> 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/en/archives/2026/09.md).

# September 2026

MaiAgent platform September 2026 release notes

## v2026.09.29 <a href="#v2026-09-29" id="v2026-09-29"></a>

***

### ✨ New Features

* **Keycloak tool identity settings**: Administrators can enable tools to use the Keycloak login identity in the conversation platform login settings and specify which hosts may receive credentials.
* **Additional model options**: The model catalog now includes OpenAI GPT-6 Sol and GPT-6 Luna. Administrators can grant access by organization and role; the Azure versions are not yet enabled.
* **Assistant token usage charts**: Assistant usage statistics now include a token usage chart showing usage by model and totals, including records with token consumption but no conversation character count.
* **Reuse presentation images**: The presentation editor now includes an image drawer for browsing images generated for the current presentation and inserting them into other slides without generating them again.
* **Customer service availability and assignment**: 1. Customer service agents can switch between available and paused, and schedule time off; 2. Automatic assignment can reassign requests when the wait times out; 3. In takeover mode, agents can take over requests themselves and configure them with no timeout; 4. Lists and personal menus display time-off status.
* **Conversation handoff settings for each platform**: Administrators can enable or disable conversation handoff for each Web Chat separately, preventing the handoff interface from appearing in every public-facing chat when enabled globally.
* **API tool parameter locations**: API tool settings let you place parameters in the path, query, body, or headers. An immediate notice appears if a path parameter in the URL is undefined or its location does not match.
* **EML email parsing**: Knowledge base parsing now supports EML email bodies and metadata such as subject, sender, and recipients. Attachments within emails are not extracted in this release.
* **Queue reminders and return to AI**: Administrators can configure follow-up reminders while visitors wait for a human agent and allow visitors to leave the queue. When enabled, visitors can return to AI responses using a button or keyword.
* **Independent context usage display**: Context usage and conversation handoff can be configured independently for each Web Chat. MaiGPT displays a usage bar so users can view current usage without enabling handoff.

### 🚀 Core Performance Improvements

* **Improved frontend resource loading**: Chat previews now load on demand, and static resource compression and caching settings have been corrected to reduce unnecessary downloads.
* **Stable background task connections**: Queries and cancellations of background tasks use separate connections to prevent operations such as creating assistants or deleting knowledge bases from remaining stuck.
* **List data transfer**: Standard API responses now support compression, reducing data transfer when opening the admin interface and loading lists. Streaming responses retain their existing behavior.
* **On-demand chat widget loading**: Floating and sidebar chat windows, as well as Ask AI in the admin interface, load chat content when first opened, reducing the load on pages where chat has not yet been used.
* **Loading tool log filters**: Tool execution logs retain the assistant used at the time. Assistant filters list only assistants from the current organization, reducing duplicate queries when loading options.
* **Admin chat resource loading**: Chat previews and large editing features now load when needed. Failed loads can be retried, and unexpected borders around input fields have been fixed.
* **Streaming display resource management**: 1. Limits retained page resources after browsing multiple Web Chat conversations; 2. Updates streaming text at a fixed cadence to reduce repeated parsing; 3. Reduces repeated connection status queries during responses while preserving existing content and interruption behavior.
* **First-response initialization**: Resources needed by the model service are preloaded when a worker process starts, reducing the extra wait for the first assistant request handled by each worker.
* **Loading large role membership lists**: When opening the member list of a large role, each identical permission combination is calculated only once, reducing wait time while keeping member and permission data consistent.
* **LINE group message processing**: Recent LINE group names, avatars, and bot names can be reused, reducing the wait caused by repeatedly querying LINE for each message.
* **Conversation platform menu loading**: Loading the menu of all available conversation platforms now takes less work to assemble each entry, benefiting organizations with many platforms.
* **Agent marketplace loading**: The Agent marketplace and assistant menus in work chat rooms load only the data needed for the list, avoiding processing every assistant’s full configuration when opening the catalog.
* **Role assistant list queries**: Role assistant lists prefetch fallback model data, reducing the extra queries needed to retrieve model information for each entry.
* **On-demand voice feature loading**: Call resources load only when voice interaction is opened or recording starts. Canceling while connecting or switching meetings stops further room creation or microphone activation.
* **Background cleanup load management**: Cleanup of expired organizations now uses separate background tasks and small batches, deferring work when the database is busy to reduce its impact on file parsing, usage logging, and email delivery.

### 😊 User Experience and Interface Improvements

* **Consistent branding in plan notifications**: The subjects and signatures of plan and quota notification emails reflect the deployed brand, keeping notifications for own-brand services consistent with their websites.
* **Consistent branding in share previews**: Own-brand Web Chat share previews can use the corresponding brand image. If no brand name is configured, default text from another brand is no longer displayed.
* **Crawler rule confirmation**: A confirmation prompt appears when advancing or previewing after entering a prompt without generating rules. Warnings and applicability limits are also displayed when the prompt cannot be converted into URL rules.
* **Skill descriptions in shortcut panels**: Skill menus in the presentation and document shortcut panels display descriptions. Users can view full descriptions and search by description to find suitable skills.
* **Collapsed Output downloads**：When Output is collapsed, it lists the three most recent file names and download buttons. Switching conversations immediately clears the previous file list, making it easier to identify the current output.
* **Qwen pricing update**：The output price for Qwen2.5 VL 72B Instruct is now 19.8 Credits per thousand Tokens. Input and cache prices remain unchanged.
* **LINE waiting feedback**：A thinking animation is enabled by default for LINE one-to-one chats, including existing channels, so users see waiting feedback before the answer is sent.
* **Tool progress summary**：1. Tool execution status shows the current task and results in the conversation language; 2. Once the answer begins, multiple tool cards collapse into a single expandable summary; 3. Failed steps are highlighted, and hidden tools are excluded from the count.
* **Scheduled result source identification**：Execution results sent to a Webhook by a schedule include the schedule identifier, allowing the same recipient to distinguish results from different schedules.
* **Long text in tool logs**：Long strings in the input, output, and request content of tool execution details wrap onto multiple lines for easier reading on narrow screens.
* **Model list ordering**：The model catalog groups models by brand and displays newer versions first. Assistant menus and the model access page use the same ordering.
* **Skill usage timestamps**：An organization’s custom skills update their last-used timestamp when an assistant expands them for use. Official shared skills do not show usage timestamps from other organizations.
* **Contact Email editing**：The forms for adding and editing contacts accept an Email address and validate its format before saving.
* **Consistent thinking levels**：1. Assistant settings and the chat input bar show only the thinking levels supported by the current model; 2. Options explain the trade-offs between depth, speed, and consumption; 3. Switching models adjusts the level to an available option, keeping defaults consistent with the settings actually sent.
* **Response stage indicators**：While waiting for an answer, the indicator shows thinking or analyzing results according to the actual stage. When tools have finished but the answer has not appeared, the indicator is shown only once.
* **Model sources and filtering**：1. The model catalog supports filtering by vendor, open weights, or closed source, and displays licensing restrictions; 2. Menus show the actual inference provider rather than only the connection method name; 3. Model information identifies the developer’s country of origin, and provider names do not expose credentials or internal host information.
* **Monitoring Hook execution details**：Monitoring record details show Hook names, results, and timestamps, including turns blocked before a model call.
* **Evaluation model information**：Single-turn and multi-turn evaluation details show the scoring model actually selected for the evaluation, making it easier to confirm evaluation settings.
* **Resource permission select-all scope**：The select-all control in the header of the role resource permission table indicates that it applies only to the current page, preventing users from assuming that permissions retained on other pages have all been revoked.
* **Unified model tool capabilities**：A single check now determines a model’s tool-calling capabilities, preserving any capability enabled by either previous setting and preventing inconsistent checks between the canvas and tool entry points.
* **Reasons for missing document text**：When an assistant cannot read a file’s text, it receives information about the parser used and whether text recognition was performed, helping distinguish text that was not recognized, processing still in progress, and parsing failures.
* **Team node budgets**：Agent Teams node Token and iteration limits take effect according to the configured settings. The settings page explains the relationship between a single team turn, delegated nodes, and the overall team limit.
* **Meeting speaker identification update**：Offline meeting transcription and audio imports using pyannoteAI now use Precision-3, retaining speaker labels in transcripts.
* **Crawler failure log entry point**：The action in completed crawler rows is renamed to failed import records, clarifying that this entry point shows failed items rather than all crawled content.
* **Database connection string updates**：After changing an external database connection string, users can save or cancel directly in the same row. The button that could only copy masked content has been removed to prevent use of a string that cannot establish a connection.
* **Custom models for conversation analysis**：Human handoff decisions, conversation analysis, and analysis reports can use the organization’s own custom AI Gateway models. Pricing information lists these traffic charge scenarios, and models from other organizations cannot be selected.
* **Consistent branding in login flows**：SSO redirects, registration, and password reset pages use the corresponding brand Logo. Images are hidden if they fail to load, and template brand links have also been removed from the admin footer.
* **Text recognition in document images**：Docling parsers configured with a multimodal model process images according to the parser prompt. By default, they primarily transcribe text in images and provide a brief description only for images without text.
* **Model retirement and replacements**：The platform can record model retirement dates and replacement models, migrate assistants when the retirement date arrives, and disable old models to reduce the risk of continued use of retired services.
* **System messages in the conversation language**：Chat system errors appear in a single language matching the conversation locale. When a response exceeds the length limit for a single message, a plain-language explanation avoids requiring users to understand internal parameters.
* **Select role members across pages**：When assigning role members, select all can select members across all pages that match the current search criteria at once, reducing the need to select each page separately.
* **Video completion messages**：The video tool prompt tells the system to return completion results automatically, without asking users to confirm the download button or incorrectly claiming that the next video has been submitted.
* **Document parsing cost information**：Uploading a file or reselecting automatic parser detection now displays a notice that the token-billed Vision Parser may be used. Usage records explain that the parsing model may differ from the assistant model.
* **Role member management permission information**：The role permissions page explains that changes to members and roles require subpermissions under Organization Overview, including when using API keys, helping administrators confirm the scope of authorization.
* **Long role name display**：Long role names wrap and display in full in role permissions and member lists, making it easier to distinguish departments or service units whose names differ at the end.
* **Satisfaction statistics criteria**：User satisfaction statistics exclude likes and dislikes on opening messages and recalculate affected historical statistics so that the figures reflect feedback on actual answers.

### 🔐 Security and Governance Enhancements

* **Web Chat conversation isolation when switching accounts**：Switching login accounts in the same browser creates a conversation belonging to the new account, preventing restoration of the previous user’s records.
* **Team knowledge base tag scope**：File browsing and reading in Agent Teams nodes apply the configured knowledge base tag scope, preventing access to documents outside that scope.
* **Agent Marketplace menu permissions**：Members with only Agent Marketplace permissions no longer see an empty customer service conversation entry. The conversation platform enable switch is controlled by platform editing permissions.
* **AI Gateway credential isolation**：When connecting to a custom model endpoint with user-provided credentials, platform model keys are never sent to the customer endpoint, even if no key is entered in those credentials.
* **Protection of original team input**：When a node’s permission to access original user input is disabled, the node receives neither current nor historical original user text and receives only permitted upstream output.
* **Connector authorization revocation permissions**：Only members with connector management permissions can revoke a contact’s connector authorization. The interface and backend validation use consistent rules.
* **Old links after platform deletion**：Deleting a conversation platform directly also disables the old URLs for its conversations, preventing continued access to records from the removed platform through deep links.
* **MCP query credential protection**：When the administration interface retrieves available MCP tools again, authentication headers are placed in the request body, keeping keys out of URLs and access logs.
* **Role assistant assignment synchronization**：Adding or removing assistant assignments for a role updates members’ chat entry points accordingly. Assignments for rented official Agents provide only chat access.
* **Response text colors and content safety**：MaiGPT can display valid text colors in answers. Text markup containing unsafe content is downgraded to plain text to prevent execution in chat.
* **Masking sensitive tool headers**：Tool logs mask sensitive request headers according to the current settings, including in existing records. Removing custom names restores their display without changing how the data is stored.
* **Presentation knowledge bases and creator verification**：The presentation tool can find readable knowledge bases selected in the shortcut panel by name and defaults to the sources selected in that panel. The creator and query permissions are determined from the verified sender.
* **Preserving permissions when copying roles**：Copying a role preserves read, edit, and delete permissions for each resource, without broadening read-only permissions or adding members of the original role.

