plan users receive reminders when their remaining quota is low or exhausted. Subscriptions without automatic renewal also receive reminders before expiration, with options to check usage, reset limits, or renew. Session status: The session list shows clearer indicators for running tasks and pending permission requests, including relevant activity in child sessions. Goal results and file downloads: Completed Goals show a final reply with delivered files in the results area. Choose a save location before downloading generated files. Fixed BYOK model order reverting after saving a drag-and-drop reorder. Fixed clipped double-digit list numbers, numbered questions being misread as headings, and incorrect rendering of single-line code blocks without a language label. Fixed missing files required by the DOCX skill in desktop installation packages. Fixed some conversation interruptions after questionnaire responses and context compaction, and improved automatic retries for model timeouts and certain network errors. ​ v3.1.0 — 2026-09-30 macOS (Apple silicon) Windows (64-bit) ​ A Refreshed MiniMax Code A new look and feel: Refreshed branding and visuals, with refined interactions for a smoother everyday experience. Agent upgrades: Improved Computer Use, Mini Apps, appearance customization, the Git Graph panel, and live Agent activity updates. Introducing M Plan: M Plan builds on Token Plan with unchanged subscription prices and text usage quotas. Select tiers include access to the full MiniMax model lineup, including H3. ​ Limited-time offers 50% off the first month: Through October 14, new M Plan subscriptions and upgrades to M Plan receive 50% off the first month. Renewals return to the regular price from the second month. Seven days of unlimited usage: From October 1 through 7, all subscribers can use M3.1-Flash-Preview without usage limits in MiniMax Code. Bonus points for Token Plan users: All Token Plan users will receive bonus points within 24 hours of this release. ​ v3.0.74 — 2026-09-28 macOS (Apple silicon) Windows (64-bit) M3.1-Flash-Preview: MiniMax’s latest text model debuts in MiniMax Code, supporting everyday development work from bug fixes to complete feature implementation. ​ Limited-time offers Double check-in points: From September 28 through October 7, log in to MiniMax Code and complete the daily check-in to receive double free points. New and existing users can participate, and points can be used with all supported text and multimodal models. Token Plan quota reset: Token Plan quotas have been reset for all users following the launch of M3.1-Flash-Preview. ​ v3.0.73 — 2026-09-18 macOS (Apple silicon) Windows (64-bit) Fixed known issues. ​ v3.0.72 — 2026-09-14 macOS (Apple silicon) Windows (64-bit) Local-to-cloud task handoff: Hand local tasks to the cloud to continue work, review files before upload, then download cloud outputs or compare and sync changes back to local files. Infinite Canvas HTML preview and annotations: Drag HTML files onto the canvas for side-by-side previews, scroll through long pages, resize preview areas, and comment on elements or selected regions for the Agent to revise. Voice input: Use the microphone in the input box to dictate ideas with real-time transcription, appending new speech to any existing text. Independent Subagent chats: Start dedicated conversations with configured Subagents, and manage each custom Agent’s name, avatar, instructions, and model in one place. In-app notifications: A new notification center gathers release updates, official activities, tutorials, and account-related reminders. ​ v3.0.71 — 2026-09-11 macOS (Apple silicon) Windows (64-bit) Updated the macOS app signing certificate to improve installation and launch compatibility. Fixed known issues. ​ v3.0.70 — 2026-09-10 macOS (Apple silicon) Windows (64-bit) Fixed login-related issues. ​ v3.0.69 — 2026-09-08 macOS (Apple silicon) Windows (64-bit) Infinite Canvas: Gather images and videos from chats and the workspace into one canvas, then search, filter, preview, batch manage, and send selected assets back into the conversation. Local and cloud sync: The desktop app now shows both local and cloud sessions, and can start cloud work, view results, and download cloud task outputs directly. H3 Max: Added the H3 Max video model with text-to-video, image-to-video, 480p/768p output, and synchronized audio generation. Custom Mini Apps: Build a personal visual workspace with natural language, then keep adjusting pages, modules, and connected data sources. AI watermark: Image generation can now disable watermarks directly, with settings available under General; the default watermark style is also improved. Git and code review: Review all uncommitted changes or only the last Agent-generated changes, open source files from review, mark files as viewed, and keep Diff display preferences across tasks. ​ v3.0.68 — 2026-08-27 macOS (Apple silicon) Windows (64-bit) Memory management: Settings now include a clearer memory management view, and new chats can choose whether to use and generate memory. Goal mode: Improved long-running task execution by reducing repeated replies, looping, and interruptions, while keeping target edits and status tracking available during a run. Browser Use: Better page reading, form filling, and issue diagnosis for spotting errors, loading problems, and failed requests. Remote Control: Added artifact download and task notification cards, plus better remote Plan, Goal, branch, and Worktree management. Settings search: Quickly find model, connection, shortcut, Worktree, and custom instruction settings by keyword. Project creation: Improved the project creation flow and related guidance. Permissions and stability: Improved permission handling, browser verification, and runtime stability. ​ v3.0.67 — 2026-08-26 macOS (Apple silicon) Windows (64-bit) Fixed known issues. ​ v3.0.66 — 2026-08-19 macOS (Apple silicon) Windows (64-bit) H3 video generation: The flagship H3 video model is now integrated. Pair it with built-in Skills, Browser Use, Remote Control, and Goal to turn ideas into videos faster. Site / Fork & Rewind / Plan: Added site management, conversation branching and rewind, plus a Plan mode that reviews before editing. MCP Servers: Added a dedicated configuration page for viewing, adding, and managing MCP Servers in one place. Text selection comments and preview: You can comment on selected text, and preview flows for markdown, SVG, LaTeX, and similar files are smoother. Worktree management: View and clean up worktrees from Settings while protecting active, uncommitted, or locked directories. Plugin Marketplace: Added WPS and WeCom plugins for common office workflows. BYOK: Configure Model ID, context window, max output tokens, and reasoning level in custom model settings. Goal: Improved long-running task continuity, timer display, and pause, edit, and delete entry points. ​ v3.0.65 — 2026-08-13 macOS (Apple silicon) Windows (64-bit) Fixed known issues. ​ v3.0.63 — 2026-08-12 macOS (Apple silicon) Windows (64-bit) Plugin Marketplace: Added a built-in plugin marketplace with initial plugins for financial data, company lookup, legal search, knowledge workflows, and Office work. Code Review: Added a built-in Review command for checking local code changes and linking findings to specific code lines. Red Skill Search: Added Red Skill search. Conversation Navigation: Long conversations now support quick jumps to target turns. Browser Use 2.0: Added multi-tab support for opening, keeping, and switching between pages in the same task. Account, Billing, and Usage: Improved account, billing, and usage displays. Windows Custom Directory: Windows now supports a custom data directory. ​ v3.0.60 — 2026-08-07 macOS (Apple silicon) Windows (64-bit) Fixed an issue with file uploads in the conversation flow. ​ v3.0.59 — 2026-08-05 macOS (Apple silicon) Windows (64-bit) Fixed several known issues. ​ v3.0.58 — 2026-08-05 macOS (Apple silicon) Windows (64-bit) Daily Check-In Rewards: Check in daily to earn 400 points. Check-ins on days 4 and 7 award 1,000 points, for up to 4,000 points in a full week. Points are valid for 30 days and can be used across MiniMax Code tasks. Text Selection Comments: Select text, code, or table content in an Agent response to quote and comment on it. Multiple selections can be annotated and sent with the next instruction. Proactive Memory: MiniMax Code can retain useful long-term context such as communication preferences, work habits, project conventions, and problem-solving experience, then use it when relevant in future tasks. Like and Dislike Responses: Rate each Agent response individually and add a reason when needed, helping us improve response quality and task execution. Usage and Billing: The usage page now shows context-window usage details broken down by messages, tools, skills, system prompts, and memory. MiniMax Code also warns when usage is near or at its limit and provides relevant next steps. Windows Office File Preview: Windows desktop now supports PPTX previews alongside DOCX and XLSX. You can preview all three common Office formats directly in the app, including conversation attachments and local files. Browser Use: The built-in Browser now includes a mobile preview mode and a visible cursor with movement tracking during Agent operations. You can choose the built-in Browser or the system default browser separately for regular web pages and local development pages, with preferences saved automatically. ​ v3.0.57 — 2026-07-30 macOS (Apple silicon) Windows (64-bit) Fixed known issues. ​ v3.0.56 — 2026-07-29 macOS (Apple silicon) Windows (64-bit) Fixed network errors related to proxies. ​ v3.0.55 — 2026-07-28 macOS (Apple silicon) Windows (64-bit) Fixed a history display issue in edge cases. ​ v3.0.54 — 2026-07-27 macOS (Apple silicon) Windows (64-bit) Browser Control: MiniMax Code can now operate continuously inside the built-in browser: reading pages, filling forms, uploading files from the current task or workspace, and checking the result after each action. For outward-facing actions such as publishing, sending, or submitting, it asks for confirmation before proceeding. On first use, you can also confirm importing cookies from Chrome to reuse existing sign-in sessions; the import does not include browsing history, cache, or saved passwords. Goal Mode: Set complex or long-running work as a Goal so MiniMax Code can keep progressing around a clear objective. You can check the current status and elapsed time at any point. When more information or permission is needed, the Agent asks at the relevant step; when the Goal is completed or can no longer continue, it returns a clear result. Feedback Center: The issue reporting and response workflow has been upgraded. You can submit feedback from the lower-left menu or from a conversation, with task context, workspace, and client version attached automatically. Screenshots and diagnostic logs can be added, and log upload can be turned off if privacy is a concern. After submission, My Feedback shows the ticket ID and handling status. File outputs in conversations: File results and interactions in conversations are clearer. Files generated or referenced by MiniMax Code are presented as actionable results, and clicking them can locate the exact file, folder, or code line in the project. Default shortcuts: New default shortcuts are available, with a unified shortcut management entry in Settings. You can search, edit, delete, or restore shortcuts, and quickly trigger common actions such as creating a task, opening a project, searching, toggling the sidebar, opening the terminal, and switching tasks from the keyboard. ​ v3.0.52 — 2026-07-20 macOS (Apple silicon) Windows (64-bit) Remote Control: Connect to your desktop Agent from your phone to check progress, send instructions, and approve permission requests on the go. Built-in Browser: Open web pages and HTML artifacts directly in the app sidebar, so you can preview results, debug pages, and update code without leaving the task. Coding / Work Modes: Coding mode keeps your project environment, Git branches, worktree tools, and other development context close at hand. Work mode