> For the complete documentation index, see [llms.txt](https://docs.maiagent.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.maiagent.ai/release-note/untitled/archives/06.md).
# June 2026
## v2026.06.29
***
### ✨ New Features
* **Quick Reply Interactive Component**: AI assistants can now display a row of quick reply buttons in conversations to guide users toward tap-to-respond interactions. Administrators can create a Quick Reply type Agent UI in the admin console, configure the action for each button (send a message, open a link, or pre-fill the input box), and choose between a static button row or context-driven dynamic generation. Rendering is supported in both Web Chat and admin console conversations.
* **Unified AI Gateway Management Page**: The AI Gateway management interface has been redesigned, moving from 3 tabs (Overview/Models/BYOK) to 4 tabs. The "Models" tab consolidates adding, editing, and deleting both platform models and custom models; the "Providers" tab replaces the original BYOK view with a card-based layout and supports managing multiple credential sets of the same type; a new "Fallback" tab lets you create fallback groups and arrange their priority order.
* **MaiGPT Embed Mode Self-Service Configuration**: A third "MaiGPT" option has been added to the embed settings dialog on the admin console conversation platform. Customers can configure the target container element, locale, title, and other options themselves, preview in real time, and copy the embed code, without needing to request documentation from the official team. An "Embed Preview" button is also provided to view the actual result in a new tab.
* **"Go Create" Resource Guidance and Draft Saving When Creating AI Assistants**: In the AI assistant creation flow, a persistent "Go Create" button has been added to the top-right corner of the Tools, Knowledge Base, and Agent UI resource selectors. Clicking it automatically saves the current form content and selected items, then navigates to the corresponding creation page; after creation completes, the system automatically returns to the assistant creation flow with all settings intact, and the newly created resource is automatically selected.
* **Scheduled Crawl Execution History**: The scheduled crawling feature now includes an execution history page where administrators can view the execution status and history of each scheduled crawl.
* **Evaluation Task Cancellation**: The evaluation feature now includes a cancellation mechanism. Administrators can stop an in-progress evaluation task at any time without waiting for all test cases to finish.
* **Map and Route Tool Display in Conversations**: The conversation interface now shows dedicated icons and status hints for the place search and transit route lookup tools ("Searching for places" / "Looking up transit routes"). This release also fixes the issue where the Canvas panel could not be reopened after closing, and adds a clickable reference block so users can revisit it at any time.
* **Organization Data Retention Countdown Alert**: Organizations whose plan has expired now see a red alert banner on the organization settings page, clearly showing the scheduled data deletion date and remaining days, reminding administrators to renew or upgrade their plan in time.
### 🚀 Performance Improvements
* **Automatic Retry for Claude Model Streaming Interruptions**: The automatic retry mechanism for interrupted Claude model streaming responses is now enabled by default. When a stalled stream is detected, the system automatically switches to non-streaming mode and retries, so conversations no longer fail due to streaming interruptions.
### 😊 User Experience and Interface Improvements
* **Support Entry Copy Improvement**: The support entry button in the top-right corner of the admin console header has been renamed from the brand name "MaiDesk" to the more intuitive "Support Center", reducing the effort needed to understand it. The brand name and purpose description remain in the hover tooltip ("Go to MaiDesk Support Center to report issues").
* **Agent Workflow Iteration and Timeout Relationship Hint**: In the AI assistant advanced settings, a relationship explanation has been added below the maximum iterations and timeout fields, explaining the "whichever comes first" mechanism and its impact on Credit consumption. When the configured combination is unreasonable (for example, less than 15 seconds of budget per iteration), a yellow warning and the suggested minimum timeout are shown immediately.
* **Real-Time Updates for Newly Added Skills and Tools in the MaiGPT "+" Menu**: After adding a skill or tool on the settings page, returning to the "+" menu in a MaiGPT conversation no longer requires refreshing the entire page. The menu now updates in real time, automatically loading the latest list in the background when opened.
### 🛠️ Bug Fixes
* **Fixed pinned conversations reappearing after new messages**: After sending a message to a pinned conversation in MaiGPT, the conversation no longer appears in both the pinned section and the general conversation list. After the fix, updates to pinned conversations happen only within the pinned section.
* **Fixed Canvas panel not auto-closing in Web Chat and redundant pre-chat form calls**: In Web Chat, the right-side panel opened by an AI reply containing Canvas content now closes automatically when the user sends a new message, no longer leaving a blank area behind. This release also fixes unnecessary API calls made when the pre-chat form is closed.
* **Fixed missing processing animation during tool execution in conversations**: When the AI calls a tool that takes a long time to run (such as PDF parsing or web crawling), the screen now correctly shows a processing loading animation so users clearly know the system is still working.
* **Fixed "+" menu submenus closing unexpectedly on click**: In the MaiGPT conversation "+" menu, clicking again on items such as "Tools" or "Skills" after hovering has opened a submenu no longer closes it. Closing is now unified to clicking outside, pressing Esc, or clicking the "+" button again.
* **Fixed knowledge base and web search tool results not displaying in MaiGPT conversations**: When using the knowledge base search or web search tools in MaiGPT conversations, the results returned by the tools now render correctly as structured lists instead of unreadable raw text.
* **Fixed Copilot incorrectly showing failure suggestions when a hook skips a message**: When a conversation platform has an on\_input hook configured and it determines a message should be skipped, Copilot no longer incorrectly marks the skip as "generation failed". Skipped suggestions are now removed silently, and no misleading failure suggestion bubbles appear.
* **Fixed FlexMessage tool responses not returned correctly when calling AI assistants via API**: When calling an AI assistant with the FlexMessage tool via the Completion API, if the AI's final call was the FlexMessage tool with no closing text, the system no longer incorrectly returns "No response received from the LLM"; the FlexMessage content is returned correctly.
## v2026.06.26
***
### ✨ New Features
* **Cross-Channel Generated Content Download Links**: AI assistants in Agent mode can now provide download links for code execution results and Canvas content across all channels, including LINE, WhatsApp, and Teams. Once a Canvas is created, third-party channels automatically receive the download link with no extra steps required.
* **Conversation History Search**: The conversation list now includes keyword search, letting you quickly find relevant conversations by entering keywords. Search results highlight matching keywords and show a preview snippet of the conversation content. Pressing Enter searches immediately, and results filter in real time as you type.
* **Unified AI Gateway Management**: Organization administrators can manage AI model providers centrally on the AI Gateway page. It supports adding custom models, managing multiple provider credential sets, and configuring model fallback groups. AI assistants can enable the fallback mechanism to automatically switch to a fallback model when the primary model is unavailable.
* **Public Transit Route Lookup**: AI assistants now support real-time public transit route lookup, providing real bus and metro route information via the Google Routes API. Results are presented in a map Canvas with multiple alternative routes, switchable via tabs, including intermediate stops and transfer information.