### 🛠️ Bug Fixes

* **File creation with tool search**：When both tool search and Code Interpreter are enabled, the assistant can still create documents directly, avoiding failures to find the enabled file creation capability.
* **Guidance when no models are available**：When a role has no available models, MaiGPT displays a no-models-available state and restricts submission, preventing errors caused by selecting an unauthorized model.
* **Permission notices when editing template assistants**：Regular members editing assistants created from the template library no longer receive erroneous permission errors from connector icon queries. Names of linked connectors remain visible.
* **Protection against repetitive output**：When an open-source model gets stuck producing repeated output, the system blocks the abnormal content and retries once. Blocked degraded output is not billed.
* **Stopping notifications after channel deletion**：Deleting a Viber or Telegram conversation platform removes message notifications that still point to it, preventing continued delivery to the deleted platform.
* **Embedded chat initialization**：The chat window embedded in a website notifies the host page after it has actually finished loading, preventing origin mismatch warnings during initial loading.
* **Protecting knowledge base search data**：Updating tags, moving files, or adjusting speaker names preserves semantic search data, preventing existing documents from becoming unavailable through semantic retrieval.
* **PDF text order correction**：When Vision Parser detects that the text order in a PDF differs from its layout, it recognizes the page image instead, arranging paragraphs and tables according to the original layout.
* **Branching Agent Marketplace conversations**：Members with Agent Marketplace permissions can create branches from their own conversations. Output no longer opens an additional misleading error dialog when the execution environment has been cleared.
* **Browser language detection**：The administration interface applies the current locale before its first display, preventing browsers from misidentifying Traditional Chinese or other language pages when switching organizations or refreshing.
* **FAQ template answers**：FAQ questions and answers populate correctly even when response template field names differ only in letter case, preventing matched content from appearing blank.
* **Azure file parsing**：Deployments using Azure storage can open knowledge base files for parsing correctly, preventing file path handling from causing parsing failures across multiple formats.
* **JSON data import**：JSON rows wrapped under a single outer name can be converted into a table. Text fields containing monetary terms are no longer mistaken for numbers and no longer cause imports to fail.
* **Settings and schedule execution fixes**: 1. Skill files can be read from the designated workspace path; 2. Recurring schedules trigger correctly after being changed to one-time schedules and retain the configured time zone; 3. Tool card templates are checked for structural validity when their content changes; 4. Model filtering on the monitoring dashboard and the display of imported field types have been fixed.
* **Form and monitoring operation fixes**: 1. Closing the key creation dialog clears the displayed key; 2. Resetting monitoring filters restores the last 90 days, and zero usage is shown as no data; 3. Evaluation metrics accept Chinese names; 4. Waiting behavior for private MCP tests and card template validation have been fixed; 5. Disabling Hook flags and using the reauthorization entry point now work correctly.
* **Sending more messages outside reply hours**: After a message is sent outside automatic reply hours, Web Chat no longer waits indefinitely for an AI reply that will never arrive, and subsequent messages can still be sent normally.
* **Meeting recording backup on private networks**: For private network deployments with an internal upload proxy configured, meeting recording backups also use the internal upload address, avoiding attempts to connect to inaccessible public storage.
* **Embedded chat mode restored**: When the chat interface is embedded in a designated section of a web page, the chat window loads correctly; the mobile version also hides the inapplicable close button.
* **Anonymous conversation ownership after login**: When users log in and continue an existing anonymous conversation, sending the next message assigns the conversation to their own contact, preserves its history, and displays the correct name.
* **Mobile link tooltips**: On touch devices, URL tooltips no longer remain visible after users tap a link in a reply and return; mouse tooltips remain available on desktop.
* **Retrying knowledge base file deletion**: Files stuck in deletion or whose deletion failed can be deleted again; files in the same batch are handled individually, and a notice is shown for files skipped because deletion is already in progress.
* **Quick reply button compatibility**: Web Chat skips quick reply actions that are exclusive to LINE and correctly displays emoji and abbreviation icons for supported buttons.
* **Conversation model and retrieval settings**: 1. Conversations without a manually selected model follow the assistant's current model, and disabled explicitly selected models fall back to an available model; 2. Reranking is skipped when it is disabled for all knowledge bases in use; 3. Embedded login credential validation has been fixed; 4. The AI Gateway model catalog includes custom models and fallback groups; 5. Post-reply Hooks can run detection.
* **Asynchronous jobs and MCP settings**: 1. Deleting files and databases no longer locks data for extended periods while dispatching background jobs; 2. Crawlers trigger correctly after being changed to one-time schedules; 3. Deleted knowledge base documents no longer appear in queries; 4. Private MCP tests have an overall wait limit, and uploaded logos are saved correctly.
* **Individual conversations with team members**: When chatting individually with a team member assistant, the system explicitly states that this mode cannot hand off to other team members, reducing cases where delegation is claimed only in text.
* **File locations and connector icons**: When a connector logo is unavailable, a default icon is used to prevent broken images in assistant and connector settings.
* **File cleanup when retrying parsing**: Reparsing, retrying, or deleting knowledge base files also cleans up old images and derived files, preventing repeated operations from leaving duplicate content.
* **Duplicate team node notices**: Adding the same assistant more than once to an Agent Teams team returns an understandable validation error instead of a system error.
* **Contact import previews**: Previewing a batch contact import does not directly modify data, and canceling leaves no imported results; selecting the option to merge conversation platforms preserves existing settings.
* **Content policy block notices**: When a Hook blocks a message and no reply text is configured, a default content policy notice is displayed to prevent blank messages in the conversation.
* **Schedule form error messages**: When a new schedule is missing required data on another tab, the form switches to that tab and displays an error; reaching the limit on enabled schedules also displays a readable notice.
* **Email validation when importing members**: Email formats are checked when batch importing members and role members to prevent the creation of invalid accounts that cannot log in.
* **Filtering evaluations by assistant**: The evaluation list sends filter criteria for the selected assistant, preventing evaluations for other assistants from appearing after an assistant has been selected.
* **Rejecting invalid batch import files**: The member, role, and member data import dialogs exclude files with invalid formats or excessive sizes, instead of displaying an error while still allowing the upload.
* **Conversation sorting after unpinning**: After a conversation is unpinned, it returns to the correct position in the recent conversation list based on its last message time, rather than jumping to the top.
* **Tool logs retained in the admin interface**: When Web Chat is configured to hide tool steps, customer service conversations in the admin interface still show complete tool logs, making it easier to trace how a reply was produced.
* **Stable display when answers finish**: MaiGPT retains the content blocks already displayed when reply streaming ends, preventing the entire answer from flickering or the reading position from shifting.
* **Preserving the new conversation display**: After MaiGPT creates a conversation from the first reply, it preserves the chat content and canvas currently displayed, avoiding a return to the skeleton loading view or closing the canvas.
* **Telegram reply waiting indicators**: Telegram messages that will not receive an AI reply no longer show a typing indicator, preventing users from assuming a reply is being generated for group messages that do not mention the assistant.
* **Assistant iteration limit**: Assistants stop at the configured maximum number of iterations instead of being forced to execute at least 25 steps; when the limit is reached, they send closing content and an explanation of the limit.
* **MCP tools with long names**: When the combined length of an MCP connector name and tool name is too long, the connector prefix is shortened while preserving the tool identifier, avoiding the model's name length limit.
* **Evaluation model compatibility**: When automated evaluations use a model service that does not support returning multiple results at once, requests are sent separately to prevent evaluation failures caused by unsupported parameters.
* **Attachment download filenames and formats**: Downloaded chat attachments retain their original upload filenames; exported Excel, Word, and PowerPoint files also include the correct format information.
* **On-premises download addresses**: For MinIO deployments with a public download endpoint configured, file links provided to browsers and external services use the public address to avoid connections to internal hosts.
* **Assignment count field validation**:The concurrent service and queue limits for customer service cannot be saved as 0 or negative numbers, preventing invalid settings from blocking all requests for transfer to a human agent.
* **Test case source filenames**:Test cases generated from the knowledge base display the original uploaded filenames instead of temporary paths or unrecognizable codes.
* **Official Agent knowledge base cleanup**:Cleaning up a rented official Agent's knowledge base removes both documents and FAQ entries; the confirmation message also clearly states that both will be deleted.
* **Browser tool limits**:The browser tool stops at the configured maximum number of actions, and failed actions are marked as failed in the execution log, making unfinished work easier to identify.
* **Skill resource loading**:When an assistant uses a skill with additional tools, it gathers the required resources before packaging them, preventing data query failures from interrupting resource loading.
* **Message order within the same second**:Questions and replies sent within the same second in Web Chat retain their original order after refreshing or resynchronizing.
* **Model request compatibility after handoff**:When using an OpenAI-compatible endpoint after a canvas or team handoff, incompatible tool exchange records are converted to text, preventing the model from rejecting requests that have no tools bound.
* **Knowledge base reranking availability**:Knowledge bases using the system's default vector store display RRF availability based on actual deployment support; automatically created knowledge bases follow the same rules.
* **Saving dynamic card fields**:When editing dynamic card fields containing additional markup, prompt text displays correctly and existing markup is preserved, preventing a simple save from corrupting the configuration.
* **Warnings for images exceeding model limits**:When the images sent in a single request exceed the model service limit, a message asks users to reduce the number of images or have an administrator adjust the limit, instead of displaying only a generic retry-later error.
* **Preserving conversation history when deleting an assistant**:When an assistant is deleted, the system handles its associated conversation platforms. Members who previously had viewing permission can still access the history in customer service conversations, preventing the deletion process from removing conversations.
* **Input bar on narrow screens**:When MaiGPT's sidebar is expanded on a narrow screen, the thinking, model selection, and send controls stay within the input bar without overlapping.
* **Mutually exclusive transcription language options**:After automatic detection is selected for live transcription or audio uploads, a specific language is no longer retained alongside it, preventing combinations that the model does not accept.
* **Knowledge base search test feedback**:When reranking is enabled and more chunks are retrieved, search tests are no longer cut short by the frontend's 20-second timeout; completed searches with no results clearly indicate that no matches were found.
* **Long-term memory for the current question**:Assistants with long-term memory enabled retrieve relevant memories using the current question, preventing the current request from being missed when searching only with previously saved messages.
* **Preserving presentation template information**:After a template is selected from the shortcut panel, the project still records the selected template even if the assistant asks a follow-up question before creating the presentation, preventing the template name from being lost.
* **On-premises canvas loading**:On-premises and white-label deployments include the runtime resources required by the canvas, preventing the React canvas from remaining stuck on the loading screen.
* **Parsing service busy message**:When a scanned PDF temporarily cannot be read because of model service rate limits, users are asked to try asking again later and can reparse the file, instead of being incorrectly told that it has no recognizable text.
* **Card memory in long conversations**:The assistant's conversation memory uses the visible text on cards instead of extensive layout data, reducing the risk of card structures filling the context in long conversations and affecting subsequent text answers.
* **Removing obsolete Google Drive tools**:Legacy Google Drive authorization tools that are no longer supported are disabled and unbound, preventing assistants from continuing to direct users to inactive authorization pages.
* **Input protection for long conversations**:Before sending a model request, the assistant checks the context size, retains the content needed for the current turn, and trims history that can be removed. It provides a clear error if the input exceeds what can be processed.
* **Knowledge base import via media links**:When knowledge base files are uploaded using the platform's stable media links, the storage path is identified correctly, preventing uploads from immediately failing with a system error.
* **Bedrock structured responses**:Structured output and forced tool calls with Bedrock GPT-6 Astra correctly send the required instructions, preventing empty results in classification, evaluation, or analysis.
* **Reading original spreadsheet files**:Assistants without a knowledge base can also pass original conversation attachments to Code Interpreter to compare multiple spreadsheets, instead of receiving only truncated text previews.
* **Downloading outputs in conversations**:Workspace output links in admin panel replies download files directly instead of being treated as internal pages and leading to a 404; a message is shown if the download fails.
* **Pre-chat form field validation**:An enabled pre-chat form must retain at least one field before it can be saved, preventing the admin panel from showing it as enabled while visitors cannot see it.
* **Citation source links**:When a file has a citation redirect link configured, clicking the cited filename opens the source page directly, even if downloads of cited files are disabled.
* **Returning to assistant settings**:After opening website conversation platform settings from assistant settings, breadcrumbs and back navigation return to the original assistant settings location instead of the standalone platform list.
* **MCP responses and model retries**:MCP tool responses preserve the original text provided by the service instead of displaying hard-to-read object contents; requests are retried when Gemini temporarily returns an empty tool call error.
* **Synchronizing customer service takeover status**:After a customer service agent takes over or transfers the conversation back to AI, the auto-reply toggle in the input area synchronizes immediately, so the correct state is visible without refreshing.
* **Presentation tool and canvas routing**:Presentation shortcut conversations use the presentation tool instead of the general canvas; tool search mode only suggests direct calls after the presentation skill has been expanded, preventing required steps from being skipped.
* **PPTX photo aspect ratios**:The built-in PPTX skill guides assistants to preserve the original image aspect ratio or crop to fit the frame, reducing squashed or stretched photos in exported presentations.
* **Waiting for long tool inputs**: While Bedrock generates a long tool input, a temporary absence of answer text no longer causes the response to be mistakenly flagged as stalled, preventing tasks such as skill creation from restarting midway.
* **Continuing work after form responses**: After MaiGPT users answer the assistant’s form about style or page count, the system preserves the canvas tool’s request and response history so the assistant continues creating content instead of merely acknowledging the answers.
* **Failure messages for crawls with no pages**: Crawls that finish without retrieving any pages are marked as failed and display the available failure reason, rather than showing completion with no content.
* **MP3 export for on-premises meetings**: The Fedora deployment image now includes MP3 encoding support, allowing meeting recording conversion and MP3 downloads to produce files correctly.
* **Parsing elongated images in PDFs**: PDF parsing now retains wide, short or tall, narrow images with sufficient area, preventing workflow screenshots from being skipped because one dimension is small.
* **Loading and querying reference documents**: MaiGPT prompts the assistant to query the reference documents selected by the user first. In a new conversation, sending is temporarily restricted until the knowledge base selections load, ensuring the first question includes the selected sources.
* **Feedback on chat connection failures**: When Web Chat cannot establish a real-time connection, it displays an initialization error and a retry option, along with a prompt to check network restrictions, instead of remaining on gray skeleton placeholders.
* **Transcription confirmation messages**: The unconfirmed-result message now appears only for real-time transcription models that report a final confirmation. Other models no longer display an incorrect warning after recording ends.
* **Messages for connector role restrictions**: When a connector is unavailable, members who cannot manage role permissions see guidance to request access and notify an administrator, rather than being directed to a settings page they cannot use.
* **Status for recording startup failures**: If the recording service refuses to start, the meeting is marked as failed and the reason is saved. Periodic checks also finalize meetings whose end notifications were lost, preventing them from remaining in the recording state indefinitely.
* **Avoiding repeated answers after tool errors**: When the model calls an incorrect tool name, the system reminds it that the previous text has already been delivered, reducing repeated full answers when correcting the tool call.
* **Plain JSON responses from AI Gateway**: When JSON format is requested, the system removes code fences added by the model, keeping the response as plain JSON that integration code can parse directly.
* **Recovery from interrupted parsing**: Interrupted knowledge base parsing tasks can automatically retry once, with the failure reason recorded if they still do not complete. Environments with the required deployment permissions also support protection against scaling down during parsing.
* **Continuity of content in team delegation**: Delegated assistants can access the current turn’s input and attachment information within the permitted scope. When tasks are completed through forms, quick replies, or image cards, the upstream assistant receives the actual results instead of an incorrect report of no response.
* **Messages when no plan is active**: When an organization has no active plan, a clear warning appears even if its plan history is empty, preventing the absence of records from obscuring the plan’s status.
* **Waiting for AI presentation edits**: Full-page rewrites now allow sufficient time for model processing, with the frontend wait limit adjusted accordingly. Additional layout adjustments are avoided when insufficient time remains.
* **Messages while waiting after uploads**: If storage does not respond within three minutes after a file finishes sending, the upload is marked as failed and users are prompted to check their network and upload again, instead of remaining stuck in the uploading state. Large files still being transmitted are not subject to this wait limit.

## v2026.09.23 <a href="#v2026-09-23" id="v2026-09-23"></a>

***

### ✨ New Features

* **Real-time voice transcript settings and billing**：Real-time voice assistants can configure the transcription model, language, and prompt for user transcripts, and record speech-to-text usage by the model actually used, making recognition settings and costs easier to manage for multilingual calls.
* **Site-specific branding**：The same platform can display different brand names, logos, and primary colors by site domain. Organizations without a custom primary color inherit the site’s brand color, keeping the login page, admin interface, and Web Chat consistent.
* **AI image editing**：AI assistants can now edit existing images directly. Requests to change colors or add or remove objects with built-in image generation tools edit the original image instead of silently generating a new one.
* **Browsing file library workspaces**：AI assistants can browse file library workspaces, read files, and identify parsing status. The interface also shows correct system folder names and read-only or write permissions, reducing unnecessary knowledge base queries.
* **File library workspace instructions and output filing**：Workspace instructions can be configured in the file library so AI assistants handle files according to team rules. Documents, images, and attachments generated in conversations can also be filed automatically with their source information retained.
* **File version history and AI writes**：The file library now includes version history, and AI assistants can create or update files with permission. Users can view each version in file details to track content changes.
* **Approval for AI file deletion**：When AI assistants organize the file library, deletion operations first submit an approval request with the source conversation. Deletion proceeds only after an administrator confirms, preventing important files from being removed directly.
* **Conversation handoff summaries**：Long conversations can generate handoff summaries and continue in a new conversation while preserving goals, completed work, and next steps. The input bar also shows context usage, making it easier to complete a handoff before reaching the limit.
* **Custom Credit usage ranges and totals**: The Credit usage page now supports custom date ranges and instantly displays income, spending, and transaction counts for filtered results. The end date includes the entire day, making reconciliation and comparisons easier.
* **Path parameters for custom API tools**: Custom API tools now support URL path parameters to integrate with common REST services that place customer IDs in URLs. AI assistants fill in the corresponding path values before sending requests.
* **gpt-realtime-2.1-mini voice model**: Real-time voice assistants now offer gpt-realtime-2.1-mini, supporting existing voice workflows and calculating Credit usage based on input, output, and cache usage.
* **Custom metrics for multi-turn evaluations**: AgentOps multi-turn evaluations now support custom metrics. Configure score names, criteria, and evaluation steps to assess complete conversations against your team’s own service standards.
* **Web Chat activation defaults and permissions**: Organizations can set whether Web Chat is enabled by default for new AI assistants. Visitors see a clear notice when a channel is disabled. Members who can edit the conversation platform can turn Web Chat on or off (the separate role permission originally announced was removed on 2026-09-24).
* **Embed MaiGPT with a member identity**: Partner systems can now embed MaiGPT using an organization member’s identity. The SDK supports automatic credential renewal, timeout protection, and organization isolation, and creates the first conversation only after identity setup is complete.
* **Image sizing in conversations**: Images in AI replies can now specify display dimensions, allowing instructional images, charts, and thumbnails to fit their content while retaining existing security restrictions and responsive layouts.
* **Final live meeting transcription results**: Meeting records now offer OpenAI gpt-live-transcribe for real-time transcription. When recording ends, the final processing results appear and the saved transcript refreshes, preventing the screen from remaining on temporary content.

### 🚀 Core Performance Optimizations

* **Parallel multi-turn evaluations and live progress**: Multiple workers can run multi-turn conversation evaluations in parallel, and the detail page updates each case’s progress in real time. Large datasets no longer need to wait for a single queue to process cases one by one.
* **Batch jobs no longer block real-time messages**: File parsing, dataset evaluations, and batch Q\&A now use dedicated queues, preventing large volumes of background jobs from consuming real-time messaging resources. Batch Q\&A also releases the next set of jobs more frequently.
* **Reliable loading after frontend updates**: Web Chat pages rebuild automatically after platform updates, and versioned assets ensure browsers load matching files, reducing brief blank screens and mismatched feature versions after an update.

### 😊 User Experience and Interface Improvements

* **Unified website crawling workflow**: Website crawling now consistently uses the new workflow. Creating tasks, viewing results, and importing content into the knowledge base no longer differ depending on whether the old or new mode is used.
* **Table message width**: Table replies in Web Chat now fit their content. Short tables no longer fill the entire message area, making reading and horizontal scrolling more natural.
* **Consistent model names and brand icons**: The language model list now displays clear model names and consistent brand icons, including GPT-6 Astra and GPT-OSS 20B, making models easier for administrators to identify and select.
* **Structured long-term AI assistant memory**: Long-term AI assistant memory now displays topics, summaries, and details, with quality checks and Skill usage tracking. Individuals can view their memories in one place in the sidebar, with data limited to the current organization.
* **Names in team execution traces**: Team execution traces now show actual node and AI assistant names, letting administrators see who handled each step and investigate collaboration workflows more easily.
* **Notice when conversation history is truncated**: When earlier messages in a long conversation are omitted due to context limits, a notice appears below the AI reply. Users can see that the model did not read the full history and provide key information again.
* **AI assistant filters on conversation platforms**: Conversation platforms and the MaiAgent role assignment interface now include AI assistant and channel filters, plus an option to hide built-in assistants, helping users find the right items in large lists.

### 🔐 Security and Governance Enhancements

* **Permission checks when binding AI assistants**: When a conversation platform or knowledge base is bound to an AI assistant, the system also checks whether the user can access that assistant, preventing association operations from bypassing existing role permissions.
* **Block malformed AI safety check results**: If an AI assistant’s input or output safety check returns incomplete or unparseable results, the system blocks the request as a precaution, rather than mistaking malformed data for permission to proceed.