simplifies technical details so you can focus on task progress and final deliverables. BYOK Model Configuration: Use your MiniMax API key, or add a custom provider using a base URL, API key, and model name. ​ v3.0.50~v3.0.51 — 2026-07-14 macOS (Apple silicon) Windows (64-bit) Fixed an issue where, after upgrading, some users could see an “LLM service error” in existing conversations, experience interrupted responses, or be unable to continue the conversation. ​ v3.0.48~v3.0.49 — 2026-07-13 macOS (Apple silicon) Windows (64-bit) ​ MiniMax Code 2.0: All-New Core Reduced the time from starting a new conversation to the first output, cutting p95/p99 first-token latency by over 90%. Reduced unexpected interruptions and silent stalls during conversations, with clearer status feedback when errors occur. Improved how conversation state and tool-call data are saved and carried over, reducing cases where conversations are unexpectedly lost or execution context is interrupted. Reduced runtime dependencies for smoother performance on low-memory devices. Streamlined the Windows installation process to reduce setup time. macOS (Apple silicon) Windows (64-bit) Fixed an issue causing service unavailability in certain proxy environments; Fixed a version data conflict between existing opencode installations and the built-in opencode version. macOS (Apple silicon) Windows (64-bit) Worktree base ref picker : Choose base branch when creating new worktrees; cross-surface ref validation before submit. Context compaction (/compact) now shows full feedback loop: thinking indicator → pending state → auto-refreshed history. TurnDiffCard rendered per user input turn; Agent Team multi-turn shows one card at turn end only; scratchpad/memory changes excluded from diff. Sidebar visual details aligned with Figma; project sort order persisted; group position stable on archive. Agent Team switched to explicit user toggle (default team, opt-in single); deep-research skill retired (requires explicit /team). Skill Hub deduplication, Workspace panel scroll fix, slash palette ghost entry fix. User images render as images instead of file cards; questionnaire internal protocol text hidden. Stream stall heartbeat detection and recovery; tool-call blackhole passive recovery; SSE heartbeat extended. Windows: Team plan EPERM rename, plugin dir colon separator, dangling junction startup crash, TLS root CA trust via NODE_EXTRA_CA_CERTS. macOS Squirrel.Mac update race fix; GUI launcher ENAMETOOLONG tolerance; daemon health-check timeout raised to 60s. MCP JSON corruption graceful recovery; cron orphan session cleanup; bash timeout description consistency. /resume framework session id misjudgment; compaction regression revert; session restore loader flicker reduction. mavis-trash hardened: flag parsing, path resolution, root/PWD bypass prevention, dash-prefixed filename protection. Telegram/IM sender verification: owner-on-first-DM gate, hard-block non-p2p events, per-plugin opt-out. Computer Use has been temporarily disabled due to compatibility and user experience concerns. We are working on improvements and plan to bring it back in a future version. ​ v3.0.37~3.0.40 — 2026-06-08 macOS (Apple silicon) Windows (64-bit) IM /stop Command : Send /stop in WeChat, Feishu, or Telegram to interrupt the Agent’s current reply — same as the desktop Stop button. IM Queue Notification : Text-only channels like WeChat now show queue position when messages are waiting behind an active reply. Cron Auto-Delivery : Cron task results are automatically pushed to all IM channels bound to the Agent — no extra configuration needed. IDE Quick Open : New IDE picker in the chat top bar — open the current project in Cursor, VS Code, WebStorm, Zed, and 20+ other editors with one click. Diff Preview on Hover : Hover over a file change row to see a quick diff preview without opening the full review panel. Skill Localization : Built-in skills show localized names and descriptions in the slash command palette, with bilingual search support. Deep Research Quick Access : Deep Research now appears as a top-level entry in both the ”+” menu and slash palette. Feedback Group Entry : Sidebar user menu now includes a QR code to join the Feishu feedback group. Redesigned usage popover showing 5-hour window usage, weekly usage, video quota, and credit balance. Sidebar history defaults to project-grouped view with status dots and collapsible project sections. Standalone Thinking toggle button next to the model selector. Unified toolbar dropdowns with auto-flip near viewport edges. Feedback now creates a ticket with screenshots, task context, and diagnostic logs attached automatically, with status available in My Feedback. Markdown links open in the system browser; anchor links scroll in-page. First-screen HTTP requests merged and cached to reduce concurrent spikes. Code blocks in long conversations render lazily — syntax highlighting only triggers when scrolled into view. Fixed CLI failing to start on Windows when install path contains Chinese characters. Fixed mavis-trash following NTFS junctions and deleting target data on Windows. Fixed “Reveal in File Explorer” not selecting the target file on Windows. Fixed GPU crash loop on Windows devices with Intel integrated graphics. Fixed deleted cron tasks still showing in the sidebar. Fixed OpenCode session state lost after restart (session not found). Fixed tooltip text unreadable in dark mode. File reads outside the workspace directory are now subject to permission mode review. ​ v3.0.34 ~ v3.0.36 — 2026-06-04 macOS (Apple silicon) Windows (64-bit) Run Location Picker : Choose where the Agent runs when starting a conversation — current project, a fresh branch, or an existing branch. Replaces the old Worktree toggle. Code Change Review : New file preview panel with diff comparison and one-click commit for reviewing Agent code changes. Sidebar Search : Search conversation history by keyword in the desktop sidebar. Environment Variable Management : Securely manage secrets and environment variables within conversations for Agent use. Deep Research : Built-in five-step research skill that auto-searches, analyzes multiple sources, and produces evidence-backed reports. Agent Team Enhancements : New task detail panel, automatic stuck-task recovery for long-running tasks, and reduced polling during decision waits. Queue Messages : Send follow-up messages while the Agent is still working — they’re queued and injected at the right time. Task History by Project : Sidebar supports grouping conversation history by project directory. Disable Global Shortcut : The global launch shortcut can now be left empty to disable it, avoiding conflicts with other apps. Redesigned home toolbar with Team mode toggle moved into the menu. Slash command panel keeps skill submenus open on hover. Multi-image preview order aligned with thumbnails. Sidebar Agent list UX unified with session history items. Fixed built-in Skill compatibility on Windows — now auto-routes through Git Bash. Fixed PowerShell corrupting UTF-8 files on Windows. Fixed enterprise proxy environment variables causing connection issues. Fixed WeChat QR binding not refreshing after expiry. Fixed missing recovery prompt when conversation context overflows. ​ v3.0.33 — 2026-05-29 macOS (Apple silicon) Windows (64-bit) Desktop Rebrand : Desktop app officially renamed to MiniMax Code. New name appears in Dock / Taskbar, with a new minimax CLI shortcut added. Enhanced Questionnaire : The multi-step questionnaire now supports close, skip, and free-text replies. Enabled by default for all users. Skill Auto-Discovery : Automatically scans and imports existing skills from Claude Code, Codex, and other AI coding tools — no manual reinstall needed. Quick Scroll Navigation : Click the “scroll to bottom” button once to jump to the previous question, double-click to go to the top. Slash command names simplified — redundant prefixes auto-stripped, case-insensitive matching. AI reply action buttons left-aligned to body text. File picker dialog centered on screen. Scheduled task auto-archive off by default. Task runs refresh on click. Sidebar Agents section collapsed by default. Project selector auto-flips when bottom space is tight. Onboarding and branding visual updates. Fixed Windows install path and shortcut issues during brand migration. Fixed permission system not handling file deletions correctly in full-authorize mode. Fixed schedule agent selector display issue. ​ v3.0.32 — 2026-05-26 macOS (Apple silicon) Windows (64-bit) Clear Default Conversation : The home page default conversation now supports a clear action. Notification toggle persists across sessions. Auto-Generated Archive Titles : Archived conversations now get a short AI-generated topic summary as the title for easier browsing. Sidebar Context Menu : Pinned and history sessions now support right-click actions. “Bookmarks” section renamed to “Pinned” for clarity. Windows Install Improvements : Auto-creates desktop shortcut, warns about system-protected directories, auto-configures firewall rules on first launch. Smart Power Management : System sleep prevention now activates only when Agent is actively running, saving battery during idle time. Settings Shortcut : Ctrl+, shortcut to open settings on Windows. Input box only auto-detects links with explicit http(s):// protocol, preventing false positives on regular text. @ file mention search results show full file path on hover. Feishu IM card interaction and notification style improvements. Long code replies now scroll horizontally. Multiple visual refinements. Fixed schedule agent selector incorrectly showing internal project agents. Fixed IM message filter rules not persisting across restarts. Fixed project selector dropdown direction in mini-window mode. Improved SSL connection compatibility for some enterprise networks. ​ v3.0.31 — 2026-05-23 macOS (Apple silicon) Windows (64-bit) More accurate permission handling for compound commands and file deletions. Fixed permission rules not applying correctly in some scenarios. ​ v3.0.30 — 2026-05-22 macOS (Apple silicon) Windows (64-bit) Skill Lifecycle Management : New “Needs Attention” and “Archived” filters in the Skill panel. Keep, archive, pin, or restore skills with one click. Unused skills are auto-archived with notifications. Batch Skill Import : Preview GitHub repo contents before installing. Select and install multiple skills from the same repo in one go. Multi-Step Questionnaire : Agent can present structured multi-step questions with single/multi-select options, free-text input, and images. Content Filter Graceful Fallback : When the model can’t complete a response, the original message returns to the input box with a friendly notice instead of an error. Rephrase and resend directly. Avatar Click-to-Upload : Click the Agent avatar on the home page to upload a custom image. Message Hover Timestamp : Hover near the copy button to see when a message was sent. Worktree Mode Toggle : Opt into isolated code development mode when starting a new conversation. Subscription Tier Display : Current subscription tier shown in user menus with one-click upgrade. Markdown heading levels auto-optimized in chat for cleaner visual hierarchy. Unified folder permissions: one authorization covers read, write, edit, and search — no repeated prompts for the same folder. Smarter permission handling for common command patterns, significantly fewer prompts in daily use. Schedule panel and IM connection panel visual polish. Fixed occasional “Service Start Failed” after desktop app restart. Fixed memory leak when closing tabs. Fixed multi-line commands being incorrectly blocked by the permission system. Fixed Agent tool calls passing inaccurate parameters in some scenarios. Windows firewall blocking now shows dedicated setup guidance. ​ v3.0.29 — 2026-05-20 macOS (Apple silicon) Windows (64-bit) Agent reflection entries now surface as cards instead of occupying the chat stream. User-level memory writes now require a reason to prevent project-specific info from leaking into personal memory. ​ v3.0.28 — 2026-05-19 macOS (Apple silicon) Windows (64-bit) In-Session Search (Cmd+F / Ctrl+F) : Search