* **90-Day Data Retention Mechanism for Organizations Without an Active Plan**: Organizations without an active plan now enter a 90-day data retention countdown. The system sends Email and in-app notification reminders 30, 7, and 1 day(s) before expiration. During the countdown, the organization can still log in to view and export data, and resuming a paid plan resets the countdown.
* **Taiwan AI RAP Breeze ASR Speech Recognition**: The meeting notes speech-to-text feature now includes the Taiwan AI RAP provider, offering two model options: Breeze-ASR-26 and whisper-Breeze-ASR-25. Administrators can enable it in the admin console for more localized speech recognition choices.
### 🚀 Performance Improvements
* **Significantly Faster Credit Statistics Backfill**: The credit statistics backfill job has been switched from row-by-row processing to database-level aggregate queries, reducing processing time from hours (an estimated 5 hours for 1.3 million records in production) to seconds.
### 😊 User Experience and Interface Improvements
* **AI Assistants Recognize Original Filenames of Uploaded Images**: When chatting with an AI assistant with uploaded images, the assistant now correctly recognizes and answers with the images' original filenames instead of internal system identifiers. Multiple images each map to their correct filename. Applies to both Agent mode and classic mode.
### 🛠️ Bug Fixes
* **Fixed PDF table numbers being misread as ranges**: After uploading a PDF with certain table formats to the knowledge base, individual numbers in tables were incorrectly merged into range formats (for example, a single amount becoming a range), causing incorrect AI assistant answers. Also fixed the issue where subsequent tables on the same page disappeared when a page contained multiple tables.
* **Fixed ODF document uploads to the knowledge base failing**: When uploading ODF documents such as .odt, .ods, and .odp to the knowledge base, the system incorrectly rejected them as unsupported formats. After the fix, ODF formats are recognized correctly and uploads no longer fail.
* **Fixed certain AI models failing during classification responses**: When using AI models that support thinking mode (such as Claude Opus) for classification or structured responses, a parameter conflict caused the entire response to fail. After the fix, the system automatically handles compatibility, and classification works normally again.
* **Fixed MCP connector timeouts interrupting conversations**: When an MCP connector's external service was unresponsive, the conversation would break and return an error. After the fix, the system automatically skips unavailable tools so the AI assistant can continue answering normally.
* **Fixed AI assistants unable to retry when video generation prompts were too long**: When a video generation prompt exceeded the length limit, the AI assistant did not receive an error message. After the fix, the system reports the actual length and the limit, allowing the AI assistant to automatically shorten the prompt and retry.
* **Fixed a persistent error on the admin console conversation platform settings page**: Opening conversation platform settings or creating a new conversation platform kept producing form initialization errors. This fix eliminates roughly 92% of frontend error events in the admin console.
* **Fixed missing API request logs for conversations containing images**: Conversations containing images failed silently when logging API requests, causing the request logs for that turn to be completely lost. The streaming response path had a similar log-loss issue. After the fix, request logs are fully preserved for all conversations.
* **Fixed the browser tool generating excessive error reports on remote disconnection**: When the AI assistant's browser tool closed the connection after the remote browser service disconnected, it generated a large number of unnecessary error events. After the fix, the system handles disconnections correctly.
* **Fixed free-plan organizations receiving unnecessary low-balance Email reminders**: Administrators of free-plan or plan-less organizations were receiving low credit balance Email reminders. After the fix, these organizations only receive in-app notifications and no longer receive Emails.
* **Fixed AWS Marketplace users unable to log back in after logging out**: Users subscribed via AWS Marketplace could log in automatically the first time, but could not log back in with Email and password after logging out. After the fix, account verification records are created correctly during setup, and subsequent logins work normally.
## v2026.06.24
***
### ✨ New Features
* **AI Presentation Generation**: Users can generate presentation cards directly during conversations. The feature supports presentation panel preview and navigation, and presentations can be triggered quickly from the chat input bar, letting users create presentations without leaving the conversation page.
* **Account Avatar Upload and Management**: Users can upload and manage their own account avatars. The feature supports drag-and-drop upload, image cropping and re-cropping, shows a visual highlight during drag-and-drop, and provides clear upload guidance in the empty state.
* **AWS Marketplace Billing Integration**: Enterprise customers subscribed via AWS Marketplace can complete account setup and billing management directly within the platform. The billing page automatically identifies the subscription source and displays the corresponding billing information and management interface.
* **Pin and Unpin Conversations**: MaiGPT users can pin important or frequently used conversations to the top of the list and unpin them at any time to restore the original order. Ideal for users who need frequent access to specific conversations.
* **Batch Disable Credit Quotas**: Organization administrators can select multiple members at once and disable their credit quotas in bulk. Ideal for management scenarios requiring large-scale adjustment of member permissions.
* **New Qwen3-ASR Speech Recognition Model**: The meeting notes feature adds Alibaba's Qwen3-ASR speech recognition model, offering Flash (fast) and Realtime (real-time) versions, expanding the speech-to-text model options.
### 😊 User Experience and Interface Improvements
* **Organization Overview Page Redesign**: The organization overview page has been fully redesigned with a more intuitive layout presenting organization status, member count, and usage summary, so administrators can grasp the full picture of the organization more quickly.
* **MaiDesk Quick Entry**: A MaiDesk entry button has been added to the header. Clicking it jumps directly to MaiDesk with login information carried over automatically, eliminating the hassle of manually switching tabs.
* **Unified New Chat Interface**: All AI assistant chat interfaces have been unified to the new design, no longer showing different conversation interface versions due to setting differences, ensuring a consistent experience.
### 🛠️ Bug Fixes
* **Fixed multiple issues with map point-of-interest display**: Point-of-interest information on the map is now more accurate, including corrected labeling for unscheduled spots, improved place recognition accuracy (now using the Google Places API), and an improved photo carousel for attractions.
* **Fixed assignment member permission issues on conversation platforms**: When administrators configured assignment on a conversation platform, members who had lost permissions previously disappeared from the list, making it impossible to save. After the fix, administrators can see and deselect members who have lost permissions and complete saving normally.
* **Fixed incorrect initial values when loading the feature settings page**: After opening a conversation platform's feature settings page, form defaults occasionally loaded incorrect values. After the fix, the settings page shows the correct initial values.
* **Fixed the billing button showing incorrectly for users without a subscription**: Users who have not yet subscribed to a plan no longer see an invalid billing management button in the prompt window.
* **Fixed inaccurate error messages in some AI assistant conversations**: When model responses were temporarily interrupted, the error message has been changed from a generic message to more precise categories, making it easier for users to decide whether to retry.
* **Fixed stability issues in the AI assistant "ask the user" feature**: Certain question data formats could previously break the feature; after the fix, the system correctly handles questions in all formats.
* **Fixed incorrect role labels in the Microsoft Teams conversation handoff flow**: When using an AI assistant via a Teams channel, the handoff boundary markers previously used the wrong role type, potentially affecting the conversation flow after handoff. After the fix, handoff boundaries use the correct role.