### 🛠️ Bug Fixes

* **Fixed inapplicable settings on non-Web Chat channels**: When managing LINE, Messenger, and other channels, visitor forms and satisfaction surveys no longer appear. The settings page now shows only features available to that channel.
* **Fixed credentials overwriting each other between on-premises object storage and AWS services**: Environments using both MinIO and AWS can configure separate keys. File access, meeting recordings, and cloud services no longer fail because credentials are mixed up.
* **Fixed prolonged unresponsiveness in Amazon Bedrock conversations**: When users ask another question after an idle period or a model read times out, the system recycles stale connections and retries automatically, reducing stalled conversations.
* **Fixed frequent reauthorization requests from Google Workspace connectors**: When multiple users or tools run at the same time, authorization refreshes complete sequentially, preventing connectors from being disabled by credential rotation conflicts.
* **Fixed inaccurate knowledge base image searches by filename**: When searching for images using Chinese or full filenames, the system prioritizes matching files and includes source information, reducing unrelated results.
* **Fixed internal action markers appearing in AI replies**: When using long-term memory, messages no longer begin with internal formats such as “Called tool” or “Asked the user”; conversations display only the actual answer.
* **Fixed false health check timeouts in on-premises environments**: The system probes only active language and embedding models, avoiding false service errors caused by preloaded cloud models that are not enabled.
* **Fixed intermittent failures when creating conversation branches**: Even when the URL contains no conversation information, the system creates a branch from the currently open conversation so users can start a new discussion from a selected message.
* **Fixed tildes incorrectly applying strikethrough or subscript formatting**: “\~” in model numbers, paths, and ordinary sentences now displays as written, without changing the style of subsequent content.
* **Fixed missing images in PowerPoint exports**: Platform images in presentations now embed correctly in PPTX files. Image formats can also be identified when file type information is incomplete, preventing “Not Found” errors or empty image frames.
* **Fixed member names disappearing after bulk reimport**: Updating existing members preserves their display names, so member lists and search results no longer show blank names.
* **Fixed loading indicators remaining active after AI replies finish**: The interface now correctly stops waiting even when the final content is a special block such as a card.
* **Fixed citation content being inaccessible when downloads are disabled**: Even when file downloads are disabled for a knowledge base, users can click citation labels to view content within the site while download restrictions remain in place.
* **Fixed reduced image generation quality in MaiGPT**: The image tool now produces results at the expected quality, preventing lower-quality images caused by missing settings.
* **Fixed inability to submit dislike feedback on small screens**: Feedback dialog buttons remain visible, and selected like/dislike reactions are correctly highlighted and preserved after message updates.
* **Fixed AI replies ending prematurely during brief network disruptions**: After resynchronizing, Web Chat continues waiting if the server is still generating content and retrieves the complete message.
* **Fixed incorrect calculation of available schedule slots**: Schedules that have exhausted their budgets no longer count toward the active schedule limit, and the correct limit is maintained when multiple administrators create schedules concurrently.
* **Fixed inconsistent knowledge base file states during reparsing, searching, and deletion**: Reparsing preserves tags, files with no text are marked as failed, semantic search follows the selected mode, and storage and chunk statistics reset to zero after all files are deleted.
* **Fixed inability to connect to Oracle 11g databases**: On-premises environments now use a compatible Oracle Instant Client mode, allowing table reads and database Q\&A to work normally.
* **Fixed thinking settings still being applied after deep thinking is disabled**: OpenAI and Azure models now use reasoning effort only when the feature is enabled and the settings are valid, preventing residual settings from causing ordinary conversations to fail.
* **Fixed replies arriving out of order when multiple requests are sent in the same conversation**: The system processes replies for each conversation sequentially to prevent interleaved content or duplicates.
* **Fixed unclear messages when deleting language models that have been used**: The admin interface explains that the model has usage records and directs administrators to disable it safely with “Is active,” replacing messages that resembled permission errors.
* **Fixed false permission errors after creating AI assistants from the template library**: Non-owner members opening the edit page with their existing permissions no longer see errors triggered by system roles or connector icon loading.
* **Fixed custom tools being unable to retrieve conversation attachments**: Even if the AI rephrases an attachment URL, the system maps it back to the original file before calling the tool, preventing interrupted uploads or analysis.
* **Fixed inconsistency between plan expiration dates and remaining-day counts**: The organization overview now shows an “Available until” date using the same basis as the remaining-day count, avoiding confusion about the service period.

## v2026.09.16 <a href="#v2026-09-16" id="v2026-09-16"></a>

***

### 🛠️ Bug Fixes

* **Fixed tool call failures caused by optional MCP connector fields**: When AI assistants call MCP connector tools such as Monday, Zapier, or Tableau, omitted optional fields are no longer sent as null values, preventing parameter validation failures from interrupting comment creation or other actions.
* **Fixed incorrect navigation positions in long conversations**: After viewing more than one page of content and loading earlier messages in the conversation platform, clicking an item in the right-hand navigation panel now stops at the selected message instead of jumping back to the top.
* **Fixed inability to click navigation items consecutively in long conversations**: In the conversation platform, users can now select different messages consecutively in the right-hand navigation panel and move to the correct positions even when the panel requires scrolling, without the second click becoming unresponsive.

## v2026.09.15 <a href="#v2026-09-15" id="v2026-09-15"></a>

***

### ✨ New Features

* **AI Assistant Template Library**: Administrators can preview, clone, and ask questions about templates in the library, then quickly create new AI assistants for their organization. The page also shows the model each template uses and whether it is available to the organization.
* **AI Assistant Long-Term Memory**: Enable persistent memory for AI assistants to extract useful facts from conversations and reuse them in later interactions. The management interface shows memory sources, while sensitive personal data is filtered out and memories are isolated by contact.
* **Conversation Quick Menu**: Configure a persistent Quick Menu in the input bar for an AI assistant, using icons and action items to guide users through common workflows.
* **MaiGPT Conversation Folders (Coming Soon)**: This feature will launch together with the Work workspace. It is not yet available in production and will be announced when it launches. Planned capabilities: Create folders in the Work sidebar to organize conversations and quickly find content by task or project as the number of conversations grows.
* **Azure Blob Storage File Storage**: On-premises and enterprise deployments can use Azure Blob Storage. The administration interface and Web Chat upload files directly using the preauthorized upload method.
* **Quoted Replies in LINE Groups**: Replying to or quoting an AI assistant's message in a LINE group now invokes that assistant and preserves the quote relationship in the reply, without requiring an additional mention.

### 🚀 Core Performance Improvements

* **Faster Viewing of Large Knowledge Base Files**: File content now loads in pages. Editing a single paragraph updates only that paragraph instead of downloading the entire file again.
* **More Reliable AI Assistant Settings Saves**: When saving an AI assistant with a large conversation history, tool association synchronization skips unchanged data and runs in the background, reducing long waits and timeouts.

### 😊 User Experience and Interface Improvements

* **Clearer Attribution for Team Replies**: When multiple member AI assistants collaborate on an answer, message headers in Web Chat, conversation testing, and the management interface show the member assistant that actually generated each reply.
* **Knowledge Base Label Filters**: Set default label conditions at the AI assistant level to filter retrievable content using clear combinations of “include any,” “include all,” and “exclude.” If the knowledge base has no labels yet, the interface clearly explains the next steps available.
* **Full Knowledge Base Retrieval Range**: Administrators can choose to retrieve knowledge base candidates without first limiting their number, allowing tools that need the full retrieval range to obtain data according to the AI assistant's settings. The interface also displays the parameters actually used.
* **Redesigned Quick Replies and Suggestion Buttons**: Suggested options now appear as floating buttons above the input bar or compact buttons below messages. Clicking a suggestion sends only the button text, without sending pending attachments or clearing the draft.
* **Expanded AgentOps Execution Record Filters**: Filter team execution traces by conversation and member. Scheduled executions are also included in the records with their source clearly labeled, and dataset lists show single-turn and multi-turn test counts separately.
* **Live Recording Limit Notifications**: A clear notice appears as recording approaches the three-hour limit. Reaching the limit finalizes the room normally instead of marking it as failed. The browser also notifies users of interruptions and explains why fallback is unavailable.

### 🔐 Security and Governance Enhancements

* **Secure Forwarding of Keycloak Identity Tokens**: Administrators can choose whether a conversation platform forwards visitors' Keycloak tokens to MCP or API tools, allowing backend services to use the original user's identity and permissions after explicit activation.
* **Stronger Knowledge Base File Access Controls**: Queries for knowledge base content by file now consistently check the current organization and file permissions. Files that do not exist or belong to other organizations do not expose information.
* **Immediate OAuth Token Revocation**: After a user logs out or revokes a refresh token, the revoked token no longer remains usable due to brief caching, reducing the risk of delayed termination of account access.

### 🛠️ Bug Fixes

* **Fixed Temporary Retrieval Gaps When Editing Knowledge Base Files**: When saving pages in PDFs or other documents, the file's new index is created before the old one is removed. At least one version remains available during reindexing, and a single save no longer processes the entire file repeatedly.
* **Fixed Web Chat Message Queuing and Interruption States**: Queue the next message while AI is replying, and press Esc to interrupt. Even if you stop before the first response chunk appears, or a tool finishes because of an error or cancellation, the interface immediately unlocks and properly clears the in-progress state.
* **Fixed Hybrid Mode Fallback for Non-Streaming Replies**: When hybrid reply mode falls back to a general answer because no template content was found, it now uses the request's delivery method. Non-streaming calls no longer return a generic error due to an uninitialized connection.
* **Fixed AI Assistant Day-of-Week Calculations**: When system-time-aware answers are enabled, the current time provided to the AI assistant now includes the correct day of the week, without requiring the model to calculate it from the date.
* **Fixed Broken Quick Menu Icons**: After a custom Quick Menu icon is uploaded, the system stores its media location and generates a valid authorized URL when it is read, allowing icons to display correctly with private media storage.
* **Fixed missing knowledge base images in LINE and other channels**: Image syntax with titles in knowledge base answers is now correctly recognized and sent through external channels such as LINE.
* **Fixed the scope of attachments automatically passed to the code execution tool**: The code execution tool now automatically receives attachments uploaded in the current conversation, rather than only those in the latest message. Image attachments can also be processed by the vision parser, and unsupported types are explicitly rejected before further processing.
* **Fixed duplicate FAQ Excel imports and stalled failed jobs**: The interface now prevents starting another FAQ import while one is already running in the same knowledge base. Jobs are also marked as failed when batch processing fails permanently, instead of remaining in progress indefinitely.
* **Fixed Text to SQL connection interruptions and table name limitations**: Queries are retried after temporary connection interruptions, with retry counts recorded correctly. MSSQL tables and tables with schema-qualified names can now be listed, inspected for structure, and queried.
* **Fixed server errors caused by overly long shared conversation titles**: Saving a shared conversation title that exceeds the length limit now returns a manageable input error instead of a server error.

## v2026.09.10 <a href="#v2026-09-10" id="v2026-09-10"></a>

***

### 🚀 Core Performance Optimization

* **Faster settings saves for large AI assistants**: For AI assistants with extensive conversation histories, saving settings such as role instructions now takes about 0.3–0.6 seconds instead of over 20 seconds. Tool status updates complete in the background within a few seconds.

### 😊 User Experience and Interface Improvements

* **Improved “Back to Latest” navigation for long answers**: When scrolling up to review earlier content in MaiGPT or Work conversations, a “Back to Latest” button appears. Click it to return to the latest question. Misleading directional prompts no longer appear when scrolling down through the current long answer.

### 🔐 Security and Governance Enhancements

* **Stronger data protection for anonymous Web Chat conversations**: Responses for anonymous conversations in public Web Chat now include only the data required by the interface. Visitor network locations, conversation platform operational statistics, and AI model and tool settings are no longer transmitted. Management features for signed-in members are unaffected.

### 🛠️ Bug Fixes

* **Fixed Gemini assistants being unable to use tools**: AI assistants connecting to Gemini models through compatible services can now use their assigned tools normally. Tool constraints and multi-turn execution no longer interrupt answers.
* **Fixed duplicate quick replies in Web Chat**: Quick reply options now appear correctly below messages, without an extra unrecognized “Tool Response” card inside the message bubble.
* **Fixed incorrect permission errors on the API key page**: Organizations without AI Gateway enabled no longer see an unrelated permission error dialog when opening API key management. The key list remains fully usable.
* **Fixed interrupted Gemini assistant answers involving multiple tools**: AI assistants connecting to Gemini models through compatible services can now use two or more tools for a single question and complete their answers. They no longer report that the system is temporarily unable to process the request after tools have already retrieved data.

## v2026.09.08 <a href="#v2026-09-08" id="v2026-09-08"></a>

***

### 🛠️ Bug Fixes

* **Fixed internal evaluation text appearing in AI replies**: Quality check results no longer appear in conversation content. Users see only the reply the AI assistant is meant to provide.