the current conversation with keyboard shortcuts. Supports keyword highlighting, prev/next navigation, and IME input. Native MCP Tool Exposure : Selected MCP tools can now be registered as model-native tools for improved call efficiency and success rate. Windows Recycle Bin Support : File deletions now go through the system Recycle Bin, making accidental deletions recoverable. Permission Review Reasons : Permission cards now display LLM review rationale, helping users understand why confirmation is needed. Chat Scroll Progress Bar : A scroll progress indicator in long conversations for quick orientation. Permission mode dropdown reordered: Ask First → Smart Authorize → Always Authorize. “Always Allow” label unified across UI and Feishu to match native OS wording. Temp directory and home directory reads auto-allowed, reducing unnecessary permission prompts. Introduced HARD / SOFT dual-tier risk registries — high-risk commands blocked immediately, medium-risk routed to LLM contextual review. Added detection for Windows delete commands, disk erasure, reverse shells, and interpreter inline deletions. ​ v3.0.27 — 2026-05-16 macOS (Apple silicon) Windows (64-bit) Credit & Usage Limit System : Clear prompts when credits are exhausted with direct links to subscription renewal — no more infinite loading. Home Page Quick Actions : Shortcut buttons below the input box to quickly start creating teams, slides, PDFs, documents, and spreadsheets. Feishu One-Click Authorization : Automatic user authorization after Feishu app registration — one scan to get the recommended permission set. Desktop Notifications : System notifications when Agent encounters errors or needs permission confirmation. Enhanced Scheduled Tasks : New hourly frequency option, /watch shortcut command, and improved task result cards. Deeplink Support : Open specific conversations directly via links. Markdown typography overhaul: headings, tables, code blocks, and blockquotes redesigned per Figma specs. Project selector improvements: Windows paths shown as folder names with full path tooltips on hover. Activity panel redesign with privacy safeguards and day-by-day navigation. Permission classifier upgraded from binary to three-state (allow / confirm / block), with 90+ common commands on the fast track. Catastrophic commands now require explicit user confirmation and cannot be bypassed. Fixed API failures causing infinite Agent restarts. Fixed SSE reconnection process leak (100+ processes observed on Windows). Fixed orphaned processes remaining after Electron quit. Fixed onboarding page reappearing after app updates. ​ v3.0.26 — 2026-05-14 macOS (Apple silicon) Windows (64-bit) Simplified Permission Modes : Permission mode selection streamlined to Smart Authorize and Always Authorize. Scoped Permission Cards : Permission prompts now only show requests from the active session tree. Compression Image : Optimized image compression pipeline for faster uploads. Fixed IM message reply errors. Multiple UI stability improvements. ​ v3.0.25 — 2026-05-13 macOS (Apple silicon) Windows (64-bit) Built-in Skills status corrected to show as installed, ensuring proper invocation. Migration guide popup visual upgrade with clearer copy. Fixed multiple Windows compatibility issues. ​ v3.0.24 — 2026-05-13 macOS (Apple silicon) Windows (64-bit) Composer Image Preview : Preview attached images before sending a message. Sidebar No-Project Indicator : Sessions without an associated project now display a dot icon for quick identification. Workspace Panel Optimization : Working Folders area supports internal scrolling without expanding the page. Keep-Awake Default : New installations default to preventing system sleep, ensuring long-running tasks are not interrupted. Unified avatar component with consistent visual style. Dark mode scan light effect fixed. Team member usage display compatibility improvements. Fixed data resumption errors when switching sessions. Fixed queued messages being incorrectly aborted. ​ v3.0.23 — 2026-05-11 macOS (Apple silicon) Windows (64-bit) Sidebar Pin & Drag-to-Reorder : Pin frequently used sessions to the top and reorder them via drag-and-drop. Session-Level Message Management : Manage and organize messages at the session level. Work Insights : Activity panel now includes work insight analysis with multimodal model awareness. Auto-Compaction Summaries : Long conversations display collapsible summary blocks when auto-compacted. Session rotation mechanism upgraded for smoother transitions. IM message filtering improvements. Daemon login state automatically cleared on logout. Fixed IM multi-message send pipeline issues. Fixed Windows installer display name. ​ v3.0.22 — 2026-05-08 macOS (Apple silicon) Windows (64-bit) Rich Text Input : Input box upgraded to a rich text editor with support for richer formatting. @File References : Reference files with @ in conversations — Agent can read and process them directly. Mini Window Defaults to Main Agent : Quick-launch mini window now defaults to the main Agent. Improved send-immediately and loading experience. Fixed multiple permission and security edge cases. ​ v3.0.21 — 2026-05-06 macOS (Apple silicon) Windows (64-bit) Unified Upload Flow : File upload pipeline unified for a more consistent experience. Agent Team waiting state display improvements. Scroll and expand/collapse interaction polish. Fixed Electron build signing symlink issues. Fixed plugin SDK dependency cold-start delay. Fixed MCP tool calls losing tokens. ​ v3.0.20 — 2026-05-01 macOS (Apple silicon) Windows (64-bit) ​ New Architecture Launch Agent Persona System : Configure independent personas for each Agent, including personality, behavior style, and domain expertise. Cross-Session Memory : Three-layer memory system (User / Agent / Project) for persistent learning across conversations. Multi-Agent Collaboration : Multiple sub-agents work in parallel, automatically splitting and assigning complex tasks. Security & Permissions : Built-in content safety and permission management — Agent asks for confirmation before executing sensitive operations. MCP Tool Integration : Connect to the MCP tool ecosystem to extend Agent capabilities. Multi-Channel Access : Integrated with Feishu, Telegram, and other IM platforms for 24/7 real-time conversations. Scheduled Tasks : Create recurring tasks with natural language — Agent executes and reports results automatically. ​ 0.6.7 · 2026-10-10 Set MCODE_BROWSER_USE_SWIFTSHADER=true to enable software WebGL rendering in the TUI browser on virtual machines without a usable GPU. Browser Use must also be enabled. This opt-in mode is intended for trusted local page testing; CPU rendering is slower and does not measure hardware-accelerated game performance. ​ 0.6.6 · 2026-10-10 config.yaml has a new requestMedia block, read on every use so changes need no restart. requestMedia.image.max_width , max_height and max_base64_bytes set the size every image entering the conversation is encoded to, for both tool results and user attachments (defaults 2000 px, 2000 px and 4.5 MiB; the byte limit is capped at 5 MiB). requestMedia.recoveryDivisor (default 2 ) sets how much of the images a rejected request keeps when it is resent. Invalid values fall back to the defaults field by field. When a provider rejects a request because it carries too many images or too many bytes (HTTP 413, request_too_large , “Request Entity Too Large”, “too many images”, per-image size limits), the turn now drops the older images and resends once instead of failing every later request the same way. The newest image is always kept. Tool result images dropped this way are saved to the Session’s reports/tool-outputs/ folder and replaced with a note giving the path, so read can load them again. Images in each request are kept within a budget: 20 images for MiniMax M3 models and 100 for other models unless the model sets max_images_per_request , and two thirds of the request body budget in bytes. The request body budget defaults by endpoint (28 MiB for the official Messages API host, 26 MiB for Google Cloud, 64 MiB for api.openai.com , 18 MiB for other gateways) unless the model sets max_request_body_bytes . MCP tool images, Computer Use screenshots and pasted or @path images are encoded like Read tool images. A shrunk image is followed by a note with its original size and the coordinate scale. Computer Use sessions no longer keep only the latest 2 screenshots per step. Compaction summaries no longer send images to the summary model; each image is described by its saved path. After a compaction, the summary names the Session’s snapshots/ and reports/tool-outputs/ folders so earlier files can be found. A stdio MCP server returning a message over 10 MiB, such as a full-size image, no longer closes the connection with Connection closed ; the limit is now 64 MiB. A 413 with a JSON body from a BYOK provider now recovers: the error keeps its HTTP status, so the image recovery above applies instead of the request failing. requestMedia.image is now the only resize applied to Read tool images, attached images and browser screenshots. Limits above 1920 or 2000 px take effect, an image within the limits is sent unchanged, and a scaled-down image, attachments included, carries one note on its original size instead of two. Attached images no longer share a 512 KiB budget per message, so one 1900×1000 image among several is no longer halved. A screenshot taken after a browser action stays within 1280 px and 1 MiB, or the configured limits when smaller. Rejections for image dimensions, and errors that report a 413 inside another status, are no longer retried as transient failures; the image recovery above runs on the first rejection. When images just returned by tools push the context over the automatic compaction threshold, the older tool result images are saved to reports/tool-outputs/ and replaced with their path, keeping the newest, instead of summarizing the whole conversation. The summary dropped those images, and the model went on to describe images it no longer had. A Bash rm -f that is moved to the trash keeps -f , so a missing path in cd project && rm -rf missing && mkdir -p output no longer fails and stops the commands after && . A redirect such as 2>&1 , even after -- , is no longer treated as a file to delete. Long-thinking models behind a gateway that holds the response until generation finishes no longer fail after about 5 minutes 15 seconds with LLM request timed out . SSE heartbeats from such a gateway are not model events, so the stream idle bound used to cut the request and its single retry. The stream idle default is now 60 minutes; the first-response bound stays 300 seconds, and a connection that stops sending any bytes is still closed after 300 seconds. MCODE_LLM_STREAM_IDLE_TIMEOUT_MS still overrides it. ​ 0.6.5 · 2026-10-08 The TUI and embedded CLI now honor the per-Session subagent concurrency limit from config.yaml , the same setting Desktop already used. The default stays 4 , and only a positive safe integer takes effect. New task and activating task_append calls are admitted immediately: when the limit is reached they return SUBAGENT_CONCURRENCY_LIMIT instead of queueing and reporting Started first. Changing the value takes effect after restarting the Runtime. When a TUI and Desktop share one data directory, both count occupancy from the same parent-child records, so each process sees the same usage. The output-speed readout no longer spikes at the start of a turn. Output buffered upstream and released in one burst is now excluded, so a burst arriving within the first tenth of the request no longer reports a rate far above what the model actually produced. ​ 0.6.4 · 2026-10-08 Start a new Session in lightweight conversation mode with --mode lightweight . It sends a small conversational system prompt and no tool schemas, and omits workspace instructions, Skills, memory blocks, and MCP schemas, so it cannot inspect or change local files. The mode is fixed when a root Session is created and cannot be combined with --session or --continue ; child Agents, /compact , ACP, and Desktop Sessions stay on the standard context. Reopening a lightweight Session