* **Fixed conversation interruptions when the AI assistant memory feature failed**: When the AI assistant's memory module fails to load, the system now automatically falls back to recent conversation history as context, so conversations are no longer interrupted by memory module failures.
## v2026.06.23
***
### ✨ New Features
* **Filter Contacts by Channel Type**: The contact list now supports filtering by channel type, including LINE, Telegram, WhatsApp, and Web Chat. It also supports Web Chat subtype filtering (web, admin console, API), making it easier for administrators to manage contacts for specific communication channels.
### 😊 User Experience and Interface Improvements
* **Credit Estimation Check for High-Cost Features**: High credit-consumption features such as video generation and deep research now pre-check whether the account balance is sufficient before execution, avoiding mid-run interruptions due to insufficient credits and reducing wasted waiting time for users.
* **AI Assistant Geolocation Tool Enabled**: The AI assistant's geolocation tool (geocoding) is now fully operational in production, allowing users to look up addresses and location information through the AI assistant.
### 🛠️ Bug Fixes
* **Fixed incorrect display of AI assistant thinking effort settings**: Some existing AI assistants' thinking effort settings were incorrectly shown as "Off" in the interface when they should have been "Default". After the fix, AI assistants without a custom thinking effort setting correctly display the "Default" state.
* **Fixed occasional conversation interruptions with Gemini 2.5 models**: AI assistants using Gemini 2.5 models occasionally had their entire conversation flow interrupted when the model returned an unexpected state. After the fix, the system properly handles such transient anomalies, and AI assistant conversations are no longer interrupted by them.
* **Fixed abnormal plan management button display when no plan is active**: When an account had no active plan, the "No active plan" dialog still showed a "Go to plan management" button, which could confuse users. After the fix, the dialog no longer shows that button.
* **Fixed occasional errors on the conversation platform settings page**: Opening the conversation platform settings page occasionally produced errors due to a form component initialization timing conflict, affecting about 557 users across 2,713 occurrences. After the fix, these errors no longer occur when opening the settings page.
* **Fixed attachment upload path validation returning a server error**: Invalid file paths during attachment upload caused the system to return an HTTP 500 server error instead of the correct HTTP 400 validation error. After the fix, the system correctly returns a validation error message.
## v2026.06.16
***
### ✨ New Features
* **Conversation Satisfaction Survey**: Web Chat adds a conversation satisfaction survey that automatically pops up after an AI reply, featuring a 5-point emoji rating and an optional text feedback field. The survey appears after the user has been idle for a specified time and is shown only once per conversation to avoid repeated interruptions. It supports custom survey prompt text and provides interfaces in five languages: Traditional Chinese, Simplified Chinese, English, Japanese, and Korean.
* **Google Account Sign-In Officially Live**: The production login page now officially supports Google account sign-in. Users can log in quickly with their Google account without registering a separate account.
* **Website Crawl Scheduling and Progress Tracking**: The knowledge base website crawling feature adds scheduling settings for periodic automatic crawls. It also adds real-time progress tracking, showing the number of pages crawled versus the total while crawling is in progress, so you no longer have to wait until completion to see results.
### 🚀 Performance Improvements
* **Faster Conversation Attachment Processing**: Attachment parsing, credit deduction, and auto top-up tasks in conversations now use a dedicated queue instead of sharing one with knowledge base uploads. When large volumes of documents are uploaded to the knowledge base, conversation attachment processing speed is no longer affected, avoiding parsing delays caused by queue congestion.
### 😊 User Experience and Interface Improvements
* **Comprehensive Upgrade of the AI Conversation Tool Call Interface**: The tool call display in the admin console conversation interface has been fully redesigned: 1. Added a web search results component showing site icons, titles, and domains grouped by query 2. Added a knowledge base retrieval node component showing retrieval scores and FAQ/file labels 3. Redesigned progress cards with expand/collapse animation and completion status indicators 4. Tool names are now shown as friendly multilingual names, covering more than 40 tools 5. Sections such as thinking process and to-do lists support full/compact/hidden display modes 6. Image generation results and video loading states render correctly
* **Tool Connection SSL Certificate Relaxation Supports Wildcard Domains**: The SSL certificate relaxation domain list for tool connections now supports wildcard formats (such as \*.example.com), covering the root domain and all subdomains at once, so you no longer need to configure each subdomain individually.
* **Unified ID Column Widths in Admin Console Lists**: The ID/UUID column widths across 20 list pages in the admin console have been unified to a consistent value, improving overall visual consistency. Longer UUIDs are truncated with an ellipsis, and clicking to copy still retrieves the full value.
### 🛠️ Bug Fixes
* **Fixed meeting notes recording being blocked by security headers**: The meeting notes recording feature could not access the microphone because the browser security headers were configured too strictly. The microphone permission policy has been adjusted to allow the site's own domain, restoring normal recording.
* **Fixed the live transcription settings page not scrolling when zoomed in**: When the browser was zoomed in (150%/200%) or on smaller screens, the meeting notes live transcription settings form extended beyond the screen but could not scroll, leaving the "Start Recording" button hidden and unclickable. After the fix, the page scrolls normally to all buttons at the bottom.
* **Fixed CJK .doc file conversion stalling when uploaded to the knowledge base**: Uploading .doc files containing Chinese, Japanese, or Korean text (Big5 encoding) to the knowledge base caused document conversion to stall indefinitely. The document conversion engine has been upgraded to a new version that resolves known CJK character handling issues, restoring normal conversion.
## v2026.06.15
***
### ✨ New Features
* **Comprehensive Upgrade of the AI Map Place Details Page**: The place details panel has been substantially upgraded to a responsive wide layout (920px on desktop), adding a photo gallery with draggable thumbnail browsing, "Overview" and "Reviews" dual-tab switching, structured place information cards (address, phone, website, business hours), and advanced skeleton loading animations, providing a richer place information browsing experience.
* **Canvas Content Embedded in the Conversation Message Flow**: Temporary visual aids generated by the AI assistant, such as Mermaid diagrams, small charts, and SVG illustrations, can now be shown directly within conversation messages without opening the right-side panel. The AI assistant automatically determines the best display position for each Canvas: temporary visual aids are embedded in the message flow, while important content still uses the right-side panel. Older conversations keep the original panel behavior and remain fully backward compatible.
* **Real-Time Audio Spectrum Visualization for Voice Calls**: The microphone button in the voice call interface adds a brand-new circular sound wave spectrum animation that fluctuates in real time with the user's speaking volume, providing more intuitive audio capture feedback. When muted, the spectrum disappears entirely, clearly indicating the microphone is off. Both action buttons (microphone/end call) are now a unified 64px, for a more balanced look.
* **Cookbook Quick Entry in the Admin Console**: The admin console top navigation bar adds a Cookbook button next to the Ask AI button, providing one-click access to the Cookbook user guide website and lowering the learning curve. The button text supports five languages.
* **Input Hook Triggered on Message Edit**: The AI assistant's custom input hooks now also trigger when a user edits an already-sent message, ensuring preprocessing logic is not skipped due to message editing and keeping hook behavior consistent.