## v2026.09.07 <a href="#v2026-09-07" id="v2026-09-07"></a>

***

### ✨ New Features

* **AI Gateway monitoring and key governance**: Organization administrators can set custom start and end times in the AI Gateway dashboard and view request counts, Token usage, fallback triggers, and costs by API key. 1. Expand each model row to see which keys drove usage during that period and identify the source of usage spikes; 2. Configure individual keys in place, including enabled status, requests per minute, Token limits, and allowed models. Keys are still created and revoked in organization settings; 3. New access diagnostics list the checkpoints each key passes or fails; 4. Model nodes on the fallback rules page display badges for the model vendor and hosting cloud.
* **Sampling parameter support detection for custom models**: When editing a custom model in AI Gateway, the interface shows whether the model supports temperature, Top-P, and Top-K, along with the detection time and determination source. If the result does not match actual behavior, click “Re-detect” to test again immediately. Detection runs automatically in the background after a new model is saved, so administrators no longer need to investigate each parameter when newer models reject non-default values.
* **Self-service organization model access settings**: Organization owners can enable or disable available models within their MaiAgent authorization scope on the “Organization Settings → Model Access Permissions” page. Models are grouped by cloud or on-premises deployment, with vendor icons, brand labels, and search filters. Models that have not been activated are marked with dashed borders and cannot be enabled. Disabling a model still used by AI assistants requires confirmation to prevent accidental disruption of live assistants.
* **New Gemini 3.1 Pro image recognition**: When building a knowledge base, Gemini 3.1 Pro Preview through Vertex AI is now available as a document image recognition (OCR) parser. Entire PDFs are submitted in one request, supporting up to 50MB or 1,000 pages. Failed requests automatically fall back to resubmission in batches of 50 pages, providing another option for speed and cost. The default parser is unchanged; switch parsers in the list when needed.
* **Return-to-AI event notifications for human handoff Webhooks**: Enterprise customers integrating external customer service systems through Webhooks now receive a “Return to AI” event when an agent hands a conversation back to AI, in addition to the existing handoff creation event. Payload fields, signatures, and retry behavior match existing events; receivers only need to route by event name. Conversation data in message Webhooks also includes handling status, progress status, the auto-reply toggle, and the assigned agent. External systems can use these fields to process only messages under human service and attribute replies to the correct agent.
* **Cache savings shown in usage statistics and reply details**: Organization administrators can now view cache writes, cache reads, and cache hit rates in ranking tables across six dimensions and trend charts on the usage statistics page. The Credit consumption card includes a new “Cache Savings” row. The Credit consumption panel in AI reply quality control details also shows the Credits saved by caching for each reply. Savings appear only when positive; the entire row is hidden if the rounded amount is zero or represents an extra cost. Credit audit exports are unchanged, and the export entry point notes that derived savings values are excluded.
* **Cross-conversation retrieval settings for AI assistants**: A “Cross-conversation Retrieval” toggle and lookback period field have been added to “Conversation Memory Settings” in the AI assistant advanced settings (1 to 365 days, default 90 days, disabled by default). When enabled, the assistant can also search messages and read context from the same user’s other conversations within the specified period. This suits customer service and personal assistant scenarios that need continuity with prior conversations. This setting is independent of vector memory and can be used in any combination with it.
* **Automatic retries for transient tool call errors**: When AI assistants encounter transient errors such as connection loss, resets, or timeouts while calling tools, the platform automatically retries up to 3 times (at intervals of 1, 2, and 4 seconds). Retries no longer depend on the model’s behavior—in real cases, some models simply reported that the same database query could not be completed after an error. Scope: 1. MCP tools are retried only when their names indicate read operations (queries, lists, or searches); write operations are not retried to avoid duplicate execution; 2. API tools retain their existing retry mechanism and accurately record retry counts; 3. Tool execution records include a new “Retry Count” field for administrators. Credits are deducted only once after final success, with no duplicate billing for retries.
* **AI assistant monitoring conversation record detail page**: AI assistant monitoring in AgentOps now has a dedicated conversation record detail page. Administrators can open it by clicking anywhere on a list row: 1. View basic information, identifiers (copyable with one click), an execution timeline, and statistics (processing time, Tokens, character count, Credits, and quality scores), with JSON, Markdown, and Excel exports; 2. Navigate to the previous or next record and share the URL directly with colleagues, with the corresponding organization selected automatically on opening; 3. Returning to the list restores filters, page number, and page size, while a new search starts at page 1.
* **Add conversation records to FAQ with one click**: When administrators find a question that the AI answered poorly in AI assistant monitoring records, they can click "Add to FAQ" in the list's actions column or the detail page's title bar and create a Q\&A in a knowledge base they can edit, all on the same screen. The question is prefilled with the user's original question. Administrators can leave the answer blank and write it themselves, or click "Use AI response" to insert the original response (with reminders about its quality score and feedback) and then edit it. Blank answers cannot be submitted, drafts are retained if saving fails, and the entry is disabled with an explanation when no editable knowledge base is available. There is no longer a need to switch pages to create an FAQ manually or import it from a spreadsheet.
* **Hook settings for personal data detection and content safety classification**: Organization administrators can configure two types of protection directly in the Hook settings form: 1. Personal data (PII) detection—after selecting a detection service such as Google Cloud DLP, choose the items to detect from general types (email addresses, phone numbers, Taiwan national ID numbers, etc.) and the service's native detailed types (IBANs, credit card numbers, etc.), shown in separate sections of the same menu; 2. Content safety classification—after selecting a classification service (currently OpenAI Moderation), choose categories to block, such as hate, violence, self-harm, and jailbreak attacks; 3. Switching services automatically clears selected items to prevent save failures caused by incompatible types. Both settings interfaces are available in five languages.
* **Official built-in Agent marketplace and organization management**: 1. Organization members can browse official built-in Agents in the marketplace, search by scenario category, distinguish enabled Agents from disabled ones, and use the preview window to replay demo conversations, insert sample questions with one click, and view sample outputs; 2. Administrators can filter built-in Agents by status, category, and keyword in organization settings, enable or disable them in bulk, and specify who can access them. Items are locked when seats are full or during a cooldown period; 3. Enabled official Agents do not appear in the list of self-built AI assistants or the marketplace's "Organization-built" tab. Organizations without marketplace access or with zero seats in their plan do not see the marketplace entry, while organizations that have rented Agents before can still access the management page to release instances.
* **MaiGPT meeting records quick entry and calendar connection**: 1. The MaiGPT blank conversation page now has a quick entry bar with four entries: presentations, documents, spreadsheets, and meeting records. Organizations with meeting records enabled can start live transcription with one click and view the timer, live transcript, and three most recent records in the panel; 2. Connect either a Google Workspace or Microsoft 365 calendar. The panel lists upcoming meetings by date; for an ongoing meeting, click "Record" to automatically populate its title and participants; 3. When recording ends, an AI summary is generated automatically and displayed directly in the panel. It can be inserted into the conversation input or opened on the full page; 4. Return from the summary view to the entry screen with one click.
* **File previews support presentations, SVG, and Office formatting**: 1. Word document previews now render with high fidelity, preserving the original table borders, fonts, and paragraph spacing. Files are processed entirely in the browser without being sent to external services; 2. Excel previews retain bold text, text and background colors, alignment, and merged cells; 3. Presentations generated by AI or uploaded to a knowledge base can be previewed page by page. Presentations that have not been converted beforehand are converted when opened, with a "Converting" indicator. Downloads still provide the original presentation file; 4. Added SVG image previews. These features apply to users opening files in MaiGPT and the information sidebar on the right of the conversation page.
* **Official LINE WORKS messaging platform support**: Businesses using LINE WORKS can now add LINE WORKS directly from the admin console's channel selection page. After entering Bot and Client credentials and pasting the private key, they can test the integration before creating the messaging platform. Once created, its settings page offers the same eight tabs as other channels: Basic settings, Feature settings, Webhook, Human handoff, Conversation tags, Conversation analytics, AI Copilot, and Permissions. Credentials are entered only during creation. The edit page does not display stored secrets and retains only read-only fields that can be checked, such as the Bot ID.
* **Viber messaging platform integration**: Viber is now available on the admin console's channel selection page. Paste the Bot Token provided by Viber to create a messaging platform. The system automatically registers the event receiving URL with Viber, so there is no need to manually paste it into Viber's admin console. A welcome message and reset command can be entered during creation, and the settings page supports feature, permission, and customer service handoff management just like other channels. Because Viber does not support retracting sent messages, cleanup options for waiting replies are not shown on this channel.
* **Login-free Web Chat access with Source ID**: When a customer's website already had its own membership system, visitors previously still had to log in again inside Web Chat. Administrators can now enable this feature on the messaging platform's Login settings tab and choose between signature verification (recommended) and no verification. The platform generates a signing secret and provides backend examples for PHP, Node, and Python, plus a one-line frontend call. When the embedding page supplies the member identifier and signature, Web Chat starts the conversation directly under that identity without showing a login screen. Failed verification records can be expanded on the same page to help troubleshoot integration issues.
* **Customize Web Chat waiting styles and tool text**: The waiting screen shown before an AI response and tool execution cards can both be tailored to your brand: 1. Messaging platform settings now include a "Response waiting style" tab. Choose narrative text or brand animation, configure multilingual text and long-wait messages for each stage (typing, thinking, analyzing, interrupted), upload brand animation assets, and translate text with AI in one click; 2. The live preview on the right is aligned pixel by pixel with the actual Web Chat and can play the entire flow. A new "Show text alongside animation" toggle controls whether text is displayed during the typing stage; 3. Tools can have categories and display text for running/completed states, and MCP connectors can also specify a default category. Tool cards in conversations prioritize this text over English technical names; 4. Web Chat applies text according to the visitor's language. Tool cards can show custom running/completed icons (automatically falling back to system icons if loading fails), and animations remain on a static frame when visitors prefer reduced motion.
* **Pre-chat data collection supports logged-in users**: Previously, when a logged-in visitor filled out a Pre-chat form, a separate unowned anonymous contact was created, leaving the conversation disconnected from the account. The system now identifies the logged-in user first: the first submission is attached directly to that account; if the account already has data, the form is skipped and the conversation starts immediately; if an earlier anonymous record is detected, the user is asked to confirm ownership and claim it before a conversation is created. Resubmitting data updates the existing contact instead of creating a new one. Behavior for anonymous visitors and messaging platforms without login enabled remains exactly the same.
* **Contact bulk import preview and merge mode**: Bulk contact imports now use two steps. After uploading a file, the interface first shows a statistics card with "X records will be created / Y records will be updated" and a summary of the current settings. No data has been written at this point, and you can go back to make changes. The import runs only after confirmation. A new messaging platform association option lets you replace existing associations with the imported ones (the default, matching previous behavior) or merge them into existing associations. If there are 0 importable records or the import fails, a message appears and the interface stays on the settings page to prevent mistakes.
* **New file library**: Previously, files could only be added to a knowledge base or used as conversation attachments. The platform had no unified file management interface, and accidental imports, deletions, or moves could not be undone. Members can now use two spaces, "My files" and "Organization shared files", to: 1. Create folders and import multiple files at once, with a warning before upload if a single file exceeds 10MB; 2. Browse in list or thumbnail view, filter by file type, search across spaces, and see each file's full location path; 3. Rename, move, or delete files in bulk, and undo mistakes with one click from version history; 4. See unauthorized actions disabled, with an explanation on hover.
* **Move files between knowledge bases**: Previously, reorganizing knowledge base content required deleting files and uploading them to another knowledge base, then waiting for parsing and vectorization again. Files can now be moved directly from the knowledge base list or an AI assistant's file list to another knowledge base using the same vectorization model, without reparsing: 1. The transfer selector supports search, pagination, and multi-file selection, and warns in advance which tags will be cleared; 2. Moves run in the background and continue even if the window is closed. Completion is indicated by a notification and a red dot on the entry; 3. The results drawer separately lists completed, skipped, and failed files, with one-click retry for failures.
* **MaiAgent role Credit quota control**: Organizations billed by Credit could previously set quotas only for members and messaging platforms, with no way to control usage by role. Now: 1. The Credit usage page includes a "Role quotas" tab where administrators can filter by status, search, enable quota control for each role, and allocate, edit, or disable quotas; 2. On the profile page under "My quotas", members can see the quota, usage, remaining balance, and expiration date for each quota-controlled role they belong to, with warnings for low balances and approaching expiration; 3. The role settings page has a read-only quota summary card, and users with Credit permissions can go directly to management. Usage is tracked independently for each role quota, and service stops as soon as any quota is exhausted.
* **Self-Service API Key Issuance and Key Source Management**: Previously, members could not issue their own personal API keys, and each person could hold only one active key. Members can now issue, view, filter, and disable personal keys on their own settings page. Organization administrators can see whether each key was issued by an administrator or by the member on the organization's API key page. The limit of one active key per person has also been removed.
* **Plan History Timeline Added to Organization Overview**: Organization members previously could only see the current plan name and expiration date. They could not tell which plans they had used or when the next switch would happen, so they often had to contact support. A timeline below the current plan card in the organization overview now shows past, current, and preactivated plans with their start and end dates, automatically switching to a vertical layout on mobile. When there are more than 5 records, only the most recent ones appear by default. Members can expand earlier records and return to today with one click.
* **Redirect to a Specified URL After Logout**: For organizations with on-premises deployments or single sign-on (SSO), the built-in admin login page usually serves no purpose, leaving users at a dead end after logout. Deployment operators can now set a logout redirect URL. Once configured, manual logout, logout due to inactivity, and opening an admin link while signed out all trigger a full-page redirect to that URL. The Back button will not return users to the admin login page. If no URL is configured, the original behavior remains.
* **Self-Service OAuth Application Registration for Organizations**: Organization settings now include an OAuth Applications page where organizations can manage applications used by MCP clients such as Claude, without asking platform staff to do it. The page provides: 1. Application creation, editing, and deletion, plus client secret regeneration; the secret appears only once, when created or regenerated. 2. Connection information, including authorization and token endpoints, with one-click copying. 3. View access for all members; only owners or members with OAuth Application Management permission can make changes. Buttons are disabled for members without permission and explain why. 4. A limit of 20 applications per organization, with the Create button disabled and an explanation shown when the limit is reached.
* **Connectors Can Pass Web Chat Context Data**: Organizations using MCP connectors can now specify allowed context data fields in connector settings. Data passed when embedding Web Chat in a website (such as a patient identification token) can then be included in authentication headers sent to the MCP server. This supports scenarios that need to pass the conversation user's identity back to a hospital or internal enterprise system. Supporting changes: 1. The settings dialog for service-account connectors can now be reopened for editing, and the list includes a settings entry. Existing headers are masked; leaving them blank preserves them. 2. The context variable help panel in the connector form now includes context data usage and examples. 3. Invalid field name formats or more than 20 fields trigger immediate feedback; unchanged fields are not overwritten.
* **Embed AI Images in MaiGPT Document Generation**: When generating presentations, Word documents, spreadsheets, or AI Slides presentations in MaiGPT, you can ask the assistant to generate covers, illustrations, or diagrams from a description and embed them directly in the document. Generated images use the existing image generation credit pricing. If the quota is insufficient or image generation fails, the document is completed using text and layout elements without interrupting generation. Administrators can specify the image generation model used for document generation in the admin interface.
* **Voice Assistant Monitoring Dashboard (Coming Soon)**: The admin console screens are not yet available in production and will be announced when they launch. Planned capabilities: AI assistant monitoring now includes a dedicated voice assistant section: 1. The overview summarizes call architecture distribution, token composition, credit breakdown, and call answering statistics. 2. Individual call records include complete transcripts and usage details. 3. New calls record per-message metrics for each conversation turn, allowing usage to be matched precisely to each turn. Older calls are explicitly marked as having no per-message metrics, without filling the gaps with estimates.
* **Organization-Level Platform Processing Fee**: 1. The platform can configure a processing fee rate for an individual organization. Each subsequent credit consumption entry for that organization incurs an additional processing fee at the same rate. The fee is deducted and logged alongside the original consumption, and is reversed when the original consumption is reversed. 2. Customers can see the rate actually applied to their organization on the billing method page. The default rate is 0, leaving billing unchanged for organizations without a configured rate. 3. This also closes a gap where consumption from direct AI Gateway calls and team workflows was missing from usage reports; future consumption is now counted correctly.
* **Resume Interrupted Multi-Turn Evaluations**: Multi-turn conversation evaluations now record results per case, saving each case as soon as it completes. If an evaluation stops halfway due to a timeout, cancellation, or system interruption, it is marked as partially completed and can resume from where it stopped. Scores for completed cases remain intact, and only failed or unexecuted cases are rerun. Credits are charged only for newly completed cases. If no cases complete, the evaluation is marked as failed instead of completed.
* **Safety Metrics and Unstable Case Markers Added to AI Assistant Evaluations**: When creating and reviewing multi-turn evaluations, teams can select safety-related metrics and identify test cases with repeatedly inconsistent results, helping them prioritize quality risks.
* **Conversation Platform Handoff Notification Settings**: Administrators can configure handoff event notifications in notification settings, automatically sending events that transfer conversations to human service to a specified external system.
* **AI Assistant Favorites Sync**: Favorites for self-built AI assistants are saved to the account, preserving frequently used lists when users switch devices or sign in again.
* **Attachment Parsing Progress Display**: After users upload an attachment in MaiGPT, the conversation displays parsing progress and processing status so they can clearly see when the file is ready for AI use.
* **Custom Tool Message Display in Web Chat**: Administrators can choose a display mode and configure output templates for different tools. Web Chat applies these settings to present more readable execution results.
* **Deep Research Output Format Selection**: Users can specify the output format before starting research. Once results generated in the background are ready, a file card provides a download without requiring users to stay and wait.
* **API Key Creation Wizard**: API key management now includes a three-step creation flow, a shared information panel, and provisioning status fields to help administrators confirm that configuration is complete.
* **AI Gateway Dashboard and Provider Setup Wizard**: Administrators can review service status, usage trends, and items needing attention from the overview, and use a guided flow to discover providers, select models, and configure fallback order.

### 🚀 Core Performance Optimizations

* **Faster Web Search Tool Responses**: When AI assistants use web search, measurements show a median time to first token of 9.75 seconds for replies with search and 3.27 seconds without. This update removes three sources of idle waiting: 1. Search connections are reused, so each search no longer establishes a new encrypted connection. 2. The timeout for a single search is reduced from 60 seconds to 15 seconds; a timeout reports failure so the assistant can use another approach instead of waiting. 3. Search summaries are expanded from 500 to 1,500 characters, reducing the extra round of fetching full web pages after a search (observed in about 55% of search replies). Longer summaries slightly increase input Token usage per reply.
* **Less Unnecessary Data Transfer When Loading Conversation Messages**: When viewing conversation messages, each page of 20 messages previously fetched the associated assistant's full prompts from the database along with every message (about 14 KB per message on average, or up to 300 KB per page for assistants with longer prompts), although the interface only used the assistant's name. Only necessary fields are now fetched, reducing measured query time per page from about 22 milliseconds to 14 milliseconds. Organizations with more conversation records and longer assistant prompts receive faster responses when paging through messages.
* **Faster Calendar Event Loading for Meeting Records**: When the meeting records panel loads Google/Microsoft calendar events, it now returns the existing calendar cache immediately and refreshes it in the background. External calendar service delays no longer slow the response: the slowest 1% of requests previously waited more than 15 seconds; they now receive cached responses marked as awaiting refresh. Members with invalid calendar authorization no longer retry a doomed authorization refresh on every request. Retries now start at 1 minute and double up to a maximum interval of 1 hour, with normal operation restored immediately after reauthorization.
* **Faster Web Chat message responses**: Previously, each message sent by a visitor in Web Chat caused the server to read the same Web Chat configuration 6 times and run an additional reverse lookup that scanned about 70,000 conversation platform records. The configuration is now resolved once per message and shared across all subsequent processing steps. Web Chat configuration reads drop from 6 to 1, and the reverse lookup, which averaged about 10 milliseconds, is eliminated entirely. The response format remains unchanged.
* **Faster conversation platform menu loading**: The conversation platform dropdown loaded when entering the admin console, also used on the contacts page and in the role assignment form, retrieves all conversation platforms readable within the organization at once. Previously, each record included more than a hundred fields unused by the menu, such as prompts, welcome messages, and themes. It now reads only the fields displayed in the menu. Tests with an organization containing about 1,087 Web Chat platforms reduced request time from 210 to 157 milliseconds and transferred about 500 KB less data. The standard paginated list is unaffected.
* **Faster new Web Chat conversation creation**: In the first request made when a visitor opens Web Chat to create a conversation, 3 permission and data queries previously scanned all roughly 70,000 records because the conversation platform table lacked the corresponding index, consuming about 22 milliseconds of database time per request. A dedicated index now lets these 3 queries find records directly. Queries for each inbound message across channels and for Web Chat contacts also benefit. Response formats and functional behavior remain unchanged.
* **Reduced memory usage for conversations with large tool outputs**: Restoring an old conversation containing large tool results in Ask AI or Web Chat (nearly 3 MB on one page in an actual customer report) caused the page to freeze for about 1 second and memory usage to keep climbing while typing until browsing history was cleared. Tool card step lists and step details now load only when first expanded. Tool results longer than 20,000 characters show only the beginning with a truncation notice. Tests with the same conversation reduced page nodes from 5,056 to 2,533 and resident memory from 255 MB to 158 MB, with a slightly shorter pause on opening. Card appearance and collapse animations remain unchanged.
* **Faster PDF vision parsing**: Knowledge bases using vision to parse PDFs previously sent a separate recognition request for every page, taking nearly 30 minutes for a 150-page document in testing. Multiple pages (10 by default) are now combined into one request. If output is truncated or content is too large, the batch is automatically halved and retried down to a single page. The per-page parsing result structure stays the same. Speed gains from batching vary by recognition model; internal tests with Claude Sonnet 5 were about 1.5 times faster.
* **Faster role member list loading**: When opening a MaiAgent role's member list or member edit form, the system previously queried the organization plan for each member individually. A role with over 3,800 members required thousands of queries, with average response times exceeding 1 second. Plan data is now prefetched once, so query counts no longer grow with member counts: queries for a role with 3,826 members drop from 3,826 to 1, and the paginated list drops from 20 to 1.
* **Faster Telegram message processing**: For organizations using Telegram as a conversation channel, each incoming message previously fetched the sender's avatar from Telegram twice, with a median processing time of about 1.6 seconds. Avatars retrieved within the past 24 hours are now reused. They are fetched again only for first-time users, missing avatars, or avatars older than 24 hours. Based on actual repeat rates, processing time is estimated to fall to about 0.6–0.9 seconds. After a user changes their avatar, the contact avatar on the conversation platform may take up to 24 hours to update; display names still update immediately with every message.
* **Faster AI assistant ranking dashboard loading**: When organizations with high conversation volumes open the assistant ranking in the Agent Ops dashboard, queries now calculate only the fields needed for rankings. Error counts use a dedicated index path, and assistant names are retrieved only for the top 50. Each load no longer reads wide fields from the entire set of conversation records or the full organization assistant list, reducing dashboard wait times.
* **Faster Messenger webhook responses**: Incoming Facebook Messenger and Instagram messages now receive an immediate platform response after signature verification, with message processing moved to the background. Contact data is no longer fetched synchronously from Facebook on receipt. Webhook response times are expected to drop from about 0.5 seconds to under 0.1 seconds, reducing the risk of Meta resending events or suspending event notifications due to slow responses.
* **Faster knowledge base file keyword searches**: When searching by keyword in a knowledge base with tens of thousands of files, redundant deduplication has been removed from pagination count calculations. Tests reduced the same query from 832 to 96 milliseconds and eliminated disk-based temporary sorting, shortening the wait after each keystroke when typing a search.
* **Faster filtering on the built-in AI assistant management page**: When switching categories or searching on the built-in AI assistant management page in organization settings, access audience data for all assistants is now retrieved in a single batch query rather than one request per enabled assistant (75 requests were observed in testing). Organizations with many enabled assistants no longer experience stalls when changing filters.
* **Lower background resource usage for embedded Ask AI**: Unnecessary layout and timer tasks pause when the page or browser tab is hidden, and state synchronizes when returning to the page, reducing CPU and memory usage during extended use.