keeps its mode and shows Lightweight in the status line. Each model request attempt now has a first-response bound and a stream idle bound, both 300 seconds by default. A request that passes a bound is aborted and reported as a retryable timeout, so the normal model-request retry runs instead of waiting out the 20-minute overall request limit; once visible output has started, the turn fails instead of retrying. Override the bounds with MCODE_LLM_FIRST_EVENT_TIMEOUT_MS and MCODE_LLM_STREAM_IDLE_TIMEOUT_MS , where 0 disables a bound. A single model request carries at most 20 images by default. Older images in the request are replaced with a text note while the stored history keeps every image, and a provider that allows fewer images per request can set max_images_per_request on the model. /compact falls back to a candidate without attachments when the provider rejects an image limit. Deterministic provider rejections, such as an image limit or an HTTP 400 invalid_request_error , are now reported immediately instead of being retried. Scrolling back through a long conversation no longer snaps the viewport back to the top of the session when output changes already-scrolled text or collapses a region while you have not typed anything. The rebuild is deferred to your next keystroke, so the screen holds your position; readers following the tail still stay at the bottom. Clearing or pausing a Goal while a response is still running now says that the current response keeps running and to press Esc to interrupt it, instead of implying the turn already stopped. /retry is only offered when the Session is idle. A response that is still running has not failed yet, so the command points at stopping it instead of reporting that there is nothing to retry. cron once --after accepts w (weeks) alongside the existing duration units. ​ 0.6.3 · 2026-10-05 Customize the main Agent’s system prompt at startup with --system-prompt or --system-prompt-file , or extend it with --append-system-prompt or --append-system-prompt-file . Available in the interactive TUI, exec , exec review , and ACP. Replacement and appended prompts can be combined. They apply only to the current process, preserve runtime rules and project instructions, and do not affect Task child Agents. ​ 0.6.2 · 2026-10-02 In a /btw side session, pressing Esc on an empty Composer no longer suggests editing the latest message; /edit is unavailable there. Switching between a side session and the main conversation no longer resets the main run’s Running/Loading timer to zero. ​ 0.6.1 · 2026-10-01 /doctor and /feedback now work inside a /btw side session, so you can diagnose or report a failed side response without leaving it. ​ 0.6.0 · 2026-10-01 /retry now works inside a /btw side session. When a side-session model request fails and can be retried, run /retry there to resend the side session’s last message without switching back to the main conversation. The official default model is now MiniMax-M3.1-Flash-Preview when no model has been selected. When a Goal is marked complete, the agent now ends with one final reply that summarizes what was done and where the deliverables are, instead of stopping right after the status update. Safety refusals from the model provider are no longer retried, including on BYOK. The original error is shown and identified as a refusal from the model provider. TLS record verification errors during a model request now trigger the automatic retry like other transient network failures. In regular mode, the screen no longer jumps back to the opening logo and replays the whole conversation while a run is in progress. The tok/s speed is no longer inflated when the provider hides thinking text; generation time now starts when thinking starts. Fixed repeated failures after context compaction during an automatic continuation. Fixed sessions that could pause unexpectedly after continuing from a questionnaire answer. ​ 0.5.10 · 2026-09-30 Your own messages now appear exactly as typed in the conversation, pending instructions, and the transcript viewer. Text such as __pycache__ , * , and a leading > is no longer reformatted as Markdown, and trailing blank lines no longer add empty rows. The TUI no longer stays stuck showing a run as in progress after Runtime has already finished it. Enter now sends a new message instead of steering the finished run, and queued messages are no longer held back. ​ 0.5.9 · 2026-09-29 The build-mode status line now reports background=N , the number of active background tasks in the current root session. Bash still owned by its foreground tool call is not counted, and the count holds at 1 after a turn settles until a fresh task list arrives, so a command that has just moved to the background is not reported as 0 . BYOK requests now apply byte limits to large media and accumulated history context, reducing failures caused by oversized requests. Long conversations load their file list and history navigation with less data, making browsing them faster. Editing and rewinding a message now works after an interrupt that left only a background reminder behind. Terminal errors from a failed run are cleared when the next run starts, so a successful retry no longer shows the earlier error again.

​ 0.5.8 · 2026-09-28 Plugin Hook messages appear in the active conversation and restored history, making Hook results visible without adding them to model context. Stopping a session also stops its background tasks and prevents their completion notifications from restarting the conversation. Switching sessions and /clear continue to preserve background tasks. Restoring a session clears terminal scrollback left by the previous process before replaying history, avoiding duplicate content. Normal startup preserves shell history. Improved output queuing on slow terminals and waits for pending output before exiting, suspending, or opening an external editor. ​ 0.5.7 · 2026-09-28 Custom provider setup can import models from /models with the API key entered during setup. Search and select a model, then save the imported models after a successful connection test; retry discovery or enter a model ID manually when needed. Added M3.1-Flash-Preview to the built-in fallback model catalog. Bash tool descriptions are displayed as plain text instead of receiving shell command syntax highlighting; actual commands retain syntax highlighting. ​ 0.5.6 · 2026-09-27 Token speed now uses actual provider output tokens and the generation window, excluding the wait for the first token and tool execution. Streaming text estimates are no longer shown; confirmed speed remains visible while waiting for subsequent requests or tools. History reloads preserve generation timing. Older messages without timing are excluded from speed statistics, and speeds below 10 token/s retain one decimal place. ​ 0.5.5 · 2026-09-25 Bash commands that automatically move to the background now have a one-hour default and maximum total timeout. Explicit background tasks also default to one hour when no timeout is specified. Foreground calls may return a task ID after 60 seconds without resetting the command timer. Reduced the dependencies loaded at startup to keep the startup bundle smaller. Improved automatic context compaction reliability: large-context models reserve room for replies and start compacting earlier, and adaptive-thinking models at high reasoning effort no longer fail repeatedly when reasoning consumes the output budget. ​ 0.5.4 · 2026-09-24 Press Ctrl+S in /model to add or remove a favorite. Favorites appear in a separate group, retain their insertion order, and persist across sessions. Bash tools describe the selected shell and the current turn’s background capabilities. Managed foreground commands wait up to 60 seconds before returning a task ID, with one command deadline covering foreground and background execution. Large Bash results preserve the beginning and end of the output, report process outcomes and output persistence separately, and keep task receipts bounded. Proxy routing remains active after loading tools that use different Undici versions, and updates follow the same proxy settings as the TUI. Scrolling back through chat remains detached from the latest output when the viewport grows or content shrinks. Closing the permission picker restores table rendering correctly. Goal tool calls outside a bound Goal turn return a recoverable error without ending the user turn. Windows managed updates validate the installed package through its Node entry point, including installation paths containing spaces and shell metacharacters. ​ 0.5.3 · 2026-09-23 Type @ to select an installed Plugin. Its source identity is preserved when sending, queueing, restoring drafts, and editing history, so plugins with the same name cannot replace one another. Custom Provider credentials support whole-value ${NAME} environment variable references. The configuration stores only the reference, and missing variables produce a clear error. Provider presets now include OpenRouter and the MiniMax provider family. /theme uses the left and right arrow keys to select light, automatic, or dark appearance. Custom theme watching reconciles once at startup to catch edits that could otherwise be missed. Saving model and Provider configuration preserves YAML comments. External Skill directories support symbolic links to other locations and watch linked directories for additions and changes. Pressing Esc before the model has produced an answer or used a tool restores the submitted message to the composer, preserving attachments and text entered while stopping. Closing autocomplete, history search, settings, or multiline input restores the released chat area correctly. Sending a message resumes following the latest output. Editing or rewinding history clears todos from discarded turns and stale duration indicators. Entering the complete /rewind command while editing still runs the local command. Large-file edit previews now bound diff size and computation time. Omitted previews show a clear notice instead of incorrectly reporting no changes. Fixed overly small context-compaction budgets. Failed automatic compaction keeps the original history and lets the model service decide whether the request fits. Compaction requests retry once when the model service returns an error. ​ 0.5.2 · 2026-09-22 /theme opens a theme panel with live preview. Pick from the built-in minimax , midnight , graphite , and aurora themes, or write your own theme file under tui/themes/*.json . Each theme provides both a dark and a light palette, and the appearance follows your terminal’s background until you lock it to light or dark. Custom theme files support partial overrides, color aliases, syntax palettes, and hot reload, so editing a theme you are already using takes effect without restarting MCode. The theme choice is stored in tui/tui-settings.json alongside tuiMode and can also be set as <theme-id>/light or <theme-id>/dark to pin one appearance. An unrecognized value falls back to the default theme instead of blocking startup. When the build-mode status line is enabled, session and turn now report the original Runtime IDs, so sessions and turns can be correlated directly. request remains a 6-character opaque token. In sessions that can read background task output, task now runs in the background by default and reports progress through the task output tools. session send now returns as soon as the message is accepted. A busy target Session queues the message or steers the running turn based on your sending preference; only sync: true waits for the target to finish. Use force_steer to always steer a running turn. Subagents report progress as they hit results, change direction, or hit blockers, and separate read-only investigation, code review, acceptance, and implementation more clearly. Turning off data contribution now takes effect without restarting MCode, and turning it back on applies on the next run. A failed optional Provider preset sync no longer blocks startup as long as the existing configuration stays intact. Long conversations stay responsive by reusing verified history and avoiding repeated token estimation. Fixed the transcript