* **Tool SSL Certificate Relaxation Allowlist Management**: A centrally managed SSL certificate relaxation domain allowlist has been added, letting you manage domains that need relaxed strict SSL validation in the admin console (for third-party services using legacy root certificates, such as TWCA). Configure once and all tools share it, with no need to enable the relaxation option tool by tool.
* **Advanced Web Crawling Settings**: The knowledge base web crawling feature adds advanced settings fields, providing finer-grained crawl control.
### 😊 User Experience and Interface Improvements
* **AI Map Filter Panel Redesign**: The AI map filtering feature has been fully upgraded with a custom dropdown design, adding dark active-state indicators, custom checkboxes, and expand/collapse animations. It supports keyboard operation (Escape to close) and automatically closes the panel after a category is selected, for a smoother experience.
* **"Other Sign-In Methods" Divider Style Improvement on the Login Page**: The "Other sign-in methods" divider on the admin console login page has been visually upgraded to a refined design with a 1px thin line and soft gray text, plus new dark mode support, for a more balanced overall look.
### 🔐 Security and Governance Enhancements
* **Comprehensive Security Header Hardening (CASA Tier 2 Compliance)**: Added multiple security header protections, including Content Security Policy, Permissions-Policy, and Cross-Origin-Embedder-Policy, and extended the HSTS validity period from 60 seconds to one year, significantly raising the system's security posture. Microphone permissions for meeting recording are confirmed unaffected.
* **Admin Console Mermaid Chart Security Update**: Upgraded the admin console Mermaid chart component to version 11.15.0, patching the Gantt chart infinite loop denial-of-service vulnerability (CVE-2026-41150) and the CSS injection vulnerability (CVE-2026-41159).
* **Document Processing XML Parsing Security Hardening**: Hardened XML parsing security in document processing (DOCX/PPTX uploads). All XML parsing now uses secure parsers, blocking XML external entity injection (XXE) and entity expansion attacks to protect the processing of user-uploaded documents.
* **Code Interpreter Container Security Hardening**: The code interpreter container has been downgraded from root to a non-privileged user, further strengthening defense in depth. False positives for development environment and database containers have been annotated as exclusions.
### 🛠️ Bug Fixes
* **Fixed MCP tool OAuth authorization failing for non-authorizing users**: For MCP tools authorized by an administrator via OAuth (such as Google Drive and other third-party connectors), guests or other users accessing them through LINE, Web Chat, and other channels could not obtain authorization, leaving the tools unusable. After the fix, users on all channels can correctly use authorized tools.
* **Fixed MaiGPT greeting showing the wrong name**: In MaiGPT mode without a welcome message, the greeting incorrectly displayed the AI assistant's name instead of the user's name. After the fix, logged-in users see the personalized greeting "Hi {name}, welcome back!", while anonymous users see "Hi, welcome back". Supported in 13 languages.
* **Fixed the admin console failing to load**: The admin console at qadmin.maiagent.ai failed to load due to a missing environment variable. After the fix, the page works normally again.
* **Fixed code interpreter connection errors during live calls**: When using the code interpreter tool in real-time environments such as voice calls, connections occasionally failed due to internal call conflicts. After the fix, the code interpreter reliably establishes connections in all usage environments.
## v2026.06.14
***
### ✨ New Features
* **Custom MaiGPT Sidebar Brand Title**: Enterprise customers can customize the brand name in the top-left corner of the MaiGPT sidebar via SDK settings, for example changing "MaiGPT" to their own brand name. When not set, the default value "MaiGPT" is shown automatically.
* **Ask AI Changed to a Right-Side Push Sidebar**: The Ask AI chat window in the admin console has changed from a centered floating window to a right-side push sidebar; when opened, page content automatically shifts left so nothing is obscured. Width is adjustable by dragging (320–800px), and the mobile version automatically displays full width.
* **Inline Display Mode for Canvas Content**: The AI assistant's Canvas feature adds an inline display mode, allowing short charts, flowcharts, and other visualizations to be shown directly in the conversation flow without switching to the right-side panel. Full documents and applications still open in the panel.
* **Desktop Conversation Read-Only State and Platform Badge**: A Desktop platform badge has been added to the Web conversation list for easy identification of conversation sources. Conversations created in Desktop automatically switch to read-only mode on Web, with an "Open in Desktop" button for quickly jumping to the Desktop application.
* **Billing Item Deduction Scenario Descriptions**: An information icon has been added next to each item name on the billing methods page; hovering shows which feature operations deduct credits for that item. Supported in five languages: Traditional Chinese, Simplified Chinese, English, Japanese, and Korean.
### 🚀 Performance Improvements
* **Faster Developer API Log Page Loading**: A dedicated query index has been added to the API log page, noticeably improving page load speed when the log volume is large.
### 😊 User Experience and Interface Improvements
* **Improved Alignment of the Create Knowledge Base Dialog**: The "Create Knowledge Base" dialog has changed from centered to top-aligned display, reducing unnecessary blank space and making form filling smoother.
* **Unified Connector Authorization Scope Dialog Style**: The connector authorization scope selection dialog has been upgraded to the system standard style, matching other dialogs in the title bar, button row, and full-screen behavior.
* **Improved Hook Settings Interface Stability**: Hook automation settings now have complete type definitions, making operations in the settings interface more stable and reliable.
### 🔐 Security and Governance Enhancements
* **Admin Console Security Response Header Hardening**: The admin console adds multiple security response headers, including HSTS (enforced encrypted connections), CSP (content security policy), X-Content-Type-Options (prevents type sniffing), and X-Frame-Options (prevents embedding attacks), meeting the Google CASA Tier 2 security assessment standard. Overly permissive cross-origin access settings have also been removed.
### 🛠️ Bug Fixes
* **Fixed incorrect field type mapping for Pre-chat form contact fields**: When configuring a Pre-chat form, phone or Email type fields could be mapped to the contact name, causing incorrect data such as phone numbers to be stored in the name field. After the fix, the contact name field can only map to text-type fields, and incompatible types show a disabled hint.
* **Fixed missing end-date data in the AgentOps dashboard**: When the date range filter on the AgentOps dashboard included today, today's data was excluded from statistics. After the fix, all data on the selected end date is correctly included.
* **Fixed incorrect image format detection in AI assistant conversations**: When uploading images to an AI assistant, if the file extension did not match the actual format (for example, a JPEG image uploaded with a .png extension), the AI reply produced an error. After the fix, the system automatically detects the actual image format and handles it correctly even with a wrong extension.
## v2026.06.11
***
### ✨ New Features
* **Contact Excel Batch Import**: The contacts module adds batch import. Just three steps: download the Excel template, select the target conversation platform, and upload the completed .xlsx file to import a large number of contacts at once. After import, the success count and skipped details are shown immediately.
* **Real-Time Progress Tracking for Test Case Generation from the Knowledge Base**: When generating test cases from knowledge base files, you can now view generation progress in real time (progress bar and completed question count). A new "Generation History" feature lets you view the history of each generation run, including each file's processing status and failure reasons.