### 😊 User experience and interface improvements

* **AI Gateway traffic fee pricing display**: For organizations using self-hosted models (BYOK) billed through AI Gateway, the traffic fee row on the billing method page now displays the formula “total tokens × organization rate,” with the tooltip showing the organization's actual percentage. “Platform fees” has been added to category filters in usage records and category tabs in usage statistics, so traffic fees are no longer excluded from pie charts and totals. The AI Gateway model list also has a new “Pricing” column showing the traffic fee formula for self-hosted models.
* **Three-level Agent marketplace filters and recently used Agents**: The official tab in the Agent marketplace has been redesigned: 1. Filters now have three levels: “tab → category group → multi-select tags,” with no selections meaning no restrictions. 2. Cards use a light background based on their category's primary color, with dedicated colors for all 14 categories. Icons fill the entire tile and match the top-right category label's color family, consistently across the marketplace, recently used section, preview window, and management rows. 3. A new “Recently used” section lists the 5 Agents most recently used in conversations; hover to continue a conversation. 4. Seat counts are included in the page title, with distinct empty-state messages for no search matches, no filter results, and no enabled Agents.
* **Assistant settings connector tab available to all organizations**: The “Connectors” tab in AI assistant settings and the “Connector binding” item in organization advanced features previously appeared only after individual activation by the platform. All organizations can now see and use the connector tab directly when an assistant has Agent mode enabled, without requesting separate activation.
* **Red confirmation buttons for irreversible actions such as deletion**: After clicking a red delete icon in the admin console, the confirmation screen previously showed a brand-blue primary button suggesting a “safe default,” contrary to the action's consequences. This update: 1. Makes confirmation buttons solid red in table row confirmation popovers and centered confirmation dialogs (deleting SIP numbers, teams, Hooks, crawler schedules, meeting records, conversations, MCP connectors, AI Gateway entries, and more); non-destructive confirmations such as logout or leaving without saving remain blue. 2. Changes the API Key revocation confirmation button to the verb “Revoke” and adds hover text on the icon button explaining the action or why revocation is unavailable. 3. Stops automatically inserting spaces into two-character Chinese buttons across the site (removing the space between the two Chinese characters), matching the spacing of surrounding text.
* **AI presentation preview panel can be closed and remembers its state**: After MaiGPT generates an AI presentation, the right preview panel now has a close button in its header. Once closed, it stays closed after sending another message, refreshing, or switching to another conversation and back. An “Open presentation” card appears at the end of the message that generated the presentation, allowing the panel to be reopened anytime. On narrow screens such as phones, the panel overlays the conversation area so the close button and input field are no longer clipped. Generating a new presentation still opens the panel automatically.
* **Fixed-width right information sidebar and resizable document previews**: The knowledge base/output file card sidebar on MaiGPT and conversation pages now has a fixed width, with no drag-to-resize handle on its left edge. Document previews opened from cards use a separate floating layer whose left edge can be dragged to adjust its width between 400 and 800 pixels. Each user's setting is saved, and resizing the preview does not affect the card sidebar itself. The sidebar collapse hover area is widened to 24 pixels, making it easier to target; the cursor now indicates clicking and shows a resize cursor only while dragging.
* **Background AI meeting summary generation with progress display**: Summaries now run in the backend as background jobs, preventing long meeting transcripts from failing due to frontend wait timeouts. The summary page displays progress as “Analyzing transcript N/M → Combining summaries N/M,” automatically reconnects after leaving or refreshing, and saves the completed result directly as a new version. Any version can also be deleted. Generation failures show a consistent retry prompt. The assistant selector lists only available assistants and uses the default template when none are available.
* **Meeting transcript loading error notices and a retranscription dialog**: If transcript loading on the meeting record detail page still fails after repeated retries, it now displays a “Transcript failed to load” error and a reload button, distinguishing loading failures from the empty state with no transcript. Retranscription now uses a dedicated dialog combining history, speech recognition provider selection, and speaker diarization settings. Fields in the offline audio import form are no longer clipped on phone-sized screens.
* **Login guidance for unauthenticated embedded Copilot views**: When AI Copilot is embedded in external systems such as Genesys Cloud, browsers isolate the embedded window’s login state from the main window. Support staff often saw only a generic “You do not have access to this conversation” error, with no guidance on what to do. Unauthenticated users now see a “Please log in first” prompt and a “Go to login” button. After logging in within the embedded window, they automatically return to the original conversation page; subsequent use in the same browser requires only that one login. Organizations using Google account single sign-on must still log in with a username and password because Google does not allow its login page to appear in an embedded window.
* **Synchronize conversation platform names when renaming an AI assistant**: After you save a new name for an AI assistant, a confirmation dialog now lists its Web Chat and connected channels such as LINE and Telegram. Select the platforms to rename together and apply the changes at once, without opening each platform’s settings. System-managed built-in platforms synchronize automatically and are excluded from selection. Assistant deletion is also fixed: selected platforms are deleted before the assistant, resolving the previous issue where success was reported even though the platforms had not been deleted. If any platform deletions fail, the dialog lists them accurately and retains the assistant for a retry.
* **Show setting names and translated language names in the top-right menu**: End customers using Web Chat’s top-right menu previously saw only values such as “Traditional Chinese” and “Dark,” making it hard to identify each setting. The interface language, speech language, and theme rows now use “Name: Value,” matching the existing speech playback and font size rows. Language lists in the top-right menu, pre-chat form, and MaiGPT settings dialog also use “Native name (name in the current interface language),” such as English (English). Language searches match native names, translated names, and language codes. When a translation is unavailable, only the native name appears, with no empty parentheses.
* **Dedicated desktop Web Chat layout and quick access to common settings**: Full-page Web Chat in desktop browsers previously just stretched the mobile layout. Content remained in a narrow central column with large blank margins, and font size controls were hard to find in the second level of the “⋮” menu. Now: 1. When the window is wider than 768px, the message list and input expand fluidly up to 1200px, with user message bubbles capped at 72% of the content area; 2. Font size, theme mode, speech playback, and interface language are always accessible in one step from the header, with visibility controlled by conversation platform settings; 3. The “⋮” menu is hidden entirely if no items remain after moving those settings out, and any remaining drawer closes automatically when the window widens; 4. Loading skeletons match the width of the loaded content area, eliminating layout jumps. Mobile and embedded floating bubble layouts remain unchanged.
* **Collapsible cards in the MaiGPT right panel**: The “Knowledge Base” and “Outputs” sections on the right side of MaiGPT and conversation pages previously used horizontal tabs. They now appear as two vertically stacked cards: 1. At most one card can be expanded at a time. Click an expanded card to collapse it; collapsed cards still show a thumbnail and a one-line description. Newly generated files are indicated by a count badge instead of automatically opening the panel; 2. Controls in the expanded knowledge base card fit in a single row. The scope filter stays visible and defaults to “All knowledge bases,” and the “Selected N/M” count updates immediately with the filter scope; 3. The right panel and conversation area share a white canvas. The divider is transparent except for its handle, making card boundaries and titles easier to identify.
* **Next-step guidance for oversized conversation attachments**: Conversation attachments are limited to 10 MB. Uploading oversized contracts, medical records, official documents, or other files no longer leaves users stuck with only a “File size exceeds the maximum limit” message. Admin conversations and MaiGPT show a guidance dialog explaining the limit and offering a “Go to knowledge base” button (the knowledge base limit remains 100 MB). Members without knowledge base access are prompted to contact an administrator. Web Chat visitors see “Please contact an administrator to add the file to the knowledge base,” along with practical alternatives such as pasting excerpts or splitting the file into smaller files; audio files also receive compression suggestions. Size checks now run before reading the file, so even files of tens of MB receive an immediate prompt. Limit values always follow system settings.
* **Group database tables by source file**: When creating a MaiAgent database by uploading an Excel file with multiple worksheets, each worksheet previously occupied a separate row with the same file name, making them hard to distinguish. The table list now groups tables by source file in expandable groups displaying the file name, knowledge base, and table count. Each table also identifies its source worksheet. Tables created during upload processing appear directly under their source file.
* **Add the “Micro” plan name**: Organizations subscribing to the Micro plan previously saw a blank plan name or an English code on plan, billing, and organization information pages. All five interface languages now display the corresponding plan name, for example, the Traditional Chinese interface now displays the Micro plan name.
* **Automatic MaiDesk login from the Support Center entry**: Clicking “Support Center” in the top-right corner of the admin interface previously required another login to the support system using an Email verification code. With no account information passed along, all tickets were assigned to the demo organization. Clicking the entry now logs you into MaiDesk automatically using your current identity and passes your actual organization, so tickets belong to the correct company. If automatic login fails, you return to the original login page and can continue using the service.
* **Site-wide brand refresh (v1.4)**: Logos, wordmarks, favicons, and mobile installation icons in the admin interface and Web Chat now use the new brand assets: 1. The admin sidebar’s default width is 240px so the new horizontal wordmark is clearly readable, with the login page logo height adjusted accordingly; 2. Favicons use a vector format to stay sharp on high-resolution screens; 3. Web Chat installation icons for iOS and Android use a white background, and link-sharing preview images are also added. White-label customers’ own assets are unaffected.
* **Align advanced feature labels in the organization overview with product specifications**: The “Advanced Features” section of the organization overview previously had two incorrect names (“Support Center” was actually “Call Center,” and “Webhook System” was actually “Hook Engine”), a label that never worked, and disabled features with no indication of their plan. Label names and order now match the product specification sheet. Disabled features show “Enterprise” with a tooltip explaining how to enable them, and Beta features show Beta indicators on labels, sidebar menu items, and page titles. Sales staff can use this page directly to explain plan differences to customers.
* **Four-state overview for Manage My Connectors**: In MaiGPT’s “Manage My Connectors,” a single “Connected/Not connected” label previously left members unable to tell whether their organization had not enabled a connector or they had not authorized it personally. Expired credentials were also invisible. The redesign: 1. Uses a list on the left and details on the right, with four states for each connector—“Authorized,” “Pending authorization,” “Reauthorization required,” and “Not enabled”—and items requiring action listed first; 2. Supports search and state filters with counts for each state; 3. Previews requested access scopes before authorization, and shows the authorized account, authorization time, granted scopes, and credential expiration afterward; 4. Lets members directly request enablement for connectors their organization has not enabled, with immediate status updates after submission; 5. Requires confirmation before disconnecting to avoid accidental clicks that force users to repeat the entire authorization flow; 6. Falls back to the first letter of the connector’s name if its icon fails to load.
* **MCP app card widths adapt to the device**: Interactive app cards in Web Chat, such as display cards with sliders or multi-column statistics tables, were previously capped at 448 pixels wide on every device, crowding content on tablets and desktops. Cards now follow the same rules as Flex messages: their content determines their width, allowing them to expand fully on tablets and desktops. On phones, the message column still constrains their width, preserving the display.
* **Tiered parsing for LlamaParse**: Knowledge base files and conversation attachments parsed with LlamaParse now use the new tiered parsing mode, defaulting to an advanced tier with cost optimization. Simple pages are automatically processed at a lower tier. Results are returned as Markdown that preserves table structure, making it easier for question-answering retrieval to find the right passages in tabular documents. Operations staff can switch parsing tiers or pin a version as needed.
* **Vision parser reads embedded images at original resolution**: 1. When parsing PDFs containing operation screenshots, embedded images are sent to the model at their original resolution alongside the page image. Details such as menu paths and red-box highlights are no longer misinterpreted because of full-page downscaling. For the same customer document, menu-path accuracy improved from one correct result in four parses to four out of four; 2. The per-page output limit is increased, so dense pages are not silently cut off in the middle of a table, and a warning is recorded if truncation does occur; 3. Each parser can set its own chunk size and overlap, preventing global chunking rules from splitting highly structured results. Parsers without these settings retain their previous behavior.
* **Response traces show the duration of each model call**: When viewing response traces for MaiGPT or Agent-mode assistants in AgentOps, you previously saw only total model tokens for the entire turn, with no indication of the time spent on each model call. The timeline now lists each call’s start time, time to first token, end time, and token counts, making it easier to identify slow steps. The record’s model processing time is also the sum of all model calls within the same turn. For turns with multiple calls, such as tool loops, statistical charts therefore use different calculation bases for old and new records.
* **More consistent multilingual interfaces**: Billing, conversation, login, AgentOps, tool, identity synchronization, and other pages now include text in all five languages, with corrected Japanese and Korean translations and date displays.
* **Improved AI assistant marketplace browsing and previews**: The self-built assistants tab now uses the same card layout as the official marketplace, displays content according to permissions, avoids repeated background loading, and shows greetings in the selected language in previews.
* **Improved text display in pre-chat forms**: Field labels and descriptions now support line breaks and clickable links, keeping longer notices, privacy policies, and instructions clear and readable.
* **External participants in meeting records**: When creating or editing meeting records, you can enter external participants who are not on the organization member list to keep attendance information complete.
* **Consistent ordering of usage and billing details**: Credit usage, transaction records, and AgentOps details now appear in the same order and include image embedding usage, making it easier to verify costs item by item.
* **Credit quota reset cycle settings**: Credit quotas at the organization, plan, and member levels can now have separate reset cycles to accommodate different budgeting policies.
* **Knowledge base similarity threshold settings**: You can now adjust the similarity threshold when editing a knowledge base to control more precisely which search results are available to AI assistants as reference material.