jumping back to the start of a conversation when a long turn ended after you had scrolled up. Previously visible content now stays in place and the composer remains at the bottom. Fixed failed commands being reported as successful when Bash ran through the Windows PowerShell wrapper; the native exit code is now preserved. Attachment paths are resolved automatically in WSL, so Windows drive letters, UNC paths, and file:// URLs can be attached directly. A finished or terminated Goal no longer offers a meaningless continuation. Repeatedly polling the same background task now raises a stop-and-wait reminder instead of spinning silently. Rewriting private configuration files on Windows no longer interrupts startup. Code review, audit, and acceptance requests are routed to Verifier instead of Explore. Product usage Skills no longer trigger for ordinary development work in the product source repo. Skills bound to the primary Agent alias are discovered and invoked correctly again. Custom Providers that only expose limited reasoning levels now start from a supported default. Fixed style-only redraws leaving a stale composer, status bar, or duplicated body text behind in Apple Terminal. Visible-region repaints now erase in place instead of scrolling old lines into the terminal’s native scrollback. ​ 0.5.1 · 2026-09-21 Reduced repeated history processing and token estimation in long conversations. Provider setup now displays the Base URL and supports editing it with ctrl+e before testing. Failed connection tests clearly indicate that changes were not saved. Fixed /btw side conversations failing to open while tools are running. Fixed multiline text pasted by terminals without bracketed-paste framing being submitted line by line. Kept the input method candidate window anchored to the composer during redraws and prevented stale history rows after content changes. Local sessions using custom providers or API keys no longer require a managed login for title handling. ​ 0.5.0 · 2026-09-20 Interactive sessions support mcode -m <provider/model> to select a model for that session. Provider setup supports explicit context and output limits, image input capability, and testing before saving with --use . ACP clients can discover enabled Skills for the attached session’s Agent and workspace alongside native slash commands. Fullscreen Transcript supports clicking and dragging the scrollbar. An optional context-meter status line item shows remaining context capacity. Reduced repeated history processing and token estimation during long conversations. Preserved terminal scrollback during redraws and kept collapsed thinking summaries and the composer visible when content shrinks. Settings updates now reject malformed configuration files without overwriting their contents. Invalid attachment markup is removed from assistant messages even when no valid attachment remains. ​ 0.4.12 · 2026-09-18 The default status line shows context capacity before usage is available, then switches to the remaining percentage as usage updates, avoiding duplicate context indicators. Terminal scrollback is preserved when session status updates trigger a redraw, preventing the history from being cleared or Windows Terminal from jumping to the beginning. ​ 0.4.11 · 2026-09-17 Pasted image paths on macOS and Linux now support escaped spaces, including long filenames and filenames containing both backslashes and spaces. The top login status now refreshes immediately after signing in or out. ​ 0.4.10 · 2026-09-17 The default status line shows the current context capacity, such as Context 1M or Context 512K . Users with a custom status line can enable context-window in /statusline . New sessions and default-model updates now preserve the selected context capacity. Messages submitted during context compaction now enter the waiting queue instead of failing to send. Completed Shell command results no longer reappear in subsequent conversation turns. ​ 0.4.9 · 2026-09-16 mcode exec and mcode exec review support --effort to choose reasoning effort for a single run. Unsupported levels fail before execution, and the override does not change the session’s saved effort. Fixed saving and restoring the default reasoning effort, so new sessions and submitted requests use the same level shown in the model picker and status line. Model selection now respects configured effort defaults and fixed levels while preserving the Thinking Off setting. ​ 0.4.8 · 2026-09-16 Explore subagents can run foreground Bash commands to inspect code and Git state, with a read-only filesystem policy when sandboxing is available. Improved content review handling with guided response regeneration and fixed replies when applicable, while retracting blocked drafts. Improved responsiveness when previewing very long thinking output, while keeping the full text available in expanded details. Unified elapsed time in the running status, turn summary, and Goal display using s , min , and h , retaining seconds for runs longer than an hour. Restored the thinking toggle for capable custom provider models without changing their existing effective thinking defaults. Fixed startup failures when custom agents conflict with built-in agent names, preserving custom configuration during migration. Improved cancellation of streaming responses from compatible providers so interrupted reads can finish promptly. Restored editable image labels when editing a previous message with double Esc or /edit . Images can be removed, restored with undo, previewed when a local file is available, and resent. Fixed attachment validation errors when sending images or files during an active run, and prevented accepted attachments from being lost when temporary draft files are cleaned up. ​ 0.4.7 · 2026-09-15 Choose terminal, coding, or general work instructions for non-interactive runs with mcode exec --prompt-mode tui|coding|work ; terminal mode remains the default. New custom provider models enable switchable thinking by default when supported. Existing model settings and explicit thinking choices are preserved. Discover models from compatible gateways that expose an OpenAI-style model list. Connection tests now use the model’s configured output limit. Improved MCP compatibility for server and tool names, isolated connections across sessions, and prevented connection tests from continuing tool discovery after a timeout. Fixed background Shell results being lost in subtasks and cleaned up remaining background work when a subtask ends. Prevented retries from replaying custom provider responses after visible output has begun. Preserved messages sent during an active run when the run ends abnormally, keeping conversation history consistent. ​ 0.4.6 · 2026-09-14 Preview pasted images above the input, including format, dimensions, file size, and filename. Move the cursor back to an image label to view it again. Terminals without image support still show image details. Press Esc to dismiss an image preview; Enter continues to send. Preview failures do not prevent sending the original image. Removed extra per-message arrows and command highlights in terminals such as iTerm2, and avoided terminal markers interfering with cursor positioning during redraws. ​ 0.4.5 · 2026-09-14 Web search now uses a built-in tool without waiting for a separate search MCP service to start. User-configured MCP servers continue to work as before. Removed the search service component that is no longer needed in the installation package. ​ 0.4.3 · 2026-09-13 Configure model-level compat in config.yaml for custom providers to declare upstream message roles, token parameters, and thinking formats when using a gateway. Explicit configuration is required for these overrides to take effect. Removed the duplicate, truncated original title from the session rename panel. The input retains the complete title and scrolls as you move the cursor to edit it. ​ 0.4.2 · 2026-09-11 Automatically load project MCP servers from .mcp.json in the session’s main working directory, with configuration sources and connection status visible in /mcp . Switch supported model context sizes, including M3 512K / 1M, with Tab / Shift+Tab in /model ; save the selection for the current session and future use. Where OpenAI Codex login is available, choose browser or device code login, with cancellation and retry for expired codes. Refresh a third-party connection’s model list in /provider with r , and edit its API Key, Base URL, models, or alias with e . Reuse an existing provider connection when adding models without entering the API Key again; add another account explicitly when needed. Fixed mismatched tool arguments when multiple OpenAI Responses tool calls stream concurrently. The side-conversation shortcut now clearly warns when no side conversation exists and directs you to run /btw first. ​ 0.4.1 · 2026-09-11 Switching models in an existing conversation warns that prompt cache reuse may be lost and additional costs may apply. Goal displays its current wait reason, including result verification. Goal pauses after three consecutive rounds without tool calls to prevent continued idle execution. Restored the terminal-specific system prompt, with capability instructions following the enabled features. Fixed Goal continuation after user intervention and late usage settlement, and allowed queued user messages to proceed while a Goal waits on dependencies. Shortcut hints follow custom keybindings, including aliases and disabled bindings. Background task details show the full command and error, wrapping long and multiline content. ​ 0.4.0 · 2026-09-11 Review local changes with /review or mcode exec review , with findings linked to file locations. Run Shell commands with ! , keep results local with !! , and use Tab to complete commands and paths. Customize the status line interactively with /statusline , apply changes immediately, and display output from a custom command. Slash commands support argument hints and continued completion; command panels use consistent titles, borders, and action bars. Headless tasks provide structured output validation and clearer progress and failure diagnostics. Improved recovery when context compaction encounters oversized input, and preserved history consistency when forking conversations. Diagnostic reports retain available files when part of the collection fails. ​ 0.3.11 · 2026-09-09 Fixed long-running tasks stopping with a 401 error after repeated sign-in credential renewals. ​ 0.3.10 · 2026-09-08 Selected options use consistent blue highlighting across TUI menus and pickers. Command descriptions consistently use English across the command menu. ​ 0.3.9 · 2026-09-07 /btw and /side sessions now follow the parent session permission mode without an extra confirmation step. Full access needs no additional approval; other modes retain their normal checks. Draft recovery cleanup skips missing files, preventing repeated no-op deletions from exhausting the safety budget and disrupting later operations. ​ 0.3.8 · 2026-09-07 /status now shows the running version, account and plan, instruction file sources, Default/Plan mode transitions, and remaining 5-hour and weekly quotas with reset countdowns. The compact status panel wraps long paths and keeps local configuration visible while account details load or when the account request fails. /logout and mcode logout print a complete browser sign-out link and open the page for the current account region and environment. On remote or headless machines, copy the link to your browser; command completion does not wait for the browser or page result. Fixed authentication when uploading automatic error reports from a signed-in account. ​ 0.3.7 · 2026-09-07 Long sessions search and render more smoothly, with cached search indexes and stable Markdown, fewer history replays during terminal resizing, and clearer tool retry summaries. The default status line warns when remaining context reaches 25% or less. Explicitly configuring context-remaining keeps valid usage visible; narrow terminals prioritize essential status. Terminal notifications support focus-aware policies and configurable delivery methods, with improved compatibility for cmux. Composer