* **New Bronci Speech Recognition Service**: The speech-to-text feature adds the Bronci ASR provider option, supporting an asynchronous transcription flow and speaker labeling. The knowledge base speech recognition provider list is now dynamically provided by the backend, so adding providers no longer requires frontend updates.
* **Support for Claude Models via Vertex AI**: AI assistants can now call Claude family models (including Claude Fable 5) through Google Cloud Vertex AI, expanding model deployment options for enterprise users.
### 😊 User Experience and Interface Improvements
* **Aggregated Tool Usage Counts in Quality Control**: In the credit details on the quality control detail page, tools called multiple times are now shown with an aggregated usage count (such as "×3") instead of repeated entries of the same name, making the details clearer and easier to read.
### 🛠️ Bug Fixes
* **Fixed Web Chat survey options occasionally failing to submit**: After clicking an AskUserQuestion survey option in Web Chat, the answer occasionally got stuck in the input box and could not be sent. Now, while the AI reply is being processed, the submit button is temporarily disabled with a hint; once processing completes, normal submission resumes automatically.
* **Fixed AI assistant tool selections lost across pages**: When editing AI assistant settings, if the assigned tools exceeded one page (more than 20), tools sorted onto unloaded pages were silently removed on save. After the fix, all selected tools are fully preserved on save, regardless of pagination.
* **Fixed .m4a audio uploads to the knowledge base being rejected**: Uploading .m4a audio files to the knowledge base was incorrectly rejected. The cause was that .m4a shares the same container format as .mp4, and the validator misjudged valid .m4a files as an incompatible format. After the fix, .m4a audio files upload normally.
* **Fixed layout shifts caused by the system announcement banner**: When the system announcement banner appeared or was closed, full-height pages in the admin console showed extra scrollbars and the entire UI shifted upward. After the fix, the page automatically adjusts its height when the banner appears or closes, with no more layout misalignment.
* **Fixed excess blank space in the Web Chat welcome message**: The Web Chat welcome message showed large unnecessary blank spacing, making the card about 36% taller than normal. After the fix, paragraph spacing returns to standard layout, consistent with the admin console preview.
* **Fixed the deep research feature hanging with no response**: When using the deep research feature, queries could hang without responding under specific concurrency conditions. After the fix, the deep research flow works normally again.
* **Fixed occasional Skill resource loading errors**: The Skill feature occasionally failed to load under high concurrency. The cause was a database query running on the wrong thread and sharing a closed connection. After the fix, Skill resource loading is stable.
* **Fixed large images interrupting AI conversations**: When the code tool converted large PDFs (such as A0/A1 engineering drawings) to PNG and sent them to the AI model, images exceeding the 8000-pixel limit interrupted the conversation. After the fix, the system automatically scales down oversized images and supports configuring image size limits per AI provider.
* **Fixed a conflict between AI extended thinking and the early-stop mechanism**: AI assistants with extended thinking enabled hit validation errors when the early-stop mechanism was triggered, interrupting the conversation. After the fix, the early-stop mechanism correctly preserves thinking blocks and conversations complete normally.
* **Fixed Excel export of logs failing with special characters**: When exporting AI assistant conversation logs to Excel, content containing invisible control characters (possibly from user-pasted content or AI output) caused the entire export to fail. After the fix, the system automatically strips invisible control characters, and the export flow is no longer interrupted.
* **Fixed the Claude Fable 5 model being unusable**: AI assistants using the Claude Fable 5 model failed to start because the underlying framework used hardcoded rules to detect whether a model supports function calling, and the new-generation model name did not match existing rules, causing a misjudgment. After the fix, Claude 5 family models work normally.
* **Fixed scanned PDF translation failures not reported promptly**: Scanned PDFs (full-page images with an invisible OCR text layer) or Type 3 font PDFs only reported errors after being submitted for translation, so users had to wait for processing to finish to learn the result. The file type is now automatically detected before translation, and files unsuitable for translation are flagged immediately with a clear explanation.
* **Fixed double scrollbars in admin console modals**: Some admin console modals (confirmation dialogs and fixed-height popups) showed double scrollbars or content extending beyond the visible screen area. After the fix, all modals display correctly at all screen sizes with no extra scrollbars.
## v2026.06.10
***
### ✨ New Features
* **Web Chat Voice Calls Auto-Enable the Microphone**: The microphone now turns on automatically once a voice call connects, so users can start talking right away without enabling it manually. A mute toggle button has also been added; when muted, the button turns red and switches icons for intuitive operation.
* **Google Social Account Sign-In**: The admin console login page adds a "Sign in with Google" button, allowing users to authenticate and log in directly with their Google account without registering a separate account. The system architecture reserves room for expansion, with future support planned for other social logins such as Facebook and Yahoo.
### 😊 User Experience and Interface Improvements
* **Smarter Real-Time Connection Error Notifications**: The connection error notification mechanism for conversation platforms and Web Chat has been improved. Auto-recoverable network blips no longer pop up an error on every reconnect; instead, a single notification is delayed until the second attempt, with no repeated interruptions afterward. Non-recoverable connection errors (such as insufficient permissions) notify the user immediately instead of being silently ignored.
* **Aggregated Tool Usage Details in Quality Control**: Tool usage details in AI assistant quality control records are now aggregated by tool type. Previously each call of the same tool type occupied its own row (for example, MCP reads appearing 14 times); they are now merged into one row with a call count, greatly improving readability.
* **Automatic Cleanup of Expired Subscription Plans**: The system now automatically scans and marks expired subscription plans daily, ensuring plan status in management reports matches actual expiration times and preventing expired plans from still showing as active.
### 🛠️ Bug Fixes
* **Fixed white-screen crashes when triggering Web Chat voice features**: Using Web Chat voice features could cause the page to crash to a white screen and become completely unusable. This was a non-deterministic load-order issue with the voice package during bundling — the same deployed version sometimes worked and sometimes crashed. After the fix, voice features load reliably with no more white screens.
* **Fixed voice call welcome message captions not displaying**: During voice calls, the AI assistant's welcome message captions occasionally did not appear on screen, affecting the call experience. After the fix, welcome captions display reliably, and voice conversation transcription is also more dependable.
* **Fixed accidental triggers with Chinese input methods in conversation options**: In free-input boxes for conversation options in conversation platforms and Web Chat, pressing arrow keys while selecting characters with a Chinese input method was treated as option navigation, and pressing Enter to confirm a character was treated as submitting the answer. After the fix, input method operations and option navigation no longer interfere with each other.
* **Fixed validation errors when editing MCP connector tools**: When editing MCP connector tool settings, leaving the custom header field blank produced a "value must be valid JSON data" error on save. After the fix, optional fields left blank save normally.
* **Fixed tools lost when saving AI assistant tool settings**: With more than 20 custom tools, adding a tool on the AI assistant tool settings page and saving silently removed assigned tools sorted onto pages that had not loaded. After the fix, all assigned tools are correctly preserved across pages.