### 🔐 Security and Governance Enhancements

* **Role-based permissions for the AI Gateway menu**: Previously, enabling AI Gateway for an organization made its menu visible to all members. Role permissions now include an “AI Gateway permission” checkbox. Roles without this permission cannot see the AI Gateway tab in the sidebar or access it directly by URL. Default roles in organizations already enabled at rollout have automatically received this permission. Administrators of organizations that enable it later must select it in role settings.
* **Input Hook checks for knowledge base and attachment content**: AI assistants with input Hooks configured (such as personal data masking) previously checked only user-entered text. Uploaded files and retrieved knowledge base passages were sent to the model without masking. In standard reply mode, attachments and knowledge base passages now pass through the same Hooks before reaching the model: content requiring masking is replaced with the masked version; blocked content is withheld from the model, while the assistant still answers using conversation history. Message records mark the retrieved passage as blocked and not sent to the model. No additional configuration is required, and the citation source panel remains unchanged.
* **AI assistant settings visibility and locking based on edit permissions**: The “AI assistant settings” button in the Agent marketplace and other entry points is now available according to the member’s edit permission for that assistant: 1. Members without edit permission cannot see the button, avoiding edits that are rejected only when saved; 2. Settings pages opened directly through bookmarks, shared links, or AgentOps initially load as read-only and recheck permissions before allowing edits; 3. For organizations whose plan does not include the assistant building module, the marketplace settings button, AI assistant menu, and Agent marketplace entry point are all hidden, and direct URL access redirects to an available page.
* **Agent marketplace shows when members lack access**: When an organization has enabled an official Agent but the current member is outside its allowed audience, the marketplace card remains in the “Enabled” section with a lock icon and a “No access permission” label, and the “Start conversation” button is disabled. Clicking the card or the button in the preview prompts the member to contact the organization administrator. Members can distinguish between an Agent that the organization has not enabled and one they have not been authorized to use, avoiding confusion with a malfunction.
* **Meeting record visibility consistently controlled by role permissions**: Meeting record visibility is now determined by role permissions for all organizations. The advanced features section of organization settings no longer shows the “Meeting privacy mode” badge. Selecting “Meeting record permission” in the role editor also selects “Meeting record management permission” by default (allowing access to all meetings). Administrators can deselect it to restrict visibility to meetings the member participated in. Saved permissions for existing roles remain unchanged.
* **Allowed embedding domain settings for Web Chat**: Previously, restricting which websites could embed Web Chat required engineers to adjust system administration settings. Administrators can now add or remove allowed embedding URLs below the custom domain field on the “Basic” tab of the conversation platform settings page. Input formats are checked immediately, and casing and trailing slashes are normalized automatically, preventing typos from causing browser blocks that are hard to diagnose. An empty list allows embedding on any website. Web Chat also reports the actual embedding page origin to the platform for comparison, and embedding from websites outside the list is rejected.
* **Separate permission to view all conversations when assigning conversation platforms to roles**: Previously, assigning a conversation platform to a MaiAgent role let its members see all conversation records for that platform on the “All conversations” page, including those of sensitive AI assistants such as HR assistants on the same platform. The “Assign conversation platforms” table in role settings now has a separate “View all conversations” checkbox column, including a header select-all control and explanatory tooltip. Administrators can authorize members to handle human handoffs and use the platform in the Agent marketplace without granting access to all conversations.
* **Web Chat embedding messages target the website origin**: When customers scanned their websites for security vulnerabilities with Fortify WebInspect, cross-window messages between the embedded Web Chat window and the page were flagged as an “Overly permissive messaging policy” because no recipient was specified, producing many findings. Both sources are now fixed: the attachment preview package no longer emits its self-detection message on page load, and the embedding code passes the embedding page origin when creating the window, so startup signals target the recipient from the first message. Testing found 5 messages without a specified target per load before the fix and 0 afterward. Legacy embedding without the parameter behaves as before.
* **Retiring o3-mini and switching to GPT-5.6**: The provider will soon stop offering o3-mini. AI assistants configured to use o3-mini and ongoing conversations still locked to that model will automatically receive replies from GPT-5.6 luna, with no manual changes needed. GPT-5.6 luna has lower input and output rates than o3-mini, so costs will decrease rather than increase after the switch. Calls through AI Gateway using the original o3-mini identifier continue to work, with responses generated by GPT-5.6 luna. The replacement no longer uses hidden reasoning, so spot-checking response quality is recommended for reasoning-intensive tasks.
* **Fixed Ask AI conversations remaining usable after logout**: Logging out of the admin console with the Ask AI sidebar open previously left the sidebar on the login page, where users could continue asking questions, and did not clear the identity in the embedded window. Logout now first closes the Ask AI window and clears the embedded identity. Web Chat simultaneously disconnects its real-time connection, clears the current conversation and history list, and returns to an uninitialized state, preventing further input on the login screen. After logging back in, Ask AI can reopen normally and receive the identity again.
* **Fixed Ask AI operations affecting the wrong organization**: For admin console users belonging to multiple organizations, Ask AI tool execution could previously affect an organization other than the current one. Creating a conversation and sending a message now include the user’s current organization identity, binding the conversation to the organization where it began. After switching organizations, Ask AI automatically switches to that organization’s most recent conversation (or creates one if none exists). The conversation history sidebar lists only that organization’s conversations, and the title bar shows the bound organization name without duplicate badges in centered-title mode. When the identity expires, the conversation automatically falls back to anonymous mode and continues without interruption.
* **Fixed AI referencing all data when tag filtering fails**: When organizations use tag filters to restrict AI references to specific knowledge base data, renaming or deleting a filter tag, or sending a query before tag data finishes loading, could make the filter empty and prevent it from properly limiting the reference scope. In MaiGPT, search tests, and contact query settings, unresolved tags are now shown in gray with a warning icon. Queries are blocked with a prompt, and users can remove invalid tags with one click before selecting replacements.
* **Hook inspection now covers tool returns in Agent mode**: For AI assistants with Hooks enabled, content returned to the model by tools such as attachments, knowledge bases, MCP, browsers, code execution, and Skills in Agent mode now passes through the existing input inspection rules without additional configuration. When blocked, only that tool result is withheld from the model; the conversation continues. Each inspection creates an execution record that includes the source tool, making it possible to trace which tool supplied sensitive content to the model.
* **Recording credentials expire immediately when a meeting ends**: Starting a meeting recording issues credentials for joining the meeting room. Previously, those credentials could still be used to rejoin after the meeting ended while they remained valid, potentially creating an empty room without recording or transcription services. The credentials are now revoked when a meeting is manually stopped or the system reports completion or failure, and reconnecting with the same credentials is rejected. A revocation failure does not affect meeting cleanup, recording file generation, or usage settlement.
* **New models automatically include billing rules**: Model catalog synchronization now creates credit billing rules for LLM, Embedding, and reranking models, with prices aligned to each provider's current list prices. Previously, newly listed models without dedicated rules fell back to generic provider rates or minimum rates, causing some advanced models to be billed at prices several times below their actual cost. Models still missing input or output prices after synchronization remain disabled and cannot be used without pricing; existing rules customized by administrators are preserved.
* **Consistent MaiGPT role permission messages**: Users without permission no longer see settings entry points they cannot use. Restricted actions display clear reasons, avoiding rejection only after users enter a feature.