keyboard hints follow the current platform. Improved shared login recovery when switching accounts, cancelling login, or starting offline, including concurrent credential renewal and delayed logout synchronization. Installation and repair guidance now preserves the selected package, version, and registry and explains how to restore native SQLite dependencies when npm 12 blocks installation scripts. ​ 0.3.6 · 2026-09-07 When MCode shares its data directory with MiniMax Desktop, official model endpoints are now resolved in memory for the current environment: reading configuration or first-run initialization no longer writes new official endpoints into the shared config file, so older Desktop versions keep working alongside the TUI. Custom (BYOK) endpoints behave as before. Legacy model recovery now also covers numbered model aliases and retired default models, and applies consistently to the model list, Goal model checks, and login checks: selections pointing at the legacy official gateway resolve to the same-name official model, then the configured default model, then the current official default. Shared default models, session model references, and migration records are preserved; only explicit user selections are written back to configuration. Installing or updating with npm 12 no longer breaks the native SQLite module: install and update authorize the MCode and better-sqlite3 lifecycle scripts and keep optional native dependencies. When native SQLite fails to load, startup prints a repair command matching the install registry, and early startup errors are reported instead of failing silently. ​ 0.3.5 · 2026-09-07 Legacy model repair now also covers resuming sessions and queued messages, not just startup: selections from the legacy custom provider switch to the same model on the official MiniMax provider, or to the official default model when the same model is unavailable, and the repair is saved to the session or default configuration. The model list no longer shows legacy entries. After a response fails or is stopped, queued messages that have not run yet are kept and shown as paused. Run /queue and press c to continue from the front of the queue; pending messages can also be edited, restored, or deleted. Sending a new message while paused offers to send it first and then continue the original queue, or to clear pending messages first; cancelling keeps your draft. Reopening the session shows the paused queue and waits for you to continue. In-TUI updates now run npm with MCode’s own Node runtime and validate the new version’s native modules with that exact runtime before activation, fixing startup failures after updating on machines whose system Node version differs. If validation fails, the previous version stays active. The macOS/Linux installer resolves the selected Node to its real path for install, launcher, and receipt, and re-checks native modules with that runtime on reruns. Broken same-version installations now enter a repair install automatically instead of being reported as healthy. ​ 0.3.4 · 2026-09-06 On startup, when the selected model comes from the legacy custom provider, the TUI automatically switches to the same model on the official MiniMax provider. /login and /logout are always available. /login refreshes credentials with the server even when already signed in and keeps you signed in if a transient refresh fails; /logout safely does nothing when already signed out. When the server permanently rejects saved credentials, the TUI clears them and falls back to device authorization instead of looping on /login errors. Temporary or configuration errors no longer discard valid credentials. Service or resource failures now report the actual failure instead of asking you to sign in again; sign-in prompts appear only when signing in is really required. Domestic accounts sign in through the updated authentication endpoint. Draft recovery failures no longer interfere with typing or other user actions. Duplicate session history manifests are recovered automatically so session history keeps loading. ​ 0.3.3 · 2026-09-05 Side conversation footers now name the view that the switch shortcut opens. Terminal.app on macOS displays the supported Ctrl+- shortcut for switching between main and side views. Signing in and out now synchronizes the running session’s credentials immediately, including repeated sign-out requests, instead of waiting for background credential notifications. Content review can recover from an expired credential and retry once with a new credential. If recovery fails, the request remains blocked and prompts you to sign in again. ​ 0.3.2 · 2026-09-04 Added /btw [question] and /side [question] to ask follow-up questions in a temporary side conversation while the main task keeps running. Use Ctrl+/ to switch views, /parent to return to the main view, or Ctrl+C in an empty side Composer to discard the side conversation. Side conversations inherit the main conversation’s committed context and show its current task status. Tools that can modify data require explicit confirmation in every permission mode. Internal packages support a global --env test|staging|prod option for the current launch, with pre as an alias for staging . The selection survives login and update restarts without changing saved configuration or the package’s update channel. Public packages do not expose this option. Markdown headings, inline code, and links have clearer visual distinctions, with more readable syntax highlighting in terminals limited to 16 colors. Fixed input-method candidate window positioning in side conversations. Queue hints now follow custom keyboard shortcuts, including multiple or unbound shortcuts. Markdown links preserve nested formatting and restore the surrounding text style correctly. PR and MR links are now associated with the Session’s actual workspace, avoiding links from another checkout. ​ 0.3.1 · 2026-09-04 TUI and Desktop now use the Shared OAuth Core for the same account session; Runtime and embedded mcode-tools consume short-lived access-token leases without reading refresh credentials. Managed mcode-tools packaging validates its version lock, package identity, lease protocol, and CLI SHA-256 before enabling the launcher, and remains fail-closed when validation fails. Added a unified /tasks center for Agent Team and Runtime background work, with a compact severity-aware summary above the Composer and drill-down into child Session transcripts. Added an opt-in machine-readable [V] status line and terminal result records for automation. The regular status line can also show the GitHub PR or GitLab MR for the current branch. Built-in Subagents can retain an explicitly enabled Web Search capability across Runtime startup and recovery. Internal test and staging builds accept an explicit global --lane <name> for OAuth integration testing. Production builds reject the option, and MCode does not infer a lane from environment variables. The Composer now identifies Prompt, Command, and Skill input before submission, shows a localized empty-state hint, and treats text following no-argument slash commands as a regular prompt. Provider setup offers broader models.dev search, clearer model browsing, and safer API-key edits. Queue and Steer presentation now stays aligned with the active Turn through pauses, cancellation, recovery, and follow-up delivery. Model retries, Goal completion, and compaction report clearer progress and actual Provider token usage. Plan Review content remains scrollable in the Transcript. Installing MCode now exposes the managed mcode-tools command without requiring a TUI startup and keeps package-owned launchers clean across updates. Invalid grep regular expressions now return a structured, actionable Search failure instead of exposing ripgrep’s raw parser error, including guidance for literal matching. Editing a Custom Provider invalidates stale connection-test status, so changed URLs, credentials, headers, and API formats are re-evaluated before their status is shown. Fixed Plugin Manager search results ignoring the Space toggle while preserving literal search spaces through Shift+Space . Fixed paused queues rejecting follow-up input, failed Steer cleanup overriding an already-started Turn, and pending Steer messages appearing at the wrong time. Fixed /compact autocomplete occasionally disappearing after Session creation, hidden model retry progress, and incomplete Goal completion or security-review guidance. Fixed new installs exposing the package-private Node runtime on PATH and tightened Windows delete command checks when shell redirection is present. ​ 0.2.7 · 2026-08-28 Installer-managed updates now publish a complete versioned release and atomically switch new Sessions to it, so running Sessions can continue on the previous version without blocking activation. Update output also shows each stage and a heartbeat while the package manager is still working. While a response is running, Alt+Enter queues a follow-up and Tab remains dedicated to Composer completion; Enter continues to steer the active Turn. Welcome highlights and /changelog now select the packaged English or Simplified Chinese changelog from the system locale, with a deterministic English fallback. TUI Sessions now use a dedicated terminal prompt profile and no longer load Desktop user.md profile content, keeping terminal behavior isolated from Desktop-only instructions. Regular and Fullscreen rendering now preserve native scrollback and scroll positions when transient rows collapse, history is remeasured, or interaction chrome changes, avoiding cleared history, large blank gaps, replayed lines, and duplicate footer frames. Permission and Questionnaire panels remain fully visible above Goal and background-task rows, and large expanded Transcript histories only format the visible viewport plus overscan. User prompts keep their full-width highlight, compact first-message spacing, and terminal semantic markers without shifting the first rendered row. ​ 0.2.6 · 2026-08-27 Bash approval prompts and transcripts now extract and highlight the actual command instead of rendering JSON wrappers and escaped characters as shell scripts. The /checkin hint now appears in the startup Tips rotation instead of the version highlights. Remaining usage credits are now displayed as whole numbers, matching the underlying credit unit. ​ 0.2.5 · 2026-08-27 /sessions is now the Session Center, with workspace or global search, switching, renaming, archiving, and restoration. Filters include id: , path: , type: , status: , and model: . Added /history for browsing the current Session input history, previewing and creating a branch, editing an earlier prompt, rewinding only the conversation, or rewinding both the conversation and files. It remains read-only while a response is running. Welcome randomly shows three highlights from the installed version. Added /changelog for viewing the complete packaged history in a read-only panel. Added a task panel above the Composer for live Agent Team and Runtime background-task status without rewriting Transcript history. Added /checkin for daily check-ins, streaks, total check-in days, and reward results inside the TUI. /model now lists and selects models from third-party Providers and guides users through Provider setup when needed. /context , /usage , and /status now open as closable interaction panels without leaving entries in the Transcript or Session history. The status bar and usage panel can show the current Session cache-read ratio and avoid misleading ratios when the available statistics are incomplete. Queue and Steer status, targets, and recovery feedback are clearer. Queued messages can be selected as Steer targets, and queue ownership remains consistent across Session switches, cancellation, and recovery. While a response is running, commands that create, switch, rename, archive, branch, rewind, edit, or compact Sessions are hidden and rejected. Safe read-only history, Queue, Steer, and Stop actions remain available. Approval prompts and Tool summaries use consistent shell-semantic command highlighting for easier reading of long commands. Coding and Work modes preserve Markdown link labels