* **Fixed PDF file parsing loading errors**: The knowledge base's PDF file upload and parsing feature was temporarily unavailable due to a system component loading issue. After the fix, PDF upload and parsing work normally again.
## v2026.06.09
***
### ✨ New Features
* **Contact Connector OAuth Authorization Management**: The contact drawer adds a "Connector Authorization" tab with a card-based interface showing each third-party service's OAuth authorization status (active/expired/revoked). Administrators can revoke authorization directly or trigger re-authorization, and expired-but-still-valid authorizations show an auto-refresh hint.
### 🚀 Performance Improvements
* **Faster Member Permission Queries**: User information loading after login is faster, with permission query time reduced from about 37ms to about 13ms. Group lookups now use subqueries to avoid cross-table scans, benefiting all pages that require permission checks.
* **Faster Conversation Platform Detail Loading**: Duplicate database queries were eliminated when loading the Web Chat settings page, improving p95 latency.
### 😊 User Experience and Interface Improvements
* **Metadata Retrieval Status Indicators in Knowledge Base AI Search**: Metadata labels in AI search results now make it easy to tell at a glance which fields have "use during retrieval" enabled. Enabled fields show a 🔍 icon, blue field names, and borders, with a new legend explaining the meaning. Hovering shows whether the field is sent to the AI for analysis.
### 🛠️ Bug Fixes
* **Fixed extra JSON shown alongside FlexMessage cards in Web Chat**: In Web Chat, messages containing FlexMessage cards displayed a block of raw JSON next to the properly rendered card. After the fix, only the rendered card is shown, with no extra text.
* **Fixed 400 errors when editing MCP connector tools**: When editing an MCP connector tool, leaving the custom header field blank triggered a "parameter validation failed" error. After the fix, blank header fields are correctly omitted and invalid values are no longer sent.
* **Fixed .xls files being rejected when uploaded to the knowledge base**: Uploading .xls Excel files to the knowledge base was incorrectly rejected because the file format detection tool misidentified valid .xls files as another format. After the fix, .xls, .doc, .ppt, and similar file types are correctly identified and uploaded.
* **Fixed external database MSSQL connection test failures**: Testing an MSSQL external database connection produced a system error because a required database driver was missing after a deployment environment update. After the fix, connection testing works normally again.
## v2026.06.08
***
### ✨ New Features
* **MaiGPT Window Mode Embed SDK**: For scenarios where the AI assistant is embedded as the primary interface, a ChatGPT-style window mode has been added. It offers a persistent left-column conversation history, Deep Research controls, a tool selector, and research progress cards, ideal for customers making MaiGPT the core UI of their product.
* **Live Browser Tool View in Web Chat**: When the AI assistant uses the browser tool to operate a web page, Web Chat users can now watch the operation in real time, with support for click interaction and keyboard input. It appears as a right-side panel on desktop and a bottom sheet on mobile.
* **Generated File Download Cards in MaiGPT Conversations**: After the AI assistant produces files with the code interpreter, a file download card appears directly below the reply message, showing the filename, size, and a one-click download button. It supports real-time streaming display and retention in message history.
* **vLLM Self-Hosted Model Support in the AI Gateway**: A vLLM option has been added under bring-your-own-key (BYOK) mode. Enterprises can connect their own vLLM servers by entering the API endpoint (required) and API key (optional) to use self-hosted models on the platform.
### 🚀 Performance Improvements
* **Faster Group AI Assistant List Loading**: Fixed missing skill and connector preloading in the group AI assistant list, eliminating the performance bottleneck of per-record queries.
* **Faster Knowledge Base Label Batch Updates**: Fixed repeated label queries per knowledge node during knowledge base label batch updates by preloading label data and injecting it into the cache, greatly reducing database queries.
### 😊 User Experience and Interface Improvements
* **Improved AI Conversation Tool Call Visualization**: In admin console AI conversations, consecutive tool calls are now merged into a collapsible visual group; expanding it shows each tool's name, input parameters, and execution results, making it easier to trace the AI's execution flow.
* **AI Monitoring Detail Panel Redesign**: The AI assistant monitoring record detail panel has been improved, reorganizing information into two sectioned cards: "Basic Info" and "Identifiers". It now displays the LLM name used at the time of the conversation, all IDs offer a copy function, and the AI assistant ID and conversation platform ID open the corresponding settings page in a new tab.
* **MCP Tool List Changed to Toggle Switches**: The MCP tool "allowed tools" list has changed from manually managed editable names to a toggle-switch mode. Tool names are read-only to prevent accidental edits, with new enable-all/disable-all shortcuts and an enabled-count display; the edit dialog opens about 60% faster.
* **Display Mode Support for Code Interpreter Tools**: Code interpreter items now support three display modes: full, compact, and hidden. In compact mode, file-write cards show only the title bar; in hidden mode, all tool-related items are filtered out entirely.
* **Web Chat Output Panel Reopenable from Messages**: After the Web Chat output file panel is closed, an "Output Files" card appears in the message; clicking it reopens the right-side panel, consistent with how the Canvas panel works.
* **Inapplicable Feature Toggles Hidden for Internal Conversation Platforms**: In settings for internal-type conversation platforms, three feature toggles that only apply to Web Chat (share conversation, MaiGPT mode, font size switching) are now automatically hidden, preventing administrators from mistakenly assuming these settings affect internal platforms.
* **Real-Time Notification for Unavailable MCP Connectors**: When an MCP tool or server is temporarily unavailable, a real-time notification appears in the AI conversation so users know the tool's current availability status.
### 🛠️ Bug Fixes
* **Fixed conversation log Excel exports missing five channels**: When exporting conversation logs, conversations from five channels — Teams, Instagram, Email, WhatsApp, and Slack — were silently dropped, producing incomplete Excel files. After the fix, conversations from all channels export correctly, and newly added channels will be included automatically.
* **Fixed extra tool response blocks and inconsistent formatting after skill execution**: After a skill ran in an AI conversation, an unexpandable extra "tool response" block appeared, file-write content leaked as plain text, and tool response code formatting was inconsistent. After the fix, a skill badge is shown, code formatting is unified, and file-write cards collapse correctly.
* **Fixed intermittent API connection errors on embedded Web Chat pages**: Web Chat used via embedding occasionally sent API requests to the wrong host, making it unusable. After the fix, the application only mounts after the configuration file fully loads, making API connections stable and reliable.
* **Fixed missing LLM API request logs in AI monitoring**: In AI assistant mode, none of the LLM API call logs were retained in the monitoring system, and Gemini model early-termination errors were misclassified. After the fix, all API logs are fully retained and error classification is more accurate.
* **Fixed tool execution errors not correctly reported to the frontend**: When an AI assistant's tool execution failed, the frontend could not distinguish success from failure. After the fix, the frontend correctly displays tool execution results via error markers.