### 🛠️ Bug fixes

* **Fixed cache and list-price billing rules for multiple models**: 1. Input/output rates for GPT-5.6 sol/terra/luna and Claude Sonnet 5 now match the current list prices of the actual supply channel; 2. The GPT-5.2 cache read rate is corrected to 0.3 times the input price, eliminating excess charges from the previously higher ratio; 3. Missing cache read/write rates have been added for 25 active platform models to prevent cached Tokens from being billed at base rates; 4. Cache rates for all platform models are standardized at 1.5 times the input price for writes and 0.3 times for reads across providers. Organization-provided (BYOK) models continue to use traffic-based billing and are unaffected.
* **Fixed the error message when vLLM tool calling is not enabled**: When running a self-hosted model with vLLM on premises, AI assistants with knowledge bases or tools failed on every response if automatic tool selection and a tool-call parser were not enabled at model service startup. Users only saw "The system is temporarily unable to process your request," which could lead them to assume the model lacked support and switch models. This situation now displays a clear message identifying the model service startup parameters that need adjustment, separate from the message for models that do not support tool calling, so administrators can address the cause directly.
* **Fixed blank model selections when an assistant's bound model is disabled**: After an organization disabled a model, AI assistants still bound to it showed a blank model selector on the edit page, making it appear that no model was configured. The selector now displays an unavailable "Model name (disabled)" option with a warning prompting users to select an available model. The warning and disabled option disappear immediately after a new model is selected.
* **Fixed disappearing assistant settings forms when a model becomes unavailable**: Three long-standing production errors are fixed together: 1. On the AI assistant settings page, the entire "Response mode settings" form disappeared when the selected model was removed, organization permissions changed, or the model list had not yet loaded. The full form now remains visible, with only output mode options temporarily hidden; 2. For users without AI assistant permissions, associated assistant labels triggered frontend errors when opening the knowledge base list. They now display names without clickable links; 3. Network timeout and disconnection errors are now reported as standard error objects, allowing the Keycloak login page to retrieve error messages correctly and error tracking to group them by message.
* **Fixed Deep Research failures with specific Claude models**: When the AI assistant model was set to Claude 4.8 Opus or Claude 5 Sonnet, Deep Research failed every time during planning because Bedrock rejected structured output. Planning now retrieves structured results through tool calling, restoring Deep Research for both models. Model calls during planning are also capped at 3; if the model repeatedly fails to produce a valid plan, an error is reported promptly rather than after thousands of retries.
* **Fixed Hook inspection failures caused by excessively long tool returns**: For AI assistants with content moderation or personal data detection Hooks configured, tool query results or retrieved content exceeding about 32,000 characters (such as large retrieval results or text attachments) caused inspection to fail as a whole and kept that content from the model, leaving responses incomplete. Long text is now automatically split for inspection, and the decisions are combined. Overlapping personal data detections at segment boundaries are also merged into complete ranges, allowing long content to pass inspection and be used in responses normally.
* **Fixed submission of personal preferences exceeding the limit**: When entering "Personal preferences" on the profile page or in the organization member edit dialog, the interface allowed up to 2,000 characters, but saving was rejected after submission. The input limit is now 1,000 characters with a character counter, and excess input is blocked while typing, preventing users from completing text they cannot save.
* **Fixed raw markup text in the Deep Research progress panel**: During Deep Research in a conversation, step details in the progress panel appeared as unreadable, unparsed markup. Progress now uses structured steps, with each tool call and result displayed using the same step cards as ordinary conversations. Previously generated conversation records are also automatically converted to a readable format.
* **Fixed Web Chat freezing on excessively long messages**: A conversation containing a single-line post of about 95,000 characters could freeze Web Chat and Ask AI for more than two minutes when opened or automatically restored, making all MaiAgent windows unusable. Messages with lines longer than 2,000 characters or total text longer than 50,000 characters now appear as collapsible plain text without full formatting parsing, allowing conversations to open normally.
* **Fixed usage chart legends overlapping the X-axis**: When viewing usage statistics in English, Japanese, or Korean, longer chart legend text wrapped and overlapped X-axis labels, making them unreadable. With cached data, the legend expanded to four items and the overlap became worse. All 12 charts across six tabs now automatically reserve bottom space based on the language and number of legend items, recalculating immediately when the language or window size changes. Titles such as "Member daily Tokens" also now use correct word spacing for each language.
* **Fixed database table management showing only the first 20 tables**: When an Excel file with multiple worksheets is uploaded to create a database, each worksheet creates a table. If there were more than 20 tables, the table management page listed only the first 20, hiding the rest and making users think files were lost. A page now displays up to 100 tables, with pagination appearing only above that limit. After uploads, creation of new tables is also checked against the total count returned by the backend.
* **Fixed later saves overwriting earlier changes when multiple users edit assistant settings**: When two tabs or two members opened the settings page for the same AI assistant, the later save submitted the entire old snapshot from when the page was opened, silently reverting fields just saved by the other party (such as a newly selected model). Before saving, the page now checks whether someone else has updated the assistant. If changes are detected, it prompts users to reload before saving, preventing overwrites.
* **Fixed AI responses referencing nonexistent knowledge base images**: When using knowledge base files containing images, AI responses sometimes referenced incorrectly copied or fabricated image links. Web Chat displayed red "Image failed to load" boxes at every step, and follow-up questions led the assistant to add another broken link. Before saving a response, each referenced knowledge base image is now checked for existence, and nonexistent images are removed entirely. This protection applies when reloading conversations and across channels such as LINE, Teams, and API; external and generated images are unaffected. A response being streamed may still briefly show a broken image, but refreshing displays the cleaned version.
* **Fixed HTML table responses appearing as plain text**: When a model output tables directly as HTML (usually with border, styling, or other attributes), security filtering in Web Chat and backend conversations degraded the entire block to plain text, leaving users with table markup instead of a table. Tables are now rebuilt using fixed templates and existing table styles, retaining only merged-cell information and discarding original attributes. They display correctly in dark mode and on mobile. Tables containing unsafe tags still fall back to text, preserving the existing security protections.
* **Fixed AI fabricating content from filenames when attachments cannot be read**: When an uploaded attachment has an unsupported format or corrupted content that prevents text extraction, AI assistants sometimes answered based on the filename, metadata, or imagined document structure, making users believe the content had been read. Assistants now explicitly inform users when attachment content is unavailable and are instructed not to infer content from filenames or expected structure. This applies to all attachment formats. As a prompt-level guardrail, it reduces but cannot completely eliminate the model's own guesses.
* **Fixed missing greetings on first entry into an assistant in admin chat**: When users first opened an AI assistant with a configured greeting in the admin internal chat view (without an existing conversation), the greeting had already been created but did not appear until they clicked “New conversation” or re-entered. This also affected the first visit after enabling an official Agent. The greeting now appears immediately on first entry, exactly once. Existing conversations and assistants without a greeting behave as before.
* **Fixed blank content across the site after leaving conversation record details**: After opening a conversation record's details in AI assistant monitoring, clicking any left-side menu turned the content area completely blank until the page was refreshed. Users can now navigate normally after leaving the details page. Automated checks have also been added to prevent similar issues.
* **Fixed test case generation from knowledge base files getting stuck in processing**: When a model call stalled during automatic test case generation from a knowledge base file, the job stayed at “Processing, progress 0” indefinitely (still running after 11 minutes in testing). Users could not tell whether to wait or give up, and had no retry option. Generation now has a 60-minute time limit. On timeout, the job is marked as failed with a readable message and the UI is notified immediately, allowing users to run it again.
* **Fixed misaligned confirmation popover icons and titles**: Warning icons in admin confirmation popovers (such as delete confirmations) sat about 3 pixels below the title text. When opened inside a clipping container, the popover width also shrank pixel by pixel for 1 to 2 seconds. Icons now align with the visual center of the title's first line (within 0.1 pixels), including when the title wraps. The popover width is fixed as soon as it opens and no longer creeps.
* **Fixed the legacy brand appearing as the default sender of system notification emails**: Deployments without a custom sender in email settings or an environment override showed the previous brand name as the default sender and subject prefix for system notification emails. Defaults now use the current MaiAgent brand and domain. Customers with custom senders are unaffected.
* **Fixed stale generated file lists and incorrect download filenames**: 1. After asking AI to generate files in MaiGPT, the generated file panel and tab count now reflect files that are currently downloadable, excluding expired files from earlier turns (for example, showing 3 files when only 1 could be downloaded); 2. Clicking an expired file displays “This file has expired. Please generate it again before downloading” and automatically refreshes the list instead of failing silently; 3. Downloads from the preview panel retain the original filename and extension instead of becoming a string of random characters.
* **Fixed syntax error graphics appearing while Canvas flowcharts are being generated**: When asking AI to generate a flowchart with Canvas in Web Chat or admin chat, streaming no longer flashes a bomb graphic and “Syntax error” message or leaves an error graphic permanently on screen until users switch to source and back or reopen the conversation. During generation, Canvas shows only a processing state and temporarily disables downloading. It renders once the content is complete.
* **Fixed incorrect line breaks in Canvas flowchart text**: When AI-generated flowchart or sequence diagram node text contains line-break markers, Web Chat Canvas no longer prints those markers as two literal characters. Admin and MaiGPT Canvas also no longer fail to display an entire sequence diagram when its notes contain line breaks. Both now break lines correctly. Web Chat Canvas also waits for content to settle before redrawing, preventing repeated flickering during streaming.
* **Fixed generated image preview panels failing to reopen**: With AI assistants that support code execution, clicking a generated image opened a preview sidebar, but after closing the sidebar, collapsing the right column, or switching conversations, clicking an image again could do nothing. Users had to manually expand the right column and click the preview's close button to recover. Every click now opens the preview normally. Previews close automatically when switching conversations or leaving the page, and files on the admin conversation record page can be previewed directly. Both Web Chat and admin are fixed.
* **Fixed parsing failures for presentation uploads to the knowledge base**: Uploading a PPTX without slide dimensions (omitted by some presentation export tools) previously failed with a “dimension validation error,” preventing indexing. The system now supplies default 16:9 dimensions before parsing, allowing presentation text to be extracted normally without affecting the original file.
* **Fixed Canvas web pages appearing as source code in shared conversations**: When a MaiGPT conversation contained an AI-generated Canvas web page, sharing it with other members previously displayed the Canvas content as a block of web page source code. Shared views now render Canvas in the same safely isolated environment as the original conversation. Text before and after Canvas retains its formatting, multiple canvases in one message each get their own card, and existing share links are also fixed.
* **Fixed Canvas charts disappearing after switching to source code**: When AI generated a web page containing charts with Canvas, switching to source code and back to preview or rerendering during generation could leave chart areas as fixed-height blank spaces, while surrounding headings, tables, and statistics cards remained normal. Every render now reloads in a fresh environment so charts display reliably.
* **Fixed selected participants not receiving view access when a meeting is created**: When creating a live transcription meeting, the “Participants” field now supports selecting multiple organization members. Selected members are automatically added as meeting members with view access after creation and appear correctly in the participants fields on list and detail pages. The creator automatically becomes the host without needing to add themselves. Member synchronization failures show a notice without interrupting recording.
* **Fixed incorrect organization labels when opening Web Chat directly**: Ask AI embeds and public Web Chat share the same URL origin. Browsers copy cached caller data when navigating in the same tab or opening a new tab from an embedded page, causing unrelated organization names to appear in the title bar when public Web Chat is opened directly and outgoing messages to be tagged as an administrator. Web Chat now clears this leftover data when opened as a top-level page. The title bar no longer shows an incorrect organization label, and public conversations retain guest identity. Embedded Ask AI is unaffected.
* **Fixed MaiGPT embeds taking over the entire page when their container is missing**: When embedding the chat interface in a specified section of a website in MaiGPT mode, a typo in the container selector or a single-page app container that had not loaded previously caused a fallback to a full-screen window with no close button, covering the entire host page. The embed now waits up to 5 seconds for the container, accommodating containers that load after the embed script. On timeout or an invalid configuration type, nothing is rendered and the browser console logs an error with instructions for fixing it. Full-screen layouts that explicitly specify the entire page as the container are unaffected.
* **Fixed ignored position settings for embedded chat windows**: Embedded Web Chat floating windows and buttons on customer websites had three positioning issues: 1. The first opening positioned the window relative to the button instead of using configured values, pushing it to the top-left if the site hid the button with styles, and immediately saving those incorrect coordinates; 2. Resizing the browser permanently saved the button position over site settings, and once users dragged the window, later site embed configuration changes stopped taking effect; 3. Positions and dimensions specified in relative font-size units were interpreted as pixels. Windows now open at the configured site position, and adjustments are saved only after user actions such as dragging or closing. Changes to site position or size settings take precedence and reset previous adjustments once. Relative font-size units are converted using the host page's root font size; unsupported units trigger a browser console notice and fall back to defaults.
* **Fixed replies getting permanently stuck after returning to Web Chat on mobile**: Mobile browsers interrupt the real-time connection after Web Chat stays in the background for a while. On returning to the foreground, the AI typing indicator could remain forever and lock the input, even though administrators had already seen the full reply. After reconnecting successfully, returning to the foreground, or a reply stalling for more than 45 seconds, Web Chat now fetches missed messages from the server and merges them with local history, clears the leftover typing indicator, and unlocks input. Reconnection also uses the latest login credentials so it can recover automatically after credentials expire.
* **Fixed disproportionate brand logo sizing in Web Chat**: Brand logos appeared out of proportion in three places in Web Chat: 1. The logo beside “Powered by” at the bottom of the conversation page filled the entire row height and did not match the 12px explanatory text. It now scales proportionally within an 80×32 frame; 2. The logo container on the login, registration, and password reset pages has been reduced from 96px to 48px, with proportional scaling added to the registration and reset pages; 3. The splash screen previously used a single 80px bitmap that became blurry when enlarged on high-resolution screens. It now uses a vertical combination of a vector logo and wordmark, while white-label customers retain their own logos. Also fixed conversation font size settings affecting login page headings and removed the overlapping “Text” label on the voice playback icon in the title bar.
* **Fixed tool cards not returning after brand animation mode**: The conversation platform's tool display mode (full/compact/hidden) is a server-side setting, but Web Chat previously saved “hidden” to browser storage during brand animation mode and retained it permanently. Once a visitor had used brand animation, tool cards did not return after switching back to narrative text mode, even after closing and reopening the browser. Changing the setting to “full” in the admin console also had no effect for that visitor. Tool display mode now applies only to the current load and always follows the admin setting, so tool cards return when switching back to narrative text mode. Tool cards remain hidden in brand animation mode, and the four-stage animation is unaffected.
* **Fixed duplicate disconnect errors during Web Chat cleanup**: When closing a conversation, logging out, or switching modes, multiple processes request that the real-time connection disconnect. A second request used to throw an error, skipping subsequent cleanup such as clearing conversation state and embedded login information. The expected “connection closed” response when unmounting MCP app cards was also exposed as an unhandled error. Duplicate disconnect requests are now safely ignored. Expected responses when unmounting MCP app cards are handled silently, while other unmount failures are still logged normally.
* **Fixed the previous conversation platform name briefly appearing when entering a new conversation**: When clicking “Start Conversation” in the official marketplace or switching to a conversation on another platform, the top selector briefly showed the previous platform's name before displaying the correct name after loading. The name is now hidden during the switch, and the new platform's name appears directly once loading completes. Refreshing the same conversation platform behaves as before.
* **Fixed unbilled image search query usage being combined with text embedding usage**: When an AI assistant had image search enabled, image embedding usage for each query was previously neither recorded nor billed and was combined with text embedding usage in the same field. Image embedding usage is now billed at the rate for the model used and included in usage statistics. Conversation records and exported details list image embedding tokens separately, while the original embedding field counts only text. Retrieval steps in the trace view display text and image usage together. This affects only AI assistants with image search enabled, and billed amounts will change accordingly. The image field shows 0 for historical records.
* **Fixed incorrect navigation from conversation platform gear icons on the team edit page**: Gear icons in the “Linked Conversation Platforms” list on the left of the team edit page previously led to incorrect destinations depending on channel type: internal channels opened the conversation page, web channels opened the public Web Chat URL in a new tab, and other channels such as API and LINE did nothing. Gear icons for all three channel categories now navigate to the platform's settings page without opening a new tab. The message icon still opens conversations as before.
* **Fixed broken image previews after uploads in conversation platform appearance settings**: After uploading an avatar or technical support logo on the conversation platform appearance settings page, the live preview on the right showed a broken image until the settings were saved. The upload list and preview window title also displayed the media URL's entire signature parameter string as the filename. Uploaded images now appear immediately in the preview using the local image, while data submitted to the backend remains unchanged. Filenames show only the actual filename. After saving or refreshing, the preview uses the URL returned by the backend.
* **Fixed embedded Google Maps being displayed as text**: When an AI assistant's response contained an embedded Google Map, Web Chat printed the entire source code as text, leaving users unable to see the map. It now allows only the place, search, directions, view, and streetview modes from Google's official Maps embed endpoints, rebuilding the embedded element from a fixed template and ignoring all other attributes from the original content. Content that does not meet these conditions still falls back to text. Missing closing tags are added automatically to avoid consuming subsequent messages.
* **Fixed suggested questions never appearing in embedded mode**: When customer websites provided Web Chat as an embed, the opening suggested questions loaded but never appeared. The embedded window initially collapsed to zero height, and the system measured content height only once when the welcome message appeared. After measuring zero, it never retried, leaving suggested questions hidden. It now continuously observes container size changes and measures again when the window expands, displaying the suggested questions. Standalone pages, cases without a welcome message, and hiding suggestions after a question is asked behave as before.
* **Fixed false warnings when embed settings contain login information**: When a website loaded embedded Web Chat with login information in its settings, the browser console reported an “unknown configuration option” warning even though the setting was valid and login worked normally. Login information is now recognized as a valid setting, eliminating the false warning.
* **Fixed Web Chat crashing immediately on launch in some browsers**: In some embedded browsers, privacy browsers, or pages where extensions modified the network information object, Web Chat crashed completely on load because its network status detection was incompatible. It now determines connection status by listening only to the browser's own online/offline events. The conditions and appearance of the offline notification remain unchanged.
* **Fixed the conversation platform type column showing “-” in usage statistics**: In the conversation platform table on the organization's usage statistics page, the type column usually displayed only “-” because the frontend did not recognize web channel subtypes (internal, API, web, Desktop). The column now matches the conversation platform quota page and correctly displays each channel's icon and name. “Desktop” has also been added to channel labels throughout the site, and Desktop channels previously labeled “Web” now display “Desktop”.
* **Fixed member search failing in organizations with more than 100 members**: In organizations with more than 100 members, the 101st member onward could not be found in any member picker regardless of the search text, including API key creation, Credit quota allocation, meeting and transcript participants, conversation tags, customer service copy lists, and MaiGPT sharing. These pickers now use server-side search and load results in batches while scrolling, allow failed loads to be retried in place, and clearly label deactivated members. Organizations with 100 or fewer members retain immediate local filtering.
* **Fixed data import failures for files containing yes/no fields**: When uploading CSV or Excel files to create a MaiAgent database, boolean fields such as true/false were incorrectly identified as monetary fields, causing table creation to fail for the entire file rather than just the affected column. Boolean fields now retain their original type, allowing files containing them to import normally.
* **Fixed files with no parsed content being marked as completed**: When corrupted, blank, or password-protected documents were uploaded to the knowledge base, files that produced no content after parsing were still marked as “Completed”. Users saw no sign of a problem, but the AI could not answer anything about the files. These files are now marked as failed with an explanation in the failure reason, immediately letting users know they need to upload them again.
* **Fixed knowledge base citation page numbers always showing page 1**: Knowledge base citation cards in the admin console and Web Chat previously always showed “1” as the page number, so users could not find the cited content in the original document. They now show the actual starting page, with page ranges for citations spanning multiple pages (for example, 1–2). Sources without pagination, such as web pages, Word documents, structured data, or speech transcripts, display “Source is not paginated” instead of a false page number.
* **Fixed timeouts when downloading knowledge base PDF translation results**: When downloading PDF translation results from the knowledge base (bilingual, translated, or original versions), downloads often timed out after 20 seconds even for files smaller than 1MB. Files now download directly from file storage without passing through the server, and retain recognizable filenames after downloading.
* **Fixed inability to manage more than 20 knowledge base citation fields**: When a knowledge base had more than 20 citation field definitions: 1. The management page had no pagination, making the 21st field onward impossible to edit, toggle, or delete, with no indication that they existed; 2. The file information window incorrectly labeled defined fields as “Removed from definitions”; 3. Search tests lost some fields' ordering, hints, and tags. The management page now provides proper pagination, while the file window and search tests fetch the complete field definitions.
* **Fixed the “No permission” notification appearing on the Credit page**: Members with Credit management permission but without billing management permission saw a “No permission” error when opening the Credit usage page, even though the page content worked normally. The page no longer requests the billing overview, eliminating the notification.
* **Fixed permission error popups in human handoff settings**: Roles with only customer service conversation permissions encountered a permission error when opening human handoff settings in the conversation platform. If disabled members remained in the assignment or notification lists, the entire configuration was rejected on save, requiring users to find and remove them manually. After the fix: 1. Human handoff settings, conversation tags, transcript speakers, and MaiGPT sharing recipients use a compact member directory accessible to any member; 2. Disabled members are clearly marked before saving and can be removed; 3. After saving, the system automatically removes disabled members and updates the interface.
* **Fixed accidental deletion of SSO settings when authentication settings fail to load**: When the organization authentication settings page failed to load the current configuration, for example due to a brief network disruption, the interface reverted to the default appearance of platform username and password login. If an administrator clicked Save at that point, the configured third-party login (SSO) and its credentials were permanently deleted. After the fix, a loading failure replaces the entire settings section with an error message and a Refresh button. Saving is blocked until the current settings have loaded successfully.
* **Fixed identical error messages for all login failures**: Users with disabled accounts or unverified Email addresses always saw “Unable to log in with the provided credentials,” leaving both users and support staff unable to identify the real cause. After the fix, the actual reason returned by the server, such as a disabled account or unverified Email, is displayed. The generic message is used only when no specific reason is available.
* **Fixed inability to configure MCP endpoints with TWCA certificates**: MCP endpoints whose certificates were issued by legacy root certificate authorities such as TWCA, common in Taiwanese healthcare, finance, and government organizations, worked when calling tools in conversations but consistently failed when fetching tools, saving MCP tools, or testing private MCP connections in the admin interface. These three configuration entry points now apply the same certificate relaxation checks used during conversation execution, based on the domain allowlist or tool-level settings. Allowlisted endpoints can be configured successfully, while others retain strict verification. MCP URL format validation has also been restricted to http/https.
* **Fixed potentially indefinite waits when generating MCP app cards**: When an AI assistant needed to retrieve an app card from an MCP server during a reply, an unresponsive server could previously block the entire automated reply indefinitely because that path had no timeout. It now follows the other paths and aborts after 30 seconds, reporting a standard tool failure so users no longer wait without an end.
* **Fixed app card height collapsing after leaving fullscreen**: In Web Chat, expanding an MCP app card, such as a high-speed rail timetable, to fullscreen and then collapsing it reset the card to the default height of 220 pixels. After the title and filter bar, no timetable rows were visible. The card now retains the height reported by the app after collapsing. For example, a timetable with 84 departures remains about 515 pixels high and shows at least 5 rows. Its height also adjusts automatically when switching time ranges, and the brief flicker during collapse no longer occurs.
* **Fixed connector management dialogs navigating the background page away**: Opening “Manage My Connectors” from the input bar menu in a MaiGPT conversation switched the background to the old connector page, leaving the current conversation. The dialog now opens directly on the current conversation page, and the input bar immediately updates connector authorization status when it closes. Also fixed a case where authorization completed in the dialog but the status remained “Pending Authorization” if the authorization window did not return a result; status now refreshes when that window closes.
* **Fixed inability to submit platform-level tool cards**: When users received an interactive card generated by a platform-level MCP tool, such as an official feedback card, clicking “Submit” always failed and displayed only a request failure message inside the card. The submission omitted the tool source identifier. Cards from platform tools and connectors now correctly include their respective sources, allowing feedback to be submitted and the card to switch to its submitted state. Connector card behavior remains unchanged.
* **Fixed blank names in the connector selection dialog**: When creating or editing an AI assistant, connectors without custom names, such as Microsoft 365, appeared in the “Select Connectors” dialog with only an icon and enabled status, making them impossible to identify. After the fix: 1. Connectors without custom names show their connector type names and can be searched by those names; 2. Failed icon loads show a placeholder, and failed list loads show an error with a retry option; 3. The search bar and selection count stay fixed at the top while only the list scrolls; 4. The list of connectors already linked to the assistant also displays icons.
* **Fixed inaccurate credit billing for document parsing**: 1. LlamaParse document parsing is now billed in credits based on the pages actually billed by the provider. Each reparse of the same document is billed, while provider cache hits that incur no charge are not billed to customers; 2. If statistics become available after knowledge base file parsing finishes, billing is automatically applied once the data is ready. Parsing charges for these files could previously be omitted; 3. LlamaParse parsing unit prices now use the actual cost and the same standard pricing formula as model fees.
* **Fixed image search tools failing to find images**: When the image search tool in a multimodal knowledge base was configured to return only images, text chunks filled the retrieval slots, consistently excluding images and producing “No results found.” Retrieval now selects only image nodes and returns the number of results configured for the tool. Both text-to-image and image-to-image searches can return images normally; text-only and mixed modes retain their existing behavior.
* **Fixed incorrect AI Gateway error status codes**: When an upstream provider rejected an AI Gateway model request because of problems in the request itself, such as invalid tool definitions or too many tools, the gateway previously always returned a 502 server error, leaving callers unaware that they needed to fix their requests. These cases now return 400 with guidance to check model parameters, tool definitions, and input size. Provider failures, timeouts, key issues, and rate limiting continue to return 502.
* **Fixed inability to save channel partner organization data**: Distributor and agent organizations configured to bill themselves previously could not save organization data in the admin interface or in workflows such as trial invitations, authentication source settings, and deactivation scheduling because the system treated a self-referencing upstream as an error. Self-billed channel partners with matching roles can now save successfully and are labeled “Self-billed” in the admin interface. The message shown when channel partners attempt to buy credits themselves is now consistently “Please contact your account manager to purchase.”
* **Fixed missing Gemini cache read billing rules**: After Gemini models were added on Vertex AI, cache read pricing rules were not created automatically, resulting in undercharging at the minimum price floor. Saving a model now automatically creates cache read rules using provider ratios. Cache status now displays “Automatic Caching (Provider-managed).” Accidentally enabling caching for these models shows an explanation rather than a full-page error, and manually entered cache prices for OpenAI and Azure models are also saved correctly.
* **Fixed missing PDF page numbers and misidentified decorative images**: When processing PDFs with the default parser, a page with no serializable text cleared chunk page numbers throughout the adjacent section, preventing citations from showing page numbers. Repeated footer logos were also treated as different images due to slight variations in cropping between pages, so decorative image filtering never took effect. Empty pages now retain their own page numbers, and images are compared using embedded PDF image objects, allowing the same logo to be recognized and filtered across pages.
* **Fixed parser toggles being reset during on-premises upgrades**: Operators in licensed on-premises environments who disabled cloud parsers such as Whisper or Azure Speech previously had to disable them again after every version upgrade because the system re-enabled them. Parser activation is now controlled by operators, just like the model catalog: new parsers added during upgrades are disabled by default, existing toggles remain unchanged, and items removed from the registry are reported without taking action. Parser toggles in SaaS environments are unaffected.
* **Fixed conversation content and tool status display issues**: AI-generated tables, bold text, and tool execution descriptions now display in full. After switching languages, descriptions for tools currently running immediately use the new language.
* **Fixed Web Chat failing to open at custom deployment paths**: When deployed under a website subpath through a reverse proxy, login, embedded pages, and related resources now automatically use the correct path.
* **Fixed file and meeting connection operation issues**: Presentation files can now be previewed directly. After a meeting connection is interrupted, the system correctly leaves the old room and recognizes recreated rooms, avoiding reconnection to an invalid session.
* **Fixed skipped questions still triggering AI replies**: Closing a skippable question now only updates the interface state, without sending extra messages or starting new AI processing. The layout for multiple questions is also more compact.
* **Fixed missing responses after sending messages**: If the real-time connection does not return an acknowledgment, a fallback delivery method is used to reduce cases where a message appears to have been sent but the conversation does not update.
* **Fixed incorrect knowledge base and file list states**: The summary card is no longer displayed when AI summaries are disabled; rejected uploads are immediately removed from the file list.

## v2026.09.04 <a href="#v2026-09-04" id="v2026-09-04"></a>

***

### ✨ New Features

* **New Work workspace (Coming Soon)**: This feature is not yet available in production and will be announced when it launches. Planned capabilities: The new Work workspace brings assistant switching, conversation history across assistants, and frequently used entry points into a single three-zone interface. The system remembers the last workspace and layout used and automatically redirects users to an available feature when they encounter a page they cannot access.
* **Role workspace permission preview (Coming Soon)**: This feature will launch together with the Work workspace. It is not yet available in production and will be announced when it launches. Planned capabilities: The MaiAgent role editor now includes workspace permission cards and a live preview. Administrators can check which feature zones a role can access before saving, reducing cases where incorrect permissions are discovered only afterward.
* **File library integration with AI workflows**: The file library can now display AI analysis status, mount knowledge bases, and bring files into the MaiGPT workspace. Parsed attachments are added directly to the work context, while long documents provide the most relevant content within the available context budget.
* **Multilingual content for built-in Agents**: The 98 built-in Agents in the Agent marketplace now display their names, descriptions, and operational content in English or Japanese according to the interface language, without mixing languages after a language switch.
* **Automatic image generation for AI Slides**: AI Slides can now generate illustrations during slide creation and insert them directly into slides, without requiring users to download and upload them again.
* **Skippable question cards**: When an AI assistant asks multiple-choice questions, optional questions can offer a Skip action, so users do not have to choose an unsuitable answer just to continue the conversation.
* **Deep Research file downloads**: Once Deep Research is complete, users can obtain downloadable report files directly, making it easier to save, share, or continue editing research results.