and targets as clickable links instead of incorrectly rendering them as inline code. Regular-mode streaming updates, theme changes, and Surface switches perform fewer full redraws. Long-Session Inspector formatting is limited to content near the current viewport. Production bundles split startup code and defer nonessential resources, reducing version-query, first-frame, and message-ready latency. Long Sessions consistently compact accumulated ToolResult content while retaining recoverable original output, reducing context usage and repeated trimming. The TUI keeps a bounded, redacted local record of abnormal exits and lifecycle failures and uploads it after sign-in. Collection or upload failures never block startup, shutdown, or normal interaction. Fixed Regular mode replaying the full Transcript during live updates, which could damage native terminal scrollback or make history jump. Fixed duplicate redraws and premature cleanup completion during shutdown, updates, returns from an external editor, and terminal restoration. Fixed terminal or autocomplete subprocess EOF incorrectly terminating the TUI or losing the first stop reason. Fixed Session history actions, queued messages, and background tasks being projected into the wrong context after asynchronous responses, recovery, or Session switches. Fixed built-in Agents without an explicit model failing to inherit the current Session model, and explicit routing for Goal verifiers and other subtasks being overwritten by Agent configuration. Fixed legacy main or mavis Agent data preventing Runtime startup. Fixed /fork offering worktree isolation when the current repository could not safely create it and failing only after selection. Fixed MiniMax Plugin Hooks receiving incorrect Tool names or input shapes, which broke Hook matching and input rewriting. Fixed built-in Matrix tools using the wrong authentication method for managed services. Fixed interrupted, repeated, or failed install/update attempts leaving incomplete versions, with recovery from the transaction record. ​ 0.2.4 · 2026-08-24 mcode acp now expands the Session control plane with Session fork, mode and configuration switching, queue/Steer, Goal, and delegation support, while preserving consistent state ownership across connection rebinding, cancellation, and Permission/Questionnaire interactions. Installed Plugin Skills can now be searched and invoked directly with / . Added support for Plugin-compatible lifecycle Hooks, with corresponding Runtime compatibility handling. Goals now support runtime budgets, automatic settlement, and independent verification; Ask waits that continue automatically now show a countdown. mcode exec now provides structured events and output for automation scenarios. Press Ctrl+U to resume the most recent Codex Session. Startup now shows a Server loading state and blocks message submission until Runtime is ready; startup animations and feature pages follow the current TUI mode layout correctly. Composer now shows action Tips while idle. Absolute image paths paste as short placeholders; Agent Team subtask rows remain stable as status updates, and progress indicators for long-running Bash and background tasks are clearer. During long Sessions, context compaction now prioritizes trimming tool results that save at least 70% of context, reducing low-value trimming; background-task reminder deduplication and read counts are more stable. Limited retained Tool Transcript content per Tool input/detail. When limits are exceeded, the beginning and end are preserved with the original byte count, preventing large payloads from occupying memory over long periods. Fixed legacy BYOK Provider, Session, and historical Custom Agent configurations interfering with migration or blocking Runtime during TUI startup. Permission Core now has a deterministic policy owner. Command-level denials remain fail-closed; Windows literal delete commands that can be safely reclaimed run against fixed targets, while complex or unverifiable commands continue to be denied. Fixed pending updates being unable to resume in npm prefixes. Fixed Session canonical history paths changing when the local timezone changes, as well as history retention boundary issues during Host failures and Fork. Fixed compaction failures after model context changes and Compact checkpoints losing Subagent state. ​ 0.2.3 · 2026-08-22 Added /export [path.md] , which exports the complete history of the current Session to Markdown. It supports workspace-relative paths, automatic suffixes for conflicting files, credential redaction, 0600 file permissions, and clickable local file links. The model selector now groups models by Provider, supports real-time search across Provider, model name, display name, variant, and provider/model , and supports Backspace, Delete, left/right cursor movement, and retaining the full model catalog. /rename without arguments opens Session Manager with the current Session preselected. Permission, Plan Review, and Questionnaire interactions now share nested-input focus, cancellation, failure feedback, and recovery paths. Ctrl+C exit confirmation now uses an armed “press again to exit” state. Cleared Drafts can be restored with Ctrl+Z, without a fixed 500 ms window. Improved startup performance by lazy-loading dependencies that are not needed at startup. The MiniMax Code product knowledge Skill is now a top-level route with trusted region context and topic-specific reference loading for account, Agent, extension, and workflow topics. Fixed history finalization after compaction within a Turn continuing to reference a stale identity, which could cause state reconciliation errors in long-running Sessions. Fixed tool_trim persistence rewriting the user message and causing the Turn to fail; original user payloads are now preserved. Fixed mcode update parsing when npm returns a single-element version array. Fixed inconsistent settlement of run results, Agent Team, and Runtime events when reconnects, duplicates, or out-of-order events arrive. ​ 0.2.2 · 2026-08-21 Added /rewind and /edit , which let you select a rollback range from history, preview the changes, edit them, and resubmit. /fork , attachments, and Draft recovery now consistently reuse the Runtime Session contract. When multiple MCode Sessions are running, /update shows the actual number of blocking processes and waits for old processes to exit safely before activating the update, instead of stopping midway at a fixed timeout. When the context budget is insufficient, Slash Review now switches automatically to a compact Git self-discovery Prompt while preserving review results and the recovery path, making long Sessions less likely to fail because of oversized input. Fixed completed subtasks accumulating additional elapsed time after an Agent Team was restored from a historical Session. Fixed terminal EOF being reported as an abnormal exit, and fixed shared cleanup timeouts leaving later resources open. Fixed an old Turn lease remaining attached to a Session after a process restart and blocking later operations. ​ 0.2.1 · 2026-08-20 /settings now shows the actual layout differences between Regular and Fullscreen. Arrow keys only move focus; the selected mode is applied after confirmation. Fixed the Fullscreen input box lagging by one frame, which made each keypress appear only after the next key was pressed. ​ 0.2.0 · 2026-08-20 Rebuilt the terminal rendering and input foundation to unify Regular and Fullscreen behavior across Markdown, Unicode, images, keyboards, mice, search, and window resizing. Permission requests now use complete approval cards that clearly show commands, files, matching rules, and risks, with separate options for one-time approval, session approval, and rejection. The approval stays open until Runtime confirms the result. /copy in SSH sessions can write to the local clipboard through the terminal protocol. Task-completion notifications are also deduplicated according to terminal capabilities. Added the sessionTitle.enabled setting to disable Runtime requests for Session title generation. Redesigned /context , /usage , and /status to present context composition, Token usage, model, reasoning effort, Compaction, and runtime status through capacity, proportions, and status hierarchy. Runtime now adjusts the output budget and Compact threshold dynamically based on the remaining context. Long sessions no longer compact too early because of the model’s full output limit, while preserving safe space for a single-turn response. When a long-running task has unfinished Todo items, the Agent checks progress again on the assistant iteration cadence, reducing the chance of remaining steps being missed. Project-level Skills now appear in TUI Slash Command completion. MCP Servers remain in a starting state during cold start instead of being incorrectly reported as failed after the first status check times out. Fullscreen scrolling, search, selection, and tail-follow are managed by one viewport. Switching between Regular and Fullscreen no longer rebuilds Runtime or loses Sessions, Drafts, attachments, or blocked interactions. Fixed the input box being clipped when Slash Command completion or a feature list was expanded in Fullscreen, and fixed the need to press an extra key before the input recovered after exiting. Fixed Fullscreen AskUser covering recent messages and namespaced AskUser Tool completion states leaking into the Transcript. Fixed the viewport jumping back to the top when Tool details were toggled during streaming. On Windows, image pasting now prefers Alt+V , while Ctrl+Z provides Editor/Draft undo without conflicting with Windows Terminal zoom. Fixed the Composer failing to accept input after an external editor launched with Ctrl+G in PowerShell exited. Fixed Runtime state refreshes in Regular mode interrupting Slash Command candidates, so the first / could show no candidates and later input could lag behind the list. Fixed long AskUser, Queue, and Plan viewer panels being clipped in Regular mode, hiding the question, current option, or document content. Unified how Settings, AskUser, and other interactions replace the Composer with Pi’s mounted-state semantics, preventing empty-frame flicker, blocked input, and stale content after an interaction ends. Fixed completed background Subagents reusing the previous Query’s collapsed state and timing, and fixed /review queued-message sources and bubbles failing to reconcile correctly. Fixed an invalid Workspace with accelerated indexing enabled from blocking the current Session from sending messages. Fixed Read-Before-Write path identity mismatches for local Windows file tools, and fixed the installer misidentifying PowerShell’s mcode.ps1 launcher resolution as a failure. Fixed /update overwriting files in place inside an npm prefix and locking native modules. Updates now use isolated staging followed by an atomic switch. ​ 0.1.6 · 2026-08-19 Fixed Matrix long-running tools such as image generation, image search and downloads, reverse image search, video, and audio failing with fetch failed under Node.js 26. ​ 0.1.5 · 2026-08-19 Added /retry , which resends a preserved message after a Provider or network failure. Unfinished requests can also continue when restoring a historical Session, and blocked Goals can be resumed directly. Added configurable status-bar sections and ordering for directory, Session, Git, model, reasoning effort, context, and other information, including the current run state. Provider, network, authentication, and rate-limit errors now explain the cause and next steps more clearly. Retry progress is shown during automatic retries, and failed-request context is preserved when switching or restoring Sessions. MCP Servers remain in a starting state during cold start instead of being incorrectly reported as failed when the first status check times out. Models and reasoning effort now take effect correctly before the first request. BYOK connectivity tests and the Codex OAuth proxy behave more accurately. The installer safely reuses a compatible system Node.js when possible, preserves the existing PATH, and supports explicitly selecting a verified download mirror for mainland China. Fixed