* **Fixed some code interpreter files displaying as plain text**: Code files generated by the code interpreter on certain execution paths (such as python-docx scripts) appeared as plain text on the frontend instead of code blocks. After the fix, all paths correctly mark content as code and display syntax highlighting.
* **Fixed the code interpreter producing blank or incomplete files**: During streamed file writing, the end instruction sometimes executed before the content was fully captured, producing blank or truncated files. After the fix, the write-end step waits for the content to be fully captured, ensuring file integrity.
* **Fixed test case generation failures for knowledge bases in offline environments**: In offline deployment environments, "generate test cases" and document parsing failed due to missing preloaded models. After the fix, the required model preloads are included, ensuring these features work in offline environments.
* **Fixed format detection failures for certain DOCX files**: Some DOCX files could not be correctly format-detected when uploaded to the knowledge base. After the fix, the detection logic is improved and correctly identifies more DOCX file variants.
## v2026.06.04
***
### ✨ New Features
* **Automatic Injection of Code Interpreter Usage Guidance**: When Code Interpreter is enabled, the system automatically injects standardized trigger rules and usage guidance into the AI assistant's system prompt, so administrators no longer need to hand-write hundreds of lines of trigger rules. Flip the switch and go, dramatically lowering the setup barrier.
* **Separate Welcome Message for Voice Assistants**: Voice assistants gain a dedicated welcome message field that can be configured separately from the text support welcome message. When left blank, the voice assistant does not open the conversation and waits for the user to speak first. Ideal for scenarios that need distinct interaction styles for voice and text support.
* **Real-Time Tool Progress Display in Conversations**: While the AI assistant executes tools or tasks, progress cards appear in the conversation in real time, showing loading, completed, and error states, so users clearly understand tool execution progress.
### 😊 User Experience and Interface Improvements
* **Admin Console Ask AI Assistant Settings Changed to Runtime Configuration**: The Ask AI assistant's enabled state, Web Chat ID, and MCP tool ID settings in the admin console can now be overridden via runtime configuration, so different deployment environments (staging/production) can switch settings directly without rebuilding.
* **Improved MaiGPT Conversation Stability**: MaiGPT AI assistants now enable the Claude streaming retry mechanism by default, automatically retrying on connection interruptions to reduce conversation dropouts. Administrators can disable this feature with one click in the admin console settings.
### 🛠️ Bug Fixes
* **Fixed long text overflowing message content**: Fixed long URLs, long strings without whitespace, and wide code blocks overflowing the display area in conversation messages and dialogs. Long text now wraps automatically, and code blocks scroll horizontally.
* **Fixed LDAP sync not taking over existing accounts**: In organizations with LDAP Bind login enabled, accounts that already existed in the system before syncing are now correctly taken over as LDAP-login accounts. After the fix, these accounts can log in with their LDAP passwords, and departed employees' accounts are automatically disabled during sync. Local accounts not in the LDAP directory are unaffected.
## v2026.06.03
***
### ✨ New Features
* **Targeted Announcement Audiences**: Organization administrators can now send announcements precisely to specified groups, specified members, or custom audiences, rather than only to everyone. Four sending modes are supported: all members, specified groups, specified members, and custom audiences (groups plus members); the management list also displays each announcement's audience labels.
### 🚀 Performance Improvements
* **Faster First Response in Knowledge Base Search**: The knowledge base search initialization flow has been improved, eliminating unnecessary deep learning framework loading delays on the first query and improving first-query response speed by about 700 milliseconds. Subsequent queries are unaffected.
### 😊 User Experience and Interface Improvements
* **Skill Trigger Condition Settings and Cache Improvements**: AI assistants now select Skills more precisely. Each Skill can be configured with a trigger condition prompt, letting the AI assistant choose the most suitable Skill based on the conversation context. Skill list loading is also faster, and Skills are now grouped into official and organization-built categories.
### 🛠️ Bug Fixes
* **Fixed Copilot analysis summaries disappearing**: When viewing conversations with Copilot, the analysis summary occasionally disappeared. After the fix, the analysis summary displays reliably, with a new editable analysis summary feature supporting viewing and restoring historical versions.
* **Fixed translation display issues in the announcement management table**: In the announcement management list, system-generated announcements (such as "Low credit balance reminder") displayed raw translation keys instead of the correct translated text. After the fix, all announcement types correctly display titles and content in the appropriate language.
* **Fixed document parsing for on-premises offline deployments**: For enterprise customers using on-premises offline deployments, document parsing features (PDF conversion, Chinese OCR recognition, PDF translation) failed because required models could not be downloaded. After the fix, all document processing models are preloaded into the deployment image, and all document parsing features work normally in offline environments.
## v2026.06.02
***
### ✨ New Features
* **Free Plan Usage Controls**: When organizations without an active paid plan use AI features, the system now shows a clear prompt window guiding administrators with billing permissions to the plan management page. A corresponding prompt is also shown on the Web Chat side, replacing the previously vague, generic "insufficient balance" message.
* **AI Assistant Conversation Memory Settings**: The AI assistant settings page adds three conversation memory controls: 1. Vector memory toggle: turn off cross-conversation memory retrieval to reduce unnecessary cost and noise 2. Tool result truncation toggle: automatically truncate oversized tool results, preventing memory space from filling up and causing the AI to forget important information 3. Maximum characters per tool result: customize the truncation threshold, defaulting to 8,000 characters
* **Real-Time Conversation Analysis Summary Display**: A "Conversation Summary" section has been added to the conversation detail panel and the Copilot sidebar, showing the summary from the latest conversation analysis. Once analysis completes, updates are pushed in real time, so support agents can grasp the conversation's key points without refreshing the page.
* **MaiGPT Code Interpreter and Official Skills**: MaiGPT has been comprehensively upgraded with a new code interpreter that can run code and process files directly in conversations. It also ships with 8 official skills (presentation creation, Excel processing, PDF generation, and more) and preloads practical tools such as Tavily search, greatly expanding the AI assistant's practical capabilities.
* **Contact Excel Batch Import**: Supports uploading Excel files (.xlsx) to batch import contact data. The system automatically validates fields, deduplicates email addresses within the organization, and provides row-by-row error reporting. Ideal for teams importing large numbers of contacts from external systems (CRM, spreadsheets).
* **New gpt-realtime-2 Model for Voice Assistants**: Voice assistants add the OpenAI gpt-realtime-2 real-time voice model option, at the same pricing tier as existing models, delivering an updated voice conversation experience.
* **Multimodal Inference Gateway**: AI assistants gain multimodal processing logic at execution time, automatically determining whether the model supports multimodal inputs such as images and audio based on its capabilities, ensuring the corresponding features are enabled only when the model supports them.
### 😊 User Experience and Interface Improvements
* **FlexMessage Button Interaction Support (Internal Q\&A)**: In internal Q\&A test chats, FlexMessage message-type buttons (such as "Cancel" and "Confirm") now correctly send the corresponding text messages, and link-type buttons open URLs normally. This makes it easy for administrators to fully test FlexMessage interaction flows in internal Q\&A.