### 🚀 Core Performance Improvements

* **More stable test dataset evaluations**: Large test dataset evaluations now distribute work on demand across multiple workers. Interrupted tasks are automatically reclaimed and retried, and completed cases save their actual usage, so long-running evaluations no longer stall because a single task is interrupted.
* **Faster visual parsing of large documents**: Long documents containing images or complex layouts are now parsed in concurrent batches. When the service is busy, processing automatically slows down and continues, reducing wait times and failures of entire documents.
* **Faster organization usage page loading**: Organization usage and quota summaries are now aggregated more efficiently, giving administrators complete statistics more quickly when they open the usage page.
* **Faster member roster loading**: When loading the complete member roster for a large organization, the system now prepares the data needed by the interface directly, reducing delays caused by processing many members individually.
* **Faster chat platform knowledge summaries**: Chat platform knowledge summaries and data for selecting all items now load faster with large numbers of documents, reducing the wait when configuring knowledge scope.
* **Faster SCIM group synchronization**: When enterprise identity directories synchronize group members, the system now processes only the members added or removed in the current update, making large group updates faster and reducing load during synchronization.
* **Faster LINE message reception**: When a LINE chat platform receives consecutive messages, it reuses user data that is still valid, reducing delays from repeated LINE queries and moving messages into processing more quickly.

### 😊 User Experience and Interface Improvements

* **Aligned login page layout**: The loading indicator on the enterprise login page is now centered, and the fields in the account identification and password steps use a consistent layout, making the interface more stable when switching between login steps.
* **Workflow reply mode retirement**: AI assistants no longer offer the option to create a new Workflow reply mode; duplicating an older assistant now switches it to standard Q\&A. Existing assistants that have not been used for a long time automatically revert to standard Q\&A, while recently active assistants retain their existing behavior.

### 🛠️ Bug Fixes

* **Fixed system event display in activity logs**: System-generated events no longer cause columns to shift in activity logs. Even when an event has no acting user, its date, action, and content remain aligned with the correct columns.
* **Fixed AI monitoring log loading**: The AI monitoring log page no longer occasionally fails to load. Data is fetched after the table is ready, preventing blank pages caused by initialization order.
* **More stable MCP App shutdown**: Closing an MCP App card or switching conversations no longer occasionally causes errors. Closed connections are cleaned up properly and no longer interfere with subsequent conversations.
* **MCP App tool timeout messages**: MCP App tools no longer remain unresponsive for extended periods. If a third-party service does not respond within 30 seconds, the system stops waiting and displays an execution failure, allowing users to retry immediately without waiting several minutes.
* **Fixed Web Chat deployment under subpaths**: Web Chat pages now open correctly when deployed under a reverse proxy subpath. The system determines the path from the actual URL, so refreshes and page navigation work correctly.
* **Fixed Skill export and reimport**: Skills can now be reimported successfully after export without losing settings. Reimport preserves when-to-use guidance and bound tools, and names automatically conform to the importable format.
* **Fixed Ask AI organization scope**: Ask AI no longer retains the previous organization's data scope after switching organizations. Tools now consistently run within the organization associated with the current conversation, preventing unrelated results from other organizations.
* **Usage statistics sorting fix**: Fixed an issue where sorting usage statistics by credits only sorted the current page. The system now sorts the complete results before pagination, keeping the order correct across pages.
* **Knowledge base search status messages**: Improved messages when knowledge base searches return no results. If relevant documents are still processing, the AI assistant identifies files that are not ready, including content mounted in the workspace, so users do not mistakenly assume the knowledge base is empty.
* **Error messages for models without tool support**: Improved error messages when a model does not support tool operations. The interface now clearly explains that the current model cannot complete the tool task and guides users to a supported model, instead of only displaying “Please try again later.”
* **Automatic fallback for image generation services**: Improved image generation reliability. If the primary service temporarily fails, the system automatically switches to a fallback service. Credit deduction failures are also reported clearly, instead of interrupting the process with an unexplained error.
* **Schedule execution limit messages**: Fixed repeated pausing after a schedule reaches its execution limit. The system now clearly indicates that all allowed executions have been used, helping users adjust their schedule settings.
* **Output file downloads in anonymous Web Chat**: Fixed an issue preventing anonymous Web Chat visitors from opening files and images generated by the code interpreter. Visitors can now view or download results directly in the original conversation without being redirected to login.
* **Monitoring timeline step fix**: Fixed missing model processing steps in the AgentOps trace timeline. Administrators can once again see each model execution step, making it easier to identify response quality and latency issues.
* **Long question fix for knowledge retrieval**: Fixed knowledge retrieval failures caused by long questions. The system preserves usable question content within service limits, allowing queries with extensive pasted background information to complete.
* **Meeting participant access fix**: Fixed incomplete participant permissions when creating meeting records from a calendar. The system now identifies internal and external participants from the invitation list, allowing everyone entitled to access to view the meeting content.

## v2026.09.02 <a href="#v2026-09-02" id="v2026-09-02"></a>

***

### ✨ New features

* **Backup recording prevents audio loss in live meetings**: Live meeting records now support local backup recording. Audio is saved in the browser before uploading, so even if the connection drops or the page is accidentally closed, recorded content uploads automatically after reconnecting. On the meeting details page, users can switch between original and backup audio and retranscribe using the backup. The list also marks meetings whose backups are syncing or incomplete.
* **Self-service management of “Connected services”**: Users can now manage third-party service authorizations themselves on both Web Chat and built-in AI assistant conversation pages. Click “Connected services” at the top of the page to view authorization status and expiration dates for services such as calendars and cloud drives, and directly authorize, reauthorize, or disconnect them (with confirmation before disconnecting). Authorization requests in conversations now appear as clear authorization cards instead of plain English URLs. This also works on mobile.
* **Automatic SSO routing on the login page**: The login page now determines the login method from the email you enter. Enterprise SSO users are directed to their company's single sign-on after entering their work email, without needing to remember a special URL; other users enter their password as usual. SSO authentication failures now show a clear error and offer retry or password login options, avoiding redirect loops.
* **Webhook channel for human handoff notifications**: Human handoff notification settings now support a Webhook channel. Enter your system's HTTPS endpoint (and optionally a signing secret so the receiver can verify the source). When a conversation handoff is assigned, the platform immediately pushes an event notification to that URL, making it easy to integrate the handoff workflow with external systems such as ticketing or on-call alerts. Failed deliveries retry automatically without affecting existing Email or notification center notifications.
* **Hierarchical channel type filtering on the contacts page**: The contacts page now offers two levels of filtering: select a channel type (web, LINE, Telegram, Facebook, etc.), then a conversation platform of that type to narrow the list instantly. The platform dropdown also shows channel labels, making it faster to find contacts across platforms.
* **Release quota restrictions when deactivating credit quotas in bulk**: When deactivating credit quotas in bulk by MaiAgent role, you can now also remove members' quota restrictions and switch them to the organization wallet, avoiding a state where members are restricted but have no quota. The confirmation dialog also uses actual statistics to show accurate counts of members whose quotas will be deactivated and whose restrictions will be removed.

### 🚀 Core performance improvements

* **Smoother navigation on the built-in Agent management page**: Switching categories or searching on the built-in Agent management page no longer repeatedly queries the audience for each Agent. Organizations with many enabled Agents can switch tabs noticeably faster, without the whole page flashing a loading screen.
* **Faster opening of conversations with large tool outputs**: Opening old conversations with large amounts of tool output no longer freezes the page, and memory use while typing has dropped significantly (measured from 255MB to 158MB). Tool step details now load only when expanded, and very long results display an initial preview marked as truncated.
* **Faster loading of knowledge base full-document views**: Full-document views in the knowledge base now assemble large documents (hundreds of paragraphs) faster. Paragraph mappings use a more efficient algorithm, with exactly the same displayed content.
* **Improved platform stability during system updates**: Server load spikes caused by large numbers of Web Chat reconnections during system deployments are now spread out substantially, reducing platform-wide slowdowns during updates. Users waiting for responses are unaffected and still receive messages immediately.

### 😊 User experience and interface improvements

* **Comprehensive improvements to conversation scrolling and reading**: Conversation scrolling in Web Chat, administration console chat, and MaiGPT has been improved throughout:
  1. Scroll up freely to revisit earlier messages while AI responses are being generated, without being forced back to the latest position;
  2. The paragraph you are reading no longer drifts upward as a response is generated, and your reading position is preserved after loading more historical messages;
  3. All conversation interfaces provide a consistent “Back to latest” button; click it to return to the newest content and resume automatic following;
  4. The opening welcome message is no longer pushed off-screen.
* **Redesigned navigation on conversation platform settings**: Conversation platform settings now use the same layout as AI assistant settings. Settings categories are grouped in the left sidebar, while basic settings, quotas, and automatic replies appear as tabs at the top. Category changes are recorded in the URL, so refreshing or sharing the link returns to the same category.
* **MaiAgent role permission changes take effect immediately**: When editing resource permissions for a MaiAgent role, selecting any permission now takes effect immediately and displays a result notification, without requiring “Update role” on each tab. Changes made across tabs are no longer partially lost.
* **Separate cache items in credit usage breakdowns**: Conversation credit usage summaries now list “LLM Cache Read” and “LLM Cache Creation” as separate items instead of grouping them under “Other.” This makes cache costs visible when using models with automatic caching, such as Gemini. Category names also wrap in narrow panels instead of being truncated.
* **Delete presentation and document templates**: When using presentation or document templates in conversations, you can now delete templates uploaded by your organization directly. Template cards include a delete button with confirmation before deletion, so accidentally uploaded templates no longer require an engineer to remove them.
* **Agent marketplace shows required connectors and setup notices**: Cards and preview windows in the Agent marketplace and built-in AI assistant management page now show icons for the third-party service connectors each assistant requires. Assistants that require an administrator to configure connections first display “Administrator connection setup required,” preventing the impression that they are ready to use out of the box.
* **Built-in Agent usage renamed to “Activation count”**: Built-in AI assistant usage descriptions on the organization management page now use “Activation count” instead of “Seats.” This clearly describes the maximum number of AI assistants that can be enabled simultaneously and avoids confusion with user licenses. Traditional Chinese, Simplified Chinese, English, Japanese, and Korean have all been updated.
* **Fully formatted voice conversation responses**: In Web Chat voice conversation mode, AI assistant text responses now display with full formatting. Lists, headings, and blockquotes render correctly and align to the left, without raw formatting symbols or misplaced centered content.
* **Consistent sidebar brand logo sizing**: Brand logos in the admin sidebar now use a fixed size and centered positioning. Logos display more consistently across brands and retain their proportions in both expanded and collapsed states.

### 🔐 Security and governance enhancements

* **Embed security hardening and integrity verification**: Customers embedding Web Chat on their websites can now enable stricter Content Security Policies (CSP) without affecting the appearance of the chat button or window. Version-pinned embed files and integrity verification information are also available to meet security scanning and audit requirements. The chat button's hover enlargement effect now works as well.
* **Restrict access to built-in Agent management**: Access to built-in Agent management now requires explicit authorization through the “Built-in Agent Management” permission under organization permissions. Unauthorized members no longer see the management entry or the marketplace's “Go to Management” link. Marketplace browsing remains available, making permission management more precise.
* **Permission checks for applying marketplace Agents**: The “Use this Agent” button on Agent marketplace detail pages now reflects permissions. Members without marketplace usage permission see a disabled button with an explanatory tooltip, instead of discovering they cannot apply the Agent after clicking. Identity zone inference has also been corrected so members with only conversation permissions are no longer misidentified as builders.

### 🛠️ Bug fixes

* **Vision Parser upgrade and blank page fix**: An issue where Vision Parser silently parsed some document pages as blank, leaving gaps in knowledge base content, has been fixed. The parser's underlying model has also been upgraded to Claude Sonnet 5, with tests in Chinese, English, Japanese, and Thai showing parsing quality comparable to the previous version. Separately, the platform has disabled MaiAgent Parser (Online) due to upstream service quota limits.
* **Fixed Web Chat crashes on some Android devices**: Opening Web Chat in certain Android apps' embedded browsers where local storage is disabled no longer crashes the entire page. Chat features now work normally.
* **Fixed login failures in on-premises deployments**: Login failures in Web Chat and the admin interface for private (on-premises) deployments have been fixed. Frontend error information in on-premises environments is also prevented from being sent to the cloud, keeping all data within the customer's environment.
* **Fixed attachment order for multiple uploads**: When uploading multiple images or files at once in Web Chat or MaiGPT, attachments now retain the order in which you added them. The AI assistant reads images in the same order as the displayed thumbnails, so asking about the “first image” returns the correct one. Files still upload simultaneously, preserving upload speed.
* **Tool narration follows the interface language**: After switching the interface to English, Japanese, or another language, default system tool narration such as “Searching the knowledge base” and “Web search completed” in Web Chat and admin conversations now follows that language. Narration in existing conversations also updates when switching languages. Administrator-customized narration is unaffected.
* **Fixed blank screens after leaving conversation log details**: Fixed an issue where navigating to another page after viewing a conversation log's details in AI assistant monitoring left the content area completely blank, requiring a page refresh to recover.
* **Fixed login requirements for read-only conversation links**: Recipients of conversation viewing links can once again view conversations that allow public read-only access without logging in. Login prompts appear only when a conversation actually requires authorization. General connection problems now show a loading failure instead of misleadingly directing users to log in.
* **Fixed outdated list previews after editing messages**: After a message is edited in the conversation platform, conversation list previews in other tabs and other support agents' views now immediately show the updated content. No manual refresh is needed, and conversation order remains unchanged.
* **Fixed inaccurate AI assistant quota displays**: The AI assistant card in the organization usage overview now shows your plan's actual limit and the number of assistants that count toward it. Organizations with purchased extra quota see the correct denominator. Legacy-mode assistants retained after downgrading are clearly identified as not counting toward the quota, and the card tells you exactly how many more assistants you can create.
* **Fixed built-in Agent usage counts changing with filters**: The usage card on the built-in Agent management page now consistently shows the organization's actual total activation count. Searching or filtering the list no longer changes the number, keeping it consistent with quota information in the bottom batch action bar.
* **Fixed partially initialized MaiGPT pages**: When opening the MaiGPT conversation page, it now completes initialization and supports normal conversations even if the model list temporarily fails to load or times out. The page no longer gets stuck partially loaded with a refresh prompt, and wait times are significantly shorter.
* **Fixed transient display errors when loading MaiGPT conversations**: When entering the MaiGPT conversation page after login, the message list no longer briefly fails to display while assistant data is still loading. Failed login form validation continues to show red field-level messages without generating unnecessary background errors.
* **Fixed unexpected workspace access points and errors**: Fixed the “File Workspace” option appearing in MaiGPT chats for organizations that had not enabled this capability. These organizations no longer see an access point that only returns a permission error when clicked. Organizations with the capability enabled can continue using it as usual.
* **Fixed scrolling on the AI Gateway overview**: Fixed the AI Gateway overview page failing to scroll down in shorter windows. Content that exceeds the viewport now scrolls correctly, allowing users to view bottom sections such as “Usage by API Key” in full.
* **Fixed misaligned activity log rows**: Columns in the activity log under organization settings now align correctly when displaying system or organization events. The user column remains blank, and view buttons and alternating row shading display correctly.
* **Fixed save failures for overly long personal preferences**: The character limit for personal preferences is now 1000 characters, matching the server limit. Users no longer enter long text only to have it rejected upon submission.
* **Fixed clipped buttons in the API key list**: Action buttons on the right of the API key management table in organization settings now remain fully visible during horizontal scrolling, making them easier to read and click.
* **Corrected the file library menu position and localized space names**: “File Library” in the left menu now appears below “Agent Marketplace” to better match the navigation flow. Space names such as “My Files” and “Organization Shared Files” now follow the interface language instead of always appearing in English.
* **Fixed draft attachment removal for guests**: After guests upload attachments in Web Chat, clicking × before sending now actually removes the attachment, so the AI assistant no longer references removed files. A notification also appears if removal fails. Also fixed an issue where removing an attachment after uploading multiple file types at once could accidentally delete another attachment.
* **Tool card display settings no longer persist incorrectly**: Fixed tool execution cards being permanently hidden across conversations after switching to a chat with brand animation. Each conversation now displays tool cards according to its settings.
* **Fixed oversized brand logos**: Brand logos in the Web Chat footer and on login, registration, and password reset pages now have appropriate dimensions. The startup screen logo also no longer appears blurry on high-resolution screens.
* **Fixed interrupted connection cleanup when closing chats**: Internal cleanup during logout or when closing Web Chat is now more reliable. Repeated disconnection no longer causes cleanup to fail midway, preventing residual conversation or authentication state.


---

# 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 by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://docs.maiagent.ai/release-note/en/archives/2026/09.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

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.