the input Draft being cleared when selecting and copying messages in PowerShell and Windows Terminal. Fixed /quit being restored as an input Draft, and fixed Session restoration incorrectly triggering Queue sends. Fixed duplicate message display after pasting images, Goal inline attachments failing to start, and elapsed execution time being lost after history pagination. Fixed Windows launcher path escaping and compatibility issues in some installation and update scenarios. ​ 0.1.4 · 2026-08-18 Added /goal for creating, viewing, and managing a Goal in the current TUI Session, including its status, progress, and elapsed time. Goal semantics are aligned with Desktop and Runtime. Added Plugin management: mcode plugin and the interactive Plugin manager can browse official and local sources, install, enable, disable, and remove Plugins. Improved /provider and /model : view and test Providers, switch between MiniMax OAuth and API Key, and select models and reasoning effort. The active model and effort are also shown in the status bar and /status . Improved ACP and ongoing collaboration: persistent Sessions can be listed, loaded, resumed, and closed; while a task is running, you can Queue follow-up messages and steer the current Draft or the next queued message. /new no longer breaks Session identity. Unified mouse scrolling, PgUp / PgDn , text selection, and copy behavior. Selections stay anchored as content scrolls and support automatic scrolling while dragging; non-TTY and compatible terminals continue to use native terminal scrollback. Added visual information for /context and /usage , including the context window, Token breakdown, session usage, and Compaction status. Tool calls, Skill completion, and Queue messages now appear more promptly. Generated-file receipts now use clear, clickable paths that open the corresponding file or directory. Permission, Questionnaire, and Compaction requests now consistently follow the Runtime contract, reducing lost or duplicated pending-request states. Fixed copy offsets while scrolling a dragged selection and delayed rendering of Queue messages. Fixed built-in Matrix tools being incorrectly removed when the mcode-tools switch was enabled. Fixed image pasting on Windows and WSL, along with compatibility issues in TUI generated-file path display. ​ 0.1.3 · 2026-08-17 ACP editor connections can reopen or continue previous Sessions after a restart. Also fixed Session mismatches that could be caused by /new . JSON output from mcode exec now includes the model actually used, making usage reporting easier. Added /copy , which copies the latest Assistant response to the system clipboard while preserving Markdown. If there is no response to copy or the clipboard is unavailable, MCode provides a clear explanation and a manual-copy path. Session titles are more accurate, and tool calls and Skill completion are clearer. macOS scroll shortcut hints now use Fn combinations that better match the actual keyboard. Conversations can now start while using a MiniMax API Key without being signed in. Fixed extra text being appended to delivered-file paths, and fixed frequent flicker in the input cursor and borders. Improved WSL sign-in and Windows PowerShell installer compatibility. Self-Evolving Skills : Automatically extracts reusable Skills from complex tasks, loaded on demand and continuously refined. Persistent Cross-Session Memory : Agent remembers feedback across sessions and improves over time. Natural Language Scheduled Tasks : Create scheduled tasks using natural language for 24/7 automated execution. Parallel Subagents : Multiple subagents run in parallel to accelerate complex task processing. MiniMax M2.7 Model : Improved tool-calling accuracy, Skills adherence, and Agent harness compatibility. Zero Setup : No server or API key required — works out of the box. Multi-Channel Support : Integrated with Feishu (Lark) for 24/7 real-time conversations. MaxClaw Settings Panel : The MaxClaw settings panel is now live, addressing the frequently reported difficulty of managing scheduled tasks. It also supports convenient management of MaxClaw’s connection channels, persona configurations, and associated skills. Personal Companion & Productivity Enhancements : Added visual shortcut buttons above the input box for one-click creation of PPTs, PDFs, and other files, as well as quick connections to platforms like WeChat and Feishu (Lark), enhancing both productivity and personalized companion experiences. Assets : All your uploaded files and AI deliverables, now organized in one place on the web. Access websites, documents, Excels, PPTs, images, and videos with ease. Download any asset, attach files to new tasks, or jump directly to the conversation where they were created. Expert : Track changes with version history and restore previous configurations. Add conversation starters for quick onboarding. Rate and discover quality experts from the community. Final deliverable : Polished the styling of all deliverables to help you find and review results faster. ​ Desktop Feature Scheduled Tasks on Desktop : Create recurring workflows in natural language from the query bar, or set them up via the Schedules page. Quick Launch : Summon MiniMax Agent anytime with Alt+A (Windows) / Option+A (Mac) to start a new chat or pick up where you left off. Browser Expert now works seamlessly even when the window is minimized. General UI polish across the board. Desktop Application : MiniMax Agent now available on Mac and Windows. Select a workspace and let AI read, analyze, and batch-process local files. Describe your task and let AI navigate, click, and submit in your browser. Experts : Create and save reusable configurations for recurring tasks with customizable instructions, models, and subagents. Custom Instructions : Set project-level instructions to guide MiniMax Agent’s behavior across all chats for consistent responses. Multi-Model Support : Assign different models to Main Agent and Subagents in Custom Mode for optimized performance. Generative UI : Automatically rendered UI elements in Pro and Custom modes for design selections and step-by-step tasks. URLs in the chatbox now display as styled links for easier identification. Right-side viewport can now be expanded or collapsed as needed. Scheduled Task : Create recurring workflows using natural language, allowing the Agent to handle tasks like information gathering or website updates while you rest. Dark Mode : Switch between Light, Dark, and System themes for reduced eyestrain. Preset Query : Access example queries to better understand MiniMax Agent’s capabilities. Custom Mode : Configure Agent behavior through custom settings to match specific workflows. Introduces subagents: specialized agents that operate independently with their own context. Subscription Management Page : Easily view account information, manage subscription plans, and check task credit consumption records. Custom Title : Edit the names of Tasks or Branch Sessions to better organize past tasks. Title Search : Quickly search by Task or Branch Session titles. Improved PPT and PDF export performance, significantly faster for large files. Project Settings : Pro Mode now supports pre-configuration and runtime configuration of Supabase and API keys. Share Chat : Now supports previewing, downloading, and packaging files. Optimized browser memory usage. Improved file preview functionality. Complex bug fix : Branch sessions only retain project files to avoid interference from historical session information. Adding new web features : Use branch sessions to manage different features independently. Context limit : Create a new branch session to continue modifying the project when approaching the context limit. ​ Improvements — File Management Supports previewing files at any time with clickable Agent references. Supports file editing and downloading at any level of granularity. Optimized file preview access speed. Team Plan : Shared credit pool and centralized member and billing management. Enhanced PPT Editing : Now supports editing every detail of PPT. Direct Code Editing : Direct code editing via click selection, enabling technical users to modify code without agent prompts. Conversational Help : Ask questions about how to use features and get instant help in chat. Image Editing : Support for image editing with natural language prompts and intelligent object modification. PPT Live Preview : See real-time previews as PPT is being generated. Instant Export : Export presentations to PDF or PPTX while the Agent continues working. Yahoo Finance Integration : Instant access to stock prices, market news, and financial analysis. Selector Editing : Click any element on complex web pages and tell Agent how to improve it. Persistent File Directory : Fixed file directory available at all stages of your workflow. Parallel Tasks : Tasks now run in parallel — less waiting time and lower credit costs. Optimized loading speed for large chat pages. Enhanced image search and understanding capabilities. Improved visual quality in reports and websites. Custom Domain : Connect your own domain to websites built by MiniMax Agent. Custom MiniMax Subdomain : Set a custom subdomain for your MiniMax space. Delete Website : Unbind or completely delete a deployed website. Stripe monthly auto-billing subscription support. Enhanced long conversation summarization. Optimized research efficiency to prevent redundant API calls. Added Mermaid diagram rendering support. ​ Feature — Enhanced Task Control Requirement Clarification : Complex tasks now include plan confirmation before execution. Testing Confirmation : Choose between Agent-performed testing vs user self-confirmation. Web Version History & Settings : Historical web pages display and watermark removal on Basic/Pro Plan. PPT cost optimization with lower-cost models and batch mode. Browser Agent upgraded for lower credit consumption. Dynamic tool loading based on task requirements. Version Control : Introduced version control for agent tasks. Restore Checkpoint : Reverts the agent to a previous state. Edit & Regenerate : Re-executes the task from a checkpoint with modified instructions. Restore Checkpoint : Reverts the agent to a previous state. Edit & Regenerate : Re-executes the task from a checkpoint with modified instructions. Stripe Integration : Seamlessly build websites with robust payment solutions. Reduced average execution time by 30% and token usage by 23%. Accelerated tasks with parallel tool call support. Added a Python linter for pre-execution error checking. Browser Agent switched to self-developed solution : Enhanced long-term memory, lowered credit consumption, improved multi-turn interactions. Enhanced PPTX Export Quality : State-of-the-art HTML-to-PPTX visual fidelity. Backend Services : Supabase integration for database, authentication, and APIs. Built-in Cron job capabilities. Automated testing for bug-free delivery. MCP Builder : Build custom MCPs from scratch or by wrapping existing tools. Add completed MCPs to the marketplace for reuse. Optimized image generation and external API call stability. Enhanced understanding of images, videos, and voice for large batches. Fixed edge cases in content concatenation tasks. Browser Agent now uses a new model with larger context window at lower cost. Fixed browser use and PDF reading issues with large files. Enhanced web development task execution efficiency and reduced credits consumption. Improved Agent’s first response speed. Optimized MCP calls for more stable invocation, especially for MiniMax Audio/Video. Improved multi-turn performance, retains sub-Agent conversation history. Optimized MCP initialization, faster first response. Fixed various tool errors (file, image search, download). Fixed message concatenation issues. Document Export : Supports PDF and Docx formats. PPT Export : Supports PDF and PPTX formats (beta). Significantly improved image search relevance. Tasks with strong visualization needs now generate HTML pages. Long task optimization with reflection process for higher completion rate. Improved web scraping success rate (anti-crawling optimized). Python visualization charts now support CJK fonts.