* **Improved MaiGPT Canvas Close Animation**: Fixed the visual issue where closing the Canvas panel in MaiGPT made the view jump to the left before disappearing; the closing transition is now smoother and more natural.
* **Unified Credit Terminology Translations**: Credit-related terminology in the admin console has been unified across languages (Traditional Chinese, Simplified Chinese, Japanese, Korean) — for example, "purchase top-up" is now "add-on purchase" and "plan gift" is now "plan quota" — keeping terminology consistent with the frontend interface.
* **Improved ASK AI Embed Mode Stability**: Improved ASK AI stability in embed mode, fixing potential crashes in private browsing or sandboxed iframe environments where local storage is inaccessible.
### 🔐 Security and Governance Enhancements
* **Knowledge Base File Upload Format Validation**: When uploading knowledge base files, the system verifies that the file's actual content (magic bytes) matches its claimed extension. For example, an .exe file disguised as a .pdf is intercepted on upload, effectively preventing malicious files from entering the system. Validation covers common format families including OLE2, OOXML, images, and PDF.
### 🛠️ Bug Fixes
* **Fixed FlexMessage dynamic fields becoming uneditable after clearing**: In the FlexMessage editor, once a button's link (URI), text, or label field was cleared, the field disappeared from the dynamic fields panel and could not be recovered. After the fix, cleared fields remain in the panel and can be filled in again at any time.
* **Fixed the conversation platform filter in the contact list**: When filtering the contact list by conversation platform, the filter condition was sent as an object format rather than a plain ID value, so the backend could not recognize it and the filter had no effect. After the fix, filtering works normally.
* **Fixed CJK characters rendering as boxes during PDF conversion**: When using the HTML-to-PDF feature, Chinese, Japanese, and Korean characters could appear as empty boxes. After the fix, the system automatically injects CJK font fallback settings, ensuring multilingual text renders correctly.
* **Fixed errors when generating test cases from multiple selected knowledge base files**: Selecting more than 6 knowledge base files for test case generation produced an error because an internal identifier exceeded its length limit. After the fix, any number of knowledge base files can be processed normally.
* **Fixed attachments being incorrectly indexed in group chats when the AI assistant was not mentioned**: In group conversations on platforms such as LINE, AI assistants configured to "reply only when mentioned" still indexed uploaded attachments into the knowledge base even when not @-mentioned. After the fix, attachments are indexed only when the AI assistant is correctly mentioned.
* **Fixed errors during agent handoff caused by messages without text content**: When a conversation contained messages with no text content, the agent handoff feature produced an error. After the fix, the system correctly handles handoffs for all message formats.
## v2026.06.01
***
### ✨ New Features
* **MaiGPT Sidebar Revamp: Desktop Collapse, Mobile Drawer, and Unified Embed Modes**: The MaiGPT sidebar has been fully revamped. The desktop version supports collapse/expand toggling — collapsed, it takes only a sliver of space; expanded, the full feature column returns. The mobile version becomes a left slide-in drawer with a hamburger button and background overlay, closing automatically after a conversation is selected. An interactive embed settings sandbox page has also been added for live previews of different embed modes, and configuration across embed modes has been unified.
* **New MaiGPT Conversation Interface Components**: The MaiGPT conversation interface adds five advanced presentation components.
* **MaiGPT Skills and Tools Menu Search**: The Skills and Tools submenus in the MaiGPT conversation "+" menu add search. With fewer items, filtering happens locally in real time (matching names and descriptions); with more items, it automatically switches to server-side search, quickly finding the Skill or Tool you need. Results are grouped into Built-in / Custom sections, with a hint message when there are no matches.
* **MaiGPT Conversation History Search**: The MaiGPT sidebar adds conversation history search — enter keywords to quickly filter and find the target conversation. Available in both the desktop version and the mobile drawer; clearing the search restores the full list.
* **Code Interpreter Output Files on the All Conversations Page**: The admin console "All Conversations" page adds a right-side Code Interpreter output panel. When opening a conversation containing code execution results, the output file panel (images, documents, and more) expands automatically on the right, matching the behavior of the single conversation detail page, making it easy for administrators to review AI-generated files directly.
* **Support for the Claude Opus 4.8 Model**: The system adds support for the latest Claude Opus 4.8 model, available via both direct Anthropic connection and AWS Bedrock deployment, offering a 1M context window and advanced reasoning capabilities. Each environment must create a model configuration separately before use.
* **Multilingual Display for Browser Tool Billing Items**: On the credit usage pricing page, the browser tool now displays its correctly localized name (such as "Browser Use Tool" in English) and billing unit (such as "per action") instead of raw codes. Supported in five languages: Traditional Chinese, Simplified Chinese, English, Japanese, and Korean.
### 😊 User Experience and Interface Improvements
* **MaiGPT Conversation Area Layout Adjustments**: In MaiGPT mode, the message column gains a maximum width limit and horizontal centering, the empty state layout displays properly, and the input box width has been fine-tuned for a better reading experience.
### 🛠️ Bug Fixes
* **Fixed the loading indicator lingering in conversations**: With certain AI assistants, the "⋯" loading animation occasionally did not disappear after the AI replied, and the actual reply appeared below as a separate bubble. After the fix, the loading indicator is correctly replaced by the AI reply content, regardless of the backend reply flow.
* **Fixed only the last screenshot being kept when pasting multiple into Ask AI**: When pasting multiple screenshots in a row into the Ask AI input box (for example, macOS screenshots), each new one overwrote the previous, so multiple images could not accumulate. After the fix, every pasted screenshot accumulates and displays correctly without overwriting the others.
* **Fixed card message buttons pressed tight against the content above**: In card-style messages with buttons (Flex Message), the first button sat directly against the bottom edge of the image above with no breathing room. After the fix, button area spacing is symmetric and visually more comfortable.
* **Fixed abnormal text backgrounds in dark mode**: In Web Chat dark mode, all text (including chat messages) showed a faint white background layer with a shadow behind it. After the fix, dark mode styles apply only to the correct input box element, and message text backgrounds are clean again.
* **Fixed insufficient height in the connector authorized members window**: The connector "Authorized Members" window did not grow with the number of members, forcing scrolling within a fixed small height when there were many members. After the fix, the window automatically expands to an appropriate height, scrolling only when it exceeds the screen, making it easy to browse large member lists.
* **Fixed inconsistent font sizes between user messages and AI replies on the conversation page**: On the admin console conversation page, the font in user message bubbles was larger than in AI replies, creating visual inconsistency. After the fix, both use the same font size for a more balanced reading experience.
* **Fixed the browser tool failing to start in some environments**: When using the browser tool, a browser-executable-not-found error occurred in certain deployment environments. After the fix, the browser tool starts correctly in all environments.
---
# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.
## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.
Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter:
```
GET https://docs.maiagent.ai/release-note/untitled/archives/06.md?ask=&goal=
```
`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.
Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.